I

VLSI
 VLSI

Edited by
Zhongfeng Wang

In-Tech
intechweb.org
Published by In-Teh

In-Teh
Olajnica 19/2, 32000 Vukovar, Croatia

Abstracting and non-profit use of the material is permitted with credit to the source. Statements and
opinions expressed in the chapters are these of the individual contributors and not necessarily those of
the editors or publisher. No responsibility is accepted for the accuracy of information contained in the
published articles. Publisher assumes no responsibility liability for any damage or injury to persons or
property arising out of the use of any materials, instructions, methods or ideas contained inside. After
this work has been published by the In-Teh, authors have the right to republish it, in whole or part, in any
publication of which they are an author or editor, and the make other personal use of the work.

© 2010 In-teh
www.intechweb.org
Additional copies can be obtained from:
[email protected]

First published February 2010

Printed in India

Technical Editor: Melita Horvat

Cover designed by Dino Smrekar

VLSI,
Edited by Zhongfeng Wang

p. cm.
ISBN 978-953-307-049-0
 V

Preface

The process of integrated circuits (IC) started its era of very-large-scale integration (VLSI)
in 1970’s when thousands of transistors were integrated into a single chip. Since then,
the transistors counts and clock frequencies of state-of-art chips have grown by orders of
magnitude. Nowadays we are able to integrate more than a billion transistors into a single
device. However, the term “VLSI” remains being commonly used, despite of some effort to
coin a new term ultralarge- scale integration (ULSI) for finer distinctions many years ago. In
the past two decades, advances of VLSI technology have led to the explosion of computer
and electronics world. VLSI integrated circuits are used everywhere in our everyday life,
including microprocessors in personal computers, image sensors in digital cameras, network
processors in the Internet switches, communication devices in smartphones, embedded
controllers in automobiles, et al.
VLSI covers many phases of design and fabrication of integrated circuits. In a complete VLSI
design process, it often involves system definition, architecture design, register transfer
language (RTL) coding, pre- and post-synthesis design verification, timing analysis, and chip
layout for fabrication. As the process technology scales down, it becomes a trend to integrate
many complicated systems into a single chip, which is called system-on-chip (SoC) design.
In addition, advanced VLSI systems often require high-speed circuits for the ever increasing
demand of data processing. For instance, Ethernet standard has evolved from 10 Mbps to
10 Gbps, and the specification for 100 Gbps Ethernet is underway. On the other hand, with
the growing popularity of smartphones and mobile computing devices, low-power VLSI
systems have become critically important. Therefore, engineers are facing new challenges to
design highly integrated VLSI systems that can meet both high performance requirement and
stringent low power consumption.
The goal of this book is to elaborate the state-of-art VLSI design techniques at multiple
levels. At device level, researchers have studied the properties of nano-scale devices and
explored possible new material for future very high speed, low-power chips. At circuit level,
interconnect has become a contemporary design issue for nano-scale integrated circuits.
At system level, hardware-software co-design methodologies have been investigated to
coherently improve the overall system performance. At architectural level, researchers have
proposed novel architectures that have been optimized for specific applications as well as
efficient reconfigurable architectures that can be adapted for a class of applications.
As VLSI systems become more and more complex, it is a great challenge but a significant task
for all experts to keep up with latest signal processing algorithms and associated architecture
designs. This book is to meet this challenge by providing a collection of advanced algorithms
VI

in conjunction with their optimized VLSI architectures, such as Turbo codes, Low Density
Parity Check (LDPC) codes, and advanced video coding standards MPEG4/H.264, et al. Each
of the selected algorithms is presented with a thorough description together with research
studies towards efficient VLSI implementations. No book is expected to cover every possible
aspect of VLSI exhaustively. Our goal is to provide the design concepts through those
selected studies, and the techniques that can be adopted into many other current and future
applications.
This book is intended to cover a wide range of VLSI design topics – both general design
techniques and state-of-art applications. It is organized into four major parts:
▪▪Part I focuses on VLSI design for image and video signal processing systems, at both
algorithmic and architectural levels.
▪▪Part II addresses VLSI architectures and designs for cryptography and error correction
coding.
▪▪Part III discusses general SoC design techniques as well as system-level design optimization
for application-specific algorithms.
▪▪Part IV is devoted to circuit-level design techniques for nano-scale devices.
It should be noted that the book is not a tutorial for beginners to learn general VLSI design
methodology. Instead, it should serve as a reference book for engineers to gain the knowledge
of advanced VLSI architecture and system design techniques. Moreover, this book also
includes many in-depth and optimized designs for advanced applications in signal processing
and communications. Therefore, it is also intended to be a reference text for graduate students
or researchers for pursuing in-depth study on specific topics.
The editors are most grateful to all coauthors for contributions of each chapter in their
respective area of expertise. We would also like to acknowledge all the technical editors for
their support and great help.

Zhongfeng Wang, Ph.D.

Broadcom Corp., CA, USA

Xinming Huang, Ph.D.

Worcester Polytechnic Institute, MA, USA
 VII

Contents

Preface V

1. Discrete Wavelet Transform Structures for VLSI Architecture Design 001

Hannu Olkkonen and Juuso T. Olkkonen

2. High Performance Parallel Pipelined Lifting-based VLSI Architectures

for Two-Dimensional Inverse Discrete Wavelet Transform 011
Ibrahim Saeed Koko and Herman Agustiawan

3. Contour-Based Binary Motion Estimation Algorithm and VLSI Design

for MPEG-4 Shape Coding 043
Tsung-Han Tsai, Chia-Pin Chen, and Yu-Nan Pan

4. Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based Discrete

Wavelet Transform for JPEG2000 069
Chih-Hsien Hsia and Jen-Shiun Chiang

5. Full HD JPEG XR Encoder Design for Digital Photography Applications 099

Ching-Yen Chien, Sheng-Chieh Huang, Chia-Ho Pan and Liang-Gee Chen

6. The Design of IP Cores in Finite Field for Error Correction 115

Ming-Haw Jing, Jian-Hong Chen, Yan-Haw Chen, Zih-Heng Chen and Yaotsu Chang

7. Scalable and Systolic Gaussian Normal Basis Multipliers

over GF(2m) Using Hankel Matrix-Vector Representation 131
Chiou-Yng Lee

8. High-Speed VLSI Architectures for Turbo Decoders 151

Zhongfeng Wang and Xinming Huang

9. Ultra-High Speed LDPC Code Design and Implementation 175

Jin Sha, Zhongfeng Wang and Minglun Gao

10. A Methodology for Parabolic Synthesis 199

Erik Hertz and Peter Nilsson

11. Fully Systolic FFT Architectures for Giga-sample Applications 221

D. Reisis
VIII

12. Radio-Frequency (RF) Beamforming Using Systolic FPGA-based Two

Dimensional (2D) IIR Space-time Filters 247
Arjuna Madanayake and Leonard T. Bruton

13. A VLSI Architecture for Output Probability Computations of HMM-based

Recognition Systems 273
Kazuhiro Nakamura, Masatoshi Yamamoto, Kazuyoshi Takagi and Naofumi Takagi

14. Efficient Built-in Self-Test for Video Coding Cores: A Case Study
on Motion Estimation Computing Array 285
Chun-Lung Hsu, Yu-Sheng Huang and Chen-Kai Chen

15. SOC Design for Speech-to-Speech Translation 297

Shun-Chieh Lin, Jia-Ching Wang, Jhing-Fa Wang, Fan-Min Li and Jer-Hao Hsu

16. A Novel De Bruijn Based MeshTopology for Networks-on-Chip 317

Reza Sabbaghi-Nadooshan, Mehdi Modarressi and Hamid Sarbazi-Azad

17. On the Efficient Design & Synthesis of Differential Clock Distribution Networks 331
Houman Zarrabi, Zeljko Zilic, Yvon Savaria and A. J. Al-Khalili

18. Robust Design and Test of Analog/Mixed-Signal Circuits

in Deeply Scaled CMOS Technologies 353
Guo Yu and Peng Li

19. Nanoelectronic Design Based on a CNT Nano-Architecture 375

Bao Liu

20. A New Technique of Interconnect Effects Equalization by using Negative Group

Delay Active Circuits 409
Blaise Ravelo, André Pérennec and Marc Le Roy

21. Book Embeddings 435

Saïd Bettayeb

22. VLSI Thermal Analysis and Monitoring 441

Ahmed Lakhssassi and Mohammed Bougataya
Discrete Wavelet Transform Structures for VLSI Architecture Design 1

X1

Discrete Wavelet Transform Structures

for VLSI Architecture Design
Hannu Olkkonen and Juuso T. Olkkonen
Department of Physics, University of Kuopio, 70211 Kuopio, Finland
VTT Technical Research Centre of Finland, 02044 VTT, Finland

1. Introduction
Wireless data transmission and high-speed image processing devices have generated a need
for efficient transform methods, which can be implemented in VLSI environment. After the
discovery of the compactly supported discrete wavelet transform (DWT) (Daubechies, 1988;
Smith & Barnwell, 1986) many DWT-based data and image processing tools have
outperformed the conventional discrete cosine transform (DCT) -based approaches. For
example, in JPEG2000 Standard (ITU-T, 2000), the DCT has been replaced by the
biorthogonal discrete wavelet transform. In this book chapter we review the DWT structures
intended for VLSI architecture design. Especially we describe methods for constructing shift
invariant analytic DWTs.

2. Biorthogonal discrete wavelet transform

The first DWT structures were based on the compactly supported conjugate quadrature
filters (CQFs) (Smith & Barnwell, 1986), which had nonlinear phase effects such as image
blurring and spatial dislocations in multi-resolution analyses. On the contrary, in
biorthogonal discrete wavelet transform (BDWT) the scaling and wavelet filters are
symmetric and linear phase. The two-channel analysis filters H 0 ( z ) and H1 ( z ) (Fig. 1) are of
the general form
H 0 ( z )  (1  z 1 ) K P ( z )
(1)
H1 ( z )  (1  z 1 ) K Q ( z )
where the scaling filter H 0 ( z ) has the Kth order zero at    . The wavelet filter H1 ( z ) has
the Kth order zero at   0 , correspondingly. P ( z ) and Q( z ) are polynomials in z 1 . The
reconstruction filters G0 ( z ) and G1 ( z ) (Fig. 1) obey the well-known perfect reconstruction
condition
H 0 ( z ) G0 ( z )  H1 ( z ) G1 ( z )  2 z  k
(2)
H 0 ( z ) G0 ( z )  H1 ( z ) G1 ( z )  0
2 VLSI

The last condition in (2) is satisfied if we select the reconstruction filters as

G0 ( z )  H1 ( z ) and G1 ( z )   H 0 ( z ) .

Fig. 1. Analysis and synthesis BDWT filters.

3. Lifting BDWT
The BDWT is most commonly realized by the ladder-type network called lifting scheme
(Sweldens, 1988). The procedure consists of sequential down and uplifting steps and the
reconstruction of the signal is made by running the lifting network in reverse order (Fig. 2).
Efficient lifting BDWT structures have been developed for VLSI design (Olkkonen et al.
2005). The analysis and synthesis filters can be implemented by integer arithmetics using
only register shifts and summations. However, the lifting DWT runs sequentially and this
may be a speed-limiting factor in some applications (Huang et al., 2005). Another drawback
considering the VLSI architecture is related to the reconstruction filters, which run in reverse
order and two different VLSI realizations are required. In the following we show that the
lifting structure can be replaced by more effective VLSI architectures. We describe two
different approaches: the discrete lattice wavelet transform and the sign modulated BDWT.

Fig. 2. The lifting BDWT structure.

4. Discrete lattice wavelet transform

In the analysis part the discrete lattice wavelet transform (DLWT) consists of the scaling
H 0 ( z ) and wavelet H1 ( z ) filters and the lattice network (Fig. 3). The lattice structure contains
two parallel transmission filters T0 ( z ) and T1 ( z ) , which exchange information via two
crossed lattice filters L0 ( z ) and L1 ( z ) . In the synthesis part the lattice structure consists of the
transmission filters R0 ( z ) and R1 ( z ) and crossed filters W0 ( z ) and W1 ( z ) , and finally the
reconstruction filters G0 ( z ) and G1 ( z ) . Supposing that the scaling and wavelet filters obey
(1), for perfect reconstruction the lattice structure should follow the condition
Discrete Wavelet Transform Structures for VLSI Architecture Design 3

Fig. 3. The general DLWT structure.

T0 R0  LW
1 0 L0 R0  TW
1 0  zk 0 
T W  L R    (3)
 0 1 1 1 T1R1  L0W1   0 z k 
This is satisfied if we state W0   L0 , W1   L1 , R0  T1 and R1  T0 . The perfect
reconstruction condition follows then from the diagonal elements (3) as
T0 ( z )T1 ( z )  L0 ( z ) L1 ( z )  z  k (4)
There exists many approaches in the design of the DLWT structures obeying (4), for
example via the Parks-McChellan-type algorithm. Especially the DLWT network is efficient
in designing half-band transmission and lattice filters (see details in Olkkonen & Olkkonen,
2007a). For VLSI design it is essential to note that in the lattice structure all computations are
carried out parallel. Also all the BDWT structures designed via the lifting scheme can be
transferred to the lattice network (Fig. 3). For example, Fig. 4 shows the DLWT equivalent
of the lifting DBWT structure consisting of down and uplifting steps (Fig. 2). The VLSI
implementation is flexible due to parallel filter blocks in analysis and synthesis parts.

Fig. 4. The DLWT equivalence of the lifting BDWT structure described in Fig. 2.

5. Sign modulated BDWT

In VLSI architectures, where the analysis and synthesis filters are directly implemented (Fig.
1), the VLSI design simplifies considerably using a spesific sign modulator defined as
(Olkkonen & Olkkonen 2008)
1 for n even
S n  (1) n   (5)
-1 for n odd
A key idea is to replace the reconstruction filters by scaling and wavelet filters using the sign
modulator (5). Fig. 5 describes the rules how H ( z ) can be replaced by H ( z ) and the sign
modulator in connection with the decimation and interpolation operators. Fig. 6
4 VLSI

Fig. 5. The equivalence rules applying the sign modulator.

describes the general BDWT structure using the sign modulator. The VLSI design simplifies
to the construction of two parallel biorthogonal filters and the sign modulator. It should be
pointed out that the scaling and wavelet filters can be still efficiently implemented using the
lifting scheme or the lattice structure. The same biorthogonal DWT/IDWT filter module
can be used in decomposition and reconstruction of the signal e.g. in video compression
unit. Especially in bidirectional data transmission the DWT/IDWT transceiver has many
advantages compared with two separate transmitter and receiver units. The same VLSI
module can also be used to construct multiplexer-demultiplexer units. Due to symmetry of
the scaling and wavelet filter coefficents a fast convolution algorithm can be used for
implementation of the filter modules (see details Olkkonen & Olkkonen, 2008).

Fig. 6. The BDWT structure using the scaling and wavelet filters and the sign modulator.

6. Design example: Symmetric half-band wavelet filter for compression coder

The general structure for the symmetric half-band filter (HBF) is, for k odd
H ( z )  z  k  B( z 2 ) (6)
where B ( z 2 ) is a symmetric polynomial in z 2 . The impulse response of the HBF contains
only one odd point. For example, we may parameterize the eleven point HBF impulse
response as h[n]  [c 0 b 0 a 1 a 0 b 0 c ] , which has three adjustable parameters. The
compression efficiency improves when the high-pass wavelet filter approaches the
frequency response of the sinc-function, which has the HBF structure. However, the
impulse response of the sinc-function is infinite, which prolongs the computation time. In
this work we select the seven point compactly supported HBF prototype as a wavelet filter,
which has the impulse response
h1[n]  [b 0 a 1 a 0 b] (7)
containing two adjustable parameters a and b. In our previous work we have introduced a
modified regulatory condition for computation of the parameters of the wavelet filter
(Olkkonen et al. 2005)
N

n
n 0
m
h1[n]  0; m  0,1,..., M  1 (8)
Discrete Wavelet Transform Structures for VLSI Architecture Design 5

This relation implies that H1 ( z ) contains Mth-order zero at z  1 (  0) , where M is the

number of vanishing moments. Writing (8) for the prototype filter (7) we obtain two
equations 2a  2b  1  0 and 20a  36b  9  0 , which give the solution a  9 /16 and
b  1/16 . The wavelet filter has the z-transform
H1 ( z )  (1  z 1 ) 4 (1  4 z 1  z 2 ) /16 (9)
having fourth order root at z=1. The wavelet filter can be realized in the HBF form
H1 ( z )  z 3  A( z 2 ) ; A( z 2 )  ( 1  9 z 2  9 z 4  z 6 ) /16 (10)
Using the equivalence
 H ( z 2 )   ( 2) H ( z ) (11)
2

the HBF structure can be implemented using the lifting scheme (Fig. 7). The functioning of
the compression coder can be explained by writing the input signal via the polyphase
components
X ( z )  X e ( z 2 )  z 1 X 0 ( z 2 ) (12)
where X e ( z ) and X o ( z ) denote the even and odd sequences. We may present the wavelet
coefficents as
W ( z )  [ X ( z ) H1 ( z )] 2  z 2 X o ( z )  A( z ) X e ( z ) (13)
A( z ) works as an approximating filter yielding an estimate of the odd data points based on
the even sequence. The wavelet sequence W ( z ) can be interpreted as the difference between
the odd points and their estimate. In tree structured compression coder the scaling sequence
S ( z ) is fed to the next stage. In many VLSI applications, for example image compression,
the input signal consists of an integer-valued sequences. By rounding or truncating the
output of the A( z ) filter to integers, the compressed wavelet sequence W ( z ) is integer-
valued and can be efficiently coded e.g. using Huffman algorithm. It is essential to note that
this integer-to-integer transform has still the perfect reconstruction property (2).

Fig. 7. The lifting structure for the HBF wavelet filter designed for the VLSI compression
coder.

7. Shift invariant BDWT

The drawback in multi-scale BWDT analysis of signals and images is the dependence of the
total energy of the wavelet coefficients on the fractional shifts of the analysed signal. If we
have a discrete signal x[ n] and the corresponding time shifted signal x[n   ] , where   [0,1] ,
there may exist a significant difference in the energy of the wavelet coefficients as a function
of the time shift. Kingsbury (2001) proposed a nearly shift invariant complex wavelet
transform, where the real and imaginary wavelet coefficients are approximately Hilbert
transform pairs. The energy (absolute value) of the wavelet coefficients equals the envelope,
6 VLSI

which warrants smoothness and shift invariance. Selesnick (2002) observed that using two
parallel CQF banks, which are constructed so that the impulse responses of the scaling
filters are half-sample delayed versions of each other: h0 [ n] and h0 [n  0.5] , the
corresponding wavelets are Hilbert transform pairs. In z-transform domain we should be
able to construct the scaling filters H 0 ( z ) and z 0.5 H 0 ( z ) . However, the constructed scaling
filters do not possess coefficient symmetry and in multi-scale analysis the nonlinearity
disturbs spatial timing and prevents accurate statistical correlations between different
scales. In the following we describe the shift invariant BDWT structures especially designed
for VLSI applications.

7.1 Half-delay filters for shift invariant BDWT

The classical approach for design of the half-sample delay filter D( z ) is based on the Thiran
all-pass interpolator
p
c  z 1
D ( z )  z 0.5   k 1
(14)
k 1 1  ck z

where the ck coefficients are designed so that the frequency response follows approximately
D ( )  e  j / 2 (15)
Recently, half-delay B-spline filters have been introduced, which have an ideal phase
response. The method yields linear phase and shift invariant transform coefficients and can
be adapted to any of the existing BDWT (Olkkonen & Olkkonen, 2007b). The half-sample
delayed scaling and wavelet filters and the corresponding reconstruction filters are
H 0 ( z )  D( z ) H 0 ( z )
H1 ( z )  D 1 ( z ) H1 ( z )
(16)
G0 ( z )  D 1 ( z )G0 ( z )
G1 ( z )  D ( z )G1 ( z )
The half-delayed BDWT filter bank obeys the perfect reconstruction condition (2). The B-
spline half-delay filters have the IIR structure
A( z )
D( z )  (17)
B( z )
which can be implemented by the inverse filtering procedure (see details Olkkonen &
Olkkonen 2007b).

7.2 Hilbert transform-based shift invariant DWT

The tree-structured complex DWT is based on the FFT-based computation of the Hilbert
transform ( H a operator in Fig. 8). The scaling and wavelet filters both obey the HBF
structure (Olkkonen et al. 2007c)
1
H 0 ( z)   z 1 B ( z 2 )
2 (18)
1
H 1 ( z )   z 1 B ( z 2 )
2
Discrete Wavelet Transform Structures for VLSI Architecture Design 7

For example, the impulse response h0 [ n]  [1 0 9 16 9 0 -1]/ 32 has the fourth order zero at
   and h1[n]  [1 0 -9 16 -9 0 1]/ 32 has the fourth order zero at   0 . In the tree structured
HBF DWT the wavelet sequences wa [n] . A key feature is that the odd coefficients of the
analytic signal wa [2n  1] can be reconstructed from the even coefficient values wa [2n] . This
avoids the need to use any reconstruction filters. The HBFs (18) are symmetric with respect
to    / 2 . Hence, the energy in the frequency range 0   is equally divided by the
scaling and wavelet filters and the energy (absolute value) of the scaling and wavelet

Fig. 8. Hilbert transform-based shift invariant DWT.

coefficients are statistically comparable. The computation of the analytic signal via the
Hilbert transform requires the FFT-based signal processing. However, efficient FFT chips
are available for VLSI implementation. In many respects the advanced method outperforms
the previous nearly shift invariant DWT structures.

7.3 Hilbert transform filter for construction of shift invariant BDWT

The FFT-based implementation of the shift invariant DWT can be avoided if we define the
Hilbert transform filter ( z ) , which has the frequency response
( )  e  j / 2 sgn( ) (19)
where sgn( )  1 for   0 and sgn( )  0 for   0 . In the following we describe a novel
method for constructing the Hilbert transform filter based on the half-sample delay filter
D( z ) (17), whose frequency response follows (15). The quadrature mirror filter D ( z ) has
the frequency response
D (   )  e  j (  ) / 2 (20)
The frequency response of the filter D ( z ) D 1 (  z ) is, correspondingly
D ( )
 e  j / 2e j (  ) / 2  e  j / 2 (21)
D (   )
Comparing (19) and using notation (17) we obtain the Hilbert transform filter as
A( z ) B(  z )
( z )  (22)
A( z ) B( z )
The corresponding parallel BDWT filter bank is
H 0 ( z )  ( z ) H 0 ( z )
H1 ( z )   1 ( z ) H1 ( z )
(23)
G0 ( z )   1 ( z )G0 ( z )
G1 ( z )  ( z )G1 ( z )
8 VLSI

By filtering the real-valued signal X ( z ) by the Hilbert transform filter results in an analytic
signal [1  j( z )] X ( z ) , whose magnitude response is zero at negative side of the frequency
spectrum. For example, an integer-valued half-delay filter D( z ) for this purpose is obtained
by the B-spline transform (Olkkonen & Olkkonen, 2007b). The frequency response of the
Hilbert transform filter designed by the fourth order B-spline (Fig. 9) shows a maximally flat
magnitude spectrum. The phase spectrum corresponds to the ideal Hilbert transformer (19).

Fig. 9. Magnitude and phase spectra of the Hilbert transform filter yielded by the fourth
order B-spline transform.

8. Conclusion
In this book chapter we have described the BDWT constructions especially tailored for VLSI
environment. Most of the VLSI designs in the literature are focused on the biorthogonal 9/7
filters, which have decimal coefficients and usually implemented using the lifting scheme
(Sweldens, 1988). However, the lifting BDWT needs two different filter banks for analysis
and synthesis parts. The speed of the lifting BDWT is also limited due to the sequential
lifting steps. In this work we showed that the lifting BDWT can be replaced by the lattice
structure (Olkkonen & Olkkonen, 2007a). The two-channel DLWT filter bank (Fig. 3) runs
parallel, which significantly increases the channel throughout. A significant advantage
compared with the previous maximally decimated filter banks is that the DLWT structure
allows the construction of the half-band lattice and transmission filters. In tree structured
wavelet transform half-band filtered scaling coefficients introduce no aliasing when they are
fed to the next scale. This is an essential feature when the frequency components in each
scale are considered, for example in electroencephalography analysis.
The VLSI design of the BDWT filter bank simplifies essentially by implementing the sign
modulator unit (Fig. 5), which eliminates the need for constructing separate reconstruction
filters. The biorthogonal DWT/IDWT transceiver module uses only two parallel filter
structures. Especially in bidirectional data transmission the DWT/IDWT module offers
several advantages compared with the separate transmit and receive modules, such as the
Discrete Wavelet Transform Structures for VLSI Architecture Design 9

reduced size, low power consumption, easier synchronization and timing requirements. For
the VLSI designer the DWT/IDWT module appears as a "black box", which readily fits to
the data under processing. This may override the relatively big barrier from the wavelet
theory to the practical VLSI and microprocessor applications. As a design example we
described the construction of the compression coder (Fig. 7), which can be used to compress
integer-valued data sequences, e.g. produced by the analog-to-digital converters.
It is well documented that the real-valued DWTs are not shift invariant, but small fractional
time-shifts may introduce significant differences in the energy of the wavelet coefficients.
Kingsbury (2001) showed that the shift invariance is improved by using two parallel filter
banks, which are designed so that the wavelet sequences constitute real and imaginary parts
of the complex analytic wavelet transform. The dual-tree discrete wavelet transform (DT-
DWT) has been shown to outperform the real-valued DWT in a variety of applications such
as denoising, texture analysis, speech recognition, processing of seismic signals and
neuroelectric signal analysis (Olkkonen et al. 2006). Selesnick (2002) made an observation
that a half-sample time-shift between the scaling filters in parallel CQF banks is enough to
produce the analytic wavelet transform, which is nearly shift invariant. In this work we
described the shift invariant DT-BDWT bank (16) based on the half-sample delay filter. It
should be pointed out that the half-delay filter approach yields wavelet bases which are
Hilbert transform pairs, but the wavelet sequences are only approximately shift invariant. In
multi-scale analysis the complex wavelet sequences should be shift invariant. This
requirement is satisfied in the Hilbert transform-based approach (Fig. 8), where the signal in
every scale is Hilbert transformed yielding strictly analytic and shift invariant transform
coefficients. The procedure needs FFT-based computation (Olkkonen et al. 2007c), which
may be an obstacle in many VLSI realizations. To avoid this we described a Hilbert
transform filter for constructing the shift invariant DT-BDWT bank (23). Instead of the half-
delay filter bank approach (16) the perfect reconstruction condition (2) is attained using the
IIR-type Hilbert transform filters, which yield analytic wavelet sequences.

9. References
Daubechies, I. (1988). Orthonormal bases of compactly supported wavelets. Commmun. Pure
Appl. Math., Vol. 41, 909-996.
Huang, C.T., Tseng, O.O. & Chen, L.G. (2005). Analysis and VLSI architecture for 1-D and 2-
D discrete wavelet transform. IEEE Trans. Signal Process. Vol. 53, No. 4, 1575-1586.
ITU-T (2000) Recommend. T.800-ISO DCD15444-1: JPEG2000 Image Coding System.
International Organization for Standardization, ISO/IEC JTC! SC29/WG1.
Kingsbury, N.G. (2001). Complex wavelets for shift invariant analysis and filtering of
signals. J. Appl. Comput. Harmonic Analysis. Vol. 10, 234-253.
Olkkonen, H., Pesola, P. & Olkkonen, J.T. (2005). Efficient lifting wavelet transform for
microprocessor and VLSI applications. IEEE Signal Process. Lett. Vol. 12, No. 2, 120-
122.
Olkkonen, H., Pesola, P., Olkkonen, J.T. & Zhou, H. (2006). Hilbert transform assisted
complex wavelet transform for neuroelectric signal analysis. J. Neuroscience Meth.
Vol. 151, 106-113.
Olkkonen, J.T. & Olkkonen, H. (2007a). Discrete lattice wavelet transform. IEEE Trans.
Circuits and Systems II. Vol. 54, No. 1, 71-75.
10 VLSI

Olkkonen, H. & Olkkonen, J.T. (2007b). Half-delay B-spline filter for construction of shift-
invariant wavelet transform. IEEE Trans. Circuits and Systems II. Vol. 54, No. 7, 611-
615.
Olkkonen, H., Olkkonen, J.T. & Pesola, P. (2007c). FFT-based computation of shift invariant
analytic wavelet transform. IEEE Signal Process. Lett. Vol. 14, No. 3, 177-180.
Olkkonen, H. & Olkkonen, J.T. (2008). Simplified biorthogonal discrete wavelet transform
for VLSI architecture design. Signal, Image and Video Process. Vol. 2, 101-105.
Selesnick, I.W. (2002). The design of approximate Hilbert transform pairs of wavelet bases.
IEEE Trans. Signal Process. Vol. 50, No. 5, 1144-1152.
Smith, M.J.T. & Barnwell, T.P. (1986). Exaxt reconstruction for tree-structured subband
coders. IEEE Trans. Acoust. Speech Signal Process. Vol. 34, 434-441.
Sweldens, W. (1988). The lifting scheme: A construction of second generation wavelets.
SIAM J. Math. Anal. Vol. 29, 511-546.
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 11

X2

High Performance Parallel Pipelined

Lifting-based VLSI Architectures
for Two-Dimensional Inverse
Discrete Wavelet Transform
Ibrahim Saeed Koko and Herman Agustiawan
Electrical & Electronic Engineering Department
Universiti Teknologi PETRONAS, Tronoh
Malaysia

1. Introduction
Two-dimensional discrete wavelet transform (2-D DWT) has evolved as an effective and
powerful tool in many applications especially in image processing and compression. This is
mainly due to its better computational efficiency achieved by factoring wavelet transforms
into lifting steps. Lifting scheme facilitates high speed and efficient implementation of
wavelet transform and it attractive for both high throughput and low-power applications
(Lan & Zheng, 2005).
DWT considered in this work is part of a compression system based on wavelet such as
JPEG2000. Fig.1 shows a simplified compression system. In this system, the function of the
2-D FDWT is to decompose an NxM image into subbands as shown in Fig. 2 for 3-level
decomposition. This process decorrelates the highly correlated pixels of the original image.
That is the decorrelation process reduces the spatial correlation among the adjacent pixels of
the original image such that they can be amenable to compression.
After transmitting to a remote site, the original image must be reconstructed from the
decorrelated image. The task of the reconstruction and completely recovering the original
image from the decorrelated image is performed by inverse discrete wavelet transform
(IDWT).
The decorrelated image shown in Fig. 2 can be reconstructed by 2-D IDWT as follows. First,
it reconstructs in the column direction subbands LL3 and LH3 column-by-column to recover
L3 decomposition. Similarly, subbands HL3 and HH3 are reconstructed to obtain H3
decomposition. Then L3 and H3 decompositions are combined row-wise to reconstruct
subband LL2. This process is repeated in each level until the whole image is reconstructed
(Ibrahim & Herman, 2008).
The reconstruction process described above implies that the task of the reconstruction can be
achieved by using 2 processors (Ibrahim & Herman, 2008). The first processor (the column-
processor) computes column-wise to combine subbands LL and LH into L and subbands HL
and HH into H, while the second processor (the row-processor) computes row-wise to
12 VLSI

combine L and H into the next level subband. The decorrelated image shown in Fig. 2 is
assumed to be residing in an external memory with the same format.
In this chapter, parallelism is explored to best meet real-time applications of 2-D DWT with
demanding requirements. The single pipelined architecture developed by (Ibrahim &
Herman, 2008) is extended to 2- and 4-parallel pipelined architectures for both 5/3 and 9/7
inverse algorithms to achieve speedup factors of 2 and 4, respectively. The advantage of the
proposed architectures is that the total temporary line buffer (TLB) size does not increase
from that of the single pipelined architecture when degree of parallelism is increased.

FDWT Compress Decompress IDWT

Decorrelat es Decorrelated decorrelat ed Re constructs
image image Transmit to a image image
remote site

Fig. 1. A simplified Compression System

LL3 HL3
LH 3 HH 3 HL 2
HL1
LH 2 HH 2

LH1 HH1

Fig. 2. Subband decomposition of an NxM image into 3 levels.

2. Lifting-based 5/3 and 9/7 synthesis algorithms and data dependency

graphs
The 5/3 and the 9/7 inverse discrete wavelet transforms algorithms are defined by the
JPEG2000 image compression standard as follow (Lan & Zheng, 2005):
5/3 synthesis algorithm
 Y (2n  1)  Y (2n  1)  2 
step1: X (2n)  Y (2n)   
 4 
(1)
 X ( 2 n )  X (2 n  2 )
step2 : X (2n  1)  Y (2n  1)   
 2 
9/7 synthesis algorithm
Step 1: Y (2n)  1 k  Y (2n)
Step 2: Y ( 2n  1)  k  Y (2n  1)
Step 3: Y (2n)  Y (2n)   (Y (2n  1)  Y (2n  1))
Step 4: Y (2n 1)  Y (2n 1)   (Y (2n)  Y (2n  2)) (2)
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 13

Step 5: X (2n)  Y (2n)   (Y (2n  1)  Y (2n  1))

Step 6: X (2n 1)  Y (2n 1)  ( X (2n)  X (2n  2))
The data dependency graphs (DDGs) for 5/3 and 9/7 derived from the synthesis algorithms
are shown in Figs. 3 and 4, respectively. The DDGs are very useful tools in architecture
development and provide the information necessary for designer to develop more accurate
architectures. The symmetric extension algorithm recommended by JPEG2000 is
incorporated into the DDGs to handle the boundaries problems. The boundary treatment is
necessary to keep number of wavelet coefficients the same as that of the original input Pixels
and as a result prevent distortion from appearing at image boundaries. The boundary
treatment is only applied at the beginning and ending of each row or column in an NxM
image (Dillin et al., 2003). In DDGs, nodes circled with the same numbers are considered
redundant computations, which will be computed once and used thereafter. Coefficients of
the nodes circled with even numbers in the DDGs are low coefficients and that circled with
odd numbers are high coefficients.

3. Scan methods
The hardware complexity and hence the memory required for 2-D DWT architecture in
general depends on the scan method adopted for scanning external memory. Therefore, the
scan method shown in Fig. 5 is proposed for both 5/3 and 9/7 CPs. Fig. 5 (A) is formed for
illustration purposes by merging together subbands LL and LH, where subband LL
coefficients occupy even row and subband LH coefficients occupy odd rows, while Fig. 5 (B)
is formed by merging subband HL and HH together.
According to the scan method shown in Fig. 5, CPs of both 5/3 and 9/7 should scan external
memory column-by-column. However, to allow the RP, which operates on data generated
by the CP, to work in parallel with the CP as earlier as possible, the A’s (LL+LH) first two
columns coefficients are interleaved in execution with the B’s (HL+HH) first two columns
coefficients, in the first run. In all subsequent runs, two columns are interleaved, one from A
with another from B.
Interleaving of 4 columns in the first run takes place as follows. First, coefficients LL0,0,
LH0,0 from the first column of (A) are scanned. Second, coefficients HL0,0 and HH0,0 from
the first column of B are scanned, then LL0,1 and LH0,1 from the second column of A
followed by HL0,1 and HH0,1 from the second column of B are scanned. The scanning
process then returns to the first Fig. 3. 5/3 synthesis algorithm’s DDGs for (a) odd and (b)
even length signals returns to the first column of A to repeat the process and so on.
14 VLSI

Y (n) 1 0 1 2 3 4 5 6 7 8 7 1 0 1 2 3 4 5 6 7 6 5

X ( 2n) 0 2 4 6 8 0 2 4 6 6

redundant
X ( 2n  1) 1 3 5 7 1 3 5 7
computatio ns
X 0 X1 X 2 X 3 X 4 X 5 X 6 X 7 X 8 X 0 X1 X 2 X 3 X 4 X 5 X 6 X 7

(a ) (b)
Fig. 3. 5/3 synthesis algorithm’s DDGs for (a) odd and (b) even length signals
Y (n) 3 2 1 0 1 2 3 4 5 6 7 8 7 6 5 3 2 1 0 1 2 3 4 5 6 7 6 5 4 3
Y 2n  1 k k k k k k k k k k k k k k k k
Y 2n  k1 k1 k1 k1 k1 k1 k1 k1 k1 k1 k1 k1 k1 k1

Y 2n 2 0 2 4 6 8 6 2 0 2 4 6 6 4

Y2n 1 1 1 3 5 7 7 1 1 3 5 7 5

x(2n) 0 2 4 6 8 0 2 4 6 6

x(2n  1) 1 3 5 7 1 3 5 7

X 0 X1 X 2 X 3 X 4 X 5 X 6 X 7 X 8 X 0 X1 X 2 X 3 X 4 X 5 X 6 X 7
(a ) (b )

Fig. 4. 9/7 synthesis algorithm’s DDGs for (a) odd and (b) even length signals

The advantage of interleaving process not only it speedups the computations by allowing
the two processors to work in parallel earlier during the computations, but also reduces the
internal memory requirement between CP and RP to a few registers.
The scan method illustrated in Fig. 5 for 5/3 and 9/7 CP along with the DDGs suggest that
the RP should scan coefficients generated by CP according to the scan method illustrated in
Fig. 6. This figure is formed for illustration purposes by merging L and H decompositions
even though they are actually separate. In Fig. 6, L’s coefficients occupy even columns,
while H’s coefficients occupy odd columns. In the first run, the RP’s scan method shown in
Fig. 6 requires considering the first four columns for scanning as follows. First, coefficients
L0,0 and H0,0 from row 0 followed by L1,0 and H1,0 from row 1 are scanned. Then the scan
returns to row 0 and scans coefficients L0,1 and H0,1 followed by L1,1 and H1,1. This
process is repeated as shown in Fig. 6 until the first run completes.
In the second run, coefficients of columns 4 and 5 are considered for scanning by RP as
shown in Fig. 6, whereas in the third run, coefficients of columns 6 and 7 are considered for
scanning and so on.
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 15

According to the 9/7 DDGs, the RP’s scan method shown in Fig. 6 will generate only one
low output coefficient each time it processes 4 coefficients in the first run, whereas 5/3 RP
will generate two output coefficients. However, in all subsequent runs, both 9/7 and 5/3
RPs generate two output coefficients.

run 1 run 2 run 1 run 2

0 1 2 3 0 1 2 3
LL 0,0 0 1 3 HL0,0 0 2 4
LH 0,0 1 HH 0,0 1
LL1,0 2 HL1,0 2
LH 1,0 3 HH 1,0 3
4 4
5 5
6 6
A B
Fig. 5. 5/3 and 9/7 CPs scan method

L0,0 H 0,0 L 0,1 H 0,1

run 1 run 2
L1,0
0 1 2 3 4 5 6
0
1
2
3
4

Fig. 6. 5/3 and 9/7 RPs scan method

4. Approach
In (Ibrahim & Herman, 2008), to ease the architecture development the strategy adopted
was to divide the details of the development into two steps each having less information to
handle. In the first step, the DDGs were looked at from the outside, which is specified by the
dotted boxes in the DDGs, in terms of the input and output requirements. It can be observed
that the DDGs for 5/3 and 9/7 are identical when they are looked at from outside, taking
into consideration only the input and output requirements; but differ in the internal details.
Based on this observation, the first level, called external architecture, which is identical for
both 5/3 and 9/7, and consists of a column-processor (CP) and a row-processor (RP), was
developed. In the second step, the internal details of the DDGs for 5/3 and 9/7 were
considered separately for the development of processors’ datapath architectures, since
DDGs internally define and specify the internal structure of the processors
In this chapter, the first level, the external architectures for 2-parallel and 4-parallel are
developed for both 5/3 and 9/7 inverse algorithms. Then the processors’ datapath
16 VLSI

architectures developed in (Ibrahim & Herman, 2008) are modified to fit into the two
proposed parallel architectures’ processors.

4.1 Proposed 2-parallel external architecture

Based on the scan methods shown in Figs. 5 and 6 and the DDGs for 5/3 and 9/7, the 2-
parallel external architecture shown in Fig. 7 (a) is proposed for 5/3 and 9/7 and combined
5/3 and 9/7 for 2-D IDWT. The architecture consists of two k-stage pipelined column-
processors (CPs) labeled CP1 and CP2 and two k-stage pipelined row-processors (RPs)
labeled RP1 and RP2. The waveforms of the two clocks f 2 and f 2 2 used in the
architecture are shown in Fig. 7 (b).
In general, the scan frequency f l and hence the period  l  1 f l of the parallel architectures
can be determined by the following algorithm when the required pixels I of an operation are
scanned simultaneously in parallel. Suppose t p and t m are the processor and the external
memory critical path delays, respectively.
If t p l  k  t m then
l  t p l  k (3)
else  l  t m

Where l = 2, 3, 4 ... denote 2, 3, and 4-parallel and t p k is the stage critical path delay of a k-
stage pipelined processor. The clock frequency f 2 is determined from Eq(3) as
f 2  2k t p (4)
The architecture scans the external memory with frequency f 2 and it operates with
frequency f 2 2 . Each time two coefficients are scanned through the two buses labeled bus0
and bus1. The two new coefficients are loaded into CP1 or CP2 latches Rt0 and Rt1 every
time clock f 2 2 makes a negative or a positive transition, respectively.
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 17

f2 2

f2

f2 2

Fig. 7. (a) Proposed 2-parallel pipelined external architecture for 5/3 and 9/7 and combined
(b) Waveform of the clocks.

On the other hand, both RP1 and RP2 latches Rt0 and Rt1 load simultaneously new data
from CP1 and CP2 output latches each time clock f 2 2 makes a negative transition.
The dataflow for 5/3 2-parallel architecture is shown in Table 1, where CPs and RPs are
assumed to be 4-stage pipelined processors. The dataflow for 9/7 2-parallel architecture is
similar, in all runs, to the 5/3 dataflow except in the first, where RP1 and RP2 of the 9/7
architecture each would generate one output coefficient every other clock cycle, reference to
clock f 2 2 . The reason is that each 4 coefficients of a row processed in the first run by RP1
or RP2 of the 9/7 would require, according to the DDGs, two successive low coefficients
18 VLSI

from the first level of the DDGs labeled Y (2n) in order to carry out node 1 computations in
the second level labeled Y (2n  1) . In Table 1, the output coefficients in Rt0 of both RP1 and
RP2 represent the output coefficients of the 9/7 in the first run.
The strategy adopted for scheduling memory columns for CP1 and CP2 of the 5/3 and 9/7
2-parallel architectures, which are scanned according to the scan method shown in Fig. 5, is
as follow. In the first run, both 5/3 and 9/7 2-parallel architectures are scheduled for
executing 4 columns of memory, two from each (A) and (B) of Fig. 5. The first two columns
of Fig. 5 (A) are executed in an interleaved manner by CP1, while the first two columns of
Fig. 5 (B) are executed by CP2 also in an interleaved fashion as shown in the dataflow Table
1 In all other runs, 2 columns are scheduled for execution at a time. One column from (A) of
Fig. 5 will be scheduled for execution by CP1, while another from (B) of Fig. 5 will be
scheduled for CP2. However, if number of columns in (A) and (B) of Fig. 5 is not equal, then
the last run will consist of only one column of (A). In that case, schedule the last column in
CP1 only, but its output coefficients will be executed by both RP1 and RP2. The reason is
that if the last column is scheduled for execution by both CP1 and CP2, they will yield more
coefficients than that can be handled by both RP1 and RP2.
On the other hand, scheduling RP1 and RP2 of 5/3 and 9/7 2-parallel architectures occur
according to scan method shown in Fig. 6. In this scheduling strategy, all rows of even and
odd numbers in Fig. 6 will be scheduled for execution by RP1 and RP2, respectively. In the
first run, 4 coefficients from each 2 successive rows will be scheduled for RP1 and RP2,
whereas in all subsequent runs, two coefficients of each 2 successive rows will be scheduled
for RP1 and RP2, as shown in Fig. 6. However, if number of columns in Fig. 6 is odd which
occur when number of columns in (A) and (B) of Fig. 5 is not equal, then the last run would
require scheduling one coefficient from each 2 successive rows to RP1 and RP2 as shown in
column 6 of Fig. 6.
In general, all coefficients that belong to columns of even numbers in Fig. 6 will be
generated by CP1 and that belong to columns of odd numbers will be generated by CP2. For
example, in the first run, CP1will first generate two coefficients labeled L0,0 and L1,0 that
belong to locations 0,0 and 1,0 in Fig. 6, while CP2 will generate coefficient H0,0 and H1,0
that belong to locations 0,1 and 1,1. Then coefficients in locations 0,0 and 0,1 are executed by
RP1, while coefficients of locations 1,0 and 1,1 are executed by RP2. Second, CP1 will
generate two coefficients for locations 0,2 and 1,2, while CP2 generates two coefficients for
locations 0,3 and 1,3. Then coefficients in locations 0,2 and 0,3 are executed by RP1, while
coefficients in locations 1,2 and 1,3 are executed by RP2. The same process is repeated in the
next two rows and so on.
In the second run, first, CP1 generates coefficients of locations 0,4 and 1,4, whereas CP2
generates coefficients of locations 0,5 and 1,5 in Fig. 6. Then coefficients in locations 0,4 and
0,5 are executed by RP1, while coefficients in locations 1,4 and 1,5 are executed by RP2. This
process is repeated until the run completes. However, in the even that the last run processes
only one column of (A), CP1 would generate first coefficients of locations 0,m and 1,m
where m refers to the last column. Then coefficients of location 0,m is passed to RP1, while
coefficient of location 1,m is passed to RP2. In the second time, CP1 would generate
coefficients of locations 2,m and 3,m. Then 2,m is passed to RP1 and 3,m to RP2 and so on.
In the following, the dataflow shown in Table 1 for 2-parallel pipelined architecture will be
explained. The first run, which ends at cycle 16 in the table, requires scheduling four
columns as follows. In the first clock cycle, reference to clock f 2 , coefficients LL0,0 and
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 19

LH0,0 from the first column of LL3 and LH3 in the external memory, respectively, are
scanned and are loaded into CP1 latches Rt0 and Rt1 by the negative transition of
clock f 2 2 . The second clock cycle scans coefficients HL0,0 and HH0,0 from the first column
of HL3 and HH3, respectively, through the buses labeled bus0 and bus1 and loads them into
CP2 latches Rt0 and Rt1 by the positive transition of clock f 2 2 . In the third clock cycle, the
scanning process returns to the second column of subbands LL3 and LH3 in the external
memory and scans coefficients LL0,1 and LH0,1, respectively, and loads them into CP1
latches Rt0 and Rt1 by the negative transition of the clock f 2 2 . The fourth clock cycle scans
coefficients HL0,1 and HH0,1 from the second column of HL3 and HH3, respectively, and
loads them into CP2 latches Rt0 and Rt1. The scanning process then returns to the first
column in subbands LL3 and LH3 to repeat the process until the first run is completed.
In cycle 9, CP1 generates its first two output coefficients L0,0 and L1,0, which belong to L3
decomposition and loads them into its output latches Rtl0 and Rtl1, respectively, by the
negative transition of clock f 2 2 . In cycle 10, CP2 generates its first two output coefficients
H0,0 and H1,0, which belong to H3 decomposition and loads them into its output latches
Rth0 and Rth1, respectively, by the positive transition of clock f 2 2 .
In cycle 11, contents of Rtl0 and Rth0 are transferred to RP1 input latches Rt0 and Rt1,
respectively. The same clock cycle also transfers contents of Rtl1 and Rth1 to RP2 input
latches Rt0 and Rt1, respectively, while the two coefficients L0,1 and L1,1 generated by CP1
during the cycle are loaded into Rtl0 and Rtl1, respectively, by the negative transition of
clock f2/2.
In cycle 21, both RP1 and RP2 each yield its first two output coefficients, which are loaded
into their respective output latches by the negative transition of clock f 2 2 . Contents of
these output latches are then transferred to external memory where they are stored in the
first 2 memory locations of each rows 0 and 1. The dataflow of the first run then proceeds as
shown in Table 1. The second run begins at cycle 19 and yields its first 4 output coefficients
at cycle 37.
20 VLSI

Ck CP CP1 & CP2 CP1 CP2 RP1 input RP2 Output latches of
f2 input output output latches input RP1
latches latches latches Rt0 Rt1 latches RP2
Rt0 Rt1 Rtl0 Rtl1 Rth0 Rt0 Rt0 Rt1 Rt0
Rth1 Rt1 Rt1
1 1 LL0,0
LH0,0
2 2 HL0,0
HH0,0
3 1 LL0,1
LH0,1
4 2 HL0,1
HH0,1
5 1 LL1,0
LH1,0
6 2 HL1,0
HH1,0
7 1 LL1,1
LH1,1
8 2 HL1,1
HH1,1
9 1 LL2,0 L0,0 L1,0
LH2,0
10 2 HL2,0 H0,0
HH2,0 H1,0
11 1 LL2,1 L0,1 L1,1 L0,0 L1,0
LH2,1 H0,0 H1,0
RUN 1

12 2 HL2,1 H0,1
HH2,1 H1,1
13 1 LL3,0 L2,0 L3,0 L0,1 L1,1
LH3,0 H0,1 H1,1
14 2 HL3,0 H2,0
HH3,0 H3,0
15 1 LL3,1 L2,1 L3,1 L2,0 L3,0
LH3,1 H2,0 H3,0
16 2 HL3,1 H2,1
HH3,1 H3,1
17 1 ------ ------- L4,0 L5,0 L2,1 L3,1
- H2,1 H3,1
18 2 ------ ------- H4,0
H5,0
19 1 LL0,2 L4,1 L5,1 L4,0 L5,0
LH0,2 H4,0 H5,0
20 2 HL0,2 H4,1
HH0,2 H5,1
RUN 2

21 1 LL1,2 L6,0 L7,0 L4,1 L5,1 X0,0 X0,1 X1,0

LH1,2 H4,1 H5,1 X1,1
22 2 HL1,2 H6,0
HH1,2 H7,0
23 1 LL2,2 L6,1 L7,1 L6,0 L7,0 X0,2 ---- X1,2
LH2,2 H6,0 H7,0 -----
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 21

24 2 HL2,2 H6,1
HH2,2 H7,1
25 1 LL3,2 ----- ------ L6,1 L7,1 X2,0 X2,1 X3,0
LH3,2 H6,1 H7,1 X3,1
26 2 HL3,2 ----- -----
HH3,2
27 1 L0,2 L1,2 ----- ---- ----- -- X2,2 ---- X3,2
- ---- -----
28 2 H0,2
H1,2
29 1 L2,2 L3,2 L0,2 L1,2 X4,0 X4,1 X5,0
H0,2 H1,2 X5,1
30 2 H2,2
H3,2
31 1 L4,2 L5,2 L2,2 L3,2 X4,2 ---- X5,2
H2,2 H3,2 -----
32 2 H4,2
H5,2
33 1 L6,2 L7,2 L4,2 L5,2 X6,0 X6,1 X7,0
H4,2 H5,2 X7,1
34 2 H6,2
H7,2
35 1 L6,2 L7,2 X6,2 ---- X7,2
H6,2 H7,2 -----
36 2
37 1 X0,3 X0,4 X1,3
X1,4
38 2
39 1 X2,3 X2,4 X3,3
X3,4
Table 1. Dataflow for 2-parallel 5/3 architecture

4.2 Modified CPs and RPs for 5/3 and 9/7 2-parallel external architecture
Each CP of the 2-parallel external architecture is required to execute two columns in an
interleave fashion in the first run and one column in all other runs. Therefore, the 5/3
processor datapath developed in (Ibrahim & Herman, 2008) should be modified as shown in
Fig. 8 by adding one more stage between stages 2 and 3 for 5/3 2- parallel external
architecture to allow interleaving of two columns as described in the dataflow Table 1.
Through the two multiplexers labeled mux the processor controls between executing 2
columns and one column. Thus, in the first run, the two multiplexers’ control signal labeled
s is set 1 to allow interleaving in execution and 0 in all other runs. The modified 9-stage CP
for 9/7 2-parallel external architecture can be obtained by cascading two copies of Fig. 8.
On the other hand, RP1 and RP2 of the proposed 2-parallel architecture for 5/3 and 9/7 are
required to scan coefficients of H and L decompositions generated by CP1 and CP2
according to the scan method shown in Fig. 6. In this scan method, all rows of even numbers
are executed by RP1 and all rows of odds numbers are executed by RP2. That is, while RP1
is executing row0 coefficients, RP2 will be executing row1 coefficients and so on. In
addition, looking at the DDGs for 5/3 and 9/7 show that applying the scan methods in Fig.
22 VLSI

6 would require inclusion of temporary line buffers (TLBs) in RP1 and RP2 of the proposed
2-parallel external architecture as follows. In the first run, the fourth input coefficient of each
row in the DDGs and the output coefficients labeled X( 2) in the 5/3 DDGs and that labeled
Y"(2), Y"(1), and X(0) in the 9/7 DDGs, generated by considering 4 inputs coefficients in each
row, should be stored in TLBs, since they are required in the next run’s computations.
Similarly, in the second run, the sixth input coefficient of each row and the output
coefficients labeled X(4) in the 5/3 DDGs and that labeled Y"(4), Y"(3), and X(2) in the 9/7
DDGs generated by considering 2 inputs coefficients in each row, should be stored in TLBs.
Accordingly, 5/3 would require addition of 2 TLBs each of size N, whereas 9/7 would
require addition of 4 TLBs each of size N. However, since 2-parallel architecture consists of
two RPs, each 5/3 RP will have 2 TLBs each of size N/2 and each 9/7 RP will have 4 TLBs
each of size N/2 as shown in Fig. 9. Fig. 9 (a) represents the 5/3 modified RP, while both (a)
and (b) represent the 9/7 modified RP for 2- parallel architecture.

se0

se1
se2

Fig. 8. Modified inverse 5/3 CP for 2-parallel External architecture

To have more insight into the two RPs operations, the dataflow for 5/3 RP1 is given in Table
2 for first and second runs. Note that stage 1 input coefficients in Table 2 are exactly the
same input coefficients of RP1 in Table 1. In the first run, TLBs are only written, but in the
second run and in all subsequent runs, TLBs are read in the first half cycle and written in the
second half cycle. In the cycle 15, Table 2 shows that coefficients H0,1 is stored in the first
location of TLB1, while coefficient H2,1 is stored in the second location in cycle 19 and so on.
Run 2 starts at cycle 29. In cycle 30, the first location of TLB1, which contains coefficients
H0,1 is read during the first half cycle of clock f 2 2 and is loaded into Rd1 by the positive
transition of the clock, whereas coefficient H0,2 is written into the same location in the
second half cycle. Then, the negative transition of clock f 2 2 transfers contents of Rd1 to
Rt2 in stage 2.
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 23

1
RW RW
se0
1

0
0
0
1
1
se1

se 2

1
RW RW
se0


0
1
se1
se 2

Fig. 9. Modified RP for 2-parallel architecture (a) 5/3 and (a) & (b) 9/7

CK RP1 input RP1output

f2 latches STAGE 2 STAGE 3 STAGE 4 latches
STAGE 1 Rt0 Rt2 Rt1 Rt0 Rt1 R0 Rt0 Rt1 Rt2 Rt0 Rt1
Rt0 Rt1 R0 TLB2
TLB1
11 L0,0 H0,0 ----- -----
13 L0,1 H0,1 L0,0 ---- H0,0 ----- -----
15 L2,0 H2,0 L0,1 ---- H0,1 X0,0 --- ---- ----- -----
RUN 1

H0,1 H0,0
17 L2,1 H2,1 L2,0 ---- H2,0 X0,2 H0,0 X0,0 X0,0 ---- ---- ----- -----
-----
19 L4,0 H4,0 L2,1 ---- H2,1 X2,0 ----- X0,2 X0,2 H0,0 ----- -----
H2,1 H2,0 X0,2 X0,0
24 VLSI

21 L4,1 H4,1 L4,0 ---- H4,0 X2,2 H2,0 X2,0 X2,0 ----- X0,0 X0,1
----- X0,2
23 L6,0 H6,0 L4,1 ---- H4,1 X4,0 ----- X2,2 X2,2 H2,0 X0,2 -----
H4,1 H4,0 X2,2 X2,0
25 L6,1 H6,1 L6,0 ---- H6,0 X4,2 H4,0 X4,0 X4,0 ----- X2,0 X2,1
----- X2,2
27 ----- ----- L6,1 ---- H6,1 X6,0 ----- X4,2 X4,2 H4,0 X2,2 -----
H6,1 H6,0 X4,2 X4,0
29 L0,2 H0,2 ---- ----- ----- ----- X6,2 H6,0 X6,0 X6,0 ----- X4,0 X4,1
- X4,2
31 L2,2 H2,2 L0,2 H0,1 H0,2 ---- ----- X6,2 X6,2 H6,0 X4,2 -----
H0,2 ----- X6,2 X6,0
33 L4,2 H4,2 L2,2 H2,1 H2,2 X0,4 H0,1 ----- ---- ---- X6,0 X6,1
H2,2 ----- X6,2 X6,2
35 L6,2 H6,2 L4,2 H4,1 H4,2 X2,4 H2,1 X0,4 X0,4 H0,1 X6,2 -----
H4,2 ----- X0,4 X0,2
37 ---- ----- L6,2 H6,1 H6,2 X4,4 H4,1 ----- X2,4 H2,1 X0,3 X0,4
H6,2 ----- X2,4 X2,2
39 ---- ----- ----- ----- ------ - X6,4 H6,1 ----- X4,4 H4,1 X2,3 X2,4
RUN 2

----- X4,4 X4,2

41 ---- ----- ----- ----- ------ - ------ ----- ----- X6,4 H6,1 X4,3 X4,4
----- X6,4 X6,2
43 ---- ----- ----- ----- ------ - ----- ----- ------ - ----- ----- --- X6,3 X6,4
----- ----- ---
Table 2. Dataflow of the 5/3 RP1 (Fig. 9a)

In Fig. 9 (a), the control signal, s, of the two multiplexers’ labeled mux is set 1 during run 1
to pass R0 of both stages 2 and 3, whereas in all other runs, it is set 0 to pass coefficients
stored in TLB1 and TLB2.

4.3 Proposed 4-parallel external architecture

To further increase speed of computations twice as that of the 2-parallel architecture, the 2-
parallel architecture is extended to 4-parallel architecture as shown in Fig. 10 (a). This
architecture is valid for 5/3, 9/7, and combined 5/3 and 9/7. It consists of 4 k-stage
pipelined CPs and 4 k-stage pipelined RPs. The waveforms of the 3 clocks f4, f4a, and f4b used
in the architecture are shown in Fig. 10 (b). The frequency of clock f4 is determined from
Eq(3) as
f 4  4k t p (5)
The architecture scans the external memory with frequency f4 and it operates with frequency
f4a and f4b. Every time clock f4a makes a negative transition CP1 loads into its input latches
Rt0 and Rt1 two new coefficients scanned from external memory through the buses labeled
bus0 and bus1, whereas CP3 loads every time clock f4a makes a positive transition. CP2 and
CP4 loads every time clock f4b makes a negative and a positive transition, respectively, as
indicated in Fig. 10 (b). On the other hand, both RP1 and RP2 load simultaneously new data
into their input latches Rt0 and Rt1 each time clock f4a makes a negative transition, whereas
RP3 and RP4 loads each time clock f4b makes a negative transition.
The dataflow for 4-parallel 5/3 external architecture is given in Table 3, where CPs and RPs
are assumed to be 3- and 4-stage pipelined processors, respectively. The dataflow table for
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 25

4-parallel 9/7 external architecture is similar in all runs to the 5/3 dataflow except in the
first run, where RPs of the 9/7 architecture, specifically RP3 and RP4 generate a pattern of
output coefficients different from that of the 5/3. RP3 and RP4 of the 9/7 architecture would
generate every clock cycle, reference to clock f4b, two output coefficients as follows. Suppose,
at cycle number n the first two coefficients X(0,0) and X(1,0) generated by RP3 and RP4,
respectively, are loaded into output latch Rt0 of both processors. Then, in cycle n+1, RP3 and
RP4 generate coefficients X(2,0) and X(3,0) followed by coefficients X(4,0) and X(5,0) in cycle
n+1 and so on. Note that these output coefficients are the coefficients generated by RP1 and
RP2 in Table 3.
The strategy used for scheduling memory columns for CPs of the 5/3 and 9/7 4-parallel
architecture, which resemble the one adopted for 2-parallel architecture, is as follow. In the
first run, both 5/3 and 9/7 4-parallel architecture will be scheduled to execute 4 columns of
memory, two from (A) and the other two from (B), both of Fig. 5. Each CP will be assigned
to execute one column of memory coefficients as illustrated in the first run of the dataflow
shown in Table 3, whereas in all subsequent runs, 2 columns at a time will scheduled for
execution by 4 CPs. One column from Fig. 5 (A) will be assigned to both CP1 and CP3, while
the other from Fig. 5 (B) will be assigned to both CP2 and CP4 as shown in the second run of
Table 3. However, if number of columns in (A) and (B) of Fig. 5 is not equal, then the last
run will consist of only one column of (A). In that case, schedule the last column’s
coefficients in both CP1 and CP3 as shown in the third run of Table 3, since an attempt to
execute the last column using 4 CPs would result in more coefficients been generated than
that can be handled by the 4 RPs.
On the other hand, scheduling rows coefficients for RPs, which take place according to scan
method shown in Fig. 6, can be understood by examining the dataflow shown in Table 3. At
cycle 13, CP1 generates its first two output coefficients labeled L0,0 and L1,0, which
correspond to locations 0,0 and 1,0 in Fig. 6, respectively. In cycle 14, CP2 generates its first
two output coefficients H0,0 and H1,0, which correspond to locations 0,1 and 1,1 in Fig. 6,
respectively. In cycle 15, CP3 generate its first two coefficients L0,1 and L1,1, which
correspond to locations 0,2 and 1,2 in Fig. 6, respectively. In cycle 16, CP4 generates its first
two output coefficients H0,1 and H1,0 which correspond to locations 0,3 and 1,3 in Fig. 6.
Note that L0,0, H0,0, L0,1, and H0,1 represents the first 4 coefficients of row 0 in Fig. 6,
whereas L1,0, H1,0, LL1,1 and H1,1, represent the first 4 coefficients of row1.
In cycle 17 and 18, the first two rows coefficients are scheduled for RPs as shown in Table 3,
while CPs generating coefficients of the next two rows, row2 and row3. Table 3 shows that
the first 4 coefficients of row 0 are scheduled for execution by RP1 and RP3, while the first 4
coefficients of row 1 are scheduled for RP2 and RP4. In addition, note that all coefficients
generated by CP4, which belong to column 3 in Fig. 6, are required in the second run’s
computations, according to the DDGs. Therefore, this would require inclusion of a TLB of
size N/4 in each of the 4 RPs to store these coefficients. The second run, however, requires
these coefficients to be stored in the 4 TLBs in a certain way as follows. Coefficients H0,1 and
H1,1 generated by CP4 in cycle 16 should be stored in the first location of TLB of RP1 and
RP2, respectively. These two coefficients would be passed to their respective TLB through
the input latches of RP1 and RP2 labeled Rt2, as shown in cycle 17 of Table 3, whereas,
coefficients H2,1 and H3,1 generated by CP4 at cycle 20 should be stored in the first location
of TLB of RP3 and RP4, respectively. These two coefficients are passed to their respective
TLB for storage through the input latches of RP3 and RP4 labeled Rt1, as shown in cycle 22
26 VLSI

of Table 3. Similarly, coefficients H4,1 and H5,1 generated by CP4 at cycle 24 should be
stored in the second location of TLB of RP1 and RP2, respectively, and so on. Note that these
TLBs are labeled TLB1 in Fig. 12.

f 4a

f 4b
f 4b f 4a

f4

f 4a  f 4 4

f 4b  f 4 4

Fig. 10. (a) Proposed 2-D IDWT 4-parallel pipelined external architecture for 5/3 and 9/7
and combined (b) Waveforms of the clocks
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 27

CK CP CPs input CPs 1 CPs 2 & RPs 1 & 3 RPs 2 & 4 RPs 1 & RPs 2 &
f4 Latches &3 4 input input 3 4
Rt0 Out Out latches latches Out Out
Rt1 latches latches RP Rt0 RP Rt0 latches latches
Rtl0 Rth0 Rt1 Rt2 Rt1 Rt2 Rt0 Rt0
Rtl1 Rth1 Rt1 Rt1
1 1 LL0,0
LH0,0
2 2 HL0,0
HH0,0
3 3 LL0,1
LH0,1
4 4 HL0,1
HH0,1
5 1 LL1,0
LH1,0
6 2 HL1,0
HH1,0
7 3 LL1,1
LH1,1
8 4 HL1,1
HH1,1
9 1 LL2,0
LH2,0
10 2 HL2,0
HH2,0
11 3 LL2,1
LH2,1
12 4 HL2,1
HH2,1
13 1 LL3,0 L0,0
LH3,0 L1,0
14 2 HL3,0 H0,0
HH3,0 H1,0
15 3 LL3,1 L0,1
1

LH3,1 L1,1
16 4 HL3,1 H0,1
RUN

HH3,1 H1,1
17 1 LL4,0 --- L2,0 1 L0,0 2 L1,0
--- L3,0 H0,0 H0,1 H1,0 H1,1
18 2 HL4,0 -- H2,0 3 L0,1 4 L1,1
---- H3,0 H0,1 H0,0 H1,1 H1,0
19 3 LL4,1 --- L2,1
--- L3,1
20 4 HL4,1 -- H2,1
---- H3,1
21 1 LL0,2 L4,0 1 L2,0 2 L3,0
LH0,2 L5,0 H2,0 ----- H3,0 -----
2

22 2 HL0,2 H4,0 3 L2,1 4 L3,1

RUN

HH0,2 H5,0 H2,1 H2,0 H3,1 H3,0

23 3 LL1,2 L4,1
28 VLSI

LH1,2 L5,1
24 4 HL1,2 H4,1
HH1,2 H5,1
25 1 LL2,2 L6,0 1 L4,0 2 L5,0
LH2,2 L7,0 H4,0 H4,1 H5,0 H5,1
26 2 HL2,2 H6,0 3 L4,1 4 L5,1
HH2,2 H7,0 H4,1 H4,0 H5,1 H5,0
27 3 LL3,2 L6,1
LH3,2 L7,1
28 4 HL3,2 H6,1
HH3,2 H7,1
29 1 LL4,2 --- L8,0 ---- 1 L6,0 2 L7,0
--- - H6,0 ----- H7,0 -----
30 2 HL4,2 -- H8,0 --- 3 L6,1 4 L7,1
---- -- H6,1 H6,0 H7,1 H7,0
31 3 ------- --- L8,1 ----
---- -
32 4 ------- --- H8,1 ---
---- --
33 1 LL0,3 L0,2 1 L8,0 2 ----- --- X0,0 --- X1,0 ---
LH0,3 L1,2 H8,0 H8,1 - ---- - --
34 2 ----- -- H0,2 3 L8,1 4 ----- --- X0,1 X1,1
----- H1,2 H8,1 H8,0 - ---- X0,2 X1,2
35 3 LL1,3 L2,2
LH1,3 L3,2
36 4 ------ --- H2,2
---- H3,2
37 1 LL2,3 L4,2 1 L0,2 2 L1,2 X2,0 --- X3,0 ---
LH2,3 L5,2 H0,2 ----- H1,2 ----- - --
38 2 ------- --- H4,2 3 L2,2 4 L3,2 X2,1 X3,1
---- H5,2 H2,2 ----- H3,2 ----- X2,2 X3,2
39 3 LL3,3 L6,2
LH3,3 L7,2
40 4 ------- --- H6,2
---- H7,2
41 1 LL4,3 --- L8,2 ---- 1 L4,2 2 L5,2 X4,0 --- X5,0 ---
--- - H4,2 ----- H5,2 ----- - --
RUN 3

42 2 ------- --- H8,2 --- 3 L6,2 4 L7,2 X4,1 X5,1

--- -- H6,2 ----- H7,2 ----- X4,2 X5,2
43 3 ------- --- ----- ---
---- --
44 4 ------- --- ------ ---
---- ---
45 1 L0,3 1 L8,2 2 ----- --- X6,0 --- X7,0 ---
L1,3 H8,2 ----- - ---- - --
46 2 ------ -- 3 ----- ---- 4 ----- --- X6,1 X7,1
---- - ----- - ---- X6,2 X7,2
47 3 L2,3
L3,3
48 4 ------ --
----
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 29

49 1 L4,3 1 L0,3 ---- 2 L1,3 --- X8,0 --- ----- ---

L5,3 - ----- -- ----- - --
50 2 ------ -- 3 L2,3 --- 4 L3,3 --- X8,1 ----- ---
---- -- ----- -- ----- X8,2 --
51 3 L6,3
L7,3
52 4 ------ --
----
53 1 L8,3 ---- 1 L4,3 --- 2 L5,3 --- X0,3 X1,3
- -- ----- -- ----- X0,4 X1,4
54 2 ------ -- 3 L6,3 --- 4 L7,3 --- X2,3 X3,3
---- -- ----- -- ----- X2,4 X3,4
55 3 ------ ---
---
56 4 ----- ---
---
Table 3. Dataflow for 4-parallel 5/3 architecture

At cycle 33, RP1 and RP2 yield their first output coefficients X0,0 and X1,0, respectively,
which must be stored in the external memory locations 0,0 and 1,0, respectively. Note that
indexes of each output coefficient indicate external memory location where the coefficient
should be stored.
The second run, which requires scheduling two columns for execution by CPs, starts at cycle
21. In cycle 33, it generates its first two output coefficients L0,2 and L1,2, which belong to
locations 0,4 and 1,4 in Fig. 6, respectively. In cycle 34, CP2 generates coefficients H0,2 and
H1,2 which belong to locations 0,5 and 1,5 in Fig. 6. In cycle 35, CP3 generates coefficients
L2,2 and L3,2, which belong to locations 2,4 and 3,4 in Fig. 6, whereas in cycle 36, CP4
generates coefficients H2,2 and H3,2 that belong to locations 2,5 and 3,5 in Fig. 6. From the
above description it is clear that these 8 coefficients are distributed along 4 rows, 0 to 3 in
Fig. 6 with each row having 2 coefficients. Table 3 shows that in cycle 37, the two coefficients
of row 0, L0,2 and H0,2, and the two coefficients of row 1, L1,2 and H1,2 are scheduled for
RP1 and RP2, respectively, while coefficients L4,2 and L5,2 generated by CP1 during the
cycle are loaded into Rtl0 and Rtl1, respectively. In cycle 38, the two coefficients of row 2,
L2,2 and H2,2, and that of row 3 L3,2 and H3,2 are scheduled for RP3 and RP4, respectively,
while coefficients H4,2 and H5,2 generated by CP2 during the cycle are loaded into Rth0
and Rth1, respectively. In cycle 53, RP1 and RP2 generate the first output coefficients of the
second run.

4.4 Column and row processors for 5/3 and 9/7 4-parallel external architecture
The 5/3 and the 9/7 processors datapath architectures proposed in (Ibrahim & Herman,
2008) were developed assuming the processors scan external memory either row by row or
column by column. However, CPs and RPs of the 4-parallel architecture are required to scan
external memory according to scan methods shown in Figs. 5 and 6, respectively. The 4-
parallel architecture, in addition, introduces the requirement for interactions among the
processors in order to accomplish their task. Therefore, the processors datapath
architectures proposed in (Ibrahim & Herman, 2008) should be modified based on the scan
methods and taking into consideration also the requirement for interactions so that they fit
30 VLSI

into the 4-parallel’s processors. Thus, in the following, the modified CPs will be developed
first followed by RPs.

4.5 Modified CPs for 4-parallel architecture

The 4 CPs of the 4-parallel architecture each is required in the first run to execute one
column at a time. That means the first run requires no modifications of the 5/3 and 9/7
datapath architectures proposed in (Ibrahim & Herman, 2008). However, in all subsequent
runs, each two processors (CP1 and CP3 or CP2 and CP4) are assigned to execute one
column together, which requires interactions between the two processors to accomplish the
required task. Therefore, both CPs 1 and 3, similarly, CPs 2 and 4 should be modified as
shown in Fig. 11 to allow interactions to take place. The two processors communicate or
interact through the paths (buses) labeled P1, P2, P3, and P4. Fig. 11 shows modified 5/3
CPs 1 and 3 which is identical to CPs 2 and 4. Fig. 11 also represents the first 3 stages of both
9/7 CPs 1 and 3 and CPs 2 and 4 and the remaining stages are identical to stages 1 to 3. Note
that since the first 3 stages of 5/3 and 9/3 are similar in structure, the 5/3 processor can be
easily incorporated into 9/7 processor to obtain the combined 5/3 and 9/7 processor for 4-
parallel architecture.
The control signal, s of the 4 multiplexers, labeled mux is set 0 in the first run to allow each
processor to execute one column and 1 in all other runs to allow execution of one column by
two processors.
In the following, the second run’s dataflow of Fig. 11, which starts at cycle 21 in Table 3 will
be described. In cycle 21, two coefficients LL0,2 and LH0,2, which correspond to the first
two coefficients of the third column in Fig. 5 (A), are read from external memory and are
loaded into CP1 first stage latches Rt0 and Rt1, respectively, by negative transition of clock
f4a to initiate the first operation. In cycle 23, the third and the fourth coefficients LL1,2 and
LH1,2 of the third column are scanned from external memory and are loaded along with Rt1
in stage1 of CP1 into CP3 first stage latches Rt0, Rt1 and Rd3, respectively, by the positive
transition of clock f4a. Note that content of Rt1 in stage 1 of CP1 takes the path labeled P1 to
Rd3.
In cycle 25, coefficients LL2,2 and LH2,2 are scanned from external memory and are loaded
along with Rt1 in stage 1 of CP3 into CP1 first stage latches Rt0, Rt1, and Rd1, respectively,
by the negative transition of clock f4a, while coefficient X(0) calculated in stage 1 of CP1 and
content of Rt1 are loaded into Rt0 and Rt1 of stage 2, respectively.
In cycle 27, coefficients LL3,2 and LH3,2 are scanned from external memory and are loaded,
along with Rt1 in stage 1 of CP1, into CP3 first stage latches Rt0, Rt1, and Rd3, respectively,
by the positive transition of clock f4a, while coefficients X(2) calculated in stage 1 and content
of Rt1 are transferred to Rt0 and Rt1 of the next stage, respectively.
In cycle 29, only one coefficient LL4,2 is scanned from external memory, which implies the
third column of Fig. 5 (A) contains odd number of coefficients, and is loaded along with Rt1
in stage1 of CP3 into CP1 first stage latches Rt0 and Rd1, respectively, by the negative
transition of clock f4a, while coefficient X(4) calculated in stage 1 and content of Rt1 are
transferred to Rt0 and Rt1 of stage 2 and that of Rt0 and Rt1 including Rt0 in stage 2 of CP3
to Rt0, Rt1, and Rt2 in stage 3 of CP1, respectively, to compute the first high coefficient
labeled X(1) in 5/3 DDGs. Note that CP1 latches Rt0 and Rt2 of stage 3 now contain
coefficients X(0) and X(2), while Rt1 contains coefficient LH0,2 (or Y(1) as labeled in the 5/3
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 31

DDGs). These 3 coefficients are required according to the DDGs to compute the high
coefficient, X(1).

f 4a f 4a f 4a

se0
0
f 4a
s 1

s se2
se1 f 4a

f 4a f 4a f 4a

f 4a f 4a f 4a

f 4a
se0 f 4a
1
s 0

s se2
se1

f 4a f 4a f 4a

Fig. 11. Modified 5/3 CPs 1 & 3 for 4-parallel architecture

32 VLSI

In cycle 31, the positive transition of clock f4a transfers Rt0 and Rt1 in stage 2 of CP3 and Rt0
in stage 2 of CP1 to stage 3 latches Rt0, Rt1, and Rt2 of CP3, respectively, to compute the
second high coefficient, X(3). While coefficient X(6) calculated in stage 1 of CP3 and
coefficient in Rt1 are loaded into Rt0 and Rt1 of stage 2. As indicated in cycle 31 of Table 3,
no coefficients are loaded into CP3 first stage latches.
In cycle 33, the negative transition of clock f4a loads the first high coefficient, X(1) calculated
in stage 3 of CP1 and Rt0, which contains X(0), into CP1 output latches Rt0 and Rt1,
respectively. Coefficients X(0) and X(1) are labeled L0,2 and L1,2 in Table 3. The same
negative transition of clock f4a transfers contents of Rt0 and Rt1 in stage 2 of CP1 and Rt0 in
stage 2 of CP3 to Rt0, Rt1, and Rt2 in stage 3 of CP1, respectively, to compute the third
coefficient, X(3) and so on.

4.6 Modified RPs for 4-parallell architecture

In section 5.2, it has been pointed out the reasons for including TLBs in the two RPs of the 2-
parallel architecture. For the same reasons, it is also necessary to include TLBs in the 4 RPs
of the 4-parallel architecture, as shown in Figs. 12 (a) and (a,b) for 5/3 and 9/7, respectively.
The processor datapath for both RP1 and RP3, which is also identical to the processor
datapath of both RP2 and RP4, are drawn together in Figs.12 (a) and (a,b) for 5/3 and 9/7,
respectively, because in the first run, both processors are required to execute together the 4
coefficients of each row. Which implies interactions between the two processors during the
computations and that take place through the paths (buses) labeled P1, P2, P3, and P4.
However, in all subsequent runs, according to the scan method shown in Fig. 6, each RP will
be scheduled to execute each clock cycle two coefficients of a row as shown in cycles 37 and
38 of Table 3. The advantage of this organization is that the total size of the TLBs does not
increase from that of the single pipelined architecture, when it is extended to 2- and 4-
parallel architecture.
In the first run, all TLBs in Fig. 12 will be written only, whereas in all other runs, the same
location of a TLB will be read in the first half cycle and written in the second half cycle with
respect to clock f4a or f4b.
The control signal, s of the six multiplexers, labeled mux in Fig. 12, is set 0 in the first run to
allow in the RP1, coefficient coming through path 0 of each multiplexer to be stored in its
respective TLB, whereas in the RP3, it allows contents of Rt2 and Rd1 in stages 1 and 3,
respectively, to be passed to the next stage. In all subsequent runs, s is set 1 to allow in the
RP1, coefficient coming through path 1 of each multiplexer to be stored in the TLB, whereas
in the RP3, it passes coefficients read from TLB1 and TLB2 to next stage.
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 33

f 4a f 4a

0
f 4a f 4a
1 1
RW RW
0 se0
N


0 N
4
1 4

0
1
se1
se2
f 4a


f 4a sco

f 4b

f 4b RW 0 f 4b
f 4b
1 0
N
4 se0 f 4b 1


0
1
1
0 N
se1 4
se2
RW


f 4b f 4b sco

Fig. 12. (a) Modified 5/3 RPs 1 and 3 for 4-parallel External Architecture

In the following, the dataflow of the processor datapath architecture shown in Fig. 12 (a)
will described in details starting from cycle 17 in Table 3. The detailed descriptions will
enable us to fully understand the behavior of the processor. In cycle 17, the negative
transition of clock f4a (consider it the first cycle of f4a) loads coefficients L0,0 and H0,0, which
represent the first two coefficients of row 0 in Fig. 6, and H0,1 into RP1 first stage latches
Rt0, Rt1, and Rt2, respectively. During the positive (high) pulse of clock f4a, coefficient in Rt2
is stored in the first location of TLB1.
In cycle 18, Table 3 shows that the negative transition of clock f4b (consider it the first cycle of
f4b) loads coefficients L0,1, H0,1, and H0,0 of row 0 into RP3 first stage latches Rt0, Rt1, and Rt2,
34 VLSI

respectively. In cycle 21, the negative transition of the second cycle of clock f4a transfers
contents of RP1 latches Rt0 and Rt1 of the first stage to stage 2 latches Rt0 and Rt1,
respectively, to compute the first low coefficient, X0,0, while loading two new coefficients
L2,0 and H2,0, which are the first two coefficients of row 2 in Fig. 6, into RP1 first stage
latches Rt0 and Rt1, respectively.
f 4a f 4a

0
f 4a f 4a
1 1
RW RW
0 se0
N


0 N
f 4b 4
1 4

0
1
se1
se2
f 4a


f 4a

f 4b

f 4b 0 f 4b
f 4b
RW 1
se0
N

0
4 N
1 4
se1
RW se2

 f 4b
f 4b

Fig. 12. (b) Modified 9/7 RPs 1 and 3 for 4-parallel External Architecture

During the second cycle of clock f4a no data are stored in TLB1 of RP1.
In cycle 22, the negative transition of the second cycle of clock f4b transfers contents of RP3
latches Rt0, Rt1 and Rt2 of the first stage to stage 2 latches Rt0, Rt1 and Rt2 to compute the
second low coefficient of row 0, X0,2, while loading two new coefficients L2,1, H2,1, and
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 35

H2,0 into RP3 first stage latches Rt0, Rt1 and Rt2, respectively. During the second cycle of
clock f4b, the positive pulse stores content of Rt1 of the first stage in the first location of TLB1
of RP3.
In cycle 25, the negative transition of the third cycle of clock f4a loads coefficient X0,0
computed in stage 2 of RP1, into Rt0 of stage 3 and transfers contents of Rt1 and Rt0 of stage
1 into Rt1 and Rt0 of stage 2 in order to compute the first low coefficient, X2,0 of row 2
labeled X(0) in the 5/3 DDGs, while loading new coefficients L4,0, H4,0, and H4,1 of row 4
in Fig. 6, into RP1 first stage latches Rt0, Rt1 and Rt2, respectively. During the third cycle of
clock f4a, the positive pulse stores Rt2 of the first stage in the second location of TLB1 of RP1.
In cycle 26, the negative transition of the third cycle of clock f4b loads coefficient X0,2
calculated in stage 2 of RP3 into both Rt0 in stage 3 of RP3 and Rd3 in stage 3 of RP1, while
content of Rt2 in stage 2 of RP3 is transferred to Rt1 of stage 3 and that of Rt0 in stage 3 of
RP1 to Rd1 in stage 3 of RP3. The same negative transition of clock f4b also transfers Rt0, Rt1,
and Rt2 of stage 1 into Rt0, Rt1, and Rt2 of stage 2, respectively, to compute the second low
coefficient, X2,2 of row 2, while loading new coefficients L4,1, H4,1, and H4,0 of row 4 into
RP3 first stage latches Rt0, Rt1, and Rt2, respectively. During the third cycle of clock f4b,
coefficient in Rt1 is not stored in TLB1 of RP3. It is important to note that Rd3 in stage 3 of
RP1, which holds coefficient X0,2, will be stored in the first location of TLB2 by the positive
pulse of the third cycle of clock f4a.
In cycle 29, the fourth cycle’s negative transition of clock f4a transfers Rt0 in stage 3 of RP1,
which contains X0,0, to Rt0 of stage 4, while X2,0 calculated in stage 2 is loaded into Rt0 of
stage 3. The same negative transition of clock f4a transfer Rt1 and Rt0 of stage 1 to Rt1 and
Rt0 of stage 2, respectively, to compute the first low coefficient of row 4, X4,0, and loads new
coefficients L6,0 and H6,0 into Rt0 and Rt1 of stage 1, respectively. During the fourth cycle
of clock f4a, coefficient in Rt2 of stage 1 is not stored in TLB1 of RP1.
In cycle 30,the negative transition of the fourth cycle of clock f4b transfers contents of Rt0,
Rt1, and Rd1 in stage 3 of RP3 to Rt0, Rt1, and Rt2 of the next stage, respectively, to compute
the first high coefficient, X0,1 of row 0. While coefficient X2,2 calculated in stage 2 of RP3 is
transferred to both Rt0 in stage3 of RP3 and Rd3 in stage 3 of RP1 and coefficient X2,0 in Rt0,
in stage 3 of RP1, is loaded into Rd1 in stage 3 of RP3, whereas Rt2 of stage 2 is transferred
to Rt1 in stage 3 of RP3. The same negative transition of clock f4b transfers contents of Rt0,
Rt1, and Rt2 of stage 1 to Rt0, Rt1, and Rt2 of stage 2 to compute the second low coefficient,
X4,0 of row 4, while loading new coefficients L6,1, H6,1, and H6,0 of row 6 into RP3 first
stage latches Rt0, Rt1, and Rt2, respectively. During the fourth cycle’s positive pulse of clock
f4b, content of Rt1 in stage 1 of RP3 will be stored in the second location of TLB1 and content
of Rt0, X2,2, in stage 3 of RP3 will be stored in the first location of TLB2, while content of
Rd3, X2,2, in stage 3 of RP1 will not be stored in TLB2 of RP1.
In cycle 33, the negative transition of the fifth cycle of clock f4a transfers content of Rt0, X0,0,
in stage 4 of RP1 to RP1 output latch Rt0, as first output coefficient and Rt0 of stage 3
holding coefficient X2,0 to Rt0 of stage 4, while coefficient X4,0 calculated in stage 2 is
loaded into Rt0 of stage 3. The same negative transition of clock f4a transfers contents of Rt0
and Rt1 of stage 1 to Rt0 and Rt1 of stage 2 to compute the first low coefficient, X6,0 of row
6, while loading new coefficients L8,0, H8,0, and H8,1 of row 8 into RP1 first stage latches
Rt0, Rt1, and Rt2, respectively. During the fifth cycle of clock f4a, the positive pulse of the
clock stores content of Rt2 in stage 1 of RP1 into the third location of TLB1.
36 VLSI

In cycle 34, the negative transition of the fifth cycle of clock f4b transfers content of Rt0, X0,2,
and the high coefficient, X0,1 computed in stage 4 of RP3 to RP3 output latches Rt0 and Rt1,
respectively, while loading contents of Rt0, Rt1, and Rd1 of stage 3 into Rt0, Rt1, and Rt2 of
stage 4 to compute the first high coefficient of row 2, X2,1. Furthermore, the same negative
transition of clock f4b also transfers coefficient X4,2 calculated in stage 2 of RP3 to both Rt0 in
stage 3 of RP3 and Rd3 in stage 3 of RP1. It also transfers coefficient X4,0 in Rt0, in stage 3 of
RP1, and content of Rt2 in stage 2 of RP3 to stage 3 of RP3 latches Rd1 and Rt1, respectively,
and contents of Rt0, Rt1, and Rt2 of stage 1 to Rt0, Rt1, and Rt2 in stage 3 of RP3, while
loading new coefficients L8,1, H8,1, and H8,0 into Rt0, Rt1, and Rt2 of stage 1. Content of
Rd3, X4,2 in stage 3 of RP1 will be stored in the second location of TLB2 by the positive
pulse of the fifth cycle of clock f4a.
In cycle 37, the second run of the RPs begins when the negative transition of the sixth cycle
of clock f4a loads two new coefficients L0,2 and H0,2, which are the fifth and sixth
coefficients of row 0, into first stage latches Rt0 and Rt1 of RP1, respectively. During the first
half (low pulse) of cycle 6, the first location of TLB1 will be read and loaded into Rd1 by the
positive transition of clock f4a, whereas during the second half cycle, content of Rt1 will be
written in the first location of TLB1.
In cycle 38, the negative transition of the sixth cycle of clock f4b loads two new coefficients
L2,2 and H2,2, the fifth and sixth coefficients of row 2, into RP3 first stage latches Rt0 and
Rt1, respectively. The first half cycle of clock f4b reads the first location of TLB1 and loads it
into Rd3 by the positive transition of the clock, whereas the second half cycle writes content
of Rt1 in the same location of TLB1.
In cycle 41, the negative transition of the seventh cycle of clock f4a transfers contents of Rt0,
Rt1, and Rd1 of stage 1 to stage 2 of RP1 latches Rt0, Rt1, and Rt2, respectively, to compute
the third low coefficient, X0,4 of row 0, while loading two new coefficients L4,2 and H4,2 of
row 4 into RP1 first stage latches Rt0 and Rt1 respectively.
Note that during run 2 all RPs execute independently with no interactions among them. In
addition, in the first run, if the first coefficient generated by stage 2 of RP3 is stored in TLB2
of RP1, then the second coefficient should be stored in TLB2 of RP3 and so on. Similarly,
TLB1, TLB3, and TLB4 of both RP1 and RP3 are handled. Furthermore, during the whole
period of run 1, the control signals of the three extension multiplexers labeled muxe0, muxe1,
and muxe2 in RP1 should be set 0, according to Table 4 (Ibrahim & Herman, 2008), whereas
those in RP3 should be set normal as shown in the second line of Table 4, since RP3 will
execute normal computations during the period. However, in the second run and in all
subsequent runs except the last run, the extension multiplexers control signals in all RPs are
set normal. Moreover, the multiplexers labeled muxco in stage 4 is only needed in the
combined 5/3 and 9/7 architecture, otherwise, it can be eliminated and Rt2 output can be
connected directly to Rt0 input of the next stage in case of 9/7, whereas in 5/3, Rt0 is
connected directly to output latch Rt0. In the combined architecture, signal sco of muxco is
set 0 if the architecture is to perform 5/3; otherwise, it is set 1 if the architecture is to
perform 9/7.
It is very important to note that when the RP executes its last set of input coefficients,
according to 9/7 DDGs for odd and even signals shown in Fig. 4 it will not yield all required
output coefficients as expected by the last run. For example, in the DDGs for odd length
signals shown in Fig. 4 (a), when the last input coefficient labeled Y8 is applied to RP it will
yield output coefficients X5 and X6. To get the last remaining two coefficients X7 and X8, the
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 37

RP must execute another run, which will be the last run in order to compute the remaining
two output coefficients. Similarly, when the last two input coefficients labeled Y6 and Y7 in
the DDG for even length signals shown in Fig. 4 (b) are applied to 9/7 RP it will yield
output coefficients X3 and X4. To obtain the remaining output coefficients X5, X6, and X7,
two more runs should be executed by RP according to the DDG. The first run will yield X5
and X6, whereas the last run will yield X7. The details of the computations that take place
during each of these runs can be determined by examining the specific area of the DDGs.

se0 se1 se2

First 0 0 0
Normal 0 1 0
Last 1 1 0
a) Odd length signals

se0 se1 se2

First 0 0 0
Normal 0 1 0
Last 0 1 0
b) Even length signals
Table 4. Extension’s control signals

5. Performance Evaluation
In order to evaluate performance of the two proposed parallel pipelined architectures in
terms of speedup, throughput, and power as compared with single pipelined architecture
proposed in (Ibrahim & Herman, 2008) consider the following. Assume subbands HH, HL,
LH, and LL of each level are equal in size. The dataflow table for single pipelined
architecture (Ibrahim & Herman, 2008) shows that 1  20 clock cycles are needed to
yield the first output coefficient. Then, the total number of output coefficients in the first run
of the Jth level reconstruction can be estimated as
N 2 J 1 (6)
and the total number of cycles in the first run is given by
2 N 2 J 1 (7)
The total time, T1, required to yield n pairs of output coefficients for the Jth level
reconstruction on the single pipelined architecture can be estimated as
 
T 1  1  2 N 2 J 1  2 n  1 2 N 2 J 1  1 
   2n t
(8)
1  N 2 J 1 p 2k
On the other hand, the dataflow for 2-parallel pipelined architecture shows that  2  21
clock cycles are required to yield the first 2 pairs of output coefficients. The total number of
paired output coefficients in the first run of the Jth level reconstruction on the 2-parallel
architecture can be estimated as
38 VLSI

3 2 N 2 J 1 (9)
and the total number of 2-paired output coefficients is given by
3 4 N 2 J 1 (10)
While, the total number of cycles in the first run is
2 N 2 J 1 (11)
Note that the total number of paired output coefficients of the first run in each level of
reconstruction starting from the first level can be written as
3 2 N , 3 2 N 2, 3 2 N 4,............., 3 2 N 2 J 1
where the last term is Eq (9).
The total time, T2, required to yield n pairs of output coefficients for the Jth level
reconstruction of an NxM image on the 2-parallel architecture can be estimated as


T 2   2  2 N 2 J 1  2(n 2  3 4 N 2 J 1 )  2  (12)

T 2  2  1 2 N 2 J 1

 n t p 2k (13)

The term 2 n 2  3 4 N 2 J 1
 in (12) represents the total number of cycles of run 2 and all
subsequent runs.
The speedup factor, S2, is then given by
T1 
1  N 2 J 1  2n t p 2k 
S2 
T2


 2  1 2 N 2 J 1  n t p 2k 
For large n, the above equation reduces to
2(1 2 N 2 J 1  n)
S2  2 (14)
(1 2 N 2 J 1  n)
Eq(14) implies that the 2-parallel architecture is 2 times faster than the single pipelined
architecture.
Similarly, the dataflow for the 4-parallel pipelined architecture shows that  4  33 clock
cycles are needed to yield the first two output coefficients. From the dataflow table of the 4-
paralell architecture it can be estimated that both RP1 and RP2, in the first run of the Jth level
reconstruction, yield ( N 2 J 1 ) 2 pairs of output coefficients, whereas both RP3 and RP4
J 1
yield N 2 pairs of output coefficients, a total of 3 2 N 2 J 1 pairs of output coefficients
in the first run. The total number of cycles in run 1 is then given by
4( N 2 J 1 ) 2 (15)
Thus, the total time, T4, required to yield n pairs of output coefficients for the Jth level
reconstruction of an NxM image on the 4-parallel architecture can be estimated as
 
T 4   4  2 N 2 J 1  2 n  3 2 N 2 J 1 2  4  

T 4   4  2N 2 J 1
 n  3 2 N 2 t J 1
p 4k (16)

T 4   4  1 2 N 2 J 1  n t 4k p (17)
J 1
The term (n  3 2 N 2 ) in (16) represents the total cycles of run 2 and all subsequent runs.
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 39

The speedup factor, S4, is then given by

T1 
1  N 2 J 1  2n t p 2k 
S4 
T4


 4  1 2 N 2 J 1  n t p 4k 
For large n it reduces to
4(1 2 N 2 J 1  n)
S4  4 (18)
(1 2 N 2 J 1  n)

Thus, the 4-parallel architecture is 4 times faster than the single pipelined architecture.
The throughput, H, which can be defined as number of output coefficients generated per
unit time, can be written for each architecture as
H (sin gle)  n ( 1  N 2 J 1  2n) t p 2k
The maximum throughput, Hmax, occurs when n is very large (n→∞), thus,
H max (sin gle)  H (sin gle) n
(19)
 2nkf p 2n  kf p

H (2  parallel )  n (  2  1 2 N 2 J 1  n) t p 2k
H max (2  parallel )  H (2  parallel ) n
(20)
 2knf p n  2kf p

H (4  parallel )  n (  4  1 2 N 2 J 1  n) t p 4k
H max (4  parallel )  H (4  parallel ) n
(21)
 4knf p n  4kf p
Thus, the throughputs of the 2-parallel and the 4-parallel pipelined architectures have
increased by factors of 2 and 4, respectively, as compared with the single pipelined
architecture.
On the other hand, the power consumption of l-parallel pipelined architecture as compared
with the single pipelined architecture can be obtained as follows. Let P1 and Pl denote the
power consumption of the single and l-parallel architectures without the external memory,
and Pm1 and Pml denote the power consumption of the external memory for the single and l-
parallel architectures, respectively. The power consumption of VLSI architecture can be
estimated as
P  C total  V02  f
where Ctotal denotes the total capacitance of the architecture, V0 is the supply voltage, and f
is the clock frequency. Then,
P1  Ctotal  V02  f1 2 , Pl  l  Ctotal  V02  f l l and
Pl l  Ctotal  V02  fl l 2 f l
 
P1 Ctotal  V02  f1 2 f1
(21)
l k 
 2  2k t  l
p
 tp 
 
40 VLSI

While, Pm1 and Pml can be estimated as

m m Pml 2  fl 2  l  k t p
Pm1  Ctotal  V02  f1 , Pml  2  Ctotal  V02  fl , and   l (22)
Pm1 f1 2k t p
m
C total , is the total capacitance of the external memory.
In summary, the above equations indicates that as degree of parallelism increases the
speedup and the power consumption of the proposed architectures, without external
memory, and the power consumption of the external memory increase by a factor of l, as
compared with single pipelined architecture.

6. Comparisons
Table 5 provides comparison results of the proposed architectures with most recent
architectures in the literature. The architecture proposed in (Lan & Zheng, 2005) achieves a
critical path of one multiplier delay using very large number of pipelined registers, 52
registers. In addition, it requires a total line buffer of size 6N, which is a very expensive
memory component, while the proposed architectures require only 4N. In (Rahul & Preeti,
2007), a critical path delay of Tm + Ta is achieved through optimal dataflow graph, but
requires a total line buffer of size 10N.
In (wang et at., 2007), by rewriting the lifting-based DWT of the 9/7, the critical path delay
of the pipelined architectures have been reduced to one multiplier delay but it requires a
total line buffer of size 5.5N. In addition, it requires real floating-point multipliers with long
delay that can not be implemented by using arithmetic shift method (Qing & Sheng, 2005).
(Qing & Sheng, 2005) has illustrated that the multipliers used for scale factor k and
coefficients  ,  ,  , and  of the 9/7 filter can be implemented in hardware using
only two adders. Moreover, the fold architecture which uses one module to perform both
the predictor and update steps in fact increases the hardware complexity, e.g., use of several
multiplexers, and the control complexity. In addition, use of one module to perform both
predictor and update steps implies both steps have to be sequenced and that would slow
down the computation process.
In the 2-parallel architecture proposed in (Bao & Yong, 2007), writing results generated by
CPs into MM (main memory) and then switching them out to external memory for next
level decomposition is really a drawback, since external memory in real-time applications,
e.g., in digital camera, is actually consist of charge-coupled devices which can only be
scanned. In addition, it requires a total line buffer of size 5N for 5/3 and 7N for 9/7 while
the proposed architectures require 2N and 4N for 5/3 and 9/7 respectively. The architecture
requires also used of several FIFO buffers in the datapath, which are complex and very
expensive memory components, while the proposed architectures require no such memory
components.
The 2-parallel architecture proposed in (Cheng et al., 2006) requires a total line buffer of size
5.5N and use of two-port RAM to implement FIFOs, whereas, the proposed architectures
require only use of single port RAM.
High Performance Parallel Pipelined Lifting-based VLSI Architectures
for Two-Dimensional Inverse Discrete Wavelet Transform 41

Architecture Multi Adders Line Buff. Computing time Critical Path

Lan 12 12 6N 2(1-4-j )NM Tm
Rahul 9 16 10N 2(1-4-j )NM Tm + Ta
Wang 6 8 5.5N 2(1-4-j )NM Tm
Ibrahim 10 16 4N 2(1-4-j )NM Tm +2Ta
Cheng (2-parallel) 18 32 5.5N (1-4-j )NM Tm +2Ta
Bao (2-parallel) 24 32 7N (1-4-j )NM Tm +2Ta
Proposed (2-parallel) 18 32 4N (1-4-j )NM Tm +2Ta
Proposed (4-parallel) 36 64 4N 1/2 (1-4-j)NM Tm +2Ta
Tm: multiplier delay Ta: adder delay
Table 5. Comparison results of several 2-D 9/7 architectures

7. Conclusions
In this chapter, two high performance parallel VLSI architectures for 2-D IDWT are
proposed that meet high-speed, low-power, and low memory requirements for real-time
applications. The two parallel architectures achieve speedup factors of 2 and 4 as compared
with single pipelined architecture.
• The advantages of the proposed parallel architectures :

i. Only require a total temporary line buffer (TLB) of size 2N and 4N in 5/3 and 9/7,
respectively.
ii. The scan method adopted not only reduces the internal memory between CPs and
RPs to a few registers, but also reduces the internal memory or TLB size in the CPs
to minimum and allows RPs to work in parallel with CPs earlier during the
computation.
iii. The proposed architectures are simple to control and their control algorithms can
be immediately developed.

8. References
Bao-Feng, L. & Yong, D. (2007). “FIDP A novel architecture for lifting-based 2D DWT in
JPEG2000,” MMM (2), lecture note in computer science, vol. 4352, PP. 373-382,
Springer, 2007.
Cheng-Yi, X.; Jin-Wen, T. & Jian, L. (2006). “Efficient high-speed/low-power line-based
architecture for two-dimensional discrete wavelet transforms using lifting scheme,”
IEEE Trans. on Circuits & sys. For Video Tech. Vol.16, No. 2, February 2006, PP 309-
316.
Dillin, G.; Georis B.; Legant J-D. & Cantineau, O. (2003). “Combined Line-based Architecture
for the 5-3 and 9-7 Wavelet Transform of JPEG2000,” IEEE Trans. on circuits and
systems for video tech., Vol. 13, No. 9, Sep. 2003, PP. 944-950.
Ibrahim saeed koko & Herman Agustiawan, (2008). “Pipelined lifting-based VLSI
architecture for two-dimensional inverse discrete wavelet transform,” proceedings
of the IEEE International Conference on Computer and Electrical Engineering,
ICCEE 2008, Phuket Island, Thailand.
42 VLSI

Lan, X. & Zheng N. (2005). ”Low-Power and High-Speed VLSI Architecturefor Lifting-
Based Forward and Inverse Wavelet Transform,” IEEE tran. on consumer
electronics, Vol. 51, No. 2, May 2005, PP. 379 –385.
Qing-ming Yi & Sheng-Li Xie, (2005). ”Arithmetic shift method suitable for VLSI
implementation to CDF 9/7 discrete wavelet transform based on lifting scheme,”
Proceedings of the Fourth Int. Conf. on Machine Learning and `Cybernetics,
Guangzhou, August 2005, PP. 5241-5244.
Rahul. J. & Preeti R. (2007). ”An efficient pipelined VLSI architecture for Lifting-based 2D-
discrete wavelet transform,” ISCAS, 2007 IEEE, PP. 1377-1380.
Wang, C.; Wu, Z.; Cao, P. & Li, J. (2007). “An efficient VLSI Architecture for lifting-based
discrete wavelet transform,” Mulltimedia and Epo, 2007 IEEE International
conference, PP. 1575-1578.
Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 43

X3

Contour-Based Binary Motion Estimation

Algorithm and VLSI Design for MPEG-4
Shape Coding
Tsung-Han Tsai, Chia-Pin Chen, and Yu-Nan Pan
Department of Electronic Engineering
National Central University, Chung-Li, Taiwan, R.O.C

1. Introduction
MPEG-4 is a new international standard for multimedia communication [1]. It provides a set
of tools for object-based coding of natural and synthetic videos/audios. MPEG-4 also enables
content-based functionalities by introducing the concept of video object plane (VOP), and
such a content-based representation is a key to enable interactivity with objects for a variety
of multimedia applications. The VOP is composed of texture components (YUV) and an
alpha component [2]-0. The texture component contains the colorific information of video
object, and the alpha component contains the information to identify the pixels. The pixels
which are inside an object are opaque and the pixels which are outside the object are
transparent. MPEG-4 supports a content-based representation by allowing the coding of the
alpha component along with the object texture and motion information. Therefore, MPEG-4
shape coding becomes the key technology for supporting the content-based video coding.
MPEG-4 shape coding mainly comprises the following coding algorithms: binary-shaped
motion estimation/motion compensation (BME/BMC), context-based arithmetic coding
(CAE), size conversion, mode decision, and so on. As full search (FS) algorithm is adopted
for MPEG-4 shape coding, most of the computational complexity is due to binary motion
estimation (BME). From the profiling on shape coding in Fig. 1, it can be seen that BME
contributes to 90% of total computational complexity of MPEG-4 shape encoder. It is well
known that an effective and popular technique to reduce the temporal redundancy of BME,
called block-matching motion estimation, has been widely adopted in various video coding
standard, such as MPEG-2 0, H.263 [5] and MPEG-4 shape coding [1]. In block-matching
motion estimation, the most accurate strategy is the full search algorithm which exhaustively
evaluates all possible candidate motion vectors over a predetermined neighborhood search
window to find the global minimum block distortion position.
44 VLSI

Fig. 1. Computational complexity of MPEG-4 shape encoder.

Fast BME algorithms for MPEG-4 shape coding were presented in several previous papers
[6]-[8]. Among these techniques, our previous work, contour-based binary motion
estimation (CBBME), largely reduced the computational complexity of shape coding [9]. It is
applied with the properties of boundary search for block-matching motion estimation, and
the diamond search pattern for further improvement. This algorithm can largely reduce the
number of search points to 0.6% compared with that of full search method, which is
described in MPEG-4 verification model (VM) [2].
In contrast with algorithm-level developments, architecture-level designs for shape coding
are relatively less. Generally, a completed shape coding method should include different
types of algorithms. In CAE part, it needs some bit-level operation. However, in binary
motion estimation part, a high speed search method is needed. With these algorithm
combinations on the shape coding, implementation should not be as straightforward as
expected and it offers some challenges especially on architecture design. Since MPEG-4
shape coding has features of high-computing and high-data-traffic properties, it is suitable
with the consideration of efficient VLSI architecture design. Most literatures have also been
presented to focus on the main computation-expensive part, BME, to improve its
performance [10]. Additionally, CAE is also an important part for architecture design and
discussed in [11]-[12]. They utilized the multi-symbol technique to accelerate the arithmetic
coding performance. As regards the complete MPEG-4 shape coding, some of these designs
utilized array processor to perform the shaping coding algorithm [13]-[15], while others
used pipelined architecture [16]. They can reach the relative high performance at the
expense of these high cost and high complexity architectures. All of them intuitively apply
the full search algorithm for easy realization on architecture design. However, the
algorithm-level achievement on the large reduction of computation complexity is attractive
and not negligible. This demonstrates that, without the supporting on an efficient algorithm,
the straightforward implementation based on full search algorithm is hard to reach a
cost-effective design.
In this paper, we proposed a fast BME algorithm, diamond boundary search (DBS), for
MPEG-4 shape coding to reduce the number of search points. By using the properties of
Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 45

block-matching motion estimation in shape coding and diamond search pattern, we can skip
a large number of search points in BME. Simulation results show that the proposed
algorithm can marvelously reduce the number of search points to 0.6% compared with that
of full search method, which is described in MPEG-4 verification model (VM)[2]. Compared
with other fast BME in [6]-[7], the proposed BME algorithm uses less search points
especially in high motion video sequences, such as ‘Bream’ and ‘Forman’. We also present
an efficient architecture design for MPEG-4 shape coding. This architecture is elaborated
based on our fast shape coding algorithm with the binary motion estimation. Since this
block-matching motion estimation can achieve the high performance based on the
information of boundary mask, the dedicated architecture needs some optimization and
consideration to reduce the memory access and processing cycles. Experimental results also
demonstrate the equal performance on full-search based approach. This paper contributes a
comprehensive exploration of the cost-effective architecture design of shaping coding, and
is organized as follows.
In Section 2 the binary motion estimation in shape coding is described. We describe the
highlights of the proposed fast BME algorithm for MPEG-4 shape coding in Section 3. The
design exploration on CBBME is described in Section 4. In Section 5, the architecture design
based on this BME algorithm is proposed. In Section 6, we present the implementation
results and give some comparisons. Finally we summarize the conclusions in Section 7.

2. BME for MPEG-4 Shape Coding

The MPEG-4 VM [2] describes the coding method for binary shape information. It uses
block-matching motion estimation to find the minimum block distortion position and sets
the position to be motion vector for shape (MVS). The procedure of BME consists of two
steps: first to determine motion vector predictor for shape (MVPS) and then to compute
MVS accordingly.
MVPS is taken from a list of candidate motion vectors. As indicated in Fig. 2, the list of
candidate motion vectors includes the shape motion vectors (MVS) from the three binary
alpha blocks (BABs) which are adjacent to the current BAB and the texture motion vectors
(MV) associated with the three adjacent texture blocks. By scanning the locations of MVS1,
MVS2, MVS3, MV1, MV2 and MV3 in this order, MVPS is determined by taking the first
encountered MV which is valid. Note that if the procedure fails to find a defined motion
vector, the MVPS is set to (0, 0).
Based on MVPS determined above, the motion compensated (MC) error is computed by
comparing the BAB indicated by the MVPS and current BAB. If the computed MC error is
less or equal to 16xAlphaTH for any 4x4 sub-blocks, the MVPS is directly employed as MVS
and the procedure terminates. Otherwise, MV is searched around the MVPS while
computing sum of absolute difference (SAD) by comparing the BAB indicated by the MV
and current BAB. The search range is ±16 pixels around MVPS along both horizontal and
vertical directions. The MV that minimizes the SAD is taken as MVS and this is further
interpreted as MV Difference for shape (MVDS), i.e. MVDS=MVS-MVPS.
46 VLSI

Fig. 2. Candidates for MVPS in shape VOP and texture VOP.

If more than one MVS minimize SAD by an identical value, the MVDS that minimizes the
code length of MVDS is selected. If more than one MVS minimize SAD by an identical value
with an identical code length of MVDS, MVDS with smaller vertical element is selected. If
the vertical elements are also the same, MVDS with smaller horizontal element is selected.
After binary motion estimation, motion compensated block is constructed from the 16x16
BAB with a border of width 1 around the 16x16 BAB (bordered MC BAB). Then,
context-based arithmetic encoding (CAE) is adopted for shape coding.

3. Proposed Contour-Based Binary Motion Estimation (CBBME) Method

The basic concept of the proposed BME method is that the contour of video objects in
current BAB should overlap that in the motion compensated BAB, which is determined by
binary motion estimation [9]. Therefore, those search positions, which contour lays apart
from the contour of video objects in current BAB, can be skipped and the reduced number
will be enormous. Moreover, based on the property that most real-world sequences have a
central biased motion vector distribution [23], we use weighted SAD and diamond search
pattern for furthermost improvement.

3.1 Definition of Boundary Pixel

In order to decide whether the pixel is on the contour of VOP, the boundary pixel is
determined by the following procedure:
 If current pixel is opaque and one of its four adjacent pixels is transparent, the
current pixel directly employed as boundary pixel.
Fig. 3 shows the correlation between current pixel and its four adjacent pixels. In the figure,
the light grey area corresponds to the pixels outside the binary alpha map. It can be seen
that the pixels outside the binary alpha map will not be taken into consideration, and the
number of adjacent pixels will change into two or three.
Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 47

Fig. 3. The correlation between current pixel and adjacent pixel. Where gray part denotes
pixels outside the VOP.

3.2 Boundary Search (BS)

According to the boundary pixels in VOP, we build a mask for BME process in MPEG-4
shape coding. Fig. 4 shows an example of boundary mask from the ‘Foreman’ sequence. In
this figure, the white area denotes the efficient search position for fast BME, and it is much
more efficient than the fast algorithm 0 illustrated in Fig. 5.

Fig. 4. Example of a boundary mask for shape coding. White area denotes boundary pixel.

A suggested implementation of the proposed Boundary search (BS) algorithm for shape
coding is processed as follows:
Step 1. Perform a pixel loop over the entire reference VOP. If pixel (x,y) is an boundary
pixel, set the mask at (x,y) to ‘1’. Otherwise set the mask at (x,y) to ‘0’.
Step 2. Perform a pixel loop over the entire current BAB. If pixel (i,j) is a boundary
pixel, set (i,j) to be “reference point”, and terminate the pixel loop. This step is
illustrated in Fig. 6(b). Therefore, there is only one reference point in current
BAB.
Step 3. For each search point within ±16 search range, check the pixel (x+i, y+j) which
is fully aligned with the “reference point” from the current BAB. If the mask
value at (x+i, y+j) is ‘1’, which means that the reference point is on the
48 VLSI

boundary of reference VOP, the procedure will compute SAD of the search
point (x, y). Otherwise, SAD of the search point (x, y) will not be computed, and
the processing continues at the next position. Fig. 6(a) shows an example of this
step. The search points in (x1, y1) and (x2, y2) will be skipped by this procedure,
while the SAD will be computed in (x3, y3).
Step 4. When all the search points within ±16 search range is done, the MV that
minimizes the SAD will be taken as MVS. Fig. 7 illustrates the overall scheme of
proposed BS algorithm for MPEG-4 shape coding.
In the worst case, the proposed BS algorithm needs (256+ (16+1)2) determinations to check
whether the pixel is a boundary pixel. For each non-skipped search point, the SAD obtained
by 256 exclusive-OR operations and 255 addition operations was taken as the distortion
measure. However, based on BS algorithm, the number of non-skipped search points was
reduced significantly, and the additional computational load due to BS algorithm was
negligible.

Fig. 5. Example of an effective search area, which has been proposed in reference 0.
Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 49

Fig. 6. The illustration of the proposed BS algorithm.

Fig. 7. Flow chart for proposed BS algorithm.

3.3 Diamond Boundary Search (DBS)

A better solution for block-matching motion estimation is to perform the search using a
diamond pattern because of center-biased motion vector distribution characteristic. This is
achieved by dividing the search area into diamond shaped zones and using half-stop
criterion [17]. Fig. 8 shows an example of diamond shaped zones in a ±5 search window,
and each number denotes the search zone in the search procedure.
50 VLSI

Fig. 8. A ±5 search window using diamond-shaped zones.

We combine the proposed BS algorithm with diamond-shaped zones, called DBS, and give
different thresholds (Thn) for each search zone for furthermost improvement. The procedure
is explained in below.
Step 1. Construct diamond-shaped zones around MVPS within ±16 search window. Se
t n=1.
Step 2. Calculate SAD for each search point in zone n. Let MinSAD be the smallest SAD
up to now.
Step 3. If MinSAD≦Thn, goto Step 4. Otherwise, set n=n+1 and goto Step 2.
Step 4. The motion vector is chosen according to the block corresponding to MinSAD.

3.4 Weight SAD (WSAD)

In MPEG-4 shape coding, the reference BAB with minimum SAD will be selected as motion
compensated BAB. However, the BAB with minimum SAD may be far away from the
original. It means that encoder should waste much more bits to code MVDS. Some previous
motion estimation algorithms for color space used the concept of the weighted SAD
(WSAD) to compensate the distortion [24]. In this paper, we proposed the similar concept of
WSAD which takes both SAD and MVDS into consideration as the distortion measure. The
WSAD is given by

WSAD=W1*SAD+W2*(|mvds_x|+|mvds_y|) (1)
SAD=ΣΣ|pi-1(i+u,j+v)-pi(i,j)| … (2)

where mvds_x is MVDS in the horizontal direction and mvds_y is MVDS in the vertical
direction. W1 and W2 denote the weighting values for SAD and MVDS, respectively. The
WSAD is evaluated in every search points and the BAB with minimum WSAD is selected as
Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 51

motion compensated BAB. Based on the experimentation, W1 and W2 are determined as 10

and 7, respectively. The improvement of WSAD on bit-rate can compensate for the
drawback when fast BME algorithm was adopted. Since the number of search points has
been reduced significantly, the computational power due to calculate the WSAD will
increase negligibly.

4. Design Exploration on CBBME

In this section, two approaches are developed based on CBBME. One is the center-biased
motion vector distribution. The other is the search range shrinking. Both of them can further
reduce the computation complexity in BME.

4.1 Center-biased motion vector distribution

The property is obvious that real-world sequences have a central biased motion vector
distribution. This can be achieved by dividing the search area into diamond shaped zones
and using half-stop criterion [17]. Diamond shaped zone has the higher center-biased
distribution. The closer the search area to the center position, the less the number of the
search zone in the search procedure.

4.2 Search Range Shrinking

With the property on Section 4.1, it is possible to reduce the default search range to a less size
of search range. Default search range is ±16 in standard and it is straightforward used in
conventional works. With the aid of WSAD in CBBME and the diamond shaped zones on
search range, we make a search range shrinking technique from ±16 to ±13 pixels in our
experimentation. Table 1 shows the simulation result of the proposed binary motion estimator
using the above two techniques. This result supports the usage of the ±13 shrinking search
range. An important contribution is that the WSAD makes some improvement on bit-rate.
Therefore it takes similar even less bits to represent the shape per VOP in the same quality.

Sequence Full Search (SR=±16) Proposed Architecture with limited

SSB (SR=±13)
Bits/VOP % Bits/VOP %
Bream 1599.80 100 1599.27 99.97
News 890.22 100 890.18 100.00
Foreman 1186.94 100 1183.12 99.68
Children 2056.03 100 2055.43 99.97
Table 1. Performance comparison of FS and proposed architecture.

Fig. 9 shows the algorithm flow of CBBME. First, we construct diamond-shaped zones around
motion vector predictor for shape (MVPS) within shrinking search range. Second, we calculate
WSAD for each search point. Let MinWSAD be the smallest SAD up to now. If MinWSAD is
smaller or equal to a threshold (Th) for each search zone, then the motion vector is chosen
according to the block corresponding to MinWSAD; otherwise, it changes to next position and
calculates the WSAD again.
52 VLSI

Construct diamond
shaped zones

Change to next yes

Skipped position
position
no

Compute WSAD

Set MVs=MVs with

MinWSAD

no
MinWSAD≦Th

yes

Find MVs

Fig. 9. Flow chart for the CBBME algorithm.

5. Architecture Design for MPEG-4 Shape Coding

The proposed MPEG-4 shape coding system, as illustrated in Fig. 10 consists of five major
modules: BAB type decision, size conversion, BME, CAE and variable length coding (VLC).
Based on the computational complexity analysis in Fig. 1, it can be seen that BME, size
conversion and CAE take a great part of the processing time. In this section we propose an
efficient architecture for these main modules with some design optimization and
consideration.

BAB (binary aplpha block)

BAB Type Size Intra/Inter

BME
Decision Conversion CAE
bitsstream
MVDs
CR
BAB type VLC

data
control Frame
AG
bitstream Memory

Fig. 10. Block diagram of MPEG-4 shape encoding.

Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 53

5.1 Binary Motion Estimation

Fast motion estimation architectures were presented in several previous papers [12]-[15].
The performance could be high but high cost and high complexity architectures are always
needed. Therefore, a novel and efficient architecture for binary motion estimation using
proposed CBBME algorithm is proposed. Fig. 11 shows the block diagram of the proposed
BME architecture. It mainly consists of a Boundary Pixel Detector, a processing element (PE)
Array, a Compare and Selection (CAS) module and two main memories for search range (SR)
and BAB buffer. Boundary Pixel Detector finds the “reference point” in current BAB and
checks whether the candidate search positions are non-skipped positions. PE Array performs
the binary motion estimation. CAS finds the minimum WSAD and its MVDS for shape
coding.

BAB SR
Shift
Buffer Buffer

●
●

Boundary MVDS
PE
Pixel CAS
Array
Detector
Fig. 11. Block diagram of the proposed BME architecture.

A. Boundary Pixel Detector

Boundary Pixel Detector is a design dedicated to our CCBME algorithm. From the analysis
in Section 4, since the search range has been reduced to ±13, the search points with one
dimension is 13+13+1=27 and induce a detected region with 27×27 search points. Including
the BAB size of 16×16, totally 42×42 pixels are used as the total search area.
According to the effect on 27×27 search points, a multiple of 3 is considered on architecture
design. In our binary motion estimator, a 3×3 PE array is used. Therefore, we separate the
search positions into a 9×9 sub-search-block (SSB) array structure, and each SSB contains 3×3
search positions for the calculation of PE array. Fig.12 illustrates the detected region and its
related array structure with totally 9x9= 81 SSBs.
54 VLSI

27

SSB SSB SSB SSB SSB SSB SSB SSB SSB

SSB SSB SSB SSB SSB SSB SSB SSB SSB

SSB SSB SSB SSB SSB SSB SSB SSB SSB

1 2 3

SSB SSB SSB SSB SSB SSB SSB SSB SSB 4 5 6

27 SSB SSB SSB SSB SSB SSB SSB SSB SSB 7 8 9

SSB SSB SSB SSB SSB SSB SSB SSB SSB SSB

SSB SSB SSB SSB SSB SSB SSB SSB SSB

SSB SSB SSB SSB SSB SSB SSB SSB SSB

SSB SSB SSB SSB SSB SSB SSB SSB SSB

Fig. 12. The 27x27 detected region composes of 9x9=81 SSBs; each SSB contains 9 checking
pixel.

In each SSB, 9 checking pixels are included. In addition, in order to detect boundary pixels
and obtain the bordered MC-BAB, a border with the width of one pixel around the 42×42
search area, called bordered search area, is applied. As depicted in Fig. 13, the pixels in the
grey area are the border pixels. This range indicates the total needed pixels in memory.
Boundary Pixel Detector first selects a boundary pixel (i,j) as “reference point” in 16×16
current BAB. Based on the reference point, a 27×27 detected region is generated as shown in
Fig. 14. Each pixel in detected region, called checking pixel, denotes whether the correlative
search position is non-skipped search position or not. If the checking pixel belongs to a
boundary pixel, the correlative search position is denoted as non-skipped search position.
Hereafter, Boundary Pixel Detector detects 9 checking pixels, which are relative to the coding
SSB in detected region. As shown in Fig. 14, the pixels in the dark area are used for detecting
boundary pixel. If all of the checking pixels in light grey area are not boundary pixels, it
means that all search positions in the SSB are skipped search positions. Therefore, the coding
SSB will not be processed in PE Array. Otherwise, the PE Array calculates WSAD of these 9
search positions in coding SSB.
Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 55

(a) (b)
Fig. 13. (a) 44×44 bordered search area, (b) 16×16 current BAB.

Fig. 14. The relation between search positions and detected region.

B. PE Array and CAS

PE array architecture which performs the binary motion estimation algorithm is shown in
Fig. 15(a). The array is the 3×3 architecture and totally consists of 9 PEs. In this architecture,
18-bits reference data (denoted as Ref[17:0]) are read from SR buffer, and 16-bits current data
are read from BAB buffer. The reference data are broadcasted to all PEs (Ref[17:2] for PE1,
PE4 and PE7; Ref[16:1] for PE2, PE5 and PE8; Ref[15:0] for PE3, PE6 and PE9 respectively),
while the current data is delayed and fed to the corresponding PE. Each PE calculates the
WSAD of a search position in coding SSB. It is noted that only the non-skipped SSB, which is
56 VLSI

detected from Boundary Pixel Detector, is processed in PE Array.

(a)

mvds_y mvds_x Reference Curren

Data t Data

W 1 W 4 W7 W 2 W 5 W8 W 3 W 6 W9
XOR
ABS ABS

Adder
tree
Add(+)
Accumulator
CAS Unit
>>1 SAD

Add(+) Dff

Wn
(b) (c)
Fig. 15. Architecture of (a) PE Array (b) PE element (c) CAS unit.
Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 57

The weighted data is calculated by adding absolute MVDS in both horizontal and vertical
directions and shifting right one bit. From the analysis in our previous paper [9], the ratio for
V2/V1 will not make large difference from 0.5 to 0.8. And in that range our WSAD is indeed
better than the result for SAD. For reducing the computational complexity, V1 and V2 in (1)
are determined as 1 and 0.5, respectively. The architecture of PE element is shown in Fig.
15(b). It produces WSAD with the sum of weighted data and SAD, where Wn means the
WSAD value for each PE element from n=1 to 9.
The architecture of Compare and Selection (CAS) module is shown in Fig. 15(c). It finds the
smallest WSAD and its MVS in coding SSB and feedbacks the smallest WSAD as the input for
the next SSB.

5.2 Size Conversion

In MPEG-4 shape coding, rate control and rate reduction are realized through size
conversion. Fig. 16 shows the block diagram of size conversion module.

BAB Down Up ACQ

●
Buffer Sample Sample Detector

SC SC
Buffer_0 Buffer_1

Fig. 16. Block diagram of size conversion module.

It consists of three major units: down-sample, up-sample and accepted quality (ACQ)
detector. The bordered BAB is read from BAB buffer and down-sampled to 4×4 and 8×8 BAB
in “SC Buffer_0” and “SC Buffer_1” respectively. Then, the 4×4 BAB is up-sampled to SC
Buffer_1, and 8×8 BAB is up-sampled to ACQ detector by up-sample unit. ACQ detector
calculates the conversion error between the original BAB and the BAB which is
down-sampled and reconstructed by up-sample unit. ACQ also needs to determine the
conversion ratio. In down-sample procedure, several pixels are down-sampled to one pixel,
while interpolated pixels are produced between original pixels in up-sample procedure. To
compute the value of the interpolated pixel, a border with the width of two around the
current BAB is used to obtain the neighboring pixels (A~L), and the unknown pixels are
extended from the outermost pixels inside the BAB. The template and pixel relationship
used for up-sampling operation can be referred as in MPEG-4 standard and [15].
Since the implementation of the down-sample and ACQ detector is relatively simple, we
only address the design of up-sample unit here. The block diagram of up-sample unit is
58 VLSI

shown in Fig. 17. Due to the window-like slicing operations, the up-sample can be easily
mapped into a delay line model. A delay line model is used to obtain the pixels in A~L,
which determine interpolated pixels (P1~P4). Based on the pixels in A~L, four 4-bits
threshold values are obtained from “Table CF”. “UP_PE” generates corresponding values
for comparison. After comparison between threshold values and the values from “UP_PE”,
four interpolated pixels (P1~P4) are stored in “Shift Register” and outputted later.

top-border
Decoder
(A~L)
Up-sampled
12 Up
Original data
data PE 20
Delay Shift
● ● Compare
Line 4 Register
Table
8 CF 16 (P1~P4)
(E~L)
left-border
Decoder
Fig. 17. Block diagram of up-sample unit.

5.3 Context Based Arithmetic Encoder (CAE)

CAE architecture mainly comprises the context generation unit and the binary arithmetic
coder [18]-[19]. Fig. 18(a) shows the block diagram of CAE module for our design.

MC Current
BAB BAB
o Template for intra CAE
Shift
Register Template for inter CAE
cx
20 bits
Table
o o o
p0 bit
Code o o o o o Inter BAB
Symbol
L R Intra BAB
o o
fw
Bit Re-
Follow Normalise
L R
Inter MC
bitstream

(a) (b)
Fig. 18. (a) Block diagram of CAE module. (b). Illustration of the Shift Register.

As mentioned before, for pixel-by-pixel processing in CAE, it basically uses the raster scan
order. Since most of the execution time is spent on the context generation in the CAE, the
“Shift Register” is used to obtain context and the related operation is illustrated in Fig. 18(b)
[20]. Data in the shift registers can be effectively reused and thus this redundant data
Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 59

accesses can be removed. In Fig. 18(b) all the rectangles are represented as registers. Pixels in
current BAB and MC-BAB are first loaded into the Shift Register, and then shifted left one
bit at every clock cycle. Registers in context box are arranged such that various contexts can
be achieved. For intra-CAE mode, the first three rows of Shift Register are used to store
current BAB. The first two rows of Shift Register are used to store current BAB and the last
three rows are used to store MC-BAB in inter-CAE mode. Therefore, the context (cx) and
coded bit (bit) are obtained from Shift Register per cycle.

6. Implementation Results and Comparisons

We use three MPEG-4 test video sequences of CIF (352×288) format for experiment: Bream,
News and Foreman. The three sequences characterize a variety of spatial and motion
activities. Each sequence consists of 300 VOP’s of arbitrary shape. A search range of ±16
pixels is used and frame-based lossless shape coding is performed for all the test sequence.
The comparisons of SAD and WSAD using full search algorithm are shown in Fig. 19. In this
figure, the correlations between bit-rate and the ratio of W2 to W1 (W2/W1) are also shown
in Fig. 19.

Bream
1662 SAD
WSAD
1660
Bitrate

1658
1638

1636

1634
0.4 0.5 0.6 0.7 0.8
w2/w1

News
912 SAD
WSAD
910
Bitrate

908

906

904
0.4 0.5 0.6 0.7 0.8
W2/W1
60 VLSI

Foreman
1216 SAD
WSAD
1214

Bitrate 1212
1200

1198

1196
0.4 0.5 0.6 0.7 0.8
W2/W1
Fig. 19. Performance comparisons of SAD and WSAD using full search algorithm.

It can be seen that the result of using WSAD as the distortion measure takes less bits than
that of using SAD except “News” in W2/W1= 0.4. This is because the “News” sequence is a
low motion video sequence, and WSAD is of no benefit in such sequence. As the result of
Fig. 19, W1 and W2, which is derived from (1), are determined as 10 and 7, respectively.
Based on the weighting values, the average number of bits to represent the shape per VOP is
shown in Table 2. The percentages compared to the results of full search, which uses SAD as
the distortion measure, are also shown in Table 2. An important contribution is that the
WSAD makes some improvement on bit-rate, and it can compensate for the inaccuracy of
motion vector when the fast search algorithm is used.

Full Search with SAD Full Search with WSAD

Sequence
Bits/VOP % Bits/VOP %
Bream 1659.35 100 1636.46 98.62
News 909.43 100 906.10 99.63
Foreman 1213.73 100 1197.55 98.67
Table 2. Performance comparison of SAD and WSAD based on full search algorithm in
bit-rate (W1=10, W2=7).

Fig. 20 shows the comparison of various search algorithms. It can be seen that the proposed
BS and DBS algorithm take less search points than FS algorithm and the algorithms
described in 0-0.
Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 61

Bream
50000
FS
Ref. [6]
40000 Ref. [7]
BS
DBS
Search Point

30000

20000

10000

0
0 20 40 60 80 100
Frame

News
20000
FS
Ref. [6]
Ref. [7]
15000 BS
DBS
Search Point

10000

5000

0
0 20 40 60 80 100
Frame

Foreman
FS
60000 Ref. [6]
Ref. [7]
BS
DBS
Search Point

40000

20000

0
0 20 40 60 80 100
Frame
Fig. 20. Performance comparisons of various search algorithms.

Table 3 shows the number of search points (SP), and Table 4 shows the average number of bits
to represent the shape per VOP. The percentages compared to the results corresponding to the
full search in MPEG-4 VM are also shown in these tables. “BS” denotes the proposed
algorithm without using diamond search pattern and “DBS” denotes the proposed algorithm
62 VLSI

using diamond search pattern. Table 5 shows the runtime simulation results of various BME
algorithms. It is noted that the BME algorithm 0 takes much more computational complexity
because of the generation of the mask for effective search area.

Sequence Full Search Ref. 0 Ref. 0 Proposed (BS) Proposed

(DBS)

SP SP % SP % SP % SP %
Bream 10,504,494 6,495,838 61.84 1,560,221 14.85 367,115 3.49 67,232 0.64
News 747,054 403,131 53.96 6,568 0.88 24,923 3.36 2,048 0.27
Foreman 9,085,527 5,093,409 56.06 1,564,551 17.22 287,272 3.16 57,214 0.63
Table 3. Total search points for various search algorithms.

Sequence Full Ref. 0 Ref. 0 Proposed (BS) Proposed (DBS)

Search
Bits/VOP Bits/VOP % Bits/VOP % Bits/VOP % Bits/VOP %
Bream 1659.35 1659.35 100.00 1683.28 101.44 1655.78 99.78 1669.58 100.62
News 909.43 909.43 100.00 902.68 99.26 910.82 100.15 908.39 99.89
Foreman 1213.73 1213.73 100.00 1219.47 100.47 1209.35 99.64 1219.23 100.45
Table 4. Average bit-rate for various search algorithms.

Compared with the full search method, the proposed fast BME algorithm (BS) needs 3.5%
search points and takes equal bit rate in the same quality. By using the diamond-shaped
zones, the proposed algorithm (DBS) needs only 0.6% search points. Compared with other
fast BME 0-0, our algorithm uses less search points, especially in high motion video
sequences, such as ‘Bream’ and ‘Foreman’.

Sequence Full Ref. 0 Ref. 0 Proposed (BS) Proposed

Search (DBS)
ms ms % ms % ms % ms %

Bream 57718.67 62776.36 108.76 7782.83 13.48 8678.02 15.04 1738.24 3.01
News 3830.13 4202.36 109.72 35.05 0.92 574.04 14.99 22.16 0.58
Foreman 44877.19 52574.24 117.15 7507.74 16.73 6487.52 14.46 1703.88 3.80
Table 5. Runtime simulation results of various search algorithms.

Table 6 shows the number of non-skipped SSB per PE. Due to the contribution of the
proposed CBBME algorithm, the number of non-skipped SSB is reduced largely. It can be
seen that the average number of non-skipped SSB is much less than the total number of SSB,
which is denoted as 81 per PE. In the worst case, the additional non-skipped SSB is usually
in the positions with large motion vector, which does not tend to be the adoptive MV.
Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 63

Sequence Average non-skipped Maximum non-skipped SSB

SSB per PE per PE
Bream 12.78 33
News 10.73 16
Foreman 10.94 24
Children 13.40 46
Table 6. Average and maximum number of non-skipped SSB per PE.

Therefore, in the binary motion estimator, the number of non-skipped SSB is limited to 32
from experimentation in maximum situation. For average situation, we set the number as
12.In our architecture, the processing cycles for various BAB types are shown in Fig. 21.

(average/maximum)
BME
& BMC
(400/780)

Inter
BAB BAB Size Intra CAE
MVPs
Decision Conv. ● CAE (303/610)
(35)
(19) (264) (303/610)

0 19 54 318 1098 1708 Time (cycle)

: VLC(3~2x)
BAB type 2, 3
BAB type 0
BAB type 4 for I -VOP
BAB type 1
BAB type 4, 5, 6
for B-,P-VOP
Fig. 21. Processing cycles for various BAB types.

Totally seven types of mode are described for BAB. The number in parentheses indicates the
latency of that module with average and maximum number of cycles. Notice that the
processing cycles of BME and CAE depend on the content of BAB. To complete one BAB
processing in the worst case scenario, our architecture requires 1708 clock cycles, including:
19 clock cycles for mode decision, 35 clock cycles for identifying MVPS, 264 clock cycles for
size conversion, 780 clock cycles for BME, and 610 clock cycles for inter CAE.
Actually, few literatures explored the architecture design of shape coding. In [10], they only
designed the BME. We can extract our data for BME part as comparison in Table 7. In terms
of the whole shape coding,

E. A. Al_Qaralleh’s [10] Proposed

Gate count 11582 10523
One BAB processing 563 780
(cycles)
Comment Partial design Completed design
(Only BME implemented)
Table 7. Comparison with BME only.

Table 8 illustrates some results with different architectures. For our design the size of BAB
64 VLSI

buffer and SR buffer are 16×16 and 44×44 respectively. The average and maximum numbers
of non-skipped SSB are determined as 12 and 32 from experiments.

DDBME [16] Natarajan’s [21] Proposed

Current BAB size 16×16 bits RAM 16×16 bits SRAM 16×16 bits SRAM
SR buffer size 16×32 bits RAM 47×32 bits SRAM 44×44 bits SRAM
Access from SR buff to 4096 5828 Average: 1348
obtain one MV (Bytes) Maximum: 3328
Latency to obtain one MV 1039 1039 Average: 360
(cycles) Maximum: 740
One BAB processing 3034 N/A 1708
(cycles) (without pipelinig) (without pipelining)
Table 8. Architecture analysis and comparison for various binary motion estimations.

Table 8 also lists the architecture comparisons between the proposed and some previous
works in [16] and [21]. In [16] it adopts the data-dispatch technique and is named as
data-dispatch based BME (DDBME). [21] is Natarajan’s architecture which is modified from
BME-based Yang’s 1-D systolic array [22]. In their design, they use extra memory, SAP module,
to process the bit shifting and bit packing for the alignment of BAB. It also results in a
computation overhead. In our design, we have used the Boundary Pixel Detector for the
alignment of boundary of BAB. Accordingly, no SAP memory is needed. Furthermore, the
proposed CBBME design needs less data transfer and latency to obtain one motion vector
compared with [16] and [21], because we consider the skipping on redundant searches.
Compared with the implementation for one BAB processing in the worst case, our design also
requires less cycles than [16] with the same base of non-pipelining work. Only 56% cycles of
[16] is needed in our approach.
Fig. 22(a) shows the synthesized gate count of each module and Fig. 22(b) shows the chip
layout using synthesizable Verilog HDL. There are 7 synchronous RAMs in the chip. Two
1600×16 bits RAMs are used for frame buffer. Two 48×22 bits RAMs are used for SR buffer.
One 32×20 bits RAM is used for SC buffer, one 32×18 bits RAM is used for MC buffer and one
32×20 bits RAM is used for BAB buffer, respectively. The chip feature is summarized in Table
9. Total gate count is 40765. The chip size is 2.4×2.4 mm2 with TSMC 0.18μm CMOS
technology and the maximum operation frequency is 53 MHz

BME(10523)
2%
Size Conversion(4453)
3% 8%
35% CAE(6836)
14%
Prob. Table for CAE(4221)

BAB Type Decision(914)

23% 15%
MVPs(749)

VLC(2459)

(a)
Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 65

(b)
Fig. 22. (a) Synthesized gate count of each module. (b) Chip layout of shape coding encoder.

Technology TSMC 0.18μm CMOS (1P6M)

Package 128 CQFP

Die size 2
2.4×2.4 mm
Core size 2
1.4×1.4 mm
Clock rate 53 MHz

Power dissipation 35mW

Gate count 40765

Frame buffer: 2×1600×16

Memory size SR buffer: 2×48×22
(bits) SC buffer: 32×20
MC buffer: 32×18
BAB buffer: 32×20
Table 9. Chip Features.

7. Conclusion
MPEG-4 has provided a well-adopted object-based coding technique. When people migrate
from compressed coding domain to object coding domain, the complexity issue on shape
coding is converged. In this paper we propose a fast binary motion estimation algorithm
using diamond search pattern for shape coding and an efficient architecture for MPEG-4
shape coding. By using the properties of shape information and diamond shaped zones, we
can reduce the number of search points significantly, resulting in a proportional reduction
66 VLSI

of computational complexity. The experimental results show that the proposed method can
reduce the number of search points of BME for shape coding to only 0.6% compared with
that of the full search method described in MPEG-4 verification model. Specifically, the fast
algorithm takes equal bit rate in the same quality compared with full search algorithm. The
proposed algorithm is simple, efficient and suitable for real-time software and hardware
applications. This architecture is based on the boundary search fast algorithm which
accomplishes the large reduction on computation complexity. We also apply the approaches
on center-biased motion vector distribution and search range shrinking for further
improvement. In this paper we report a comprehensive exploration on each module of
shape coding encoder. Our architecture completely elaborates the advantages of the
proposed fast algorithm with a high performance and regular architecture. The result shows
that our design can reduce the memory access and processing cycles largely. The average
number of clock cycles for one binary alpha block processing is only 1708, which is far less
than other designs. The system architecture is implemented by synthesizable Verilog HDL
with TSMC 0.18μm CMOS technology. The chip size is 2.4 × 2.4 mm2 and the maximum
operation frequency is 53 MHz.

8. Acknowledgements
This work was supported by the CIC and the National Science Council of Taiwan, R.O.C.
under Grant NSC97-2220-E-008-001.

9. References
B. Natarajan, V. Bhaskaran, and K. Konstantinides, “Low-complexity block-based motion
estimation via one-bit trasforms,” IEEE Trans. Circuits Syst. Video Technol., vol. 7, pp.
702 -707, Aug. 1997.
D. Yu, S. K. Jang, and J. B. Ra, “A fast motion estimation algorithm for MPEG-4 shape
coding,” IEEE Int. Conf. Image Processing, vol. 1, pp. 876-879, 2000.
E. A. Al_Qaralleh, T. S. Chang, and K. B. Lee, “An efficient binary motion estimation
algorithm and its architecture for MPEG-4 shape encoding,” IEEE Trans. Circuits
Syst. Video Technol., vol. 16, no. 17, pp. 859-868, Jul. 2006.
G. Feygin, P. Glenn and P. Chow, “Architectural advances in the VLSI implementation of
arithmetic coding for binary image compression,” Proc. Data Compression Conf.
(DCC’94), pp. 254 -263, 1994.
G. Sullivan and T. Wiegand, “Rate-Distortion optimization for video compression”, IEEE
Signal Processing Magazine, pp. 74-90, Nov. 1998.
H. C. Chang, Y. C. Chang, Y. C. Wang, W. M. Chao, and L. G. Chen, “VLSI architecture
design for MPEG-4 shape coding,” IEEE Trans. Circuits Syst. Video Technol., vol. 12,
pp. 741-751, Sep. 2002.
ISO/IEC 13818-2, “Information technology-generic coding of moving pictures and associated
audio information-Video,” 1994.
ISO/IEC JTC1/SC29/WG11 N 2502a, “Generic coding of audio-visual objects: Visual
14492-2,” Atlantic City Final Draft IS, Dec. 1998.
ISO/IEC JTC1/SC29/WG11 N 3908, “MPEG-4 video verification model version 18.0,” Jan.
2001.
Contour-Based Binary Motion Estimation Algorithm and VLSI Design
for MPEG-4 Shape Coding 67

J. L. Mitchell and W. B. Pennebaker, “Optimal hardware and software arithmetic coding

procedures for the Q-Coder,” IBM J. Res. Devel., vol. 32, pp. 717-726, Nov. 1998.
Jo Yew Tham, Surendra Ranganath, Maitreya Ranganath, and Ashraf Ali Kassim, “A novel
unrestricted center-biased diamond search algorithm for block motion estimation,”
IEEE Trans. Circuits Syst. Video Technol., vol. 8, pp. 169-177, Aug. 1998.
K. B. Lee, J. Y. Lin and C. W. Jen, “A Multisymbol Context-Based Arithmetic Coding
Architecture for MPEG-4 Shape Coding,” IEEE Trans. Circuits Syst. Video Technol.,
vol. 15, pp. 283-295, Feb. 2005.
K. B. Lee, J. Y. Lin, and C. W. Jen, “A fast dual symbol context-based arithmetic coding for
MPEG-4 shape coding,” IEEE International Symposium on Circuits and Systems
(ISCAS2004), pp. 317-320, 2004.
K. M. Yang, M. T. Sun, and L. Wu, “A family of VLSI design for the motion compensation
block-matching algorithm,” IEEE Trans. Circuits Syst. Video Technol., vol. 36, pp. 1317
-1325, Oct. 1989.
K. Panusopone, and X. Chen, “A fast motion estimation method for MPEG-4 arbitrarily
shaped objects,” IEEE Int. Conf. Image Processing, vol. 3, pp. 624-627, 2000.
L. K. Liu, and E. Feig, “A block-based gradient descent search algorithm for block motion
estimation in video coding,” IEEE Trans. Circuits Syst. Video Technol., vol. 6, pp.
419-422, Aug. 1996.
M. Tourapis, O. C. Au, and M. L. Liou, “Highly efficient predictive zonal algorithms for fast
block-matching motion estimation,” IEEE Trans. Circuits Syst. Video Technol., vol. 12,
pp. 934-947, Oct. 2002.
N. Brady, “MPEG-4 standardized methods for the compression of arbitrarily shaped video
objects,” IEEE Trans. Circuits Syst. Video Technol., vol. 9, pp. 1170-1189, Dec. 1999.
Recommendation H.263: Video coding for low bit-rate communication, ITU-T H.263, 1998.
S. Dutta and W. Wolf, “A flexible parallel architecture adapted to block-matching motion
estimation algorithms,” IEEE Trans. Circuits Syst. Video Technol., vol. 6, no 1, pp.
74-86, Feb. 1996.
S. H. Han, S. W. Kwon, T. Y. Lee, and M. K. Lee, “Low power motion estimation algorithm
based on temporal correlation and its architecture,” IEEE Internal Symposium on
Signal Processing and its Applications (ISSPA), vol. 2, pp.647-650, Aug. 2001.
T. H. Tsai and C. P. Chen, “A fast binary motion estimation algorithm for MPEG-4 shape
coding,” IEEE Trans. Circuits Syst. Video Technol., vol. 14, pp. 908-913, Jun. 2004.
T. Komarek and P. Pirsch, “Array architectures for block-matching algorithms,” IEEE Trans.
Circuits Syst., vol. 36, pp.1301-1308, Oct. 1989.
T. M. Liu, B. J. Shieh and C. Y. Lee, “An efficient modeling codec architecture for binary
shape coding,” IEEE International Symposium on Circuits and Systems (ISCAS2002),
vol. 2, pp. 316 -319, 2002.
68 VLSI
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 69

X4

Memory-Efficient Hardware Architecture

of 2-D Dual-Mode Lifting-Based Discrete
Wavelet Transform for JPEG2000
Chih-Hsien Hsia and Jen-Shiun Chiang
Department of Electrical Engineering, Tamkang University
Taipei, Taiwan

1. Introduction
Discrete wavelet transform (DWT) has been adopted in a wide range of applications,
including speech analysis, numerical analysis, signal analysis, image coding, pattern
recognition, computer vision, and biometrics (Mallat, 1989). It can be considered as a multi-
resolution decomposition of a signal into several components with different frequency
bands. Moreover, DWT is a powerful tool for signal processing applications, such as
JPEG2000 still image compression, denoising, region of interest, and watermarking. For real-
time processing it needs small memory access and low computational complexity.
Implementations of two-dimensional (2-D) DWT can be classified as convolution-based
operation (Mallat, 1989) (Marino, 2000) (Vishwanath et al., 1995) (Wu. & Chen, 2001) and
lifting-based operation (Sweldens, 1996). Since the convolution-based implementations of
DWT have high computational complexity and large memory requirements, lifting-based
DWT has been presented to overcome these drawbacks (Sweldens, 1996) (Daubechies &
Sweldens, 1998). The lifting-based scheme can provide low-complexity solutions for
image/video compression applications, such as JPEG2000 (Lian et al., 2001), Motion-
JPEG2000 (Seo & Kim, 2007), MPEG-4 still image coding, and MC-EZBC (Ohm, 2005) (Chen
& Woods, 2004). However, the real-time 2-D DWT for multimedia application is still
difficult to achieve. Hereafter, efficient transformation schemes for real-time application are
highly demanded. Performing 2-D (or multi-dimensional) DWT requires many
computations and a large block of transpose memory for storing intermediate signals with
long latency time. This work presents new algorithms and hardware architectures to
improve the critical issues in 2-D dual-mode (supporting 5/3 lossless and 9/7 lossy coding)
lifting-based discrete wavelet transform (LDWT). The proposed 2-D dual-mode LDWT
architecture has the merits of low transpose memory, low latency, and regular signal flow,
making it suitable for VLSI implementation. The transpose memory requirement of the NN
2-D 5/3 mode LDWT is 2N and that of 2-D 9/7 mode LDWT is 4N.
Low transpose memory requirement is of a priority concern in spatial-frequency domain
implementation. Generally, raster scan signal flow operations are popular in NN 2-D DWT,
and under this approach the memory requirement ranges from 2N to N2 (Diou et al., 2001)
70 VLSI

(Andra et al., 2002) (Chen & Wu, 2002) (Chen, 2002) (Chiang & Hsia, 2005) (Jung & Park,
2005) (Vishwanath et al., 1995) (Huang et al., 2005) (Mei et al., 2006) (Huang et al, 2005) (Wu
& Lin, 2005) (Lan ., 2005) (Wu. & Chen, 2001) in 2-D 5/3 and 9/7 modes LDWT. In order to
reduce the amount of the transpose memory, the memory access must be redirected. In our
approach, the signal flow is revised from row-wise only to mixed row- and column-wise,
and a new approach, called interlaced read scan algorithm (IRSA), is used to reduce the
amount of the transpose memory. By the IRSA approach, a transpose memory size is of 2N
or 4N (5/3 or 9/7 mode) for an NN DWT. The proposed 2-D LDWT architecture is based
on parallel and pipelined schemes to increase the operation speed. For hardware
implementation we replace multipliers with shifters and adders to accomplish high
hardware utilization. This 2-D LDWT has the characteristics of high hardware utilization,
low memory requirement, and regular signal flow. A 256×256 2-D dual-mode LDWT was
designed and simulated by VerilogHDL, and further synthesized by the Synopsys design
compiler with TSMC 0.18μm 1P6M CMOS process technology.

2. Survey of 2-D LDWT Architecture

Among the variety of DWT algorithms, LDWT provides a new approach for constructing
biorthogonal wavelet transforms and also provides an efficient scheme for calculating
classical wavelet transforms (Sweldens, 1996) (Chen & Wu, 2002) (Andra et al., 2000) (Andra
et al., 2002) (Diou et al., 2001) (Chen, 2002) (Chiang & Hsia, 2005) (Tan & Arslan, 2001)
(Huang et al., 2005) (Huang et al., 2002) (Mei et al., 2006) (Weeks & Bayoumi, 2002)
(Varshney et al., 2007) (Huang et al., 2004) (Tan & Arslan, 2003) (Jiang & Ortega, 2001) (Jung
& Park, 2005) (Chen, 2004) (Lian et al, 2001) (Seo & Kim, 2007) (Huang et al., 2005) (Wu &
Lin, 2005) (Lan et al., 2005) (Wu. & Chen, 2001). Factoring the classical wavelet filter into
lifting steps can reduce the computational complexity of the corresponding DWT by up to
50% (Daubechies & Sweldens, 1998). The lifting steps can be implemented easily, which is
different from the direct finite impulse response (FIR) implementations of Mallat’s
algorithm (Daubechies & Sweldens, 1998). Andra et al. (Andra et al., 2000) (Andra et al.,
2002) proposed a block-based simple four-processor architecture that computes several
stages of the DWT at a time. Diou et al. (Diou et al., 2001) presented an architecture that
performs LDWT with a 5/3 filter by interleaving technique. Chen et al. (Chen & Wu, 2002)
proposed a folded and pipelined architecture for a 2-D LDWT implementation, with
memory size of 2.5N for an NN 2-D DWT. This lifting architecture for vertical filtering is
divided into two parts, each consisting of one adder and one multiplier. Since both parts are
activated in different cycles, they can share the same adder and multiplier to increase the
hardware utilization and reduce the latency. However, this architecture also has high
complexity due to the characteristics of the signal flow. Chen et al. (Chen, 2002) proposed a
flexible folded architecture for 3-level 1-D LDWT to increase the hardware utilization.
Chiang et al. (Chiang & Hsia, 2005) proposed a 2-D DWT folded architecture to improve the
hardware utilization. Jiang et al. (Jiang & Ortega, 2001) presented a parallel processing
architecture that models the DWT computation as a finite state machine and efficiently
computes the wavelet coefficients near the boundary of each segment of the input signal.
Lian et al. (Lian et al, 2001) and Chen et al. (Chen, 2004) used a 1-D folded architecture to
improve the hardware utilization of 5/3 and 9/7 filters. The recursive architecture is a
general scheme to implement any wavelet filter that is decomposed into lifting steps in
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 71

smaller hardware complexity. Jung et al. (Jung & Park, 2005) presented an efficient VLSI
architecture of dual-mode LDWT that is used by lossy or lossless compression of JPEG2000.
Marino (Marino, 2000) proposed a high-speed/low-power pipelined architecture for the
direct 2-D DWT by four-subband transforms performed in parallel. The architecture of
(Huang et al., 2002) implements 2-D DWT with only transpose memory by using recursive
pyramid algorithm (PRA). In (Vishwanath et al., 1995) it has the average of N2 computing
time for all DWT levels. However, they use many multipliers and adders. Varshney et al.
(Varshney et al., 2007) presented energy efficient single-processor and fully pipelined
architectures for 2-D 5/3 lifting-based JPEG2000. The single processor performs both row-
wise and column-wise processing simultaneously to achieve the 2-D transform with 100%
hardware utilization. Tan et al. (Tan & Arslan, 2003) presented a shift-accumulator
arithmetic logic unit architecture for 2-D lifting-based JPEG2000 5/3 DWT. This architecture
has an efficient memory organization, which uses a small amount of embedded memory for
processing and buffering. Those architectures achieve multi-level decomposition using an
interleaving scheme that reduces the size of memory and the number of memory accesses,
but have slow throughput rates and inefficient hardware utilization. Seo et al. (Seo & Kim,
2007) proposed a processor that can handle any tile size, and supports both 5/3 and 9/7
filters for Motion-JPEG2000. Huang et al. (Huang et al, 2005) proposed a generic RAM-based
architecture with high efficiency and feasibility for 2-D DWT. Wu et al. (Wu & Lin, 2005)
presented a high-performance and low-memory architecture to implement a 2-D dual-mode
LDWT. The pipelined signal path of their architecture is regular and practical. Lan et al. (Lan
et al., 2005) proposed a scheme that can process two lines simultaneously by processing two
pixels in a clock period. Wu et al. (Wu. & Chen, 2001) proposed an efficient VLSI architecture
for direct 2-D LDWT, in which the poly-phase decomposition and coefficient folding are
adopted to increase the hardware utilization. Despite these efficient improvements to
existed architectures, further improvements in the algorithm and architecture are still
needed. Some VLSI architectures of 2-D LDWT try to reduce the transpose memory
requirements and communication between the processors (Chen & Wu, 2002) (Andra et al.,
2000) (Andra et al., 2002) (Diou et al., 2001) (Chen, 2002) (Chiang & Hsia, 2005) (Tan &
Aslan, 2002) (Jiang & Ortega, 2001) (Lian et al., 2001) (Jung & Park, 2005) (Chen, 2004)
(Huang et al., 2005) (Daubechies & Sweldens, 1998) (Marino, 2000) (Vishwanath et al., 1995)
(Taubman & Marcellin, 2001) (Marcellin et al., 2000) (Mei et al., 2006) (Varshney et al., 2007)
(Huang et al., 2004) (Tan & Arslan, 2003) (Seo & Kim, 2007) (Huang et al., 2005) (Wu & Lin,
2005) (Lan et al., 2005) (Wu. & Chen, 2001), however these hardware architectures still need
large transpose memory.

3. Discrete wavelet transform and lifting-based method

This section briefly reviews the use of DWT in the coding engine of JPEG2000 (Taubman &
Marcellin, 2001). The classical DWT employs filtering and convolution to achieve signal
decomposition (Mallat, 1989) (Marino, 2000) (Vishwanath et al., 1995) (Wu. & Chen, 2001).
Meyer and Mallat found that the orthonormal wavelet decomposition and reconstruction
can be implemented in the multi-resolution signal analysis framework (Mallat, 1989). The
multi-resolution analysis is now a standard method for constructing the orthonormal
wavelet-bases. JPEG2000 adopts this characteristic to transform an image in the spatial
domain into the frequency domain.
72 VLSI

3.1 Classical DWT

DWT performs multi-resolution decomposition of the input signals (Mallat, 1989). The
original signals are first decomposed into two subspaces, called the low- (low-pass) and
high-frequency (high-pass) subbands. The classical DWT implements the decomposition
(analysis) of a signal by a low-pass digital filter H and a high-pass digital filter G. Both
digital filters are derived using the scaling function and the corresponding wavelets. The
system downsamples the signal to decimate half of the filtered results in decomposition
processing. The Z-transfer functions of H(z) and G(z) based on four-tap and non-recursive
FIR filters with length L are represented as follows:

H(z)= h0 + h1z-1 + h2z-2+ h3z-3, (1)

G(z)= g0 + g1z-1 + g2z-2+ g3z-3. (2)

The reconstruction (synthesis) process is implemented using an up-sampling process.

Mallat’s tree algorithm or pyramid algorithm (Mallat, 1989) can be used to find the multi-
resolution decomposition DWT. The decomposition DWT coefficients at each resolution
level can be calculated as follows:

for (j=1 to J)
for (i=0 to N/2j-1)
{
k 1
X Hj ( n )=  G ( z ) X J  1 H ( 2 n  i )
(3)
i0

k 1
X Lj ( n )=  H ( z ) X J  1 G ( 2 n  i )
(4)
i0

where j denotes the current resolution level, k the number of the filter tap, X Hj (n) the nth
high-pass DWT coefficient at the jth level, X Lj (n) the nth low pass DWT coefficient at the
jth level, and N the length of the original input sequence. Fig. 1 shows a 3-level 1-D DWT
decomposition using Mallat’s algorithm.
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 73

Fig. 1. 3-level 1-D DWT decomposition using Mallat’s algorithm.

The downsampling operation is then applied to the filtered results. A pair of filters are
applied to the signal to decompose the image into the low-low (LL), low-high (LH), high-
low (HL), and high-high (HH) wavelet frequency bands. Fig. 2 illustrates the basic 2-D DWT
operation and the transformed result which is composed of two cascading 1-D DWTs. The
image is first analyzed horizontally to generate two subimages. The information is then sent
into the second 1-D DWT to perform the vertical analysis to generate four subbands, and
each with a quarter of the size of the original image. Considering an image of size N×N, each
band is subsampled by a factor of two, so that each wavelet frequency band contains
N/2×N/2 samples. The four subbands can be integrated to generate an output image with
the same number of samples as the original one.
Most image compression applications can reapply the above 2-D wavelet decomposition
repeatedly to the LL subimage, each time forming four new subband images, to minimize
the energy in the lower frequency bands.
74 VLSI

Fig. 2. The 2-D analysis DWT image decomposition process.

Fig. 3. Block diagram of the LDWT.

3.2 LDWT algorithm

The lifting-based scheme proposed by Daubechies and Sweldens requires fewer
computations than the traditional convolution-based approach (Sweldens, 1996)
(Daubechies & Sweldens, 1998). The lifting-based scheme is an efficient implementation for
DWT; it can easily use integer operations and avoid the problems caused by the finite
precision or rounding. The Euclidean algorithm can be used to factorize the poly-phase
matrix of a DWT filter into a sequence of alternating upper and lower triangular matrices
and a diagonal matrix. The variables h(z) and g(z) in (5) respectively denote the low- and
high-pass analysis filters, which can be divided into even and odd parts to generate a poly-
phase matrix P(z) as in (6).

g(z)=ge(z2)+ z-1go(z2),

h(z)=he(z2)+z-1ho(z2). (5)
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 75

h ( z ) g e ( z )
P( z)   e
 ho ( z ) g o ( z ) 
. (6)

The Euclidean algorithm recursively finds the greatest common divisors of the even and
odd parts of the original filters. Since h(z) and g(z) form a complementary filter pair, P(z) can
be factorized into (7):
m 1 si( z )   1 0k 0 . (7)
P(z)   
0

1   ti( z )

10

1/ k 
i 1 
where si(z) and ti(z) are Laurent polynomials corresponding to the prediction and update
steps, respectively, and k is a nonzero constant. Therefore, the filter bank can be factorized
into three lifting steps.
As illustrated in Fig. 3, a lifting-based scheme has the following four stages:
1) Split phase: The original signal is divided into two disjoint subsets. Significantly, the
variable Xe denotes the set of even samples and Xo denotes the set of odd samples. This
phase is also called lazy wavelet transform because it does not decorrelate the data but only
subsamples the signal into even and odd samples.
2) Predict phase: The predicting operator P is applied to the subset Xo to obtain the wavelet
coefficients d[n] as in (8).

d[n]=Xo[n]+P(Xe[n]). (8)

3) Update phase: Xe[n] and d[n] are combined to obtain the scaling coefficients s[n] after an
update operator U as in (9).

s[n]=Xe[n]+U(d[n]). (9)

4) Scaling: In the final step, the normalization factor is applied on s[n] and d[n] to obtain the
wavelet coefficients. For example, (10) and (11) describe the implementation of the 5/3
integer lifting analysis DWT and are used to calculate the odd (high-pass) and even
coefficients (low-pass), respectively.

d *[ n]  X (2 n  1)   X (2n)  X (2 n  2) / 2  . (10)

s *[ n ]  X (2n )   d (2 n  1)  d (2n  1)  2 / 4  . (11)

Although the lifting-based scheme has low complexity, its long and irregular signal paths
cause the major limitation for efficient hardware implementations. Additionally, the
increasing number of pipelined registers increases the internal memory size of the 2-D DWT
architecture. The 2-D LDWT uses a vertical 1-D LDWT subband decomposition and a
horizontal 1-D LDWT subband decomposition to find the 2-D LDWT coefficients. Therefore,
the memory requirement dominates the hardware cost and architectural complexity of 2-D
LDWT. Fig. 4 shows the 5/3 mode 2-D LDWT operation. The default wavelet filters in
JPEG2000 are dual-mode (5/3 and 9/7 modes) LDWT (Taubman & Marcellin, 2001). The
lifting-based steps associated with the dual-mode wavelets are shown in Figs. 5 and 6,
76 VLSI

respectively. Assuming that the original signals are infinite in length, the first lifting stage is
first applied to perform the DWT.

(a)

Original image
Original
s0 d0 s1 d1 s2
Pixels

1-D 5/3 lifting-based High-Pass

L0 L1
DWT Output

Low-Pass
Output H0 H1 H2
High
L H Frequency H0 H1 H2 H3 H4
Pixels

H.F Part
HH-Band
HH0 HH1
Output

1-D 5/3 lifting-based HL-Band

HL0 HL1 HL2
DWT Output
Low
Frequency L0 L1 L2 L3 L4
Pixels
L.F Part

LH-Band
LH0 LH1
1-D 5/3 lifting-based Output
DWT

LL HL LL-Band
LL0 LL1 LL2
Output
LH HH

(b)
Fig. 4. 5/3 mode 2-D LDWT operation. (a) The block diagram flow of a traditional 2-D DWT.
(b) Detailed processing flow.
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 77

Fig. 5 shows the lifting-based step associated with the wavelet algorithm. The original
signals including s0, d0, s1, d1, s2, d2, … are the input pixel sequences. If the original signals
are infinite in length, then the first-stage lifting is applied to update the odd index data s0,
s1, …. In (12), the parameters 1/2 and Hi denote the first stage lifting parameter and
outcome, respectively. Equation (12) shows the operation of the 5/3 integer LDWT (Wu &
Lin, 2005) (Martina & Masera, 2007) (Hsia & Chiang, 2008).

Hi = [(Si+Si+1)×α+di]×K0,

Li = [(Hi+Hi-1)×β+Si]×K1, (12)

where α= 1/2, β=1/4, and K0= K1= 1.

Fig. 5. 5/3 LDWT algorithm.

Fig. 6. 9/7 LDWT algorithm.

78 VLSI

Together with the high-frequency lifting parameter, α, and the input signal we can find the
first stage high-frequency wavelet coefficient, Hi. After Hi is found, Hi together with the
low-frequency parameter, β, and the input signals of the second stage low-frequency
wavelet coefficients, Li, can be found. The third and fourth stages lifting can be found in a
similar manner.
Similar to the 1-level 1-D 5/3 mode LDWT, the calculation of a 1-level 1-D 9/7 mode LDWT
is shown in (13).

ai = [(Si+Si+1)×α+di],
bi = [(ai+ai-1)×β+Si],
Hi = [(bi+bi+1)×γ+ai]×K0,
L i= [(Hi+Hi-1)×δ+bi]×K1. (13)

where α= −1.586134142, β= −0.052980118, γ= +0.882911075, δ= +0.443506852, K0=

1.230174104, and K1= 1/K0.
The calculation comprises four lifting steps and two scaling steps.

3.3 Boundary extension treatment for LDWT

The finite-length signal processed by DWT leads to the boundary effect. The JPEG2000
standard enhances the symmetric extension pixel at the edge, as shown in Table 1. An
appropriate signal extension is required to maintain the same number of the wavelet
coefficients as in the original signal. The embedded signal extension algorithm (Tan &
Arslan, 2001) can be used to compute the boundary of the image.
Since both the extended signal and the lifting structure are symmetrical, all the intermediate
and final results of the lifting-based DWT are also symmetrical with regard to the boundary
points, and the boundary extension can be performed without additional computational
complexity. The inverse lifting structure is easily derived from Table 1 (Tan & Arslan, 2001).
The boundary extension signal reflects the signal to ileft samples on the left and iright samples
on right. Table 1 shows the extension parameters ileft and iright for the reversible transform
5/3 mode and the irreversible transform 9/7 mode.

i0 ileft(5/3) i1 iright(5/3) i0 ileft(9/7) i1 iright(9/7)

even 2 odd 1 even 4 odd 3
* i0: first sample index; i1: last sample index.
Table 1. Boundary extension to the left and to the right for JPEG2000.

4. Interlaced read scan algorithm (IRSA)

In recent years, many 2-D LDWT architectures have been proposed to meet the
requirements of on-chip memory for real-time processing. However, the hardware
utilization of these architectures needs to be further improved. In DWT implementation, a 1-
D DWT needs very massive computation and therefore the computation unit takes most of
the hardware cost (Chen & Wu, 2002) (Andra et al., 2000) (Andra et al., 2002) (Diou et al.,
2001) (Chen, 2002) (Chiang & Hsia, 2005) (Chen, 2004) (Huang et al., 2005) (Daubechies &
Sweldens, 1998) (Marino, 2000) (Vishwanath et al., 1995) (Taubman & Marcellin, 2001)
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 79

(Varshney et al., 2007) (Huang et al., 2004) (Tan & Arslan, 2003) (Seo & Kim, 2007) (Huang et
al., 2005) (Wu & Lin, 2005) (Lan et al., 2005) (Wu. & Chen, 2001). A 2-D DWT is composed of
two 1-D DWTs and a block of transpose memory. In the conventional approach, the size of
the transpose memory is equal to the size of the processed image signal. Fig. 7(a) shows the
concept of the proposed dual-mode LDWT architecture, which consists of signal
arrangement unit, processing element, memory unit, and control unit, as shown in Fig. 7(b).
The outputs are fed to the 2-D LDWT four-subband coefficients, HH, HL, LH, and LL. The
proposed architecture is described in detail in this section, and we focus on the 2-D dual-
mode LDWT.
Compared to the computation unit, the transpose memory becomes the main overhead in
the 2-D DWT. The block diagram of a conventional 2-D DWT is shown in Fig. 4. Without
loss of generality, the 2-D 5/3 mode LDWT is considered for the description of the 2-D
LDWT. If the image dimension is NN, during the transformation we need a large block of
transpose memory (order of N2) to store the DWT coefficients after the computation of the
first stage 1-D DWT decomposition. The second stage 1-D DWT then uses the stored data to
compute the 2-D DWT coefficients of the four subbands (Chen & Wu, 2002) (Andra et al.,
2000) (Andra et al., 2002) (Diou et al., 2001) (Chen, 2002) (Varshney et al., 2007) (Huang et
al., 2004) (Tan & Arslan, 2003) (Seo & Kim, 2007) (Huang et al., 2005) (Wu & Lin, 2005) (Lan
et al., 2005) (Wu. & Chen, 2001). The computation and the access of the memory may take
time and therefore the latency is long. Since the memory size of N2 is a large quantity, here
we try to use the approach, interlaced read scan algorithm (IRSA), to reduce the required
transpose memory to an order of 2N or 4N (5/3 or 9/7 mode).

(a)

(b)
Fig. 7. The system block diagram of the proposed 2-D DWT. (a) 2-D dual-mode LDWT. (b)
Block diagram of the proposed system architecture.
80 VLSI

x(i,j): original image, i = 0~5 and j = 0~5

b(i,j): high frequency wavelet coefficient of 1-D LDWT
c(i,j): low frequency wavelet coefficient of 1-D LDWT
HH: high-high frequency wavelet coefficient of 2-D LDWT
HL: high-low frequency wavelet coefficient of 2-D LDWT
LH: low-high frequency wavelet coefficient of 2-D LDWT
LL: low-low frequency wavelet coefficient of 2-D LDWT
Fig. 8. Example of 2-D 5/3 mode LDWT operations.

Without loss of generality, let us take a 66-pixel image to describe the 2-D 5/3 mode LDWT
operation and IRSA. Fig. 8 shows the operation diagram of the 2-D 5/3 mode LDWT
operations of a 66 image. In Fig. 8, x(i,j), i = 0 to 5 and j = 0 to 5, represents the original
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 81

image signal. The left most two columns are the left boundary extension columns, and the
right most column is the right boundary extension column. The details of the boundary
extension were described in the previous section. The left half of Fig. 8 shows the first stage
1-D DWT operations. The right half of Fig. 8 shows the second stage 1-D DWT operations
for finding the four subband coefficients, HH, HL, LH, and LL. In the first stage 1-D DWT,
three pixels are used to find a 1-D high-frequency coefficient. For example, x(0,0), x(0,1), and
x(0,2) are used to find the high-frequency coefficient b(0,0), b(0,0) = [x(0,0) + x(0,2)]/2 +
x(0,1). To calculate the next high-frequency coefficient b(0,1), we need pixels x(0,2), x(0,3),
and x(0,4). Here x(0,2) is used to calculate both b(0,0) and b(0,1) and is called the overlapped
pixel. The low-frequency coefficient is calculated using two consecutive high-frequency
coefficients and the overlapped pixel. For example, b(0,0) and b(0,1) are together with x(0,2)
to find the low-frequency coefficient c(0,1), c(0,1) = [b(0,0) + b(0,1)]/4 + x(0,2). The calculated
high-frequency coefficients, b(i,j), and low-frequency coefficients, c(i,j), are then used in the
second stage 1-D DWT to calculate the four subband coefficients, HH, HL, LH, and LL.
In the second stage 1-D DWT of Fig. 8, the first HH coefficient, HH(0,0), is calculated by
using b(0,2), b(0,1), and b(0,0), HH(0,0) = [b(0,0) + b(0,2)]/2 + b(0,1). The other HH
coefficients can be computed in the same manner using three column consecutive b(i,j)
signals. For two column consecutive HH coefficients it has an overlapped b(i,j) signal. For
example b(0,3) is the overlapped signal for computing HH(0,0) and HH(0,1). To compute HL
coefficients, it needs two column consecutive HH coefficients and an overlapped b(i,j)
signal. For example, HL(0,1) is computed from HH(0,0), HH(0,1), and b(0,3), HL(0,1) =
[HH(0,0) + HH(0,1)]/4 + b(0,3). The LH coefficients are computed from the c(i,j) signal, and
each LH coefficient needs the calculation of three c(i,j) signals. For example, LH(0,1) is
computed from c(0,2), c(0,3), and c(0,4), LH(0,1) = [c(0,2) + c(0,4)]/2 + c(0,3). For two
column consecutive LH coefficients it has an overlapped c(i,j) signal. For example, c(0,3) is
the overlapped signal for computing LH(0,0) and LH(0,1). To compute LL coefficients, it
needs two column consecutive LH coefficients and an overlapped c(i,j) signal. For example,
LL(0,1) is computed from LH(0,0), LH(0,1), and c(0,2), LL(0,1) = [LH(0,0) + LH(0,1)]/4 +
c(0,2). The detail calculation equations for the four subband coefficients are summarized in
the following equations:

1 1 2
HH(h,v) = x(2h+1,2v +1)+(1/4)  x(2h+2s,2v+2t)+(-1/2) x(2h+|s|,2v+|-1+s|). (14)
s =0 t =0 s =-1

HL(h,v) = (1/4)[HH(h,v -1)+HH(h,v)]+b(h,2v)

(15)
= (1/4)[HH(h,v-1)+HH(h,v)]+(-1/2)[x(2h,2v)+x(2h +2,2v)]+x (2h +1,2v).

LH(h,v) = (1/4)[HH(h-1,v)+HH(h,v)]+(-1/2)[x(2h,2v)+x(2h,2v +2)]+x(2h,2v +1). (16)

LL(hv
, ) = (1/4)[LH(hv
, -1)+LH(hv
, )]+c(h,2v)
= (1/4)[LH(hv, -1)+LH(hv
, )]+(1/4)[b(h-1,2v)+b(h,2v)]+x(2h,2v)
= (1/4)[LH(hv
, -1)+LH(hv
, )]+(1/4)[(-1/2)x(2h-2,2v)+x(2h-1,2v)+(-1)x(2h,2v)+x(2h+1,2v)+(-1/2)x(2h+2,2v)]+x(2h,2v).
(17)

The parameters in the above equations are defined as follows:

82 VLSI

h: horizontal row, v: vertical column, x: original image, b: high-frequency wavelet coefficient

of 1-D DWT, c: low-frequency wavelet coefficient of 1-D DWT, 1/2: Prediction parameter,
and 1/4: Updating parameter.
From the description of the operations of the 2-D 5/3 mode LDWT we find that each 1-D
high-frequency coefficient, b(i,j), is calculated from three image signals, and one of the image
signal is overlapped with the previous b(i,j). The 1-D low-frequency coefficient, c(i,j), is
calculated from two row consecutive b(i,j)’s and an overlapped pixel. The HH, HL, LH, and
LL coefficients are computed from b(i,j)’s and c(i,j)’s. If we can change the scanning order of
the first stage 1-D LDWT and the output order of the second stage 1-D LDWT, during the 2-
D LDWT operation we need only to store the b(i,j)’s to the transpose memory (First-In First-
Out, FIFO size of N) and the overlapped pixels to the internal memory (R4+R9 size of N).
For an NN image, the transpose memory block can be reduced to only a size of 2N as
shown in Fig. 9. IRSA is based on this idea, and it can reduce the requirement of the
transpose memory significantly. The block diagram of IRSA with several pixels of an image
is shown in Fig. 9. In Fig. 9, the numbers on top and left represent the coordinate indexes of
a 2-D image. In order to increase the operation speed, IRSA scans two pixels in the
consecutive rows a time. IN1 (Initial in x(0,0)) and IN2 (Initial in x(1,0)) are the scanning
inputs at beginning. At the first clock, the system scans two pixels, x(0,0) and x(1,0), from
IN1 and IN2, respectively. At the second clock, IN1 and IN2 read pixels x(0,1) and x(1,1),
respectively. At clock 3, IN1 and IN2 read pixels x(0,2) and x(1,2), respectively. After IN1
and IN2 have read three pixels, the DWT processor tries to compute two 1-D high-frequency
coefficients, b(0,0) and b(0,1), and these two high-frequency coefficients are stored in the
transpose memory for the subsequent computation of the low-frequency coefficients. Pixels
x(0,2) and x(1,2) are stored in the internal memory for the subsequent computation of the 1-
D high-frequency coefficients.
At clock 4, the DWT processor scans pixels on row 2 and row 3, and IN1 and IN2 read pixels
x(2,0) and x(3,0), respectively. At clock 5, IN1 and IN2 read pixels x(2,1) and x(3,1),
respectively. At clock 6, IN1 and IN2 read pixels x(2,2) and x(3,2), respectively. At this
moment the DWT processor tries to compute the two high-frequency coefficients, b(2,0) and
b(3,0), upon pixels x(2,0) to x(2,2) and x(3,0) to x(3,2) respectively and these two high-
frequency coefficients are stored in the transpose memory for the subsequent computation
of the low-frequency coefficients. Pixels x(2,2) and x(3,2) are stored in the internal memory
for the subsequent computation of the high-frequency coefficients. Then (at clock 7) the
DWT processor scans the subsequent two rows to read three consecutive pixels in each row
and compute the high-frequency coefficients. The coefficients are stored in the transpose
memory and pixels x(4,2) and x(5,2) are stored in the internal memory. This procedure will
continue to read three pixels and compute the high-frequency coefficients and store the
coefficients to the transpose memory and store pixels x(2,j) and x(2,j+1) to the internal
memory in each row until the last row.
Then the DWT processor scans row 0 and row 1 and makes N1 and N2 read pixels x(0,3)
and x(1,3), respectively. At the next clock, IN1 and IN2 read pixels x(0,4) and x(1,4),
respectively. The DWT processor uses pixels x(0,3), x(0,4), and x(0,2), which was stored
previously, to compute the high-frequency coefficient b(0,1). Simultaneously the DWT
processor uses pixels x(1,3), x(1,4), and x(1,2), which was stored previously, to compute the
high-frequency coefficients b(1,1). As soon as b(0,1) and b(1,1) are found, b(0,0), b(0,1), and
x(0,2) are used to generate the low-frequency coefficient c(0,1); b(1,0), b(1,1), and x(1,2) are
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 83

used to generate the low-frequency coefficient c(1,1). The computed high-frequency

coefficients are then stored in the transpose memory, and pixels x(0,4) and x(1,4) replace
pixels x(0,2) and x(1,2) to be stored in the internal memory. IN1 and IN2 then read the data
in rows 2 and 3 to process the same operations until the end of the last pixel. The detail
operations are shown in Fig. 10.
The second stage 1-D DWT works in the similar manner as the first stage 1-D DWT. In the
HH and HL operations, when three column consecutive b(i,j)’s are found in the first stage 1-
D DWT, an HH coefficient can be computed. As soon as two column consecutive HH
coefficients are found, the two HH coefficients and the overlapped b(i,j)’s can be combined
to compute an HL coefficient. Similarly, when three column consecutive c(i,j)’s are found in
the first stage 1-D DWT, an LH coefficient can be computed. As soon as two LH coefficients
are found, the two LH coefficients and the overlapped c(i,j) are used to compute an LL
coefficient. The detailed operations for the second stage 1-D DWT are shown in Fig. 11.

Fig. 9. IRSA of the 2-D LDWT.

84 VLSI

Fig. 10. The detail operations of the first stage 1-D DWT.

(a)
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 85

(b)
Fig. 11. The detailed operations of the second stage 1-D DWT. (a) The HF (HH and HL) part
operations. (b) The LF (LH and LL) part operations.

5. VLSI architecture and implementation for the 2-D dual-mode LDWT

The IRSA approach has been discussed in the previous section, and the architecture of IRSA
is described in this section. We can manipulate the control unit to read off-chip memory. In
IRSA, two pixels are scanned concurrently, and the system needs two processing units. For
the 2-D LDWT processing, the pixels are processed by the first stage 1-D DWT first. The
outputs are then fed to the second stage 1-D DWT to find the four subband coefficients, HH,
HL, LH, and LL. There are two parts in the architecture, the first stage 1-D DWT and the
second stage 1-D DWT. Here we concentrate on the 2-D 5/3 mode LDWT.

5.1 The first stage 1-D LDWT

The first stage 1-D LDWT architecture consists of the following units: signal arrangement
unit, multiplication and accumulation cell (MAC), multiplexer (MUX), and FIFO register.
The block diagram is shown in Fig. 12.
The signal arrangement unit consists of three registers, R1, R2, and R3. The pixels are input
to R1 first, and subsequently the content of R1 is transferred to R2 and then R3, and R1
keeps reading the following pixels. The operation is like a shift register. As soon as R1, R2,
and R3 get signal data, MAC starts operating. The signal arrangement unit is shown in Fig.
13. In Fig. 13 MAC operates at the clock with gray circles.
86 VLSI

Fig. 12. The architecture of the first stage 1-D DWT.

Fig. 13. The operation of the signal arrangement unit (for example, IRAS signal in N1).
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 87

Fig. 14. The block diagram of MAC.

For the low-frequency coefficients calculation we need two high-frequency coefficients and
an original pixel. Internal register R4 is used to store the original even pixel (N1) and
internal register R9 is used to store the original odd pixel (N2). We can simply shift the
content of R3 to R4 after the MAC operation. The FIFO is used to store the high-frequency
coefficients to calculate the low-frequency coefficients. Register R5 has two functions: 1) to
store the high-frequency coefficients for the low-frequency coefficient calculation, 2) to be
used as a signal buffer for MAC. MAC needs time to compute the signal, and the output of
MAC cannot directly feed the result to the output or the following operation may be
incorrect due to the synchronization problems. R5 acts as an output buffer for MAC to
prevent the error in the following operations. In the 5/3 integer lifting-based operations,
MAC is used to find the results of the high-frequency output, (a1+a3)/2 + a2, and the low
frequency output, (a1+a3)/4+a2. There are two multiplication coefficients, 1/2 and 1/4. To
save hardware, we can use shifters to implement the 1/2 and 1/4 multiplications.
Therefore the MAC needs adders, complementer, and shifters. The MAC block diagram is
shown in Fig. 14, where a1, a2, and a3 are the inputs, ‛‛’’ the 2’s complement converter, and
‛‛>>’’ the right shifter.

5.2 The second stage 1-D LDWT

Similar to the first stage 1-D DWT, the second stage 1-D DWT consists of the following
units: signal arrangement unit, MAC, and MUX, as shown in Fig. 15. Due to the parallel
architecture, two outputs are generated concurrently from the first stage 1-D DWT, and
these two outputs must be merged in the second stage 1-D DWT. The signal arrangement
unit processes the signal merging; Fig. 16 shows the processing diagram of the signal
arrangement unit. At beginning, signals H0 and H1 are from IN1 and IN2 and these two
signals are stored in R3 and R4, respectively. At the next clock, H0 and H1 are moved to R1
and R2 respectively, and concurrently new signals H3 and H4 from IN1 and IN2 are stored
to R3 and R4 respectively. The signal arrangement unit operates repeatedly to input signals
for the second stage 1-D DWT.
88 VLSI

Fig. 15. The block diagram of the second stage 1-D LDWT.

Fig. 16. Signal merging process for the signal arrangement unit.

5.3 2-D LDWT architecture

In our IRSA operation, IN1 and IN2 read signals of even row and odd row in a zig-zag
order, respectively. The detail is shown in Fig. 17. The block diagram of the proposed 2-D
LDWT is shown in Fig. 19(b). It consists of two stages, the first stage 1-D DWT and the
second stage 1-D DWT. This architecture needs only a small amount of transpose memory.
Let us consider a 44 image. The signal processing of the first stage 1-D DWT is shown in
Fig. 18. The pixels from the even rows are processed by the upper 1-D DWT, and the pixels
from the odd rows are processed by the lower 1-D DWT. Each 1-D DWT generates one set of
22 high-frequency coefficients and one set of 22 low-frequency coefficients, respectively.
The generated coefficients are fed to the second stage 1-D DWT under the direction of the
arrow head. The input for each second stage 1-D DWT becomes a set of 24 signals. The
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 89

signal processing of the second stage 1-D DWT is shown in Fig. 19. The 24 signals in each
second stage 1-D DWT are then processed, and then HH, HL, LH, and LL are generated and
each has 22 signal data.
The complete architecture of the 2-D LDWT is shown in Fig. 19. The complete 2-D LDWT
consists of four parts, two sets of the first stage 1-D DWT, two sets of the second stage 1-D
DWT, control unit, and MAC unit.

(a)

(b)
Fig. 17. The input signal sequences. (a) IN1 read signal of even row in zig-zag orders. (b) IN2
read signal of odd row in zig-zag orders.
90 VLSI

(a)

(b)
Fig. 18. The signal process of the two stage LDWT. (a) First stage 1-D LDWT. (b) Second
stage 1-D LDWT.

According to (12) and (13), the proposed IRSA architecture can also be applied to the 9/7
mode LDWT. Fig. 20 illustrates the approach. From Figs. 10 and 11 in Section 3, the original
signals (denoted as black circles) for both 5/3 and 9/7 modes LDWT can be processed by the
same IRSA for the first stage 1-D DWT operation. The high-frequency signals (denoted as grey
circles) and the correlated low-frequency signals together with the results of the first stage are
used to compute the second stage 1-D DWT coefficients. Compared to the 9/7 mode LDWT
computation, the 5/3 mode LDWT is much easier for computation, and the registers
arrangement in Figs. 12 and 15 is simple. For 9/7 mode LDWT implementation with the same
system architecture of 5/3 mode LDWT, we have to do the following modifications: 1) The
control signals of the MUX in Figs. 12 and 15 must be modified. We have to rearrange the
registers for the MAC block to process the 9/7 parameters. 2) The wavelet coefficients of the
dual-mode LDWT are different. The coefficients are α= 1/2 and β=1/4 for 5/3 mode LDWT,
but the coefficients are α= −1.586134142, β= −0.052980118, γ= +0.882911075, and δ=
+0.443506852 for 9/7 mode LDWT. For calculation simplicity and good precision, we can use
the integer approach proposed by Huang et al. (Huang et al., 2004) and Martina et al. (Martina
& Masera, 2007) for 9/7 mode LDWT calculation. Similar to the multiplication implementation
by shifters and adders in the 5/3 mode LDWT, we can adopt the shifters approach proposed
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 91

in (Huang et al., 2005) further to implement the 9/7 mode LDWT. 3) According to the
characteristics of the 9/7 mode LDWT, the control unit in Fig. 19(b) must be modified
accordingly.

(a)

(b)
Fig. 19. The complete 2-D DWT block diagram. (a) DSP diagram of the 2-D LDWT. (b) System
diagram of the 2-D LDWT.
92 VLSI

Fig. 20. The processing procedures of 2-D dual-mode LDWTs under the same IRSA
architecture.

The multi-level DWT computation can be implemented in a similar manner by the high
performance 1-level 2-D LDWT. For the multi-level computation, this architecture needs
N2/4 off-chip memory. As illustrated in Fig. 21, the off-chip memory is used to temporarily
store the LL subband coefficients for the next iteration computations. The second level
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 93

computation requires N/2 counters and N/2 FIFO’s for the control unit. The third level
computation requires N/4 counters and N/4 FIFO’s for the control unit. Generally in the jth
level computation, we need N/2j-1 counters and N/2j-1 FIFO’s.

Fig. 21. The multilevel 2-D DWT architecture.

6. Experimental results and comparisons

The 2-D dual-mode LDWT considers a trade-off between low transpose memory and low
complexity in the design of VLSI architecture. Tables 2 and 3 show the performance
comparisons of the proposed architecture and other similar architectures. Compression
results indicate that the proposed VLSI architecture outperforms previous works in terms of
transpose memory size, requiring about 50% less memory than the JPEG2000 standard
(Chen, 2004) architecture. Moreover, the 2-D LDWT is frame-based, and its implementation
bottleneck is the huge transpose memory. Less memory units are needed in our architecture
and the latency is fixed on (3/2)N+3 clock cycles. Our architecture can also provide an
embedded symmetrical extension function. The proposed IRSA approach has the
advantages of memory-efficient and high-speed. The proposed 2-D dual-mode LDWT
adopts parallel and pipelined schemes to reduce the transpose memory and increase the
operation speed. The shifters and adders replace multipliers in the computation to reduce
the hardware cost. Chen et al. (Chen & Wu, 2002) proposed a folded and pipelined
architecture to compute the 2-D 5/3 lifting-based DWT, and they used transpose memory
size of 2.5N for an NN 2-D DWT. This lifting architecture for vertical filtering with two
adders and one multiplier is divided into two parts, and each part has one adder and one
multiplier. Because both parts are activated in different cycles, they can share the same
adder and multiplier. It can increase the hardware utilization and reduce the latency.
However, according to the characteristics of the signal flow, it will increase the complexity
at the same time.
A 256×256 2-D LDWT was designed and simulated with VerilogHDL and further
synthesized by the Synopsys design compiler with TSMC 0.18μm 1P6M CMOS standard
process technology. The detailed specs of the 256×256 2-D LDWT are listed in Table 4.
94 VLSI

5/3 LDWT Ours Diou Andra Chen & Chen, Chiang & Mei Huang Wu &
architecture et al., et al., Wu, 2002 Hsia, et al., et al, Lin, 2005
2001 2002 2002 2005 2006 2005
Transpose 2N 3.5N 3.5N 2.5N 3N N2/4+5N 2N 3.5N 3.5N
memory1
(bytes)
Computation (3/4 --- (N2/2)+ N2 (N2/2)+N N2 (N2/ --- 10+(4/3)
time2 )N2+ N+5 +5 2)+N N2[1-
(3/2 (1/4)]+2N
)N [1-(1/2)]
+7
Adders 8 12 8 6 5 4 8 --- ---
Multipliers 0 6 4 4 0 0 0 --- 6
1 Transpose memory size is used to store frequency coefficients in the 1-L 2-D DWT.
2 In a system, computing time represents the time used to compute an image of size N×N.
3 Suppose the image is of size N×N.

Table 2. Comparisons of 2-D architectures for 5/3 LDWT.

9/7 LDWT Ours Andra Jung & Chen, Vishwanath Huang et Huang Wu & Lin, Lan et Wu. &
architecture et al., Park, 20041 et al., 1995 al., 2005 et al, 2005 al., 2005 Chen,
2002 2005 2005 2001
Transpose 4N N2 12N N2/4+L 22N 14N 5.5N 5.5N --- N2+4N+
memory N+L 4
(bytes)
Computatio (3/4)N2 4N2/3+ N2 N2/2~(2 N2 --- --- 22+(4/3)N2[1 --- 2N2/3
n time +(3/2) 2 /3)N -(1/4)]+6N[1-
N +7 (1/2)]
Adders 16 8 12 4L 36 16 16 8 32 16
Multipliers 0 4 9 4L 36 12 10 6 20 16
1 L:
the filter length.
Table 3. Comparisons of the 2-D architectures for 9/7 LDWT.

Chip specification N = 256, Tile size = 256256

Gate count 29,196 gates

Power supply 1.8V
Technology TSMC 0.18m 1P6M (CMOS)
On-Chip memory size (Transpose 2-D 5/3 DWT: 512 bytes
+ Internal) 2-D 9/7 DWT: 1,024 bytes
Latency (3/2)N+3 = 387 clock cycles

Computing time (3/4)N2+(3/2)N+7 = 49,543 clock cycles

Maximum clock rate 83 MHz

Table 4. Design specification of the proposed 2-D DWT.

Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 95

7. Conclusions
This work presents a new architecture to reduce the transpose memory requirement in 2-D
LDWT. The proposed architecture has a mixed row- and column-wise signal flow, rather
than purely row-wise as in traditional 2-D LDWT. Further we propose a new approach,
interlaced read scan algorithm (IRSA), to reduce the transpose memory for a 2-D dual-mode
LDWT. The proposed 2-D architectures are more efficient than previous architectures in
trading off low transpose memory, output latency, control complexity, and regular memory
access sequence. The proposed architecture reduces the transpose memory significantly to a
memory size of only 2N or 4N (5/3 or 9/7 mode) and reduces the latency to (3/2)N+3 clock
cycles. Due to the regularity and simplicity of the IRSA LDWT architecture, a dual mode
(5/3 and 9/7) 256256 2-D LDWT prototyping chip was designed by TSMC 0.18m 1P6M
standard CMOS technology. The 5/3 and 9/7 filters with different lifting steps are realized
by cascading the four modules (split, predict, update, and scaling phases). The prototyping
chip takes 29,196 gate counts and can operate at 83 MHz. The method is applicable to any
DWT-based signal compression standard, such as JPEG2000, Motion-JPEG2000, MPEG-4
still texture object decoding, and wavelet-based scalable video coding (SVC).

8. References
Andra, K.; Chakrabarti, C. & Acharya, T. (2000). A VLSI architecture for lifting-based
wavelet transform, IEEE Workshop on Signal Processing Systems, (October 2000) pp.
70-79.
Andra, K.; Chakrabarti, C. & Acharya, T. (2002). A VLSI architecture for lifting-based
forward and inverse wavelet transform, IEEE Transactions on Signal Processing, Vol.
50, No.4, (April 2002) pp. 966-977.
Chen, P.-Y. (2002). VLSI implementation of discrete wavelet transform using the 5/3 filter,
IEICE Transactions on Information and Systems, Vol. E85-D, No.12, (December 2002)
pp. 1893-1897.
Chen, P.-Y. (2004). VLSI implementation for one-dimensional multilevel lifting-based
wavelet transform, IEEE Transactions on Computer, Vol. 53, No. 4, (April 2004) pp.
386-398.
Chen, P. & Woods, J. W. (2004). Bidirectional MC-EZBC with lifting implementation, IEEE
Transactions on Circuits and Systems for Video Technology, Vol. 14, No. 10, (October
2004) pp. 1183-1194.
Chen, S.-C. & Wu, C.-C. (2002). An architecture of 2-D 3-level lifting-based discrete wavelet
transform, VLSI Design/ CAD Symposium, (August 2002) pp. 351-354.
Chiang, J.-S. & Hsia, C.-H. (2005). An efficient VLSI architecture for 2-D DWT using lifting
scheme, IEEE International Conference on Systems and Signals, (April 2005) pp. 528-
531.
Christopoulos, C.; Skodras, A. N. & Ebrahimi, T. (2000). The JPEG2000 still image coding
system: An overview, IEEE Trans. on Consumer Electronics, Vol. 46, No. 4,
(November 2000) pp. 1103-1127.
Daubechies, I. & Sweldens, W. (1998). Factoring wavelet transforms into lifting steps, The
Journal of Fourier Analysis and Applications, Vol. 4, No.3, (1998) pp. 247-269.
96 VLSI

Diou, C.; Torres, L. & Robert, M. (2001). An embedded core for the 2-D wavelet transform,
IEEE on Emerging Technologies and Factory Automation Proceedings, Vol. 2, (October
2001) pp. 179-186.
Habibi, A. & Hershel, R. S. (1974). A unified representation of differential pulse code
modulation (DPCM) and transform coding systems, IEEE Transactions on
Communications, Vol. 22, No. 5, (May 1974) pp. 692-696.
Hsia, C.-H. & Chiang, J.-S. (2008). New memory-efficient hardware architecture of 2-D dual-
mode lifting-based discrete wavelet transform for JPEG2000, IEEE International
Conference on Communication Systems, (November 2008) pp. 766-772.
Huang, C.-T.; Tseng, P.-C. & Chen, L.-G. (2002). Efficient VLSI architecture of lifting-based
discrete wavelet transform by systematic design method, IEEE International
Symposium Circuits and Systems, Vol. 5, (May 2002) pp. 26-29.
Huang, C.-T.; Tseng, P.-C. & Chen, L.-G. (2004). Flipping structure: An efficient VLSI
architecture for lifting-based discrete wavelet transform, IEEE Transactions on Signal
Processing, Vol. 52, No. 4, (April 2004) pp. 1080-1089.
Huang, C.-T.; Tseng, P.-C. & Chen, L.-G. (2005). VLSI architecture for lifting-based shape-
adaptive discrete wavelet transform with odd-symmetric filters, Journal of VLSI
Signal Processing Systems, Vol. 40, No. 2, (June 2005) pp.175-188.
Huang, C.-T.; Tseng, P.-C. & Chen, L.-G. (2005). Analysis and VLSI architecture for 1-D and
2-D discrete wavelet transform, IEEE Transactions on Signal Processing, Vol. 53, No.
4, (April 2005) pp. 1575-1586.
Huang, C.-T.; Tseng, P.-C. & Chen, L.-G. (2005). Generic RAM-based architecture for two-
dimensional discrete wavelet transform with line-based method, IEEE Transactions
on Circuits and Systems for Video Technology, Vol. 15, No. 7, (July 2005) pp. 910-919.
ISO/IEC 15444-1 JTC1/SC29 WG1. (2000). JPEG 2000 Part 1 Final Committee Draft Version 1.0,
Information Technology.
ISO/IEC JTC1/SC29/WG1 Wgln 1684 (2000). JPEG 2000 Verification Model 9.0.
ISO/IEC 15444-1 JTC1/SC29 WG1. (2000). Motion JPEG2000, ISO/IEC ISO/IEC 15444-3,
Information Technology.
ISO/IEC JTC1/SC29 WG11. (2001), Coding of Moving Pictures and Audio, Information
Technology.
Jiang, W. & Ortega, A. (2001). Lifting factorization-based discrete wavelet transform based
architecture design, IEEE Transactions on Circuits and Systems for Video Technology,
Vol. 11, No. 5, (May 2001) pp. 651-657.
Jung, G.-C. & Park, S.-M. (2005). VLSI implement of lifting wavelet transform of JPEG2000
with efficient RPA (recursive pyramid algorithm) realization, IEICE Transactions on
Fundamentals, Vol. E88-A, No. 12, (December 2005) pp. 3508-3515.
Kondo, H. & Oishi, Y. (2000). Digital image compression using directional sub-block DCT,
International Conference on Communications Technology, Vol. 1, (August 2000) p p. 985
-992.
Lan, X.; Zheng, N. & Liu, Y. (2005). Low-power and high-speed VLSI architecture for lifting-
based forward and inverse wavelet transform, IEEE Transactions on Consumer
Electronics, Vol. 51, No. 2, (May 2005) pp. 379-385.
Li, W.-M.; Hsia, C.-H. & Chiang, J.-S. (2009). Memory-efficient architecture of 2-D dual-
mode lifting scheme discrete wavelet transform for Moion-JPEG2000, IEEE
International Symposium on Circuits and Systems, (May 2009) pp. 750-753.
Memory-Efficient Hardware Architecture of 2-D Dual-Mode Lifting-Based
Discrete Wavelet Transform for JPEG2000 97

Lian, C.-J.; Chen, K.-F.; Chen, H.-H. & Chen, L.-G. (2001). Lifting based discrete wavelet
transform architecture for JPEG2000, IEEE International Symposium on Circuits and
Systems, Vol. 2, (May 2001) pp. 445-448.
Mallat, S. G. (1989). A theory for multi-resolution signal decomposition: The wavelet
representation, IEEE Transaction on Pattern Analysis and Machine Intelligence, Vol. 11,
No. 7, (July 1989) pp. 674-693.
Mallat, S. G. (1989). Multi-frequency channel decompositions of images and wavelet models,
IEEE Transactions on Acoustics, Speech and Signal Processing, Vol. ASSP-37, No. 12,
(December 1989) pp. 2091-2110.
Marcellin, M. W.; Gormish, M. J. & Skodras, A. N. (2000). JPEG2000: The new still picture
compression standard, ACM Multimedia Workshops, (September 2000) pp. 45-49.
Marino, F. (2000). Efficient high-speed/low-power pipelined architecture for the direct 2-D
discrete wavelet transform, IEEE Transactions on Circuits and Systems II, Vol. 47, No.
12, (December 2000) pp. 1476-1491.
Martina, M. & Masera, G. (2007). Folded multiplierless lifting-based wavelet pipeline, IET
Electronics Letters, Vol. 43, No. 5, (March 2007) pp. 27-28.
Mei, K.; Zheng, N. & van de Wetering, H. (2006). High-speed and memory-efficient VLSI
design of 2-D DWT for JPEG2000, IET Electronics Letter, Vol. 42, No. 16, (August
2006) pp. 907-908.
Ohm, J.-R. (2005). Advances in scalable video coding, Proceedings of The IEEE, Invited Paper,
Vol. 93, No.1, pp. 42-56, (January 2005) pp. 42-56.
Richardson, I. (2003). H.264 and MPEG-4 Video Compression, John Wiley & Sons Ltd.
Seo, Y.-H. & Kim, D.-W. (2007). VLSI architecture of line-based lifting wavelet transform for
Motion JPEG2000, IEEE Journal of Solid-State Circuits, Vol. 42, No. 2, (February 2007)
pp. 431-440.
Sweldens, W. (1996). The lifting scheme: A custom-design construction of biorthogonal
wavelets, Applied and Computation Harmonic Analysis, Vol. 3, No. 15, (1996) pp.186-
200.
Tan, K.C.B. & Arslan, T. (2001). Low power embedded extension algorithm for the lifting
based discrete wavelet transform in JPEG2000, IET Electronics Letters, Vol. 37, No.
22, (October 2001) pp.1328-1330.
Tan, K.C.B. & Arslan, T. (2003). Shift-accumulator ALU centric JPEG 2000 5/3 lifting based
discrete wavelet transform architecture, IEEE International Symposium on Circuits
and Systems, Vol. 5, (May 2003) pp. V161-V164.
Taubman, D. & Marcellin, M. W. (2001). JPEG2000 image compression fundamentals, standards,
and practice, Kluwer Academic Publisher.
Varshney, H.; Hasan, M. & Jain, S. (2007). Energy efficient novel architecture for the lifting-
based discrete wavelet transform, IET Image Process, Vol. 1, No. 3, (September 2007)
pp.305-310.
Vishwanath, M.; Owens, R. M. & Irwin, M. J. (1995). VLSI architecture for the discrete
wavelet transform, IEEE Transactions on Circuits and Systems II, Vol. 42, No. 5, (May
1995) pp. 305-316.
Weeks, M. & Bayoumi, M. A. (2002). Three-dimensional discrete wavelet transform
architectures, IEEE Transactions on Signal Processing, Vol. 50, Vo.8, (August 2002) pp.
2050-2063.
98 VLSI

Wu, B.-F. & Lin, C.-F. (2005). A high-performance and memory-efficient pipeline
architecture for the 5/3 and 9/7 discrete wavelet transform of JPEG2000 codec,
IEEE Transactions on Circuits and Systems for Video Technology, Vol. 15, No. 12,
(December 2005) pp. 1615-1628.
Wu, P.-C. & Chen, L.-G. (2001). An efficient architecture for two-dimensional discrete
wavelet transform, IEEE Transactions on Circuits and Systems for Video Technology,
Vol. 11, No. 4, (April 2001) pp. 536-545.
Full HD JPEG XR Encoder Design for Digital Photography Applications 99

X5

Full HD JPEG XR Encoder Design

for Digital Photography Applications
Ching-Yen Chien*, Sheng-Chieh Huang*, Chia-Ho Pan,
and Liang-Gee Chen
DSP/IC Design Lab, Graduate Institute of Electronics Engineering
and Department of Electrical Engineering, National Taiwan University
*Sense/TCM SOC Lab, Department of Electrical and Control Engineering,
National Chiao-Tung University
Taiwan, R.O.C.

1. Introduction
Multimedia applications, such as radio, audio, camera phone, digital still camera, camcoder,
and mobile broadcasting TV, are more and more popular in our life as the progress of image
sensor, communication, VLSI manufacture, and image/video coding standards. With rapid
progress of image sensor, display devices, and computing engines, image coding standards
are used in the digital photography application everywhere. It has been merged together
with our life such as camera phone, digital still camera, blog and many other applications.
Many advanced multimedia applications require image compression technology with
higher compression ratio and better visual quality. High quality, high compression rates of
digital image and low computational cost are important factors in many areas of consumer
electronics, ranging from digital photography to the consumer display equipments such as
digital still camera and digital frame. These requirements usually involve computationally
intensive algorithms imposing trade-offs between quality, computational resources, and
throughput.
For high quality of digital image applications, the extension of color range has becoming
more important in the consumer product. In the past, the digital cameras and the display
equipments in the consumer market typically had 8 bits information per channel. Today the
condition is quite different. In the consumer market, digital cameras and the desktop display
panels also have at least 12 bits of information per channel. If the information per channel of
digital image is still compressed into 8 bits, 4 or more bits of information per channel are lost
and the quality of the digital image is limited. Due to the improvement of the display
equipments, the JPEG XR is designed for the high dynamic range (HDR) and the high
definition (HD) photo size. JPEG XR which is already under organized by the ISO/IEC Joint
Photographic Experts Group (JPEG) Standard Committee is a new still image coding
standard and derived from the window media photo (Srinivasan et al., 2007; Srinivasan et
al., 2008; Schonberg et al., 2008). The goal of JPEG XR is to support the greatest possible level
100 VLSI

of image dynamic range and color precision, and keep the device implementations of the
encoder and decoder as simple as possible.
For the compression of digital image, the Joint Photographic Experts Group, the first
international image coding standard for continuous-tone natural images, was defined in
1992 (ITU, 1992). JPEG is a well-known image compression format today because of the
population of digital still camera and Internet. But JPEG has its limitation to satisfy the rapid
progress of consumer electronics. Another image coding standard, JPEG2000 (ISO/IEC,
2000), was finalized in 2001. Differed from JPEG standard, a Discrete Cosine Transform
(DCT) based coder, the JPEG2000 uses a Discrete Wavelet Transform (DWT) based coder.
The JPEG2000 not only enhances enhances the compression, but also includes many new
features, such as quality scalability, resolution scalability, region of interest, and
lossy/lossless coding in a unified framework. However, the design of JPEG2000 is much
complicated than the JPEG standard. The core techniques and computation complexity
comparisons of these two image coding standard are shown in (Huang et al., 2005).
For satisfaction of the high quality image compression and lower computation complexity,
the new JPEG XR compression algorithm is discussed and implemented with the VLSI
architecture. JPEG XR has high encoding efficiency and versatile functions. The XR of JPEG
XR means the extended range. It means that JPEG XR supports the extended range of
information per channel. The image quality of JPEG XR is nearly equal to JPEG 2000 with
the same bit-rate. The computation complexity is much lower than JPEG2000 as shown in
Table 1.
The efficient system-level architecture design is more important than the module design
since system-level improvements make more impacts on performance, power, and memory
bandwidth than the module-level improvements. In this chapter, the new JPEG XR
compression standard is introduced and the analysis and architecture design of JEPG XR
encoder are also proposed. Comparison was made to analyze the compression performance
among JPEG2000 and JPEG XR. Fig. 1 is the peak signal-to-noise ratio (PSNR) results under
several different bitrates. The test color image is 512x512 Baboon. The image quality of JPEG
XR is very close to that of JPEG2000. The PSNR difference between JPEG XR and JPEG2000
is under 0.5dB. Fig. 2 shows the subjective views at 80 times compression ratio. The block
artifact of JPEG image in Fig. 2 is easily observed, while the JPEG XR demonstrates
acceptable qualities by implementing the pre-filter function. For the architecture design, a
4:4:4 1920x1080 JPEG XR encoder is proposed. From the simulation and analysis, entropy
coding is the most computationally intensive part in JPEG XR encoder. We first proposed a
timing schedule of pipeline architecture to speed up the entropy encoding module. To
optimize memory bandwidth problem and maximize the silicon area efficiency, we also
proposed a data reuse skill to solve this problem. The data reuse skill can reduce 33%
memory bandwidth form the memory access. The hardware design of JPEG XR encoder has
been implemented by cell-based IC design flow. The encoder design is also verified by
FPGA platform. This JPEG XR chip design can be used for digital photography applications
to achieve low computation, low storage, and high dynamical range features.

Technologies Operations(GOPs)
JPEG2000 4.26
JPEG XR 1.2
Table 1. Machine cycles comparison between JPEG XR and JPEG 2000.
Full HD JPEG XR Encoder Design for Digital Photography Applications 101

PSNR BABOON512

60

55

50

45 JPEG XR(O=1)
JPEG 2000
40 JPEG XR(O=0)
JPEG XR(O=2)
35

30

25

20 bpp
0 1 2 3 4 5 6 7

Fig. 1. PSNR Comparison between different image standard.

(a) (b)
Fig. 2. Image Quality Comparison with (a) JPEG XR and (b) JPEG.
102 VLSI

Fig. 3. JPEG XR Coding Flow.

The coding flow of JPEG XR is shown in the Fig. 3. The JPEG XR has lower computation cost
in each module and simple coding flow while maintaining similar PSNR quality at the same
bitrate as compared with the coding flow of JPEG2000. Hence, the JPEG XR is very suitable
to be implemented with the dedicated hardware to deal with HD photo size image for the
HDR display requirement. The JPEG XR encoder architecture design is presented in the
following section.
This paper is organized as follows. In Section 2, we present the fundamentals of JPEG XR.
Section 3 describes the characteristics of our proposed architecture design of JPEG XR
encoder. Section 4 shows the implementation results. Finally, a conclusion is given in
Section 5.

2. JPEG-XR
The JPEG XR image compression standard has many options for different purposes. In the
following section, the fundamentals of JPEG XR are introduced.

2.1 Image Partition and Windowing

Figure 4 shows the partition and windowing example of JPEG XR standard. As the different
tile size makes the different compression result, the compression results under the same
quantization with different tile size for four test 512x512 benchmark images are discussed in
(Pan et al., 2008). When the tile function is used, the compressed image size is increased with
the tiles number when the small tiles size has been chosen.
When the image has been divided into the different tiles, each tiles are processed
independently which are more suitable for the design of hardware implementation and
memory buffer size. The pixels acrossing the boundary of the tiles have to be processed
without any data dependency. The data dependancy must be changed accordingly, and the
scan order has to be rebuilt as well. This characteristic overcome the error impact of data
independency when the error occurs. Although the division of tiles is helpful for the
hardware implementation and data reserving in error environment, but it decreases the data
compression efficiency. However, it depends on the tradeoff between hardware design
considerations and the data compression efficiency.
Full HD JPEG XR Encoder Design for Digital Photography Applications 103

MB aligned
Tiles
image

Input
image

Macroblock

Blocks

Fig. 4. Example of Image Partitions and Windowing.

IMG_HDR INDEXTBL TILE1 TILE2

MB_1 MB_2 MB_3

Spatial mode

DC LOWPASS HIGHPASS FLEXBITS

Frequency mode

Fig. 5. Layout of Bitstream.

2.2 Operation Modes

In the Fig. 5, there are two modes of bitstream in JPEG XR which are named as spatial mode
and frequency mode. In the spatial mode, a single tile packet carries the bitstream in
macroblock raster order scanning from left to right and top to bottom. In the frequency
mode, the bitstream of each tile is carried in multiple tile packets, where each tile packet
carries transform coefficients of each frequency band in the tile. The freqency mode
transmits the DC coefficients at first. When the user receive the bitstreams, images are
decoded progressively. The frequency mode is very suitable for limited bandwidth
applications such as web browsing.
104 VLSI

Fig. 6. The 1st part and 2nd part process of PCT.

2.3 Color Conversion

The purpose of color conversion is to reflect the color sensivity according to the
characteristics of human eyes. The color space is converted from RGB to YUV. Then, the
important energies are concentrate on Y channel. The color conversion is reversible, in other
words, this color conversion is a lossless conversion. The color conversion equation is
V=B-R
 V  
U = - R - G +   (1)
  2 
U 
Y = G +    offset
2
where offset is 128.

2.4 Pre-filter
There are three overlapping choices of the pre-filter function: non-overlapping, one-level
overlapping and two-level overlapping. (Pan et al., 2008) discuss different trade-offs of three
overlapping choices. The non-overlapping condition is used for fastest encoding and
decoding. It is efficient in low compression ratio mode or lossless mode. However, this
mode potentially introduces blocking effect in low bitrate images. The one-level overlapping
function, compared to the non-overlapping, has higher compression ratio but it needs
additional time for the overlapping operation. The two-level overlapping has the highest
computation complexity. The PSNR and objective image quality of two-level overlapping
are better than the other two at low bitrate area. The pre-filter function is recommended for
both high image quality and further compression ratio considerations at high quantization
levels. For high image quality issue, the pre-filter function eliminate the block effect which is
sensitive of visual quality.

2.5 Photo Core Transform

Photo Core Transform (PCT) function transforms the pixel values after pre-filter computing
to the lowest frequency coefficient DC, low frequency coefficients AD, and high frequency
coefficients AC. There are two parts process for the PCT as shown in Fig. 6. The macroblock
(MB) is partitioned into 16 4x4 blocks. In each part process, each 4x4 block is pre-filtered and
then transformed by 4x4 PCT. A 2x2 transform is applied to four coefficients by four times
Full HD JPEG XR Encoder Design for Digital Photography Applications 105

for each 4x4 block in first part process. The low frequency coefficient of these four
coefficients is processed to the top-left coefficient. After first part process, the DC coefficients
of 16 4x4 blocks can be collected as a 4x4 DC block. The second part process is for the 4x4
DC block from the first part process. The second part 2D 4x4 PCT is built by using the three
operators: 2x2 T_h, T_odd and T_odd_odd. After second part process, the 16 DC coefficients
are processed as DC and AD coefficients.

2.6 Quantization
The quantization is the process of rescaling the coefficients after applying the transform
process. The quantization uses the quantized value to divide and round the coefficients after
PCT transformed into an integer value. For the lossless coding mode, the quantized value =
1. For the lossy coding mode, the quantized value > 1. The quantization of JPEG XR use
integer operations. The advantage of integer operation keeps the precision after scaling
operations and only uses shift operation to perform the division operations. The
quantization parameter is allowed to differ across high pass band, low pass band, and DC
band. It varies to different values according to the sensitivity of human vision in different
coefficient bands. Fig. 7 shows a example that pixels transformed by 2 stage PCT and
quantized by the quantization process.

Fig. 7. Example of PCT.

Fig. 8. DC prediction model.

106 VLSI

(a) (b)
Fig. 9. Example of (a) Prediction of AD coefficients. (b) Prediction model and AC
coefficients.

2.7 Prediction
There are three directions of DC prediction: LEFT, TOP, and LEFT and TOP. As shown in
Fig. 8, the JPEG XR prediction rules of DC coefficient use the DC coefficient of left MB and
top MB to decide the direction of DC prediction. The DC predicted direction can be decided
by the pseudo code described in Fig. 8. After comparing the H_weight and the V_weight,
the predicted direction can be decided. At the boundary MBs of image, the purple MBs only
predict from left MB and the gray MBs only predict from top MB.
The AD/AC block can be predicted from its TOP or LEFT block as Fig. 9. If the predicted
direction is LEFT, the prediction relationship example of AD coefficients is shown as Fig.
9(a). The AD predicted direction follows DC prediction model as described in Fig. 8. The
computation after prediction judgment of AC is a little similar to AD that can reduce the
coefficient value of the block. The Fig. 9(b) shows the prediction relationship example of AC
coefficients when the prediction direction is TOP.

2.8 Entropy Encoding

The adaptive scan is used for entropy coding which is based on the latest probability of non-
zero coefficients. Fig. 10 shows an example that the previous scan order is changed to
update scan order based on the probability. The most probable non-zero coefficients are
scanned firstly and the probability is counted by numbers of the non-zero coefficients. If the
non-zero probability is larger than the previous scanned coefficient, the scan order is
exchanged. By doing so, the non-zero coefficients are collected together and processed in an
orderly fashion.
After the coefficients of current block are scanned by adaptive scan order, JPEG XR uses two
entropy coding schemes. Fig. 11 demonstrates an example of coding a 4x4 block when
ModelBits is 4 bits. The coefficients of 4x4 block are scanned by adaptive scan order first,
then the ModelBits is updated by the total overhead coefficients in each band per channel.
The coefficients which can be represented under the ModelBits are encoded by the FlexBits
table. For the extra bit such as the 17 can not be represented in four bits, the encoding block
Full HD JPEG XR Encoder Design for Digital Photography Applications 107

will increase extra bit for the new Run-Level Encode (RLE) coding. The RLE function is
added into the bitstream length for the overhead coefficient. The Levels, the Runs before
nonzero coefficients, and the number of overhead coefficients are counted for RLE. Then the
RLE block is encoded firstly while different size of Huffman table is used to make the bit
allocation in optimization. After processing the RLE algorithm, the RLE results and FlexBits
are packetized.

3. Architecture Design of JPEG XR

In the following section, a JPEG XR encoder architecture design is presented. For the system
requirement, the consideration of functional block pipelining, and the architecture design of
the pre-filter, transform, quantization, prediction, and the entropy coding modules are all
discussed in the following sections.

Fig. 10. Adaptive scan order updating example.

Fig. 11. Adaptive scan and Run-Level encode example.

108 VLSI

Fig. 12. Pipeline stage consideration of JPEG XR.

Fig. 13. Data flow of JPEG XR stage 1.

3.1 System Architecture and Functional Block Pipelining

The pipeline stages of this JPEG XR encoder design can be divided into three steps, as
shown in Fig. 12. Since the color conversion, pre-filter and the PCT modules are computed
with the 4x4 block matrix style with no feedback information, they are arranged into the
same stage at the beginning. The prediction unit is used as second stage for different
direction comparisons with the feedback processed pixels for the DC/AD/AC region. Then,
the entropy encoding module with high data dependency are divided as the third stage.

3.2 Color Conversion, Pre-filter, PCT, and Quantization

Fig. 13 shows the data flow of stage 1. A 32 bits external bus can transmit 2 R/G/B
coefficients in one clock cycle. The pre-filter and PCT/quantization modules process the
coefficients after color conversion. This paper uses a memory reused method to reduce the
memory access bandwidth for the color conversion, pre-filter, and PCT module. The black
line in Fig. 14 is the boundary of MB. When the blue range is to be processed after the yellow
range, the pre-filter function only have to deal with the amount of new data instead of the
entire block data into the register. This is due to the presence of previously stored column
coefficients of blocks in the registers and the capability to be utilized as the coefficients in the
yellow range. Therefore, the memory buffer size and the memory access for pre-filter are
reduced. Otherwise, the additional SRAM will be needed to store the data of the entire block
from off-chip memory and the execution time will be increased. The detail simulations is
discussed in (Pan et al., 2008).
Full HD JPEG XR Encoder Design for Digital Photography Applications 109

Fig. 14. Memory reused method.

Fig. 15. DC block insertion on the block pipeline.

Fig. 16. Architecture of PCT, and Quantization.

Because the functions of stage 1 are computing with the 4x4 block style with no feedback
information, the three pipeline architecture for stage1 is used: color conversion (CC), pre-
filter, PCT (include quantization). For the memory allocation, the color conversion requires 6
blocks per column, the pre-filter requires 5 blocks, and the PCT including quantization uses
4 blocks to execute the function. At the end of PCT in Fig. 15, DC block can be processed to
reduce one pipeline bubble when the next new pipeline stage starts. Hence, the well
arranged timing schedule eliminated the access of additional clock cycle to process the DC
block.
The architecture for PCT including Quantization is shown in Fig. 16. Additional registers are
implemented to buffer the related coefficients for the next pipeline processing. The left
multiplex selects the inputs for two parts PCT process. Initially, the input pre-filtered
coefficients are selected to process the PCT algorithm. Then, the yellow block (DC) will be
processed after the 16 blocks have been computed. The quantization stage de-multiplex the
DC, low pass band AD and high pass band AC coefficients to suitable process element. The
processed data are arranged into the quantized coefficients of Y, U, V SRAM blocks for
prediction operation.
110 VLSI

Top AD block

SRAM 3
1440x4 bytes

Pipeline
Buffer
1

Reg
Pipeline
0 Stage 3
Buffer
SRAM 2
DC/AD/AC block 768x4 bytes Output
Input

Reg
SRAM 1
Stage 1 768x4 bytes
CBP Stage 3
Predictor Buffer
Output

Fig. 17. Prediction architecture.

Fig. 18. Adaptive encode architecture.

3.3 Prediction
The quanitzed data stored in Y, U, V SRAM blocks are processed with the subtract operation
as the prediction algorithm. Three SRAM blocks are used in the this design. One 1440x4 byte
SRAM is used to buffer the 1 DC and 3 AD coefficients for the TOP AD prediction
coefficients in the prediction judgement, so that the regeneration of these data are
unnecessary when they are selected in the prediction mode. In Fig. 17, two 768x4 bytes
SRAMs are used to save the quantized coefficients of current block and the predicted
coefficients for current block.

3.4 Entropy Encoding

There are three complex data dependency loops in the entropy encoding module as shown
in Fig. 18. The first one is the adaptive-scan-order block used to refresh the scan order. The
second is the updated ModelBits block, which decides how many bits are necessary to
represent a coefficient. Third, the adaptive Huffman encode block to choose the most
efficient Huffman table.
Full HD JPEG XR Encoder Design for Digital Photography Applications 111

Fig. 19. Timing scheduling of entropy encoding.

Gate Level
Algorithm Analysis DfT Considerations Timing Driven P&R
Simulation

Post Layout Gate

C Simulation Syntest for Bist Power Compiler
Level Simulation

No
Spec. Architecture Design Logical synthesis Satisify? Meet Spec. ?
No
Yes Yes

HDL Design
Scan chain insertion Power/Floor Plane LPE
Simulation

Post Layout
Logical synthesis Tetra-MAX Apollo basic flow
Simulation

No No No No
Meet Spec. ? Coverage enough? Highest Utilization? Meet Spec. ?

Yes Yes Yes Yes

Fabrication

Fig. 20. Chip design flow.

In order to increase the throughput and decrease the timing of critical patch, we use three
sub-pipeline stages architecture to implement the entropy encoding module as described in
(Chien et al., 2009). Because of the feedback patch, stage 1 includes the modules of feedback
path and the feedback information generator. Stage 2 includes the modules that encode the
RLE information to symbol-based data. The bit-wise packetizer module of stage 3 processes
the symbol-based data bit by bit. By doing this, we can increase the throughput about 3
times by well arranged pipeline timing schedule as shown in Fig. 19.
The design of adaptive scan block counts the numbers of the non-zero coefficients to decide
whether the scan order should be exchanged or not. After the processing of the adaptive
scan block, the coefficients which can be represented under ModelBits are coded by the
FlexBits table. The other coefficients change to generate the Flexbits and Level. After the
processing of the RLE algorithm, the Level and Run choose the suitable Huffman table to
generate the RLE codeword. The RLE results are translated into the codewords by different
Huffman tables. The Huffman encoder in JPEG XR is different from the other standards. It is
composed of many small Huffman tables and can adaptively choose the best option in these
tables for the smaller codesize for Run and Level. Many codewords are produced after the
Huffman encoding. In order to increase the throughput, the codeword concentrating
112 VLSI

architecture is proposed. Linked to the output of the above operation, the whole RLE
codeword and RLE codesize will be produced to the packetizer. The packetizer architecture
is modified from the (Agostini et al., 2002) architecture by combining the RLE codeword and
the FlexBits for generating the JPEG XR compressed file. More detail design is described in
(Pan et al., 2008).

4. Implementations
This design is implemented to verify the proposed VLSI architecture for JPEG XR encoder.
And it is also verified by FPGA platform. The detail information about implementation
result of each module by the FPGA prototype system is shown in (Pan et al., 2008). It is used
to test the conformance bitstreams for the certification.
A three-stage MB pipelining of 4:4:4 lossless JPEG XR encoder was proposed to process the
capacity and hardware utilization. In our design, the extra registers are used to increase the
pipeline stages for achieving the specification, such as the color conversion, PCT/
quantization and the adaptive encode block. And the on-chip SRAM blocks are used to store
the reused data processed with the prediction module to eliminate the memory access. For
the entropy encoding module, the timing schedule and pipelining is well designed. The
proposed architecture of entropy encoding module increases the total throughput about
three times. We use 0.18um TSMC CMOS 1P6M process to implement the JPEG XR encoder.
Our design flow is standard cell based chip design flow. The design flow of our design is
shown as Fig 20. Test consideration is also an important issue in chip design. Therefore, the
scan chain and the built-in self-test (BIST) are considered in our chip. The chip synthesis
layout is shown as Fig. 21. The implementation results are shown as the Table 2. The power
dissipation distribution in shown as Fig. 22.

Fig. 21. Chip synthesis layout.

Full HD JPEG XR Encoder Design for Digital Photography Applications 113

Fig. 22. Chip power dissipation distribution.

Technology TSMC 0.18um CMOS 1P6M

Core Size 3.18 mm x 3.18 mm (9.3025 mm2)
Die Size 4.64 mm x 4.64 mm (20.47 mm2)
Gate Count 651.7 K
Work Clock Rate 62.5 MHz
Power Consumption 368.7 [email protected]，1.8V
On-Chip Memory 256x32 SRAM x 6 (single port)
768x32 SRAM x 2 (single port)
480x32 SRAM x 3 (single port)
Total: 144,384 Bits = 18,047 Bytes
Processing 34.1 Mega pixels within one second
Capability 5.45 fps for 4:4:4 HDTV(1920x1080) @62.5MHz
32.9 fps for 4:4:4 VGA(640x480) @62.5MHz
112 fps for 4:4:4 CIF(352x288) @62.5MHz
Input Pad 37
Output Pad 34
Table 2. Chip specification.

5. Conclusion
Compared with the JPEG2000, the coding flow of the JPEG XR is simple and has lower
complexity in the similar PSNR quality at the same bit rate. Hence, the JPEG XR is very
suitable for implementation with the dedicated hardware used to manage HD photo size
images for the HDR display requirement. In this paper, we initially analyzed the
comparison of JPEG XR with other image standard, and then a three-stage MB pipelining
was proposed to process the capacity and hardware utilization. We also made a lot of efforts
on module designs. The timing schedule and pipelining of color conversion, pre-filter, PCT
& quantization modules are well designed. In order to prevent accessing the coefficients
from off-chip memory, an on-chip SRAM is designed to buffer the coefficients for the
prediction module with only some area overhead. The pre-filter and PCT function was
designed to reduce 33.3% memory access from off-chip memory.For the entropy coding, we
designed a codeword concentrating architecture for the throughput increasing of RLE
algorithm. And the adaptive encode and packetizer modules efficiently provide the coding
information required for packing the bitstream. Based on this research result, we contribute
114 VLSI

a VLSI architecture for 1920x1080 HD photo size JPEG XR encoder design. Our proposed
design can be used in those devices which need powerful and advanced still image
compression chip, such as the next generation HDR display, the digital still camera, the
digital frame, the digital surveillance, the mobile phone, the camera and other digital
photography applications.

6. References
B. Crow, Windows Media Photo: A new format for end-to-end digitalimaging, Windows
Hardware Engineering Conference, 2006.
C.-H. Pan; C.-Y. Chien; W.-M. Chao; S.-C. Huang & L.-G. Chen, Architecture design of full
HD JPEG XR encoder for digital photography applications, IEEE Trans. Consu. Elec.,
Vol. 54, Issue 3, pp. 963-971, Aug. 2008.
C.-Y. Chien; S.-C. Huang; C.-H. Pan; C.-M. Fang & L.-G. Chen, Pipelined Arithmetic
Encoder Design for Lossless JPEG XR Encoder, IEEE Intl. Sympo. on Consu. Elec.,
Kyoto, Japan, May 2009.
D. D. Giusto & T. Onali. Data Compression for Digital Photography: Performance
comparison between proprietary solutions and standards, IEEE Conf. Consu. Elec.,
pp. 1-2, 2007.
D. Schonberg; S. Sun; G. J. Sullivan; S. Regunathan; Z. Zhou & S. Srinivasan, Techniques for
enhancing JPEG XR / HD Photo rate-distortion performance for particular fidelity
metrics, Applications of Digital Image Processing XXXI, Proceedings of SPIE, vol. 7073,
Aug. 2008.
ISO/IEC JTC1/SC29/WG1. JPEG 2000 Part I Final Committee Draft, Rev. 1.0, Mar. 2000.
ITU. T.81 : Information technology - Digital compression and coding of continuous-tone still
images. 1992.
L.V. Agostini; I.S. Silva & S. Bampi, Pipelined Entropy Coders for JPEG compression,
Integrated Circuits and System Design, 2002.
S. Groder, Modeling and Synthesis of the HD Photo Compression Algorithm, Master Thesis,
2008.
S. Srinivasan; C. Tu; S. L. Regunathan & G. J. Sullivan, HD Photo: a new image coding
technology for digital photography, Applications of Digital Image Processing XXX,
Proceedings of SPIE, vol. 6696, Aug. 2007.
S. Srinivasan; Z. Zhou; G. J. Sullivan; R. Rossi; S. Regunathan; C. Tu & A. Roy, Coding of
high dynamic range images in JPEG XR / HD Photo, Applications of Digital Image
Processing XXXI, Proceedings of SPIE, vol. 7073, Aug. 2008.
Y.-W. Huang; B.-Y. Hsieh; T.-C. Chen & L.-G. Chen, Analysis, Fast Algorithm, and VLSI
Architecture Design for H.264/AVC Intra Frame Coder, IEEE Trans. Circuits Syst.
Video Technol., vol. 15, no. 3, pp. 378-401, Mar. 2005.
The Design of IP Cores in Finite Field for Error Correction 115

X6

The Design of IP Cores in Finite Field

for Error Correction
Ming-Haw Jing, Jian-Hong Chen, Yan-Haw Chen,
Zih-Heng Chen and Yaotsu Chang
I-Shou University
Taiwan, R.O.C.

1. Introduction
In recent studies, the bandwidth of communication channel, the reliability of information
transferring, and the performance of data storing devices become the major design factors in
digital transmission /storage systems. In consideration of those factors, there are many
algorithms to detect or remove the noisefrom the communication channel and storage media,
such as cyclic redundancy check (CRC) and errorcorrecting code (Peterson & Weldon, 1972;
Wicker, 1995). The former, a hush function proposed by Peterson and Brown (Peterson &
Brown, 1961), is utilized applied in the hard disk and network for error detection; the later is
a type of channel coding algorithms recover the original data from the corrupted data
against various failures. Normally, the scheme adds redundant code(s) to the original data
to provide reliability functions such as error detection or error correction. The background
of this chapter involves the mathematics of algebra, coding theory, and so on.
In terms of the design of reliable components by hardware and / or software
implementations, a large proportion of finite filed operations is used in most related
applications. Moreover, the frequently used finite field operations are usually simplified and
reconstructed into the hardware modules for high-speed and efficient features to replace the
slow software modules or huge look-up tables (a fast software computation). Therefore, we
will introduce those common operations and some techniques for circuit simplification in
this chapter. Those finite field operations are additions, multiplications, inversions, and
constant multiplications, and the techniques include circuit simplification, resource-sharing
methods, etc. Furthermore, the designers may use mathematical techniques such as group
isomorphism and basis transformation to yield the minimum hardware complexities of
those operations. And, it takes a great deal of time and effort to search the optimal designs.
To solve this problem, we propose the computer-aided functions which can be used to
analyze the hardware speed/complexity and then provide the optimal parameters for the IP
design.
This chapter is organized as follows: In Section 2, the mathematical background of finite
field operations is presented. The VLSI implementation of those operations is described in
Section 3. Section 4 provides some techniques for simplification of VLSI design. The use of
116 VLSI

computer-aided functions in choosing the suitable parameters is introduced in Section 5.

Finally, the result and conclusion are given.

2. The mathematic background of finite field

Elements of a finite field are often expressed as a polynomial form over GF(q), the
characteristic of the field. In most computer related applications, the Galois field with
characteristic 2 is wildly used because its ground field, GF(2), can be mapped into bit-0 and
bit-1 for digital computing. For convenience, the value within two parenthesises indicates
that the coefficients for a polynomial in descending order. For example, the polynomial,
x 6  x 5  x 3  1 , is represented by {1101001} in binary form or {69} in hexadecimal form. So
does an element   GF( 2 m ) is presented as symbol based polynomial.

2.1 The common base representations

2.1.1 The standard basis

If an element   GF( 2 m ) is the root of a degree m irreducible polynomial f (x ) , i.e.,
f ( )  0 , then the set 1,  1 ,  2 ,  ,  m1  forms a basis, is called a standard basis, a
polynomial basis or a canonical basis (Lidl & Niederreiter, 1986). For example, construct
E  GF( 2 4 ) with the degree 4 irreducible polynomial f ( x )  x 4  x  1 , suppose f ( )  0 ,
that is,  4    1 and E    { 0} as Table 1.

element 3 2 1 0 element 3 2 1 0
0 0 0 0 0  7
0 1 1 1
0 0 0 0 1 8 1 1 1 0
1 0 0 1 0  9
0 1 0 1
2 0 1 0 0  10 1 0 1 0
3 1 0 0 0  11
1 1 0 1
4 1 0 0 1  12
0 0 1 1
5 1 0 1 1  13 0 1 1 0
6 1 1 1 1  14
1 1 0 0
Table 1. The standard basis expression for all elements of E  GF( 2 4 )

2.1.2 The normal basis

For a given GF( 2 m ) , there exists a normal basis  ,  2 ,  2 , ,  2 . 
m 1
Let   bi 2 be
2 m 1 i

i 0

represented in a normal basis, and the binary vector b0 , b1 ,bm1  is used to represent the
coefficients of  , denoted by   b 0 , b 1 ,  bm1  . Since  2  1   2 by Fermat’s little theorem
m 0

(Wang et al., 1985),  2  bm1 2  b0 2    bm2 2  bm1 , b0 ,  bm2 

0 1 m1

or  2  bmi , bmii ,  , bm1 , b0 , b1 ,  , bmi1  . That is, the squaring operations ( 2 i th power
i

operations) can be constructed by cyclic rotations in software or by changing lines in

The Design of IP Cores in Finite Field for Error Correction 117

hardware, which is with low complexity for practical applications (Fenn et al., 1996).

2.1.3 Composite field

For circuit design, using a composite field to execute some specific operations is an effective
method, for example, the circuit of finite field inversion obtained in composite filed has the
minimum complexity. The famous example is found in most hardware designs of AES VLSI
(Hsiao et al., 2006; Jing et al., 2007), in which the S-box is a non-linear substitution for all
elements in GF( 2 8 ) can be designed with a less area complexity by several isomorphism
composite fields such as GF(( 2 2 ) 4 ) , GF(( 2 4 ) 2 ) , and GF((( 2 2 )2 )) (Morioka & Satoh, 2003). In
this section, we introduce the process to construct a composite field and the basis
transformation between a standard basis and a basis in composite field.
Let GF ( 28 ) be represented in a standard basis with relation polynomial
f ( x )  x 8  x 4  x 3  x 2  1 ( f (x ) is primitive) and f ( )  0 such that   GF( 2 8 ) and
   r is a primitive element in the ground field GF( 2 4 ) , where r  2 8  1 2 4  1  17 . We
construct the composite field GF(( 2 4 )2 ) over the field GF( 2 4 ) using the irreducible
polynomial q(x ) with degree 2 over GF( 2 4 ) , which is given as follows

q( x )  ( x   )( x   2 )  x 2  (   2 )x   17  x 2   34 x   17 . (1)
4 4

Such that    17 is an element of GF( 2 2 ) . In order to represent the elements of the ground
field GF( 2 2 ) , we use the term in q(x ) as the basis element, which is    17 . An element A is
expressed in GF(( 2 4 ) 2 ) as

A  a0  a1 . (2)

where aj  GF( 2 4 ) . We can express aj in GF( 2 4 ) using    17 as the basis element

aj  a j 0  a j 0  a j 0  2  a j 0 3  a j 0  a j 1 17  a j 1 34  a j 1 51 . (3)

where a ji  GF( 2 ) for j  0 ,1 and i  0 ,1,2 ,3 . Therefore, the representation of A in the

composite field is obtained as

A  a0  a1  ( a 00  a 01 17  a 02 34  a 03 51 )  ( a 10  a 11 18  a 12 35  a 13 52 ) . (4)

Next, substitute the terms  17 i  j for j  0 ,1 and i  0 ,1,2 ,3 by the relation polynomial
f ( x )  x  x  x  x  1 as follows:
8 4 3 2

 17   7   4   3 ,  34   6   3   2   1 ,  51   3   1 ,
(5)
 18   5   3   2  1,  35   7   4   3   2 ,  52   4   2 .

By substituting the above terms in expression Equation (4), we obtain the representation of
118 VLSI

A in the standard basis ( 1,  ,  1 ,  7 ) as

A   0  a 1 1  a 2 2  a 3 3  a 4 4  a 5 5  a 6 6  a7 7 . (6)

The relationship between the terms ah for h  0 ,1,  ,7 and a ji for j  0 ,1 and i  0 ,1,2 ,3
determines a 8 by 8 conversion matrix T (Sunar et al., 2003). The first row of the matrix T
is obtained by gathering the constant terms in the right hand side of Equation (4) after the
substitution, which gives the constant coefficients in the left hand side, i.e., the term a0 . A
simple inspection shows that  0  a00  a11 . Therefore, we obtain the 8  8 matrix T and this
matrix gives the representation of an element in the binary field GF( 2 8 ) given its
representation in the composite field GF(( 2 4 )2 ) as follows:

 a0   1 0 0 0 0 1 0 0  a00 
 a  0 0 1 1 1 0 0 0  a01 
 1 
 a 2  0 0 1 0 0 1 1 1  a02 
    
 a 3   0 1 1 1 0 1 1 0  a03 
.
 a  0 (7)
1 0 0 0 0 1 1  a10 
 4   
 a 5  0 0 0 0 0 1 0 0  a11 
 a  0 0 1 0 0 0 0 0   a12 
 6 
 a7  0 1 0 0 0 0 1 0   a13 

The inverse transformation, i.e., the conversion from GF( 2 8 ) to GF(( 2 4 ) 2 ) , requires
computing the T 1 matrix. We can use Gauss-Jordan Elimination to derive the T 1 matrix as
follows:

 a 00  1 0 0 0 0 1 0 0 a0 
 a  0
 01   0 1 0 1 1 1 0   a 1 
 a 02  0 0 0 0 0 0 1 0 a2 
    
 a 03   0 0 0 1 0 1 1 1  a 3 
. (8)
 a  0 1 0 1 0 1 0 1  a 4 
 10    
 a 11  0 0 0 0 0 1 0 0  a5 
 a  0 0 1 0 1 1 1 1  a 6 
 12   
 a 13  0 0 0 0 1 0 0 1  a7 

2.1.4 The basis transformation between standard basis and normal basis
The normal basis is with some good features in hardware, but the standard basis is used in
popular designs. Finding the transformation between them is an important topic (Lu, 1997),
we use GF( 2 4 ) as an example to illustrate that. Suppose GF( 2 4 ) is with the relation
p( x )  x 4  x 3  1 which is a primitive polynomial. Let p( )  0 such that
B1   0 ,  1 ,  2 ,  3  form a standard basis. Let    3 and the set  1 ,  2 ,  4 ,  8  is linear
The Design of IP Cores in Finite Field for Error Correction 119

independent such that B2   1 ,  2 ,  4 ,  8  forms a normal basis. There exists a matrix T such
that B2T  T  B1T and B1T  T 1  B2T . The matrixes T and T 1 are listed as follows.
0 0 1 1 0 0 1 1
0 1 0 1 1 0 1 0 
T , T 1  
0 1 1 1 0 1 1 0
   
1 1 1 1 1 1 1 0 
. (9)
  0
8
0 1 1   3
  0
3
0 1 1  8 
 4  0
    1 0 1  2   2  1
    0 1 0   4 
,
 2  0 1 
1 1   1
  1  0 1 1 0   2 
 1   0   0   
   1 1 1 1     1 1 1 0   1 

2.2 The basic operation in finite field

2.2.1 Addition and subtraction

For a finite field with characteristic 2, addition and subtraction are performed by the bitwise
XOR operator. For example, let a( x )  x 4  x 2  x 1  1 , b( x )  x 4  x 3  x 1  1 , and c(x ) be the
summation of two polynomials, thus, c( x )  a( x )  b( x )  2 x 4  x 3  x 2  2 x 1  2  x 3  x 2 or
perform in binary form {10111} + {11011} = {01100} .

2.2.2 Multiplication and inversion

The multiplication in a finite field is performed by multiply two polynomials modulo a
specific irreducible polynomial. For example, consider the finite field E  GF( 2 4 ) which is
with the relation p( x )  x 4  x  1 and let p( )  0 thus  0 ,  1 ,  2 ,  3  forms a standard
basis. Suppose a , b , c  E and a   3  1 , b   2    1 , and c is the product of them. Thus
c  a  b   3  1   2    1   5   4   3   2    1 , refer to Table 1, we have the product
result as c   2       1   3   2    1   3   . For every nonzero element
  E  GF( 2 m ) , one has  2   or  1   2 2 equivalently (Dinh et al., 2001). Therefore, the
m m

division for finite field can be performed by the multiplicative inversion. For example,
consider the inversion in GF( 2 8 ) ,  1   2 2 , and one can obtain this as Fig. 1.
8

2.2.3 Square operation

Consider an element A  a0  a 1 x 1    a m1 x m1  E where ai  GF( 2 ) for 0  i  m , the square
operation for the characteristic 2 finite field is: A 2  a0  a1 x 1    am1 x m1  . For ai  GF( 2 ) ,
2

we have ai  ai and thus A 2  a0  a1 x 2    am1 x 2 ( m1 ) . Besides, those items with power not
2

less m can be expressed by standard basis. Thus, we can perform the square operation by
some finite field additions, i.e., XOR gates. For instance, let E  GF( 2 4 ) constructed by
f ( x )  x 4  x  1 , an element A  a 0  a 1 x 1  a 2 x 2  a 3 x 3  E , A 2  a0  a1 x 2  a2 x 4  a3 x 6 . Two
terms x 4 and x 6 can be substituted by x  1 and x 3  x according to Table 1. We have
A 2  a 0 x 0  a 1 x 2  a 2 ( x  1)  a 3 ( x 3  x ) or A 2  ( a0  a 2 )  ( a 2  a 3 )x  a1 x 2  a 3 x 3 . The same
120 VLSI

property is also suitable for the power 2 i operation, such as A 2 , A 2 , , A 2 .

2 3 m1

3. The hardware designs for finite field operations

3.1 Multiplier
Finite field multiplier is the basic component for most applications. Many designers choose
the one with standard basis for their applications, because the standard basis is easier to
show the value by the bit-vector in digital computing. As follows, we introduce two most
used types of finite field multipliers, one is the conventional multiplier and another is the
bit-serial one.

3.1.1 Conventional multiplier

As the statement in Section 2.2.2, let A , B, C  GF( 2 m ) are represented with standard basis

 
m1 m 1
and C  A B , where A i 0
a i i , B i 0
b i i , and the product

P m 1

i 0
 
a i i 
m 1

i 0
 
b i i 
2 m 1

i 0
pi i . Note that every element in GF( 2 m ) is with the
relation f ( x ) described in Section 2.1.1, such that the terms with order greater than m,
 m , m1 , , 2 m1 , can be substituted by the linear combination of standard basis
{1,  1 , ,  m1 } . Thus, we can observe that there are m 2 and gate and about m  O(m) XOR
gates in the substitution for high-order terms.

3.1.2 Massey-Omura multiplier

Here, we introduce the popular version named the bit-serial type of Massey-Omura
multiplier. It is based on the normal basis, and the transformation between standard basis
and normal basis is introduced in Section 2.1.4. Let A, B, C  GF( 2 m ) are represented with

  
m1 m 1 m 1
normal basis and C  A  B , where A  a i 2 , B  bi 2 , and C  c i 2 . Denote
i i i

i 0 i 0 i 0

the coefficient-vector of A , B , and C by a , b , and c , and the notation a( i ) means A 2 , we

i

have:

  2 2  2 2   2 2   b0 
0 0 0 1 0 m1

 2 2   
  2 2   2 2    b1   a  M  b T ,
1 0 1 1 1 m1

C  A  B  a0 , a1 , , am1   (10)

        
 2 2   
  2 2   2 2  bm1 
m 1 0 m1 1 m1 m1

where M  M 0  M 1 2    M m1 2 , such that

m1

c m1i  a  M m1i  b T  a ( i )  M m1  (b ( i ) )T . (11)

Using Equation (11), the bit-serial Massey-Omura multiplier can be designed as following:
The Design of IP Cores in Finite Field for Error Correction 121

a (i )
Shift-register A AND-XOR
b (i ) cm1i
Plane

Shift-register B
Fig. 1. The Massey-Omura bit-serial multiplier

In Fig. 1, the two shift-register perform the square operation in normal basis, and the
complexity of and-xor plane is about O(m) and relative to the number of nonzero element
in M m1i . Therefore, Massey-Omura multiplier is suitable to the design of area-limited
circuits.

3.2 Inverse
In general the inverse circuit is usually with the biggest area and time complexity among
other operations. There are two main methods to implement the finite field inverse, that is,
multiplicative inversion and inversion based on composed field. The first method
decomposes inversion by multiplier and squaring, and the optimal way for decomposing is
proposed by Itoh and Tsujii (Itoh & Tsujii, 1988). The later one is based on the composed
field and suited for area-limited circuits, which has been widely used in many applications.

3.2.1 Multiplicative inversion

From Fermat's theorem, for any nonzero element   GF( 2 m ) holds  2 1  1 . Therefore,
m


m 1
multiplicative inversion is equal to  2 2
. Based on this fact  1   2 2   2 , Itoh and
m m i

i 1

Tsujii reduced the number of required multiplications to O(log m) , which is based on the


b 1
decomposition of integer. Suppose m  1  n 0
an  2 n , where a n  GF( 2 ) and a b1  1
denoted the decimal number [1ab2  a1 a0 ]2 , we have the following facts:

2 m 1  1  ( 2 2  1 )  2 [ a a1 a0 ]2
 2[ a a1 a0 ]2
1
b1
b2 b2

 (2 2
 1)( 2 2
 1)  2 [ ab  2 a1 a0 ]2
 2[ a a1 a0 ]2
1 . (12)
b1 b2
b 2

 (2 2b2
 1)( 2  1)( 2  1)  2 21 10 [ ab  2 a1 a0 ]2
 2[ a b2 a1 a0 ]2
1
2[ a a1 a0 ]
 1  ab2 ( 2 2  1)2 [ a a1 a0 ]2
 2[ a a1 a0 ]2
1
b 2
b 2 b 3 b 3

. (13)
 a b 2 ( 2 2b 2
 1)( 2  1)( 2  1)2 [ a 21 10 b3 a1 a0 ]2
 2[ a b3 a1 a0 ]2
1
2 m1  1  ( 2 2  1)( 2 2  1)( 2 1  1)  2 [ a a1 a0 ]2
 2[ a a1 a0 ]2
1
b2 1 0
b 2 b2

 (( 2 2b2
 1)2 )( 2 2b  2 2b 3
 1)( 2  1)( 2  1)  2 [ a 21 20 b3
a1 a0 ]2

a b 2 ( 2 2b3
 1)( 2  1)( 2  1)2 21 20 [ ab  3a1 a0 ]2
2 [ ab  3a1 a0 ]2
1
. (14)
 (( 2 2  1)2 2  ab2 )( 2 2  1)( 2 2  1)( 2 2  1)  2 [ a a1 a0 ]2
1
b2 b2 b3 1 0
b3

 ((((( 2 2  1)2 2  ab2 )( 2 2  1)2 2  ab3 ))( 2 2  1)2 2  a2 )

b2 b2 b3 b3 2 2

( 2 2  1)2 2  a1 )( 2 2  1)2 2  a0
1 1 0 0
122 VLSI

This algorithm requires N M  len.(m  1)  wt.(m  1)  2 multipliers, and

N P  len.(m  1)  wt.(m  1)  1 square circuits, where len.(m  1) the length of binary
representation of m  1 and wt.(m  1) is the number of nonzero bit in the representation.
For instance, if m8 then m17 , N M  len.(7 )  wt.(7 )  2  3  3  2  4 and
N P  len.(7 )  wt.(7 )  1  3  3  1  5 . For the latency of circuit, it takes
( log 2 (m  1))TM  ( log 2 ( m  1)  1)TS , where TM (resp. TS ) is the latency of multiplier (resp.
squaring circuit). We list some results of this algorithm as Table 2.

m area latency m area latency

5 2 NM +3 NP 2 TM +3 TP 11 4 NM +5 NP 4 TM +5 TP
6 3 NM +4 NP 3 TM +4 TP 12 5 NM +6 NP 4 TM +5 TP
7 3 NM +4 NP 3 TM +4 TP 13 4 NM +5 NP 4 TM +5 TP
8 4 NM +5 NP 3 TM +4 TP 14 5 NM +6 NP 4 TM +5 TP
9 3 NM +4 NP 3 TM +4 TP 15 5 NM +6 NP 4 TM +5 TP
10 4 NM +5 NP 4 TM +5 TP 16 6 NM +7 NP 4 TM +5 TP
Table 2. The list of Itoh and Tsujii algorithm

3.2.2 Composite field inversion

The use of composite field provides an isomorphism for GF( 2 m ) , while m is not prime.
Especially, if m is even, then inverse using composite field is with very low complexity.
Consider the inverse in GF(( 2 m /2 )2 ) where m is even. Suppose A , B  GF(( 2 m /2 )2 ) constructed
by an irreducible polynomial P( x )  p1 x  p0 , where p 0 , p1  GF( 2 m /2 ) . Let A  a 1 x  a 0 and
B  b1 x  b0 , where a 1 , a0 , b1 , b0 , p 1 , p 0  GF( 2 m /2 ) . Assume that B is the inverse of A, thus
A  B  1 or ( a1 x  a 0 )  (b 1 x  b0 )  1 modulo P(x ) . After the distribution, one has
A  B  ( a1 b1 p 1  a1 b0  a0 b1 )x  ( a1 b1 p 0  a0 b 0 )  1 . Therefore, a1 b1 p 1  a1 b0  a0 b1  0 and
a 1 b 1 p 0  a0 b 0  1 . Let   ( a0  a0 a1 p1  p0 a1 ) , one has b1  a 1  and b 0  ( a 0  a 1 p 1 ) , which
2 2 1 1

is design as Fig. 2. Obviously, one can observe the inversion in GF( 2 m ) is executed by
several operations which are all in GF(( 2 m /2 )2 ) , thus the total gated count used can be
reduced.

A 2  p0
A1
a1  p1

b1
a0
b0
A2

Fig. 2. The circuit for composite field inversion

The Design of IP Cores in Finite Field for Error Correction 123

4. Some techniques for simplification of VLSI

4.1 Finding common sharing resource in various design levels
Sharing resource is a common method to reduce the area cost. This skill can be used in
different design stages. For example, consider the basis transformation in Section 2.1.4, the
element of normal basis is obtained by the linear combination of standard basis as follows:

 8  1  0 ,  4  2  0 ,  2  2  1  0 ,  1  3  2  1  0 . (15)

It takes 7 XOR gates for the straightforward implementation. However, if one calculate the
summation t   2   1   0 firstly, then  2  t and  1   3  t . Therefore, the number of
XOR gates is reduced to 5. Although it is effective in the bit-level, this idea is also effective in
other design stages. Consider another example in previous section, when we form those
components   ( a0  a0 a1 p1  p0 a1 ) and b 0  ( a 0  a 1 p 1 )1 , it takes 3 2-input adders in two
2 2

expressions. Suppose we form the component a0  a1 p 1 firstly, thus the number of 2-input
adder is reduced from 3 to 2 (   ( a0 ( a0  a1 p1 )  p0 a1 ) ). Therefore, the resource-sharing idea
2

is suitable to different design stages.

4.2 Finding the optimal parameters of components

Another technique used to simplify circuits for finite field operations is change the original
field to another isomorphism. Although these methods are equal in mathematics, it provides
different outcomes in VLSI designs. There are two main methods to be realized.

4.2.1 Change the relation polynomial

Consider the implementations of hardware multiplier/inverse in GF( 2 8 ) using FPGA, we
gather area statistics of multiplier/inverse by using different irreducible polynomials ( f (x ) )
and draw the line chart as Fig. 3 and Fig. 4, where the X axis indicates various irreducible
polynomials in decimal representation and the Y axis is the number of needed XOR gates. In
Fig. 3, one can observe the lowest complexity of area and delay is with f (x ) is 45. The
maximum difference of XOR number (resp. delay) between two polynomials is 50 (resp. 2).
Therefore, choosing the optimal parameters has great influence in complexity in VLSI. The
same phenomenon is also been observed in Fig. 4, the maximum difference is 196 XOR gates.

185 183 7
180
175
Count (XOR) .

170
Delay (XOR)

165 6
160
155
150 5
145 143
140
135 133
130 4
27 29 43 45 57 63 77 95 99 101 105 113119 123 135139 141 159163 169 177189 195 207 215221 231 243245 249
Fig. 3. The statistic of area for multiplier v.s. f (x )
124 VLSI

800 7
784
750 6
Area (XOR)

700 5

f(x) Weight
650 630 4

600 3
588

550 2
27 29 43 45 57 63 77 95 99 101 105113 119 123 135139 141159163169177189195207215221 231243245 249
Fig. 4. The statistic of area for inverse v.s. f (x )

4.2.2 Using composite field

In Section 2.1.3, we illustrate the transformation between a finite field represented by
standard basis and a composite field. The most applications for composite field are to
design the inverse, for instance, the S-box in AES algorithm (Morioka & Satoh, 2003). As we
know, the main component in S-box is the finite field inverse of GF( 2 8 ) . Here, we
implement the S-box by the multiplicative inversion described in Section 3.2.1 and by using
composite field GF(( 2 4 ) 2 ) described in 3.2.2 as Table 3 by using the Altera FPGA Stratix
2S1020C4 device. Obviously, the later method is with more advantages for both area and
time complexity than that of previous one.

Delay (ns) Throughput

Method LE/ALUT CLK (MHz)
(MHz)
mult. inverse 210 23.240 43.029 344.232
composite field 82 20.219 49.458 395.664
Table 3. The results for S-box using multiplicative inversion and using composite field

5. Using computer-aided functions to choose suitable parameters

According to the explanations in Section 4, we can realize the related VLSI IPs using various
parameters to bring the benefits for lower area or time complexity. However, there exist so
many isomorphisms in using finite filed, it seems that there are so many procedures and
variations to choose the parameters and hard to find a better ones. As a result, our group
developed a software tools which is the computer-aided design (CAD) to help engineers to do
the tedious analysis and search. This section will introduce the methods to apply the
isomorphism transformations between GF( 2 8 ) and GF(( 2 4 ) 2 ) illustrated in Section 4.1 and 4.2
step by step.
Firstly, list all irreducible and primitive polynomials in two fields as shown in Table 4 and 5,
respectively. In this table, all irreducible and primitive polynomials are represented in
hexadecimal form and we omit the most significant bit. For example, in Table 4, one chooses
1B that means (00011011)2 or x  x  x  x  1 ; in Table 5, suppose the primitive element
8 4 3

  GF( 2 4 ) , one chooses 18 that means (00011000)2 or x  1x   . 2 3

The Design of IP Cores in Finite Field for Error Correction 125

GF ( 2 8 ) irreducible polynomials
1B 1D 2B 2D 39
3F 4D 5F 63 65 69 71 77 7B 87 8B
#=30
8D 9F A3 A9 B1
BD C3 CF D7 DD E7 F3 F5 F9
primitive polynomials
#=16 1D 2B 2D 4D 5F 63 65 69 71 87 8D A9 C3 CF E7 F5
Table 4. The irreducible and primitive polynomials in GF( 2 ) 8

GF(( 2 4 )2 ) irreducible polynomials

18 19 1A 1B 1C 1D 1E 1F 21 22 25 26 29 2A 2D 2E
31 33 34 36 39 3B 3C 3E 41 42 44 47 48 4B 4D 4E
51 53 55 57 59 5B 5D 5F 62 63 64 65 6A 6B 6C 6D
72 73 74 75 78 79 7E 7F 81 83 84 86 88 8A 8D 8F
#=120
92 93 96 97 9A 9B 9E 9F A1 A2 A4 A7 A9 AA AC AF
B4 B5 B6 B7 BC BD BE BF C1 C3 C5 C7 C8 CA CC CE
D4 D5 D6 D7 D8 D9 DA DB E2 E3 E6 E7 E8 E9 EC ED
F1 F2 F5 F6 F8 FB FC FF
primitive polynomials
19 1B 1D 1E 22 25 29 2D 2E 33 34 39 3B 3E 42 44
55 59 5B 5D 62 63 64 65 6B 6D 72 73 74 75 79 7E
#=60
83 84 8D 92 93 9B 9E A2 A4 A9 B4 B5 BD BE C3 C5
CE D4 D5 D9 DB E2 E3 E9 ED F2 F5 FB
Table 5. The irreducible and primitive polynomials in GF(( 2 4 )2 )

Secondly, the CAD searches for all possible combinations by the proposed algorithm as
shown in Table 6. This algorithm regards as a function used to find transformation matrices
as shown in Table 7. After we gather all results, we can choose the better parameters from
the list of analyzed results for hardware design of new IP.

Chose relation polynomial 1D for GF ( 2 )  p( x )  x 8  x 4  x 3  x 2  1 .

8

Let p( )  0 , such that   GF( 2 8 ) can be expressed by binary form as

( 7 ,  6 ,  5 ,  4 ,  3 ,  2 ,  1 ,  0 )
Step Find a GF(( 2 4 ) 2 ) irreducible polynomial.
1: Select an irreducible polynomial in ground field GF( 2 4 ) is
f 1 ( x )  x  x  1 . Let  be the root of f 1 ( x ) , thus f 1 ( )      1  0 .
4 4

Select an irreducible polynomial in GF(( 2 4 ) 2 ) is 18 f 2 ( x )  x 2  x   3

and let  be the root of f 2 ( x ) , f 2 ( )   2     3  0 .
Step Assume a generator  in GF(( 2 4 )2 ) and generate all none-zero elements
2: of GF(( 2 4 )2 ) . For any element in GF(( 2 4 )2 ) can be expressed in binary
form as ( 13 ,  12 ,  11 ,  10 ,  03 ,  01 ,  02 ,  03 ) .
Assume the T matrix = ( 7
)T ( 6 )T  ( 0 )T 88 , we have
126 VLSI

 13   7 
   
 12   6
 11   5 
   
 10   ( 7 )T 4
  ( 6 )T  ( ) 88  
0 T


 03   3
 02   2 
   
 01   1
 00  81  0  81
Step Compute the T 1 matrix= ( 7 )T ( 6 )T  ( 0 )T  .
1
88

3:
Put all none-zero elements GF ((2 ) ) in the following equation and check
4 2

if they all hold, i.e., [ i ]T  T 1 [ i ]T , where 0  i  254 .

Step If Step 3 does not hold, return to Step 2 and choose another generator; If
4: Step 3 holds, then the T and T 1 matrices are found.
Table 6. The proposed algorithm for searching transformation matrices

input p( x )  x 8  x 4  x 3  x 2  1 p( x )  GF( 2 8 )
f 1 (x)  x  x  1
4
f 1 ( )   4    1  0
f 2 (x)  x 2  x   3 f 1 ( x ), f 2 ( x )  GF(( 2 4 ) 2 )
output 0 0 1 0 0 0 0 0 1 1 0 1 0 1 0 0
0 1 1 0 0 1 0 0  0 1 0 1 1 0 1 0 
 
0 0 0 1 1 0 1 0 1 0 0 0 0 0 0 0
   
1 0 0 1 0 0 0 0 1 1 0 0 0 1 0 0
T    , T 1
 0 0 0 1
0 1 1 1 1 0 0 0 0 1 1 0
   
0 1 0 1 0 1 0 0 1 0 0 1 1 0 1 0
1 1 0 0 1 1 0 0  1 1 1 1 0 0 1 0 
 
1 1 0 1 1 1 1 1 88 0 0 1 1 0 1 0 1 88
Table 7. The result of transformation matrices between GF( 2 ) and GF (( 2 ) ) . 8 4 2

Fig. 5. The interface of CAD tool

The Design of IP Cores in Finite Field for Error Correction 127

Because various parameters provide VLSI’s outputs with huge variation and it’s seem
impossible to run all parameters, we should provide engineers a CAD tool to obtain the
analyzed algorithm and results. In Fig. 5, a designer can use a CAD with Windows interface to
find better parameters of S-box. In this CAD, it provides the complexity information of the
multipliers or the inverse in GF( 2 ) . In this figure, designer chooses the fourth result, and the
8

estimative complexity of inverse is shown in the top right of the figure; that the choice of
multiplier is shown under the inverse information. Therefore, the CAD tool helps designer to
choose the better parameters efficiently.

6. Summary
In this chapter, we introduce the common concepts of finite field regarding its applications in
error correcting coding, cryptography, and others, including the mathematical background,
some important designs for multiplier and inversion, and the idea to utilize the computer-
aided functions to find better parameters for IP or system design. The summary of this chapter
is as follows:
1. Introducing the basic finite field operations and their hardware designs: Those common
operations include addition, multiplication, squaring, inversion, basis transformation, and so
on. The VLSI designs of those operations may be understood through the mathematical
background provided in Section 2. From the mathematical background, one should realize the
benefits and the processes of the transformation between two isomorphic finite fields.
2. Using some techniques to simplify the circuits: We have introduced some useful techniques
to reduce the area cost in VLSI design, such as the resource-sharing method, utilization of
different parameters, or use some isomorphic field to substitute the used field. The first
technique is widely used in various design stages. The later two techniques depend on the
parameters used. Different parameters lead to different hardware implementation results.
However, it seems infeasible to analyze all possible parameters manually.
3. Using the composite field inversion: Composite field inversion is used in the finite field
inversion due to its superiority in hardware implementation. The main idea is to consider the
use of intermediate fields and decompose the inversion on the original field into several
operations on smaller fields. This method has been used in the AES S-box design to minimize
the area cost.
4. Calculating the transformation matrices between isomorphic finite fields. It is well known
that finite fields of the same order are isomorphic, and this implies the existence of
transformation matrices. Finding the optimal one is important in the investigation of the VLSI
designs. Two methods are presented; one is to change the relation polynomial, and the other is
to use the composite field. An algorithm to calculate the transformation matrices is provided in
Section 5, and it can be used to find the optimal one.
5. Using the computer-aided design to search for better parameters: A good hardware CAD
tool provides fast search and enough information for designer, because it brings fast and
accurate designs. In Section 5, the computer-aided function based on the proposed algorithms
is one of the examples. When the order of the finite field gets large, the number of isomorphic
field increases rapidly. This makes it almost impossible to do the exhausting search, and the
proposed CAD can be used to support engineers to get the best choices.
128 VLSI

7. Conclusion
In this chapter, we use the concept of composite fields for the CAD designs, which can
support the VLSI designer to calculate the optimal parameters for finite field inversion.

8. Acknowledgments
This work is supported in part by the Nation Science Council, Taiwan, under grant NSC 96-
2623-7-214-001.

9. References
Dinh, A.V.; Palmer, R.J.; Bolton, R.J. & Mason, R. (2001). A low latency architecture for
computing multiplicative inverses and divisions in GF(2m). IEEE Transactions on
Circuits and Systems II: Analog and Digital Signal Processing, Vol. 48, No. 8, pp. 789-
793, ISSN: 1057-7130
Fenn, S.T.J.; Benaissa, M. & Taylor, D. (1996). Fast normal basis inversion in GF(2m).
Electronics Letters, Vol. 32, No. 17, pp. 1566-1567, ISSN: 0013-5194
Hsiao, S.-F.; Chen, M.-C. Chen & Tu, C.-S. (2006). Memory-free low-cost designs of
advanced encryption standard using common subexpression elimination for
subfunctions in transformations. IEEE Transactions on Circuits and Systems I: Regular
Papers, Vol. 53, No. 3, pp. 615–626, ISSN: 1549-8328
Itoh, T. & Tsujii, S. (1988). A fast algorithm for computing multiplicative inverses in GF(2m)
using normal basis. Information and Computing, Vol. 78, No. 3, pp. 171-177, ISSN:
0890-5401
Jing, M.-H.; Chen, Z.-H.; Chen, J.-H. & Chen, Y.-H. (2007). Reconfigurable system for high-
speed and diversified AES using FPGA. Microprocessors and Microsystems, Vol. 31,
No. 2, pp. 94-102, ISSN: 0141-9331
Lidl, R. & Niederreiter, H. (1986). Introduction to finite fields and their applications, Cambridge
University Press, ISBN: 9780521460941
Lu, C.-C. (1997). A search of minimal key functions for normal basis multipliers. IEEE
Transactions on Computers, Vol. 46, No. 5, pp.588–592, ISSN: 0018-9340
Morioka, S. & Satoh, A. (2003). An optimized S-box circuit architecture for low power AES
design. Revised Papers from the 4th International Workshop on Cryptographic Hardware
and Embedded Systems, Lecture Notes in Computer Science, Vol. 2523, pp. 172–186,
ISBN: 3-540-00409-2, August, 2002, Redwood Shores, California, USA
Peterson, W.W. & Brown, D.T. (1961). Cyclic Codes for Error Detection, Proceedings of the
IRE, Vol. 49, No. 1, pp. 228-235, ISSN: 0096-8390
Peterson, W.W. & Weldon, E.J. (1972). Error-Correcting Codes, The MIT Press, 2 edition,
Cambridge, MA, ISBN: 3540004092
Sunar, B.; Savas, E. & Koc, C.K., (2003). Constructing composite field representations for
efficient conversion. IEEE Transactions on Computer, Vol. 52, No. 11, pp. 1391-1398,
ISSN: 0018-9340
Wang, C.C.; Truong, T.K.; Shao, H.M.; Deutsch, L.J.; Omura, J.K.; & Reed, I.S. (1985). VLSI
architecture for computing multiplications and inverses in GF(2m). IEEE
Transactions on Computers, Vol. 34, No. 8, pp. 709-716, ISSN: 0018-9340
The Design of IP Cores in Finite Field for Error Correction 129

Wicker, S.B. (1995). Error Control Systems for Digital Communication and Storage, Prentice Hall,
ISBN: 0-13-308941-X, US
130 VLSI
Scalable and Systolic Gaussian Normal Basis Multipliers
over GF(2m) Using Hankel Matrix-Vector Representation 131

X7

Scalable and Systolic Gaussian Normal Basis

Multipliers over GF(2m) Using Hankel
Matrix-Vector Representation
Chiou-Yng Lee
Lunghwa University of Science and Technology
Taoyuan County, Taiwan

1. Introduction
Efficient design and implementation of finite field multipliers have received high attention
in recent years because of their applications in elliptic curve cryptography (ECC) and error
control coding (Denning, 1983; Rhee, 1994; Menezes, Oorschot & Vanstone, 1997). Although
channel codes and cryptographic algorithms both make use of the finite field GF(2m), the
field orders needed differ dramatically: channel codes are typically restricted to arithmetic
with field elements which are represented by up to eight bits, whereas ECC rely on field
sizes of several hundred bits. The majority of publications concentrate on finite field
architectures for relatively small fields suitable for implementation of channel codes. In
finite field GF(2m), multiplication is one of the most important and time-consuming
computations. Since cryptographic applications (Menezes, Oorschot & Vanstone, 1997) are
the Diffie-Hellman key exchange algorithm based on the discrete exponentiation over
GF(2m), the methods of computing exponentiation over GF(2m) based on Fermat’s theorem
are performed by the repeated multiply-square algorithm. Therefore, to provide the high
performance of the security function, the efficient design of high-speed algorithms and
hardware architectures for computing multiplication is required and considered.
There are three popular basis representations, termed polynomial basis (PB), normal basis
(NB), and dual basis (DB). Each basis representation has its own advantages. The normal
basis multiplication is generally selected for cryptography applications, because the
squaring of the element in GF(2m) is simply the right cyclic shift of its coordinates. NB
multiplication depended the selection of key function is discovered by Massey and Omura
(1986). For the elliptic curve digital signature algorithm (ECDSA) in IEEE Standard P1363
(2000) and National Institute of Standards and Technology (NIST) (2000), Gaussian normal
basis (GNB) is defined to implement the field arithmetic operation. The GNB is a special
class of normal basis, which exists for every positive integer m not divisible by eight. The
GNB for GF(2m) is determined by an integer t, and is called the type-t Gaussian normal
basis. However, the complexity of a type-t GNB multiplier is proportional to t (Reyhani-
Masoleh, 2006), small values of t are generally chosen to ensure that the field multiplication
is implemented efficiently.
132 VLSI

Among various finite field multipliers are classified either as a parallel or serial
architectures. Bit-serial multipliers (Reyhani-Masoleh & Hasan, 2005; Lee & Chang, 2004)
require less area, but are slow that is taken by m clock cycles to carry out the multiplication
of two elements. Conversely, bit-parallel multipliers (Lee, Lu & Lee, 2001; Hasan, Wang &
Bhargava, 1993; Kwon, 2003; Lee & Chiou, 2005) tend to be faster, but have higher hardware
costs. Recently, various multipliers (Lee1, 2003; Lee, Horng & Jou, 2005; Lee, 2005; Lee2, 2003)
focus on bit-parallel architectures with optimal gate count. However, previously mentioned
bit-parallel multipliers show a computational complexity of O(m2) operations in GF(2);
canonical and dual basis architectures are lower bounded by m2 multiplications and k2 - 1
additions, normal basis ones by k2 multiplications and 2k2 - k additions. A multiplication in
GF(2) can be realized by a two-input AND gate and an adder by a two-input XOR gate. For
this reason, it is attractive to provide architectures with low computational complexity for
efficient hardware implementations.
There have digit-serial/scalable multiplier architectures to enhance the trade-off between
throughput performance and hardware complexity. The scalable architecture needs both the
element A and B are separated into n=[m/d] sub-word data, while the digit-serial
architecture only requires one of the element to separate sub-word data. Both architectures
are used by the feature of the scalability to handle growing amounts of work in a graceful
manner, or to be readily enlarged. In (Tenca & Koc, 1999), a unit is considered scalable
defined that the unit can be reused or replicated in order to generate long-precision results
independently of the data path precision for which the unit was originally designed.
Various digit-serial multipliers are recently developed in (Paar, Fleischmann & Soria-
Rodriguez, 1999; Kim & Yoo, 2005; Kim, Hong and Kwon, 2005; Guo & Wang, 1998; Song &
Parhi, 1998; Reyhani-Masoleh & Hasan, 2002). Song and Parhi (1998) proposed MSD-first
and LSD-first digit-serial PB multipliers using Horner’s rule scheme. For partitioning the
structure of two-dimension arrays, efficient digit-serial PB multipliers are found in (Kim &
Yoo, 2005; Kim, Hong & Kwon, 2005; Guo & Wang, 1998). The major feature of these
architectures is combined with both serial and parallel algorithms.
For large word lengths commonly found in cryptography, the bit-serial approach is rather
slow, while bit-parallel realization requires large circuit area and power consumption. In
elliptic curve cryptosystems, it strongly depends on the implementation of finite field
arithmetic. By employing the Hankel matrix-vector representation, the new GNB
multiplication algorithm over GF(2m) is presented. Utilizing the basic characteristics of MSD-
first and LSD-first schemes (Song & Parhi, 1998), it is shown that the proposed GNB
multiplication can be decomposed into n(n+1) Hankel matrix-vector multiplications. The
proposed scalable GNB multipliers are including one dd Hankel multiplier, two registers
and one final reduction polynomial circuit. The results reveal that, if the selected digital size
is d  4 bits, the proposed architecture has less time-space complexity than traditional digit-
serial systolic multipliers, and can thus obtain an optimum architecture for GNB multiplier
over GF(2m). To further saving both time and space complexities, the proposed scalar
multiplication algorithm with Hankel matrix-vector representation can also be realized by a
scalable and systolic architecture for polynomial basis and dual basis of GF(2m).
The rest of this paper is structured as follows. Section 2 briefly reviews a conventional NB
multiplication algorithm and a Hankel matrix-vector multiplication. Section 3 proposes the
two GNB multipliers, based on Hankel matrix-vector representation, to yield a scalable and
systolic architecture. Section 4 introduces the modified GNB multiplier. Section 5 analyzes
Scalable and Systolic Gaussian Normal Basis Multipliers
over GF(2m) Using Hankel Matrix-Vector Representation 133

our proposed GNB multiplier in the term of the time-area complexity. Finally, conclusions
are drawn in Section 6.

2. Preliminaries
2.1 Gaussian normal basis multiplication
The finite field GF(2m) is well known to be viewable as a vector space of dimension m over
m 1
GF(2). A set N  { ,  2 , ,  2 } is called the normal basis of GF(2m), and  is called the
normal element of GF(2m). Let any element AGF(2m) can be represented as
m 1
A a 
i 0
i
2i
 (a 0 , a1 ,  , a m 1 ) (1)

where aiGF(2), 0 ≤ i ≤ m−1, denotes the ith coordinate of A. In hardware implementation,

the squaring element is performed by a cyclic shift of its binary representation. The
multiplication of elements in GF(2m) is uniquely determined by the m cross products
i m 1 j
 2   ij 2 , ijGF(2). M={ij} is called a multiplication matrix. Let A=(a0, a1,…, am−1)
j 0

and B=(b0, b1,…, bm−1) indicate two normal basis elements in GF(2m), and C=(c0, c1,…,
cm−1)GF(2m) represent their product, i.e., C=AB. Coordinate ci of C can then be represented
by
ci  A ( i ) M ( B ( i ) ) T (2)
where A(i)
denotes a right cyclic shift of the element A by i positions. To compute the
multiplication matrix M, one can see in (IEEE Standard P1363, 2000; Reyhani-Masoleh &
Hasan, 2003). When the multiplication matrix M is found, the NB multiplication algorithm is
described as follows:
Algorithm 1: (NB multiplication) (IEEE Standard P1363, 2000)
Input: A=(a0, a1,…, am−1) and B=(b0, b1,…, bm−1) GF(2m)
Output: C=(c0, c1,…, cm−1)=AB
1. initial: C=0
2. for i = 0 to m − 1 {
3. ci  AMB T
4. A=A(1) and B=B(1)
5. }
6. output C=(c0, c1,…, cm−1)
Applying Algorithm 1, Massey and Omura (1986) first proposed bit-serial NB multiplier.
The complexity of the normal basis N, represented by CN, is the number of nonzero ij
values in M, and determines the gate count and time delay of the NB multiplier. It is shown
in (Mullin, Onyszchuk, Vanstone & Wilson, 1988/1989) that CN for any normal basis of
GF(2m) is greater or equal to 2m−1. In order to an efficient and simple implementation, a
normal basis is chosen that CN is as small as possible. Two types of an optimal normal basis
(ONB), type-1 and type-2, exist in GF(2m) if CN =2m−1. However, such ONBs do not exist for
all m.
Definition 1. Let p=mt+1 represent a prime number and gcd(mt/k,m)=1, where k denotes the
multiplicative order of 2 module p. Let  be a primitive p root of unity. The type-t Gaussian
134 VLSI

m 2m (t 1)m
normal basis (GNB) is employed by      2   2 ...  2 to generate a normal basis N
for GF(2m) over GF(2).
Significantly, GNBs exist for GF(2m) whenever m is not divisible by 8. By adopting Definition
1, each element A of GF(2m) can also be given as
m 1 m m ( t 1)
A  a0  a1 2    am 1 2  a0 (   2     2 )
(3)
2 2 m 1 2 m ( t 1) 1 2 m 1 2 2 m 1 mt 1
 a1 (      )    am 1 (     2 )
From Equation (3), the type-t GNB can be represented by the set
2 m 1 2m 2 m 1 2 2 m 1 2 m ( t 1) 2 m ( t 1) 1 2 mt 1
{ ,  2 , ,  , , , ,  , ,  , , ,  } . Since  is a primitive p
root of unity, we have
 i 1 if i  p 1
 i   (4) .
 1 if i  p 1
Thus, the GNB can then alternate to represent the redundant basis R={, 2,…, p-1}. The field
element A can be alternated to define the following formula:
A  aF (1)  aF ( 2) 2    aF ( p 1) p 1 (5)
where
F(2i2mj mod p)=i for 0 ≤ i ≤ m−1 and 0 ≤ j ≤ t−1.
Example 1. For example, Let A=(a0,a1,a2,a3,a4) be the NB element of GF(25), and let =+10 be
used to generate the NB. Applying the redundant representation, the field element A can be
represented by A=a0+a12+a33+a24+a45+a46+a27+a38+a19+a010.
Observing this representation, the coefficients of A are duplicated by t-term coefficients of
the original normal basis element A=(a0, a1,…, am−1) if the field element A presents a type-t
normal basis of GF(2m). Thus, by using the function F(2i2mj mod p)=i, the field element
A  ( a F (1) , a F ( 2) ,  , a F ( p 1) ) with the redundant representation can be translated into the
following representation, A  ( a0 ,, a0 , a1 , a1 , a2 , , a2 , am 1 , , am1 ) . Therefore, the
 
t
redundant basis is easy converted into the normal basis element, and is without extra
hardware implementations.
Let A=(a0, a1,…, am−1) and B=(b0, b1,…, bm−1) indicate two normal basis elements in GF(2m),
and C=(c0, c1,…, cm−1) represent their product, i.e., C=AB. Coordinate ci of C can then be
calculated as in the following formula: (IEEE Standard P1363, 2000)
p2
c0  G ( A, B)  a
j 1
F ( j 1)bF ( p  j ) (6)

According to Algorithm 1, the GNB multiplication algorithm for even t is modified as

follows.
Algorithm 2: (GNB multiplication for t even) (IEEE Standard P1363, 2000)
Input: A=(a0,a1,…,am−1) and B=(b0,b1,…,bm−1)GF(2m)
Output: C=(c0,c1,…,cm−1)=AB
1. initial : C=0
2. for k=0 to m−1 {
3. ck= G(A,B)
4. A=A(1) and B=B(1)
Scalable and Systolic Gaussian Normal Basis Multipliers
over GF(2m) Using Hankel Matrix-Vector Representation 135

5. }
6. Output C=(c0,c1,…,cm−1)

2.2 Bit-parallel systolic Hankel multiplier

This section briefly introduces a bit-parallel systolic Hankel multiplier in (Lee & Chiou,
2005).
Definition 2. An mm matrix H is known as the Hankel matrix if it satisfies the relation
H(i,j)=H(i+1,j−1), for 0  i  m−2 and 1  j  m−1, where H(i,j) represents the element in the
intersection of row i and column j, such that
 h0 h1  hm 3 hm  2 hm 1 
 
 h1 h 2   hm 1 hm 
 h h3    hm 1 
H 2 .
       
h h   h2 m  4 h2 m 3 
 m  2 m 1 
 hm 1 hm  h2 m  4 h2 m 3 h2 m  2 
A Hankel matrix H is entirely determined by its last column and first row, and thus
depends on having 2m−1 parameters, i.e., h=(h0, h1,…, hm−2, hm−1, hm,…, h2m−3, h2m−2). The
entries of H are constant down the diagonals parallel to the main anti-diagonal. Let C=hB be
the product of h and B, where B=(b0,b1,…,bm−1). The coordinate ci of C is given by
m 1
ci  h
j 0
j  ib j (7)

Definition 3. Let i,j,m be these positive integers with 0  i,j  m−1, one can define the
following function (i,j):
 j i
  2  for i  j  even
 (i, j )  
j  i 1
   for i  j  odd
 2
where <q> denotes q mod m.
Let i denote a fixed integer in the complete set {1,2,…,m−1}, one verifies that the map
k=(i,j). For instance, Table 1 presents the relationship between j and k with k=(i,j), where
m=7, i=2, and 0  j 6. Therefore, by substituting j=(i,j) into Equation (6), the product C can
be denoted as
m 1  m 1 
C    h 
 ( i , j )  i b ( i , j ) 
i
(8)
i 0  j 0 

j 0 1 2 3 4 5 6
k=(i,j) 6 5 0 4 1 3 2
Table 1. The relationship between j and k for i=2
136 VLSI

Example 2: Let A=(a0, a1, a2, a3, a4) and h =(h0, h1, h2, h3, h4, h5, h6, h7, h8) represent two vectors
and H represents the Hankel matrix defined by h; and let C =(c0, c1, c2, c3, c4) be the product
of hA. By applying Equation (8), the product can be derived as follows.
1 x x2 x3 x4
a0 h0 a4 h5 a4 h6 a3 h6 a3 h7
a4 h4 a0 h1 a3 h5 a4 h7 a2 h6
a1h1 a3 h4 a0 h2 a2 h5 a4 h8
a3 h3 a1h2 a2 h4 a0 h3 a1h5
 a2 h2 a2 h3 a1h3 a1h4 a0 h4
c0 c1 c2 c3 c4
Figure 1 depicts the bit-parallel systolic Hankel multiplier given by Example 1. The
multiplier comprises 25 cells, including 14 U-cells and 11 V-cells. Every Ui,j cell (Figure 2(a))
contains one 2-input AND gate, one 2-input XOR gate and three 1-bit latches to realize
ci=ci+a(i,j)h(i,j)+i. Each Vi,j cell (Figure 2(b)) is formed by one 2-input AND gate, one 2-input
XOR gate and four 1-bit latches to realize the operation of ci=ci+a(i,j)h(i,j)+i, or
ci=ci+a(i,j)h(i,j)+i+m. As stated in the above cell operations, the latency needs m clock cycles,
and the computation time per cell is needed by one 2-input AND gate and one 2-input XOR
gate.

  h 0 h5 a4 h1 h6 a3
h2 h7
a0 0 0 0 0 0

U00  U01 V02  U03  V04 

h4 a2

U10  V11  U12  V13 U14 

h3 h8
a1

V20  U21  V22  U23  V24 

a1
h3

U30  V31  U32  V33  U34 

a2 h4

V40  U41 V42  U43  U44 

c0 c1 c2 c3 c4

Fig. 1. The bit-parallel systolic Hankel multiplier [10]

Scalable and Systolic Gaussian Normal Basis Multipliers
over GF(2m) Using Hankel Matrix-Vector Representation 137

  ci  ci 
a<(j‐i)/2 >/h<‐(i+j+1)/ 2>+i h <(j‐i)/2>+i/a<‐(i+j+1)/2>  a<-(i+j+1)/2> h <‐(i+j+1)/2>+i 
h<‐(i+j+1)/2>+i+m 

ci=ci+a(i,j) h(i,j)+ c i=ci+a(i,j) h(i,j)+i+

(a) (b)
Fig. 2. (a) The detailed circuit of U-cell; (b) The detailed circuit of V-cell

3. Proposed scalable and systolic GNB multiplier architectures

p 1 p 1
Let A  
i0
a F (i) i and B  b
i0
F (i)
i with aF(0)=bF(0)=0 and aF(i), bF(i)GF(2) for 1 i  p−1

denote two type-t GNB elements in GF(2m), where  represents the root of xp+1. Assume that
the chosen digital size is a d-bit, and n=p/d, both elements A and B can also be expressed as
follows.
n 1 n 1
A 
i0
Ai  di
, B B
i0
i
di

where
d 1 d 1
Ai   j0
a F ( di  j )  j
, Bi  b
j0
F ( di  j ) 
j
.

Based on the partial multiplication for determining AB0, the partial product can be denoted
by
AB0= A0B0 + A1B0d+…+ An−1B0d(n−1) (9)
Each term AiB0 of degree 2d−2 is the core computation of Equation (9). In a general
multiplication, let us define that AiB0 is formed by
AiB0=Si+Did, for 0  i  n−1. (10)
where
Si=si,0+si,1+...+si,d−1d−1
Di=di,0+di,1+...+di,d−1d−2
j
si,j= a
k 0
F (id  k )bF ( j  k ) , for 0  j  d−1

d 1
di,j= a
k  j 1
F ( id  k )bF ( d  j  k ) , for 0  j  d−2

Therefore, Equation (9) can be re-expressed as

AB0=( S0+D0d)+ ( S1+D1d) d +...+ ( Sn−1+Dn−1d) d(n−1)
138 VLSI

= S0+( S1+D0) d +...+ ( Sn−1+Dn−2) d(n−1)+ Dn−1dn

= C0+C1d +...+ Cn−1d(n−1) + Cndn (11)
where
C0= S0,
Ci = Si+Di−1, for 1  i  n−1
Cn= Dn−1
Si=(si,0, si,1 ,..., si,d−1) in Equation (10) can be translated with the following matrix-vector
 si , 0   0 0  0 a F ( id )   bF ( d 1) 
    
 si ,1   0 0  a F ( id ) a F ( id 1)   bF ( d  2 ) 
           (12)
    
 si , d  2   0 a F ( id )  a F ( id  d  3 ) a F ( id  d  2 )   bF (1) 
s   a F ( id  d 1)   bF ( 0 ) 
 i , d 1   a F ( id ) a F ( id 1)  a F ( id  d  2 )
Similarly, Di−1=(di−1,0, di−1,1 ,..., di−1,d−2, 0) can also be translated with the following matrix-
vector
 d i 1, 0   a F ( d ( i 1) 1)  a F ( id  2 ) a F ( id 1) 0   b F ( d 1) 
    
 d i 1,1   a F ( d ( i 1)  2 )  a F ( id 1) 0 0  bF ( d  2 ) 
           (13)
    
 d i 1, d  2   a F ( id 1)       b F (1) 
 0  
   0    0   b F ( 0 ) 
According to Equations (12) and (13), Ci=Si+Di1 in Equation (11) can obtain
 cF (di)   aF (d (i 1)1) aF (d (i 1)2)  aF (id ) bF (d 1) 
    
 cF (di1)   aF (d (i 1)2) aF (d (i 1)3)  aF (id 1) bF (d 2) 
          
     (14)
cF (d (i 1)1)   aF (di) aF (di1)  aF (id d 1)  bF (0) 
 Hni B0
From the above matrix-vector multiplication, the Hankel vector hn−i=(aF(d(i−1)+1), aF(d(i−1)+2),…,
aF(id), aF(id+1),…, aF(d(i+1)−1)) is defined by the dd Hankel matrix Hn−i. Hence, the computation
AB0 using the Hankel matrix-vector representation can be computed as follows:
AB0=hnB0+ hn1B0d+…+h0B0dn (15)
Therefore, AB0 can be dismembered into (n+1) Hankel multiplications, and can be
performed by the following algorithm:
Algorithm 3: PM(A,B0)
p1 d 1
Input: A
a
i0
F(i)
i and B0  b
i0
F ( i )
i

p 1
Output: C  c
i0
F (i )
i
=AB0 mod (p+1)

1. convert the Hankel vector hni=(aF(d(i−1)+1), aF(d(i−1)+2),…, aF(d(i+1)−1)), for 0 i  n, from A.

2. C=(cF(0), cF(1),…, cF(p−1))=(0,…,0)
3. for i=0 to n {
4. Xi=hn−iB0
5. }
Scalable and Systolic Gaussian Normal Basis Multipliers
over GF(2m) Using Hankel Matrix-Vector Representation 139

6. C= X mod (p+1)
7. return C=(cF(0), cF(1),…, cF(p−1))
Algorithm 3 for determining AB0 includes two core operations, namely the Hankel
multiplication and the reduction polynomial p+1, as illustrated in Figure 3. The proposed
partial multiplier architecture in Figure 3 can be calculated using the following procedure.
Step 1: From Equation (14), we can see that Hankel matrix Hn−i is defined by all coefficients
in A. Here A=(aF(0), aF(1),…, aF(p−1)) is firstly converted to the Hankel vector hn−i=(aF(d(i−1)+1),
aF(d(i−1)+2),…, aF(d(i+1)−1)) and its result is stored in the register Hn−i.
Step 2: Applying the bit-parallel systolic Hankel multiplier as shown in Figure1, Figure 3
shows AB0 computation. Each Hankel multiplications in Step 4 of Algorithm 3, the result of
AB0 is stored in the register X.
Step 3: After (n+1) Hankel multiplications, the result needs to perform the reduction
polynomial p+1.
Generally, the computation of ABi for 0in-1 can be obtained by the following formula:
ABi=hnBi+ hn1Bid+…+ h0Bidn (16)
The above equation indicates that each ABi computation can be dismembered into (n+1)
Hankel multiplications. As mentioned above, two GNB multipliers are described in the
following subsections.

  Bi
d 2d‐ 1 H0

Bit‐parallel  H1
systolic Hankel 
Hi
A
multiplier 
Hn
d
X register
p+d
p
C=X mod(  +1)
p
AB0
Fig. 3. The proposed scalable and systolic architecture for computing AB0

3.1 LSD-first scalable systolic multiplier

The product C= AB mod (p+1) using LSD-first multiplication algorithm can be represented
as
C=AB0 mod (p+1) +AB1d mod (p+1)+…+ ABn−1d(n−1) mod (p+1)
= A(0)B0 +A(d)B1+…+ A(d(n−1))Bn−1 (17)
p 1
where A(id )  A id mod( p  1)  A( d (i 1)) d mod( p  1)  a j 0
F ( j  id )
j
.

Applying Equations (16) and (17), the proposed LSD-first scalable systolic GNB
multiplication is addressed as follows:
140 VLSI

Algorithm 4. (LSDGNB scalable multiplication)

Input: A  (a0 , a1,, am1) and B  (b0 , b1 ,, bm1 ) are two normal basis elements in GF(2m)
Output: C= (c0 , c1 , , c m 1 ) =AB
1. Initial step:
1.1 A  (a F ( 0) , a F (1) , , a F ( p 1) )  ( a0 , a1 , , a m 1 )
1.2 B  (b F ( 0) , bF (1) , , bF ( p 1) )  (b0 , b1 , , bm 1 )
n 1 d 1
1.3 B  i 0
Bi di , where Bi  b
j 0
F ( id  j )
j
and n=p/d

1.4 C=0
2. Multiplication step:
2.1 for i=0 to n−1 do
2.2 C=C +PM(A,Bi) (where PM(A,Bi) as referred to Algorithm 3)
2.3 A=Ad mod (p+1)
2.4 endfor
3. Basis conversion step:
3.1 C  (c0 , , c0 , , cm 1 , , cm 1 )  (c F ( 0 ) , c F (1) , , c F ( p 1) )
 
t
4. Return (c0 , c1 , , cm 1 )
The proposed LSDGNB scalable multiplication algorithm is split into n-loop partial
multiplications. Figure 4 depicts the LSDGNB multiplier based on the proposed partial
multiplier in Fig.3. Both NB elements A and B are initially transformed into the redundant
basis given by Equation (4), and are stored in both registers A and B, respectively. In round
0 (see Figure 4), the systolic array in Figure 3 is adopted to compute C=A(0)B0, and the result
is stored in register C. In round 1, the element A must be cyclically shifted to the right by d
digits. The result produced in the systolic array is added to register C in the round 0. The
first round, which estimates the latency, requires d+n clock cycles. Each subsequent round
computation requires a latency of n+1 clock cycles. Finally, the entire multiplication requires
a latency of d+n(n+1) clock cycles. The critical propagation delay of every cell is the total
delay of one 2-input AND gate, one 2-input XOR gate and one 1-bit latch.
Scalable and Systolic Gaussian Normal Basis Multipliers
over GF(2m) Using Hankel Matrix-Vector Representation 141

A(id)=A(i‐ 1 )dd mod(p+1)

B0
d p
B1
ABi computation as seen 
Bi in Fig.3 
Bn‐ 1

C
p

Basis conversion
m
C
Fig. 4. The proposed LSD-first scalable systolic GNB multiplier over GF(2m)

3.2 MSD-first scalable multiplier

From Equation (17), the product C can also be re-written by
C=(…(ABn−1 mod (p+1)) d+ABn−2 mod (p+1) ) d+…) d +AB0 mod (p+1) (18)
Therefore, the MSD-first scalable systolic multiplication is addressed using the following
algorithm.
Algorithm 5. (MSDGNB scalable multiplication)
Input: A and B are two normal basis elements in GF(2m)
Output: C=AB
1. initial step:
1.1 A  (a F ( 0) , a F (1) , , a F ( p 1) )  ( a0 , a1 , , a m 1 )
1.2 B  (b F ( 0) , bF (1) , , bF ( p 1) )  (b0 , b1 , , bm 1 )
n 1 d 1
1.3 B  i 0
Bi di , where n=p/d and Bi  b
j 0
F ( id  j )
j

1.4 C=0
2. multiplication step:
2.1 for i=1 to n do
2.2 C=Cd mod (p+1)+PM(A,Bn−i), where PM(A,Bn−i) as referred to Algorithm 3
2.3 endfor
3. basis conversion step:
3.1 C  (c0 , , c0 , , cm 1 , , cm 1 )  C  (c F ( 0) , c F (1) , , c F ( p 1) )
 
t
4. return (c0 , c1 , , cm 1 )
Algorithm 5 presents the MSD-first scalable multiplication, and Figure 5 presents the entire
GNB multiplier architecture. As compared to both GNB multiplier architectures, the
LSDGNB multiplier before each round computation, the element A must be performed by
A(id)=A(i−1)dd mod(p+1). The MSDGNB multiplier after each round computation, the result C
142 VLSI

must be performed by C(id)=C(i−1)dd mod(p+1). Notably, both operations A(id) and C(id)
represent a right cyclic shift to id positions. Hence, the two proposed architectures have the
same time and space complexity.

A
Bn‐ 1
d p
Bn‐ 2
ABi computation as seen 
Bi
in Fig.3 
B0

C (id)=C (i‐ 1)d d mod( p+1) C

p

Basis conversion
m
C
Fig. 5. The proposed MSD-first scalable systolic GNB multiplier over GF(2m)

4. Modified Scalable and systolic GNB multiplier over GF(2m)

In the previous section, two scalable GNB multipliers are including one dd Hankel
multiplier, four registers and one final reduction polynomial circuit. For the whole
multiplication scheme, both circuits demand n(n+1) Hankel multiplications. To reduce time-
and space-complexity, this section will develop another version of the GNB scalable
multiplication scheme.
Let A and B in GF(2m) be represented by the redundant basis representation. If the selected
n 1
digital size is d digits, then element B can be represented as B  B
i 0
i
di
, where

d 1
Bi  b
j 0
F ( id  j )
j
and n=(mt+1)/d. From Equation (14), using LSD-first multiplication

algorithm, the partial product can be modified by

C0= AB0 mod (p+1)
=AbF(0)+AbF(1)+…+AbF(d1)d1 mod (p+1)
=A(0)bF(0)+A(1)bF(1)+…+A(d1)bF(d1)
=c0,F(0)+c0,F(1)+…+ c0,F(p1)p1 (19)
Scalable and Systolic Gaussian Normal Basis Multipliers
over GF(2m) Using Hankel Matrix-Vector Representation 143

p 1 d 1
where, A( j)  A j mod( p 1)  
i 0
aF (i  j) i , for 0  j  d−1 and c0, F (i )  b
j 0
F ( j ) aF (i  j ) , for 0  i 

p−1.
In (Wu, Hasan, Blake & Gao, 2002), it is shown that, from the GNB representation in
(Reyhani-Masoleh & Hasan, 2005) to convert the normal basis, the minimum representation
of A has a Hamming weight equal to or less than mt/2 if m is even, and (mt−t+2)/2 if m is
odd. Assume that coordinate numbers of the partial product C0 in Equation (19) are selected
by q=dk consecutive coordinates to satisfy the corresponding normal basis representation,
where q ≥ mt/2 if m is even, and q ≥ (mt−t+2)/2 if m is odd. Then, the partial product AB0 can
be calculated by
AB0 = hk1B0+hk2B0d +…+h0B0d(k1).
Similarly,
AB1d= hkB1+hk1B1d +…+ h1B1d(k1)
AB22d= hk+1B2+hkB2d +…+ h2B2d(k1)

ABn1 (n1)d= hk+n2Bn1+hk+n3Bn1d +…+ hn1Bn1d(k1)
Thus, the modified multiplication requires only nk Hankel multiplication. As stated above,
the modified LSD-first scalable multiplication is addressed as follows:
Algorithm 6. (modified LSDGNB scalable multiplication)
Input: A  (a0 , a1,, am1) and B  (b0 , b1,, bm1) are two normal basis elements in GF(2m)
Output: C= (c0 , c1 , , c m 1 ) =AB
1. initial step:
1.1 A  (a F ( 0) , a F (1) , , a F ( p 1) )  ( a0 , a1 , , a m 1 )
1.2 B  (b F ( 0) , bF (1) , , bF ( p 1) )  (b0 , b1 , , bm 1 )
n 1 d 1
1.3 B  B
i 0
i
di
, where n=p/d and Bi  b
j 0
F ( id  j )
j

k 1 d 1
1.4 C  C 
i 0
i
di
 0 , where k=q/d and Ci  c
j 0
F ( id  j )
j

1.5 All Hankel vector hi’s for 0 ≤ i ≤ n+k−1 are converted from the redundant basis
representation of A.
2. multiplication step:
2.1 for i=0 to k−1 do
2.2 for j=0 to n−1 do
2.3 Ck1i=Ck1i+Hi+jBj
2.4 endfor
2.5 endfor
3. basis conversion step:
3.1 C  (c0 , , c0 , , cm 1 , , cm 1 )  (c F ( 0) , c F (1) , , c F ( q 1) )
4. return (c0 , c1 , , cm 1 )
Applying Algorithm 6, Figure 6 shows a LSDGNB multiplier using the redundant
representation. The circuit includes two registers, one d × d Hankel multiplier, and one
144 VLSI

summation circuit. In the initial step, two registers H and B are converted by Steps 1.5 and
1.3, respectively. As for the register Hi represent (2d1)-bit latches; and the register Bi is d-bit
latches. The operation of a d × d Hankel matrix-vector multiplier has been described in the
previous section. In Figure 6, the MUX is responsible for shifting register H. The SW is in
charge of shifting the outcome of the GNB multiplication. As mentioned above, the total
Hankel multiplications can be reduced from n(n+1) to nk, where k=mt/2d if m is even and
k=(mt−t+2)/2d if m is odd. By the configuration of Figure 6, the modified multiplier is
without the final reduction polynomial circuit. It is revealed that the modified multiplier has
lower time- and space-complexity as compared to Figures 4 and 5.

  2d‐1 d

H0 d x d B0
Hankel 
H1 Multiplier B1

d
Hn‐1 Bn‐1

MUX d‐bit latches
1
0
SW
Hn ctr 
1 0
0
0 C0
Hj n‐1 C 1
0 Ci
Hn+k‐ 2 1 C k‐1
0
Fig. 6. The modified LSD-first scalable systolic GNB multiplier over GF(2m)

5. Time and Area Complexity

Various bit-parallel systolic NB multipliers are only discussed on type-1 and 2 ONBs of
GF(2m), as in (Lee & Chiou, 2005 ; Kwon, 2005 ; Lee, Lu & Lee, 2001). As is well known, a
type-1 ONB is built from an irreducible AOP, while a type-2 ONB can be constructed from a
palindromic representation of polynomials of length 2m. However, both ONB types exist
about 24.5% for m< 1000, as depicted in IEEE Standard P1363 (2000). For the ECDSA
(Elliptic Curve Digital Signature Algorithm) applications, NIST (2000) has recommended
five binary fields. They are GF(2163), GF(2233), GF(2283), GF(2409) and GF(2571). Their ONB
multipliers are implemented with unscalable architectures, and the latency needs m+1 clock
cycles. The space complexity of the previous architectures is proportional to m2. As m
becomes large, their hardware implements need very large area. Hence, the NIST
architectures have limited very applications in cryptography. However, the proposed
LSDGNB and MSDGNB multipliers do not have this problem.
Scalable and Systolic Gaussian Normal Basis Multipliers
over GF(2m) Using Hankel Matrix-Vector Representation 145

Table 2 compares the circuits of the proposed scalable multipliers with those of the other un-
scalable (bit-parallel) multipliers. Table 3 lists the proposed multipliers with the total
latency. According to this table, the proposed multipliers for a type-2 GNB save about 40%
latency as compared to Kwon's (2003) and Lee & Chiou's (2005) multipliers, and those for a
type-1 GNB save about 60% latency as compared to Lee-Lu-Lee's multipliers (2001). Since
the selected digital size d must minimize the total latency, the proposed multipliers then
have low hardware complexity and low latency.

Multipliers Kwon Lee-Lu-Lee Lee-Chiou LSD-GNBM in MSD-GNBM

(2003) (2001) (2005) Figure 4 in Figure 5
Basis Type-II Type-I ONB Type-II Gaussian NB Gaussian NB
ONB ONB
Architecture bit-parallel bit-parallel bit-parallel scalable scalable
Total
Complexity
2-input XOR 2m2+m m2+2m+1 2m2+m d2+d+p d2+d+p
2-input AND 2m2+m m2+2m+1 2m2 d2 d2
1-bit latch 5m2 3m2+6m+1 7m2 3.5d2+5p+3nd 3.5d2+5p+3nd
1x2 SW 0 0 p p
Computation TA+2TX TA+TX TA+TX TA+TX TA+TX
time per cell
Latency m+1 m+1 m+1 d+n(n+1) d+n(n+1)
Note: the value d is the selected digital size; p=mt+1 is a prime number; n=p/d; TX denotes 2-input
XOR gate delay; TA denotes 2-input AND gate delay
Table 2. Comparison of various systolic normal basis multipliers of GF(2m)

Fig. 7. Comparisons of the time-area complexity for various digit-serial multipliers over
GF(2²³³)
146 VLSI

Type-2 GNB Type-1 GNB

the proposed Kwon reduced the proposed Lee-Lu- reduced

architecture (2003) latency architecture Lee latency
(2001)
m d Minimum latency Compared m d Minimum latency Compared
latency to Kwon latency to Lee-Lu-
(2003) Lee (2001)
146 46 88 147 40% 100 29 41 101 59.40%
155 57 87 156 44% 106 31 43 107 59.80%
158 58 88 159 45% 130 30 50 131 61.80%
173 64 94 174 46% 138 31 51 139 63.80%
174 64 94 175 46% 148 34 54 149 63.80%
179 66 96 180 46.60% 162 37 57 163 65%
183 67 97 184 47% 172 39 59 173 65.90%
186 68 98 187 47.60% 178 40 60 179 66.50%
189 69 99 190 47.90% 180 41 61 181 66.30%
191 59 101 192 47.40% 196 44 64 197 67.50%
194 60 102 195 47.70% 210 48 68 211 67.80%
209 65 107 210 49% 226 51 71 227 68.70%
Table 3. Lists the total latency for type-1 and type-2 GNB multipliers over GF(2m)

Fig. 8. Comparisons of transistor count for various digit-serial multipliers over GF(2²³³)

By applying the cut-set systolization techniques (Kung, 1988), various digit-serial systolic
multipliers are recently reported in (Kim, Hong & Kwon, 2005; Guo & Wang, 1998), which are
identical of n processing elements (PE) to enhance the trade-off between throughput
performance and hardware complexity. Each PE requires a maximum propagation delay of
Tmax=(d−1)(TA+TiX+TM)+TA+TiX, where TA, TiX and TM denote the propagation delays through a
2-input AND gate, an i-input XOR gate and a 2-to-1 multiplexer, respectively. The maximum
propagation delay in each PE is large if the selected digital size d is large. However, the
proposed scalable systolic architectures do not have such problems, since the propagation
delay of each PE is independent of the selected digital size d. Applying Horner’s rule, Song
and Parhi (1998) suggested MSD-first and LSD-first digit-serial multipliers. Various digit-serial
multipliers use only one of the input signals A and B to separate n=m/d sub-word data. Our
Scalable and Systolic Gaussian Normal Basis Multipliers
over GF(2m) Using Hankel Matrix-Vector Representation 147

proposed architectures separate both input signals into n sub-word data, in which one of the
input element is translated into Hankel vector representation. The proposed LSD-first and
MSD-first scalable multiplication algorithms require n(n+1) Hankel multiplications, and the
modified multiplication algorithm only demands nk Hankel multiplications, where k=mt/2d
if m is even and k= (mt−t+2)/2d if m is odd. Using a single Hankel multiplier to implement
our proposed scalable multipliers, we have O(d2) space complexity, while other digit-serial
multipliers require O(md) space complexity, as seen in Tables 4 and 5.
For comparing the time-area complexity, the transistor count based on the standard CMOS
VLSI realization is employed for comparison. Therefore, some basic logic gates: 2-input XOR,
2-input AND, 12 SW, MUX and 1-bit latch are assumed to be composed of 6, 6, 6, 6 and 8
transistors, respectively (Kang & Leblebici, 1999). Some real circuits (STMicroelectronics,
http://www.st.com) such as M74HC86 (STMicroelectronics, XOR gate, TX=12ns (TYP.)),
M74HC08 (STMicroelectronics, AND gate, TA=7ns (TYP.)), M74HC279 (STMicroelectronics, SR
Latch, TL=13ns (TYP.)), M74H257 (STMicroelectronics, Mux, TM=11ns (TYP.)) are employed for
comparing time complexity in this paper. In the finite field GF(2233), Figures 7 and 8 show that
our proposed scalable multipliers compare to the corresponding digit-serial multipliers (Kim,
Hong & Kwon, 2005; Guo & Wang, 1998; Reyhani-Masoleh & Hasan, 2002). As the selected
digital size d ≥ 4, the proposed scalable multipliers have lower time-area complexity than two
reported digit-serial PB multipliers (Kim, Hong & Kwon, 2005; Guo& Wang, 1998) (as shown
in Figure 7). When the selected digital size d ≥ 8, the modified scalable multiplier has lower
time-area complexity than the corresponding digit-serial NB multiplier (Reyhani-Masoleh &
Hasan, 2002). For comparing a transistor count, Figure 8 reveals that our scalable multipliers
have low space complexity as compared to the reported digit-serial multipliers (Kim, Hong &
Kwon, 2005; Guo & Wang, 1998; Reyhani-Masoleh & Hasan, 2002).

Multipliers Guo & Wang (1998) Kim, Hong & Kwon Figure 5 Figure 6
(2005)
Basis polynomial polynomial Gaussian normal Gaussian
normal
Architecture digit-serial digit-serial scalable scalable
Total Complexity
2-input XOR
2-input AND 2ed² 2ed² d²+d+p d²+d
1-bit latch e(2d²+d) e(2d²+d) d² d²
1x2 SW 10d+5Pde 10d+1+4.5Pd+Pe 3.5d²+5p+3nd Q
MUX 0 0 p d
2ed 2ed 0 1
Critical Path Pipelined: Pipelined: TA+TX TA+TX
d(TA+2TX+2TM)/(P+ d(TA+TX+2TM)/(P+1)
1) Non-Pipelined:
Non-Pipelined: TA+TX+(d1)(TA+TX+2T
TA+3TX+(d1)(TA+2 M)
TX+2TM)
Latency Pipelined: 3+P)e Pipelined: (3+P)e d+n(n+1) d+nk
Non-Pipelined: 3e Non-Pipelined: 3e
Note: k =h/d, e=m/d; Q=3.5d2+nd+(2d1)(n+k1); TM: denotes 2-by-1 MUX gate delay; P+1
number of pipelining stages inside each basic cell (Kim, Hong & Kwon, 2005; Guo & Wang, 1998)
Table 4. Comparison of various digit-serial systolic multipliers over GF(2m)
148 VLSI

Multipliers Type1-DSNB1 Massey-Omura Figure 5 Figure 6

(Reyhani-Masoleh & (Reyhani-Masoleh &
Hasan , 2002) Hasan , 2002)
Architecture digit-serial digit-serial scalable systolic scalable
nonsystolic nonsystolic systolic
Total Complexity
2-input XOR (d+1)(m1) d(2m2) d²+d+p d²+d
2-input AND d(m1)+m d(2m1) d² d²
1-bit latch 0 0 3.5d²+5p+3nd Q
1x2 SW 0 0 p d
MUX 0 0 0 1
Critical Path TA+(1+log2 m)TX TA+(1+log2 m)TX TA+TX TA+TX
Latency 1 1 d+n(n+1) d+nk
Table 5. Comparison of various digit-serial multipliers for optimal normal basis of GF(2m)

6. Conclusions
This work presents new multiplication algorithms for the GNB of GF(2m) to realize LSD-
first and MSD-first scalable multipliers. The fundamental difference of our designs from
other digit-serial multipliers described in the literature is based on a Hankel matrix-vector
representation to achieve scalable multiplication architectures. In the generic field, the GNB
multiplication can be decomposed into n(n+1) Hankel multiplications. To use the
relationship from the GNB to NB, we can modify the LSD-first scalable multiplication
algorithm to decrease the number of Hankel multiplication from n(n+1) into nk, where
k=mt/2d if m is even and k=(mt-t-2)/2d if m is odd. Our analysis shows that, in finite field
GF(2233), if the selected digital size d ≥ 8, the proposed scalable multipliers then have lower
time-area complexity as compared to existing digit-serial multipliers for polynomial basis
and normal basis of GF(2m). Since our proposed scalable multiplication algorithms have
highly flexible and are suitable for implementing all-type GNB multiplications. Finally, the
proposed architectures have good trade-offs between area and speed for implementing
cryptographic schemes in embedded systems.

7. References
Denning, D.E.R.(1983). Cryptography and Data Security, Reading, MA: Addison-Wesley.
Rhee, M.Y. (1994). Cryptography and Secure Communications, McGraw-Hill, Singapore.
Menezes, A. Oorschot, P. V. & Vanstone, S. (1997). Handbook of Applied Cryptography, CRC
Press, Boca Raton, FL.
Massey, J.L. & Omura, J.K. (1986). Computational method and apparatus for finite field
arithmetic,” U.S. Patent Number 4,587,627.
Reyhani-Masoleh, A. & Hasan, M.A. (2005). Low complexity word-level sequential normal
basis multipliers. IEEE Transactions on Computers, Vol. 54, No.2.
Lee, C.Y. & Chang, C.J. (2004). Low-complexity linear array multiplier for normal basis of
type-II, IEEE International Conference on Multimedia and Expo, Vol. 3, pp. 1515-1518.
Lee, C.Y., Lu, E.H., & Lee, J.Y. (2001). Bit-Parallel Systolic Multipliers for GF(2m) Fields
Defined by All-One and Equally-Spaced Polynomials,” IEEE Transactions on
Computers, Vol. 50, No. 5, pp. 385-393.
Scalable and Systolic Gaussian Normal Basis Multipliers
over GF(2m) Using Hankel Matrix-Vector Representation 149

Hasan, M.A., Wang, M.Z. & Bhargava, V.K. (1993). A modified Massey-Omura parallel
multiplier for a class of finite fields,” IEEE Transactions on Computers, Vol. 42, No.10,
pp. 1278-1280.
Kwon, S. (2003). A low complexity and a low latency bit parallel systolic multiplier over
GF(2m) using an optimal normal basis of type II. Proceedings of 16th IEEE Symp.
Computer Arithmetic, pp. 196-202.
Lee, C.Y. & Chiou, C.W. (2005). Design of low-complexity bit-parallel systolic Hankel
multipliers to implement multiplication in normal and dual bases of GF(2m). IEICE
Transactions on Fundamentals, vol. E88-A, no.11, pp. 3169-3179.
IEEE Standard P1363 (2000). IEEE Standard Specifications for Public-Key Cryptography.
National Inst. of Standards and Technology,(2000). Digital Signature Standard, FIPS
Publication 186-2.
Reyhani-Masoleh, A. (2006). Efficient algorithms and architectures for field multiplication
using Gaussian normal bases. IEEE Transactions on Computers, Vol. 55, No. 1, pp.34-
47.
Lee1, C.Y. (2003). Low-Latency Bit-Parallel Systolic Multiplier for Irreducible xm+xn+1 with
gcd(m,n)=1. IEICE Transactions on Fundamentals, Vol. E86-A, No.11, pp. 2844-2852.
Lee, C.Y., Horng, J.S. & Jou, I.C. (2005). Low-complexity bit-parallel systolic Montgomery
multipliers for special classes of GF(2m). IEEE Transactions on Computers, vol. 54,
no.9, pp. 1061-1070.
Lee, C.Y. (2005). Systolic architectures for computing exponentiation and multiplication over
GF(2m) using polynomial ring basis. Journal of LungHwa University, vol. 19, pp.87-98.
Lee2, C.Y. (2003). Low complexity bit-parallel systolic multiplier over GF(2m) using
irreducible trinomials. IEE Proceeding Computer, and Digital Technical, Vol. 150, pp.
39-42.
Paar, C., Fleischmann, P. & Soria-Rodriguez, P. (1999). Fast arithmetic for public-key
algorithms in Galois fields with composite exponents. IEEE Transactions on
Computers, vol. 48, no.10, pp. 1025-1034.
Kim, N.Y. & Yoo, K.Y. (2005). Digit-serial AB2 systolic architecture in GF(2m). IEE Proceeding
Circuits Devices Systems, Vol. 152, No. 6, pp. 608-614.
Kang, S.M. & Leblebici, Y. (1999). CMOS Digital Integrated Circuits Analysis and Design,
McGrawHill.
Logic selection guide: STMicroelectronics. <http://www.st.com > .
Kim, C.H., Hong, C.P. & Kwon, S. (2005). A digit-serial multiplier for finite field GF(2m).
IEEE Transactions on VLSI, Vol. 13, No. 4, pp. 476-483.
Guo, J.H. & Wang, C.L. (1998). Digit-serial systolic multiplier for finite fields GF(2m). IEE
Proc.-Comput. Digit. Tech., Vol. 145, No. 2, pp. 143-148, March.
Kung, S.Y. (1988). VLSI array processors, Englewood Cliffs, NJ: Prentice-Hall.
H. Wu, M.A. Hasan, I.F. Blake and S. Gao, “Finite field multiplier using redundant
representation. IEEE Transactions on Computers, Vol. 51, No. 11, pp.1306-1316, Nov.
2002.
Mullin, R.C., Onyszchuk, I.M., Vanstone, S.A. & Wilson, R. M. (1988/1989). Optimal Normal
Bases in GF(pn). Discrete Applied Math., vol. 22, pp.149-161.
Reyhani-Masoleh, A. & Hasan, M.A. (2003). Fast normal basis multiplication using general
purpose processors. IEEE Transactions on Computers, Vol. 52, No. 11, pp. 1379-1390.
150 VLSI

Song, L. & Parhi, K.K. (1998). Low-energy digit-serial/parallel finite field multipliers. Journal
of VLSI Signal Processing , Vol.19, pp.149-166.
Tenca, A.F. & Koc, C.K. (1999). A scalable architecture for Montgomery multiplication.
Proceedings of Cryptographic Hardware and Embedded System (CHES 1999), No. 1717 in
Lecture Notes in Computer Science, pp. 94-108.
Reyhani-Masoleh, A. & Hasan, M.A. (2002). Efficient digit-serial normal basis multipliers
over GF(2M). IEEE International Conference on Circuits and Systems.
High-Speed VLSI Architectures for Turbo Decoders 151

X8

High-Speed VLSI Architectures

for Turbo Decoders
Zhongfeng Wang1 and Xinming Huang2
1Broadcom Corporation, 5300 California Avenue, Irvine, CA 92617, USA,
2Department of ECE, Worcester Polytechnic Institute, Worcester, MA 01609, USA.

Turbo code, being one of the most attractive near-Shannon limit error correction codes, has
attracted tremendous attention in both academia and industry since its invention in early
1990’s. In this chapter, we will discuss high-speed VLSI architectures for Turbo decoders.
First of all, we will explore joint algorithmic and architectural level optimization techniques
to break the high speed bottleneck in recursive computation of state metrics for soft-input
soft-output decoders. Then we will present area-efficient parallel decoding schemes and
associated architectures that aim to linearly increase the overall decoding throughput with
sub-linearly increased hardware overhead.

Keywords: Turbo code, MAP algorithm, parallel decoding, high speed, VLSI.

1. Introduction
Error correction codes are an essential component in digital communication and data
storage systems to ensure robust operation of digital applications, wherein, Turbo code,
invented by Berrou (1993), is among the two most attractive near-optimal (i.e., near-Shannon
limit) error correction codes. As a matter of fact, Turbo codes have been considered in
several new industrial standards, such as 3rd and post-3rd generation cellular wireless
systems (3GPP, 3GPP2, and 3GPP LTE), Wireless LAN (802.11a), WiMAX (broadband
wireless, IEEE 802.16e) and European DAB and DVB (digital audio broadcasting and digital
video broadcasting) systems.

One key feature associated with Turbo code is the iterative decoding process, which enables
Turbo code to achieve outstanding performance with moderate complexity. However, the
iterative process directly leads to low throughput and long decoding latency. To obtain a
high decoding throughput, a large amount of computation units have to be instantiated for
each decoder, and this results in a large chip area and high power consumption. In contrast,
the growing market of wireless and portable computing devices as well as the increasing
desire to reduce packaging costs have directed industry to focus on compact low-power
circuit implementations. This tug-of-war highlights the challenge and calls for innovations
on Very Large Scale Integration (VLSI) design of high-data rate Turbo decoders that are
both area and power efficient.
152 VLSI

For general ASIC design, there are two typical ways to increase the system throughput: 1)
raise the clock speed, and 2) increase the parallelism. In this chapter, we will tackle the high
speed Turbo decoder design in these two aspects. Turbo code decoders can be based on
either maximum-a-posterior probability (MAP) algorithm proposed in Bahl (1974) (or any
variants of approximation) or soft-output Viterbi algorithm (SOVA) proposed in Hagenauer
(1989) (or any modified version). However, either algorithm involves recursive computation
of state metrics, which forms the bottleneck in high speed integrated circuit design since
conventional pipelining techniques cannot be simply applied for raising the effective clock
speed. Look-ahead pipelining in Parhi (1999) may be applicable. But the introduced
hardware overhead can be intolerable. On the other hand, parallel processing can be
effective in increasing the system throughput. Unfortunately, direct application of this
technique will cause hardware and power consumption to increase linearly, which is against
the requirement of modern portable computing devices.

In this chapter, we will focus on MAP-based Turbo decoder design since MAP-based Turbo
decoder significantly outperforms SOVA-based Turbo decoders in terms of Bit-Error-Rate
(BER). In addition, MAP decoders are more challenging than SOVA decoders in high speed
design (Wang 2007). Interested readers are referred to Yeo (2003) and Wang (2003c) for high
data-rate SOVA or Turbo/SOVA decoder design. The rest of the chapter is organized as
follows. In Section 2, we give background information about Turbo codes and discuss
simple serial decoder structure. In Section 3, we will address high speed recursion
architectures for MAP decoders. Both Radix-2 and Radix-4 recursion architectures are
investigated. In Section 4, we present area-efficient parallel processing schemes and
associated parallel decoding architectures. We conclude the chapter in Section 5.

2. Background of Turbo Codes

ys
LLR(k )
1 SISO
Input yp
Lex (k )
buffer
Load Interleave Wtbk
y2p memory

Address
Generator

Fig. 1. A serial Turbo decoder architecture.

A typical Turbo encoder consists of two recursive systematic convolutional (RSC) encoders
and an interleaver between them (Wang 1999). The source data are encoded by the first RSC
encoder in sequential order while its interleaved sequence is encoded by the second RSC
encoder. The original source bits and parity bits generated by two RSC encoders are sent out
in a time-multiplexed way. Interested reader are referred to Berrou (1993) for details. Turbo
code usually works with large block sizes for the reason that the larger the block size, the
better the performance in general. In order to facilitate iterative decoding, the received data
High-Speed VLSI Architectures for Turbo Decoders 153

of a whole decoding block have to be stored in a memory, whose size is proportional to the
Turbo block size. Hence, Turbo decoders usually require large memory storage. Therefore
serial decoding architectures are widely used in practice.

A typical serial Turbo decoder architecture is shown in Fig. 1. It has only one soft-input soft-
output (SISO) decoder, which works in a time-multiplexed way as proposed in Suzuki
(2000). Each iteration is decomposed into two decoding phases, i.e., the sequential decoding
phase, in which the data are processed in sequential order, and the interleaved decoding phase,
in which the source data are processed in an interleaved order. Both probability MAP
algorithm proposed in Berrou (1993) and SOVA proposed in Hagenauer (1989) can be
employed for the SISO decoding.

The serial Turbo decoder includes two memories: one is used for received soft symbols,
called the input buffer or the receiver buffer, the other is used to store the extrinsic
information, denoted as the interleaver memory. The extrinsic information is feedback as the
a priori information for next decoding. The input buffer is normally indispensable. With
regard to the interleaver memory, either two ping-pong buffers can be used to complete one
Load and one Write operations required to process each information bit within one cycle or
one single-port memory can be employed to fulfil the two required operations within two
clock cycles. A memory-efficient architecture was presented by Wang (2003a) and Parhi,
which can process both Read and Write operations at the same cycle using single-port
memories with the aid of small buffers.

Turbo decoder works as follows. The SISO decoder takes soft inputs (including the received
systematic bit ys and the received parity bit y 1p or y 2p ) from the input buffer and the a
priori information from the interleaver memory. It outputs the log likelihood ratio LLR (k ) ,
and the extrinsic information, Lex (k ) , for the k-th information bit in the decoding sequence.
The extrinsic information is sent back as the new a priori information for next decoding. The
interleaving and de-interleaving processes are completed in an efficient way. Basically the
data are loaded according to the current decoding sequence. For instance, the extrinsic
information is loaded in sequential order at sequential decoding phase while being loaded
in the interleaved order at the interleaved decoding phase. After processing, the new
extrinsic information is written back to the original places. In this way, no de-interleave
pattern is required for Turbo decoding.

3. High Speed Log-MAP Decoder Design

3.1 Maximum A Posterior (MAP) algorithm
As discussed in Section 2, practical Turbo decoders usually employ serial decoding
architectures, such as Suzuki (2000), for area-efficiency. Thus, the throughput of a Turbo
decoder is highly limited by the clock speed and the maximum number of iterations to be
performed. To facilitate iterative decoding, Turbo decoders require soft-input soft-output
decoding algorithms, among which the probability MAP algorithm proposed in Bahl (1974)
is widely adopted for its excellent performance.
154 VLSI

The MAP algorithm is commonly implemented in log domain, thus called Log-MAP (Wang
1999). The Log-MAP algorithm involves recursive computation of forward state metrics
(simply called  metrics) and backward state metrics (simple called  metrics). The log-
likelihood-ratio is computed based on the two types of state metrics and associated branch
metrics (denoted as  metrics). Due to different recursion directions in computing and
metrics, a straightforward implementation of Log-MAP algorithm will not only consume
large memory but also introduce large decoding latency. The sliding window approach was
proposed by Viterbi (1998) to mitigate this issue. In this case, pre-backward recursion
operations are introduced for the warm-up process of real backward recursion. For clarity,
we denote the pre-backward recursion computing unit as unit and denote the real
(effective or valid) backward recursion unit as unit. The details of Log-MAP algorithm
will not be given in the chapter. Interested readers are referred to Wang (1999).

The timing diagram for typical sliding-window-based Log-MAP decoding is shown in Fig.
2, where SB1, SB2, etc, stand for consecutive sub-blocks (i.e., sliding windows), the branch
metrics computation was computed together with pre-backward recursion, though not
shown in the figure for simplicity and clarity.

The structure of a typical serial Turbo decoder based on log-MAP algorithm is shown in Fig.
3, where the soft-output unit is used to compute LLR and extrinsic information (denoted as
Lex), the interleaver memory is used to store the extrinsic information for next (phase of)
decoding. It can be seen from the figure that both the branch metric unit (BMU) and soft-
output unit (SOU) can be pipelined for high speed applications. However, due to recursive
computation, three state metrics computation units form the high-speed bottleneck. The
reason is that the conventional pipelining technique is not applicable for raising the effective
processing speed unless one MAP decoder is used to process more than one Turbo code
blocks or sub-blocks as discussed in Lee (2005). Among various high-speed recursion
architectures in the literature such as Lee (2005), Urard (2004), Boutillon (2003), Miyouchi
(2001) and Bickerstaff (2003), the designs presented in Urard (2004) and Bickerstaff (2003)
are most attractive. In Urard (2004), an offset-add-compare-select (OACS) architecture is
proposed to replace the traditional add-compare-select-offset (ACSO) architecture. In
addition, the look-up table (LUT) is simplified with only 1-bit output, and the computation
of absolute value is avoided through introduction of the reverse difference of two competing
path (or state) metrics. An approximate 17% speedup over the traditional Radix-2 ASCO
architecture was reported. With one-step look-ahead operation, a Radix-4 ACSO
architecture can be derived. Practical Radix-4 architectures such as Miyouchi (2001) and
Bickerstaff (2003) always involve approximations in order to achieve higher effective
speedup. For instance, the following approximation is adopted in Bickerstaff (2003):
max* (max*(A, B), max*(C, D))=max*(max(A, B), max(C, D)), (1)
where
| A B|
max*(A, B)= max(A,B)+log(1+ e ). (2)

This Radix-4 architecture can generally improve the processing speed (equivalent to twice of
its clock speed) by over 40% over the traditional Radix-2 architecture, and it has de facto the
highest processing speed among all existing (MAP decoder) designs found in the literature.
High-Speed VLSI Architectures for Turbo Decoders 155

However, the hardware will be nearly doubled compared to the traditional ACSO
architecture presented in Urard (2004).

In this section, we will first present an advanced Radix-2 recursion architecture based on
algorithmic transformation, approximation and architectural level optimization, which can
achieve comparable processing speed as the state-of-the-art Radix-4 design while having
significantly lower hardware complexity. Then we discuss an improved Radix-4 architecture
that is 32% faster than the best existing approach.

Fig. 2. The timing diagram for Log-MAP decoding.

Fig. 3. A Log-MAP Turbo decoder structure.

156 VLSI

3.2 An Advanced High-Speed Radix-2 Recursion Architecture for MAP Decoders

For convenience in later discussion, we first give a brief introduction to MAP-based Turbo
decoder structure. As shown in Fig. 3, the branch metrics unit (BMU) takes inputs from the
receiver buffer and the interleaver memory. The outputs of BMU are directly sent to the pre-
backward recursion unit (i.e., unit). The previously stored branch metrics for consecutive
sliding windows are input to the forward recursion unit (i.e.,  unit) and the effective
backward recursion unit (i.e., unit), respectively. The soft output unit (SOU) that is used
to compute the log likelihood ratio (LLR) and the extrinsic information (Lex) takes inputs
from the previously stored metrics, the currently computed  metrics and the previously
stored branch metrics (). The SOU starts to generate soft outputs after the branch metrics
have been computed for the first two sliding windows. It can be observed that the high-
speed bottleneck of a Log-MAP decoder lies in the three recursive computation units since
both BMU and SOU can be simply pipelined for high-speed applications. Thus, this section
is dedicated to the design of high-speed recursive computation units, i.e.,  and  units
shown in Fig. 3.

It is known from Log-MAP algorithm that all the three recursion units have similar
architectures. So we will focus our discussion on design of  units. The traditional design for
 computation is illustrated in Fig. 4, where the ABS block is used to compute the absolute
value of the input and the LUT (i.e., look-up table) block is used to implement a nonlinear
x
function log(1+ e ), where x > 0. For simplicity, only one branch (i.e., one state) is drawn.
The overflow approach (Wu 2001) is assumed for normalization of state metrics as used in
conventional Viterbi decoders.

Fig. 4. The original recursion architecture: Arch-O.

It can be seen that the computation of the recursive loop consists of three multi-bit additions,
the computation of absolute value and a random logic to implement the LUT. As there is
only one delay element in each recursive loop, the traditional retiming technique in Denk
(1998) can not be used to reduce the critical path.
High-Speed VLSI Architectures for Turbo Decoders 157

Fig. 5. An advanced Radix-2 fast recursion architecture.: Arch-A

In this work, we propose an advanced Radix-2 recursion architecture shown in Fig. 5. Here
we first introduce a difference metric for each competing pair of states metrics (e.g.,  and
in Fig. 4) so that we can perform the front end addition and the subtraction operations
simultaneously in order to reduce the computation delay of the loop. Secondly, we employ a
generalized LUT (see GLUT in Fig. 5) that can efficiently avoid the computation of absolute
value instead of introducing another subtraction operation as in Urard (2004). Thirdly, we
move the final addition to the input side as with the OACS architecture in Boutillon (2003)
and then utilize one stage carry-save structure to convert a 3-number addition to a 2-number
addition. Finally, we make an intelligent approximation in order to further reduce the
critical path.

The following equations are assumed for the considered recursive computation shown in
Fig. 5:

 0[k  1]  max*( 0[k ]   0[k ], 1[k ]   3[k ]), (3)

 2[k  1]  max*( 0k ]   3[k ], 1[k ]   0[k ]),

where max* function is defined in (2).

In addition, we split each state metrics into two terms as follows:

 0[ k ]   0 A[ k ]   0 B [ k ],
 1[ k ]   1 A[ k ]   1 B [ k ],
 2 [ k ]   2 A[ k ]   2 B [ k ]. (4)
158 VLSI

Similarly, the corresponding difference metric is also split into two terms:
 01 A [k ]   0 A [k ]  1 A [k ],
 01B [k ]   0 B [k ]  1B [k ]. (5)
In this way, the original add-and-compare operation is converted as an addition of three
numbers, i.e.,
( 0   0 )  (1   3 )  ( 0   3 )   01A   01B
(6)
where  0   3 is computed by BMU, the time index [k] is omitted for simplicity. In
addition, the difference between the two outputs from two GLUTs, i.e.,  01B , can be
neglected. From extensive simulations, we found that this small approximation doesn’t
cause any performance loss in Turbo decoding with either AWGN channels or Raleigh
fading channels. This fact can be simply explained in the following. If one competing path
metrics (e.g., p 0   0   0 ) is significantly larger than the other one (e.g., p1   1   3 ),
the GLUT output will not change the decision anyway due to their small magnitudes. On
the other hand, if the two competing path metrics are so close that adding or removing a
value from one GLUT may change the decision (e.g., from p0>p1 to p1>p0), picking any
survivor (p0 or p1) should not make big difference.

At the input side, a small circuitry shown in Fig. 6 is employed to convert an addition of 3
numbers to an addition of 2 numbers, where FA and HA represents full-adder and half-
adder respectively, XOR stands for exclusive OR gate, d0 and d1 correspond to the 2-bit
output of GLUT. The state metrics and branch metrics are represented with 9 and 6 bits,
respectively in this example. The sign extension is only applied to the branch metrics. It
should be noted that an extra addition operation (see dashed adder boxes) is required to
integrate each state metric before storing it into the  memory.

[8] [5] ] ] [2] [2] [1] d1 [0] d0

[1] [0]

XOR HA HA FA FA

Fig. 6. A carry-save structure in the front end of Arch-A.

The generalized LUT (GLUT) structure is shown in Fig. 7, where the computation of
absolute value is eliminated by including the sign bit into 2 logic blocks, i.e., Ls2 and ELUT,
where the Ls2 function block is used to detect if the absolute value of the input is less than
2.0, and the ELUT block is a small LUT with 3-bit inputs and 2-bit outputs. It can be derived
that Z  S b7   b4  b3  S (b7  b4b3 ) . It was reported in Gross (1998) that using two output
values for the LUT only caused a performance loss of 0.03 dB from the floating point
simulation for a 4-state Turbo code. The approximation is described as follows:

If |x|<2, f(x) = 3/8; else f(x) = 0; (7)

High-Speed VLSI Architectures for Turbo Decoders 159

where x and f(x) stands for the input and the output of the LUT, respectively. In this
approach, we only need to check if the absolute value of the input is less than 2 or not,
which can be performed by the Ls2 block in Fig. 7. A drawback of this method is that its
performance would be significantly degraded if only two bits are kept for the fractional part
of the state metrics, which is generally the case.

Fig. 7. The structure of GLUT used in Arch-A.

|x| 0.0 0.50 1.0 1.50

f(x) ¾ 2/4 ¼ ¼
Table 1. The proposed LUT approximation

In our design, both the inputs and outputs of the LUT are quantized with 4 levels. The
details are shown in Table 1. The inputs to ELUT are treated as a 3-bit signed binary
number. The outputs of ELUT are ANDed with the output of Ls2 block. This means, if the
absolute value of the input is greater than 2.0, the output from the GLUT is set as 0.
Otherwise the output from ELUT will be the final output.

The ELUT can be implemented with combinational logic for high speed applications. Its
computation latency is smaller than the latency of Ls2 block. Therefore, the overall latency
of the GLUT is almost the same as the above-discussed simplified method, where the total
delay consists of 1 MUX delay and the computation delay of Ls2.

After all the above optimization, the critical path of the recursive architecture is reduced to 2
multi-bit additions, one 2:1 MUX operation and 1-bit addition operation, which saves nearly
2 multi-bit adder delay compared to the traditional ACSO architecture. We will show
detailed comparisons in subsection 3.4.

3.3 An Improved Radix-4 Architecture for MAP Decoders

In the following, we discuss an improved Radix-4 recursion architecture. The computation
for  0[ k  2] is expressed as follows:
160 VLSI

 0[k  2]  max*( 0[k  1]   0[k  1], 1[k  1]   3[k  1])

 max*(max*(0[k ]   0[k ], 1[k ]   3[k ])   0[k  1],

max*( 2k ]   2[k ],  3[k ]   1[k ])   3[k  1]),

(8)

where

1[k  1]  max*( 2[k ]   2[k ],  3[k ]   1[k ]). (9)

In Bickerstaff (2003), Lucent Bell Labs proposed the following approximation:

 0[k  2]  max*(max( 0[k ]   0[k ], 1[k ]   3[k ])   0[k  1],

max( 2k ]   2[k ],  3[k ]   1[k ])   3[k  1]).

(10)

 0 [k ]
 0 [k ]   0 [k  1]

 3 [k ]   0 [k  1]

1 [ k ]  0 [ k  2]

 2 [k ]
 3[k ]   2 [k  1]
 3[k ]   0 [k  1]

 3 [k ]
Fig. 8. The Radix-4 architecture proposed by Lucent Bell Labs: Arch-L.

This approximation is reported to have a 0.04 dB performance loss compared to the original
Log-MAP algorithm. The architecture to implement the above computation is shown in Fig.
8 for convenience in later discussion. As it can be seen, the critical path consists of 4 multi-
bit adder delay, one generalized LUT delay (Note: the LUT1 block includes absolute value
computation and a normal LUT operation) and one 2:1 MUX delay.

Similarly, we can take an alternative approximation as follows:

 0[k  2]  max(max*( 0[k ]   0[k ], 1[k ]   3[k ])   0[k  1],

max*( 2k ]   2[k ],  3[k ]   1[k ])   3[k  1]).

(11)
High-Speed VLSI Architectures for Turbo Decoders 161

 0 A [k ]
 0 B [k ]  0 B [k  2]
 0 [k ]   0 [k  1]

 3[k ]   0 [k  1]
1 A [k ]
1B [k ]
 0 A [ k  2]
 2 A [k ]
 2 B [k ]
 3[ k ]   2 [k  1]

 3 [k ]   1[k  1]
 3 A [k ]
 3B [ k ]

Fig. 9. The improved Radix-4 recursion arch.: Arch-B.

Intuitively, Turbo decoder employing this new approximation should have the same
decoding performance as using equation (7). While directly implementing (11) does not
bring any advantage to the critical path, we intend to take advantages of the techniques that
we developed in subsection 3.2. The new architecture is shown in Fig. 9. Here we split each
state metric into two terms and we adopt the same GLUT structure as we did before. In
addition, a similar approximation is incorporated as with Arch-A. In this case, the outputs
from GLUT are not involved in the final stage comparison operation. It can be observed that
the critical path of the new architecture is close to 3 multi-bit adder delay. To compensate
for all the approximation introduced, the extrinsic information generated by the MAP
decoder based on this new Radix-4 architecture should be scaled by a factor around 0.75.

3.4 Performance Comparisons

For quantitative comparison in hardware complexity and processing speed, we used TSMC
0.18 um standard cells to synthesize one  unit for 5 different recursion architectures, i.e., 1)
the traditional ACSO architecture: Arch-O, 2) the reduced-precision Radix-2 recursion
architecture presented in Urard (2004): Arch-U, 3) the Radix-4 architecture proposed by
Lucent: Arch-L, 4) the advanced Radix-2 recursion architecture presented in this work:
Arch-A, and 5) the improved Radix-4 recursion architecture: Arch-B. All the state metrics
are quantized as 9 bits while the branch metrics are represented using 6 bits. The detailed
synthesis results are listed in Table 2.

It can be observed that the proposed Radix-2 architecture has comparable processing speed
as the Radix-4 architecture proposed by Bell Labs with significantly lower complexity, while
the improved Radix-4 architecture is 32% faster with only 9% extra hardware. This amount
of hardware overhead should be negligible compared to an entire Turbo decoder. It can also
be seen that the new Radix-4 architecture achieves twice speedup over the traditional Radix-
2 recursion architecture.
162 VLSI

Max clock Relative Relative

Frequency (Mhz) Area Processing Speed
Arch-O 241 1.0 1.0
Arch-U 355 1.14 1.39
Arch-L 182 1.82 1.51
Arch-A 370 1.03 1.54
Arch-B 241 1.99 2.0
Table 2. Comparison for various recursion architectures

We have performed extensive simulations for Turbo codes using the original MAP and
using various approximations. Fig. 10 shows the BER (bit-error-rate) performance of a rate-
1/3, 8-state, block size of 512 bits, Turbo code using different MAP architectures. The
simulations were undertaken under the assumption of AWGN channel and BPSK signaling.
A maximum of 8 iterations was performed. More than 40 million random information bits
were simulated for both Eb/No=1.6 dB and Eb/No=1.8dB cases. It can be noted from Fig.
10 that there is no observable performance difference between the true MAP algorithm and
two approximation methods associated with the proposed recursion architectures while the
approximation employed in Urard (2004) caused approximately 0.2 dB performance
degradation in general.

Fig. 10. Performance comparisons between the original MAP and some approximations.
High-Speed VLSI Architectures for Turbo Decoders 163

We argue that the proposed Radix-4 recursion architecture is optimal for high-speed MAP
decoders. Any (significantly) faster recursion architecture (e.g., a possible Radix-8
architecture) will be at the expense of significantly increased hardware. On the other hand,
when the target throughput is moderate, these fast recursion architectures can be used to
reduce power consumption because of their dramatically reduced critical paths.

In general, the throughput of a serial Turbo decoder is significantly limited by recursive

computation and VLSI technology used to implement it. On the other hand, increasing the
clock frequency would cause the dynamic power dissipation to increase linearly. Therefore,
it is not a good option to increase the clock frequency for high throughput applications.

4. Efficient Parallel Turbo Decoding Architectures

In this section, we first introduce an area-efficient parallel Turbo decoding scheme, where a
data frame is partitioned into several equivalent segments with proper overlap between
adjacent segments, and sliding-window-based Turbo decoding is then applied to each
segment. The goal is to maintain the same memory requirement as serial decoding case
while increasing the complexity of the logic components only, thus linearly increase the
system throughput with sub-linearly increased hardware.

The major challenge lies in real implementation. As each memory (receiver memory or
extrinsic information memory) needs to support multiple data (for multiple component soft-
input soft-output decoders) at the same cycle, memory access conflicts are inevitable unless
M-port (M = parallelism level) memory is used, which contradicts the original low-
complexity design objective. Two different solutions are proposed in this work. First of all, if
interleave patterns are free to design, we can adopt dividable interlavers, which inherently
ensure no memory access conflict after proper memory partition. For practical applications
wherein Turbo code interleave patterns are fixed, e.g., 3GPP and 3GPP2, a more generic
solution is introduced in this chapter. By introducing some small buffers for data and
addresses as well, we are able to avoid memory access conflict under any practical random
interleavers. Combining all the proposed techniques, it is estimated that 200 Mb/s Turbo
decoder is feasible with current CMOS technology with moderate complexity.

4.1 Area-efficient parallel decoding schemes

Parallel processing has been widely used in low power and high throughput circuit design.
Multiple serial Turbo decoders can be concatenated in a parallel way (see Fig. 11 A) or in a
serial way. In either case, both the area and power are increased linearly as the throughput
increases. The area-efficient parallel Turbo decoding schemes were first proposed in Wang
(2001). The idea is to let multiple SISO decoders work on the same data frame
simultaneously. In this case, only the hardware of the SISO decoder part needs to be
increased while the total hardware in the memory part remains unchanged on a large extent
(see Fig. 11 B). As the memory part usually dominates the overall hardware of a Turbo
decoder, the area-efficient parallel Turbo decoding architectures are expected to achieve
multiple times the throughput of a serial decoding decoder with a small fraction of
hardware overhead. Unfortunately, the real implementation issues were not covered in the
original work.
164 VLSI

Fig. 11. Traditional vs. area-efficient parallel Turbo decoding.

a) A simple area-efficient 2-level parallel Turbo decoding scheme.

Fig. 12. b) Multi-level parallel Turbo decoding scheme based on sliding-window approach.

A simple area-efficient 2-level parallel Turbo decoding scheme is shown in Fig. 12 a), where
a sequence of data is divided into two segments with equivalent length. There is an overlap
with length of 2D between two segments, where D equals the sliding window size if Log-
MAP (or MAX-Log-MAP) algorithm is employed in the SISO decoder or the survivor length
if SOVA is employed, as proposed in Wang (2003c). The two SISO decoders work on two
different data segments (with overlap) at the same time. Fig. 12 b) shows ( a portion of) a
sliding-window-based area-efficient multi-level parallel Turbo decoding data flow, where
High-Speed VLSI Architectures for Turbo Decoders 165

SISO decoders responsible for two adjacent data segments are processing data in reverse
directions in order to reuse some of the previously computed branch metrics and state
metrics. For other parallel decoding schemes with various trade-offs, interested reader are
referred to Wang (2001) and Zhang (2004).

In this section, we will focus on area-efficient two-level parallel decoding schemes. It can be
observed from Fig. 12.a) that two data accesses per cycle are required for both the receiver
buffer and the interleaver memory (assuming two ping-pong buffers are used for the
interleaver memory). Using a dual-port memory is definitely not an efficient solution since a
normal dual-port memory consumes as much area as two single-port memories with the
same memory size.

4.2 Area-efficient parallel decoding architectures

In order to maintain the total hardware of these memories, we choose to partition each
memory into multiple segments to support multiple data accesses per cycle.

Memory partitioning can be done in various ways. An easy and yet effective way is to
partition the memory according to a number of least significant bits (lsb’s) of the memory
address. To partition a memory into two segments, it can be done according to the least
significant bit (lsb). For the resultant two memory segments, one contains data with even
addresses and the other contains data with odd addresses.

The memory partitioning can also be done in a fancy way, e.g., to partition the memory into
2 segments, let the 1st memory segment contain data with addresses b2b1b0 ={0, 2, 5, 7}, and
the 2nd segment contain data with addresses b2b1b0 ={1, 3, 4, or 6}, where b2b1b0 denotes
the 3 lsb’s. Depending on the applications, different partitioning schemes may lead to
different hardware requirements.

For the two memory segments, it is possible, in principle, to support two data accesses
within one cycle. However, it is generally impossible to find an efficient partitioning scheme
to ensure the target addresses (for both Load and Write operations) are always located in
different segments at each cycle because the Turbo decoder processes the data in different
orders during different decoding phases. Consider a simple example, given a sequential
data sequence {0, 1, 2, 3, 4, 5, 6, 7}, the interleaved data sequence is assumed as {2, 5, 7, 3, 1, 0,
6, 4}. If we partition the memory into two segments according to the sequential decoding
phase, i.e., one segment contains data sequence {0, 1, 2, 3} and the other has {4, 5, 6 7}.
During the sequential phase, SISO-1 works on data set {0, 1, 2, 3} and SISO-2 works on {4, 5,
6, 7}. So there is no data conflict (note: the overlap between two segments for parallel
decoding is ignored in this simple example). However, during the interlaved decoding
phase, SISO-1 will process data in set {2, 5, 7, 3} and SISO-2 will process data in set {1, 0, 6,
4}, both in sequential order. It is easy to see that, at the first cycle, both SISO decoders
require data located in the first segment (1st and 2nd data in the input data sequence). Thus a
memory access conflict occurs. More detailed analysis can be found in Wang (2003b). The
memory access issue in parallel decoding can be much worse when multiple rates and
166 VLSI

multiple block sizes of Turbo codes are supported in a specific application, e.g., in 3GPP
CDMA systems.

In principle, the data access conflict problem can be avoided if the Turbo interleaver is
specifically designed. A generalized even-odd interleave is defined below:
 All even indexed data are interleaved to odd addresses,
 All odd indexed data are interleaved to even addresses.
Back to the previous example, an even-odd interleaver may have an interleaved data
sequence as {3, 6, 7, 4, 1, 0, 5, 2}.

With an even-odd interleaver, we can partition each memory into two segments: one
contains even-addressed data and the other with odd-addressed data.

Fig. 13. Data processing in either decoding phase with an even-odd interleaver.

As shown in Fig. 13, the data are processed in an order of {even address, odd address, even ,
odd, …} in Phase A (i.e., the sequential phase) and an order of {odd address, even address,
odd, even, ..} in Phase B (i.e., the interleaved phase). If we let SISO-1 and SISO-2 start
processing from different memory segments, then there would be no data access conflict
during either decoding phase. In other words, it is guaranteed that both SISO decoders
always access data located in different memory segments.

More complicated interleavers can be designed to support multiple (M>=2) data accesses
per cycle. The interested readers are referred to Wang (2003c), He (2005), Giulietti (2002),
Bougard (2003), and Kwak (2003) for details.

In real applications, the Turbo interleaver is usually not free to design. For example, they are
fixed in WCDMA and CDMA 2000 systems. Thus, a generic parallel decoding architecture is
desired to accommodate all possible applications. Here we propose an efficient memory
arbitration scheme to resolve the problem of data access conflict in parallel Turbo decoding.

The fundamental concept is to partition one single-port memory into S (S>=2) segments and
use B (B >=2) small buffers to assist reading from or writing data back to the memory, where
S does not have to be the same as B. For design simplicity, both S and B are normally chosen
to be a power of 2. S is better chosen to be a multiple of B. In this chapter, we will consider
High-Speed VLSI Architectures for Turbo Decoders 167

only one simple case, i.e., S=2, B=2. For all other cases of B=2, they can be easily extended
from the illustrated case. However, for cases with B>2, the control circuitry will be much
more complicated.

For the receiver buffer, only the Load operation is involved while bother Load and Write
operations are required for the interleaver memory. So we will focus our discussion on the
interleaver memory.

Fig. 14. The area-efficient parallel decoding architecture.

With regard to the interleaver memory, we assume two ping-pong buffers (i.e., RAM1 and
RAM2 shown in Fig.14) are used to ensure that one Load and one Write operations can be
completed within one cycle. Each buffer is partitioned into two segments: Seg-A contains
even addressed data and Seg-B contains odd-addressed data. For simplicity, we use Seg-A1
to represent the even-addressed part of RAM1, Seg-B1 to represent the odd-addressed part
of RAM1. The similar representations are used for RAM2 as well.

An area-efficient 2-parall Turbo decoding architecture is shown in Fig. 14. The interleaver
address generator (IAG) generates two addresses for two SISO decoders respectively at each
cycle. They must belong to one of the following cases: (1) two even addresses, (2) two odd
addresses and (3) one even and one odd address. In Case 1, two addresses are put into Read
Address Buffer 1 (RAB1), In Cases 2, two addresses are put in Read Address buffer 2
(RAB2). In Cases 3, the even address goes to RAB1 and the odd address goes to RAB2.
168 VLSI

A small buffer called Read Index Buffer (RIB) is introduced in this architecture. This buffer
is basically a FIFO (first-in first-out). Two bits are stored at each entry. The distinct four
values represented by the two bits have following meanings:
 00: 1st and 2nd even addresses for SISO-1 and SISO-2 respectively,
 11: 1st and 2nd odd addresses for SISO-1 and SISO-2 respectively,
 01: the even address is used for SISO-1 and the odd address for SISO-2,
 10: the even address is used for SISO-1 and the odd address for SISO-2.
After decoding for a number of cycles (e.g., 4 ~ 8 cycles), both RAB1 and RAB2 will have a
small amount of data. Then the data from the top of RAB1 and RAB2 are used respectively
as the addresses to Seg-A1 and Seg-B1 respectively. After one cycle, two previously stored
extrinsic information symbols are loaded to Load Data Buffer 1 (LAB1) and Load Data
Buffer 2 (LDB2). Then the data stored in RIB will be used to direct the outputs from both
load buffers to feed both SISO decoders. For example, if the data at the top of RIB is 00, then
LDB1 output two data to SISO-1 and SISO-2 respectively. The detailed actions of address
buffers and data buffers controlled by RIB for loading data are summarized in Table 3.

Value of Action of address buffers Action of data buffers

RIB

0 RAB1 load nothing, RAB2 LDB1 output nothing, LDB2 out 2

loads 2 addresses, 1st for Seg- data, 1st for SISO1 and 2nd for
1 and 2nd for Seg-2 SISO2
1 RAB1 load 1 address for Seg- LDB1 output 1 for SISO1 and
1, RAB2 load 1 address for LDB2 out 1 for SISO2
Seg-2
2 RAB2 load nothing, RAB1 LDB1 output 2 data, 1st for SISO1
load 2 addresses, 1st for Seg-1 and 2nd for SISO2
and 2nd for Seg-2
3 RAB1 load 1 address for Seg- LDB1 output 1 for SISO2 and
1, RAB2 load 1 for Seg-2 LDB2 out 1 for SISO1
Table 3. Summary of loading data operations.

B4 B5 B4

A3 A3 A5 A3

B2 B2 A4 B2 A4 B2

A2 B3 A2 B3 A2 B3 A2

A1 B1 A1 B1 A1 B1 A1 B1 A1 B1

t=1 t=2 t=3 t=4 t=5

(a) load addresses to RAB1 and RAB2
High-Speed VLSI Architectures for Turbo Decoders 169

B5 B4

A5 A3 A5 A3

A4 B2 A4 B2 A4 B2

B3 A2 B3 A2 B3 A2 B3 A2

A1 B1 A1 B1 A1 B1 A1 B1 A1 B1

t=2 t=3 t=4 t=5 t=6

(b) load data to RDB1 and RDB2

t=6 t=7 t=8 t=9 t=10

B5 B4 B5 B4 B5 B4 B5 B4 B5

A5 A3 A5 A3 A5 A3 A5 A5

A4 B2 A4 B2 A4 A4

B3 A2 B3 A2 B3

A1 B1

(c) output data to SISO1 and SISO2

Fig. 15. Procedure of loading data from memory to SISO decoders.

A simple example is shown in Fig. 15 to illustrate the details of loading data from memory
to SISO decoders, where, for instance, A3 (B5) denotes the address of data to be loaded or
the data itself corresponding to memory segment A (segment B) for the 3rd (5th) data in the
processing sequence. Fig. 15 (a) shows the details of feeding two dada addresses to two
address buffers at every cycle. Fig. 15 (b) shows the details of feeding one data to each data
buffer at every cycle. The details of outputting the required two data to both SISO decoders
are shown in Fig. 15 (c). As can be seen, both SISO1 and SISO2 can get their required data
that may be located in the same memory segment at the same cycle (e.g., t=7 in this
example). A tiny drawback of this approach is that a small fixed latency is introduced.
Fortunately this latency is negligible compared with the overall decoding cycles for a whole
Turbo code block in most applications.

In general (e.g., when sliding-window-based MAP algorithm is employed for SISO

decoders), the write sequence for each SISO decoder is the delayed version of the read
sequence. So the write index buffer and write address buffers can be avoided by using delay
lines as shown in Fig. 14.

At each cycle, two extrinsic information symbols are generated from the two SISO decoders.
They may both feed Write Data Buffer 1 (WDF1), or both feed Write Data Buffer 2 (WDF2),
or one feeds WDF1 and the other feeds WDF2. The actual action is controlled by the
delayed output from RIB.
170 VLSI

Similar to loading data, after the same delay (e.g., 4 ~ 8 cycles) from the first output of either
SISO decoder, the data from WDB1 and WDB2 will be written back to Seg-B1 and Seg-B2
respectively. In the next decoding phase, RAM-1 and RAM-2 will exchange roles. So some
extra MUXs must be employed to facilitate this functional switch.

5.3 Implementation issues and simulation results

The integer numbers shown in Fig. 14 indicated the data flow associated with each buffer.
For example, {0, -1, -2} is shown along the output line of LDB1, which means LDB1 may
output 0, 1, or 2 data at each cycle. It can be observed that all data and address buffers work
in a similar way. One common feature of these buffers is that they have a constant data flow
(+1 or –1) at one end (Load or Write) while having an irregular data flow ({0, -1, -2} or {0, +1,
+2}) at the other end (Write or Load). On the average, incoming and outgoing data are
largely balanced. So the buffer sizes can be very small.

The RIB can be implemented with a shift register as it has a regular data flow in both ends
while a 3-port memory suffices to fulfill the functions of the rest buffers.

It has been found from our cycle-accurate simulations that it is sufficient to choose a buffer
length of 25 for all buffers if the proposed 2-parallel architecture is applied in either
WCDMA or CDMA2000 systems. The maximum Turbo block size for WCDMA system is
approximately 5K bits. The lowest code rate is 1/3. Assume both the received soft inputs
and the extrinsic information symbols are expressed as 6 bits per symbol.
 The overall memory requirement is 5K*(2+3)*6=150K bits.
 The total overhead of small buffers is approximately 25*3*2*(3*6+2*6) +
25*3*13*2~= 6K bits, where the factor 3 accounts for 3 ports of memory, 13
represents each address contains 13 binary bits.

It can be seen that the overhead is about 4% of total memory. With CDMA2000 systems, the
overhead occupies an even smaller percentage in the total hardware because the maximum
Turbo block size is several times larger.

For WCDMA systems, it is reasonable to assume the total area of a Log-MAP decoder
consumes less than 15% of total hardware of a serial Turbo decoder. Thus, we can achieve
twice the throughput with the proposed 2-parallel decoding architecture while spending
less than 20% hardware overhead. If SOVA is employed in the SISO decoder, the overhead
could be less than 15%. In either case, the percentage of the overhead will be even smaller in
CDMA2000 systems.

Assume 6 iterations are performed at the most, the proposed 2-level parallel architecture, if
implemented with TSMC 0.18 um technology, can achieve a minimum decoding throughput
of 370 M*2/(6*2) > 60 Mbps. If state-of-the-art CMOS technology (e.g., 65nm CMOS) is used,
we can easily achieve 100Mbps data-rate with the proposed 2-parallel decoding architecture.

If an 4-parallel architecture is employed, over 200Mbps data rate can be obtained, though a
direct extension of the proposed 2-paralell architecture will significantly complicate the
control circuitry.
High-Speed VLSI Architectures for Turbo Decoders 171

It is worthwhile to note, when the target throughput is moderately low, the proposed area-
efficient parallel Turbo decoding architecture can be used for low power design. The reason
is that the former architecture can work at a much lower clock frequency and the supply
voltage can thus be reduced significantly.

6. Conclusions
Novel fast recursion architecture for Log-Map decoders have been presented in this chapter.
Experimental results have shown that the proposed fast recursion architectures can increase
the process speed significantly over traditional designs while maintaining the decoding
performance. As a more power-efficient approach to increase the throughput of Turbo
decoder, area-efficient parallel Turbo decoding schemes have been addressed. A hardware-
efficient 2-parallel decoding architecture for generic applications is presented in detail. It has
been shown that twice the throughput of a serial decoding architecture can be obtained with
an overhead of less than 20% of an entire Turbo decoder. The proposed memory
partitioning techniques together with the efficient memory arbitration schemes can be
extended to multi-level parallel Turbo decoding architectures as well.

7. References
3rd Generation Partnership Project (3GPP), Technical specification group radio access
network, multiplexing and channel coding (TS 25.212 version 3.0.0),
http://www.3gpp.org.
3rd Generation Partnership Project 2 (3GPP2), http://www.3gpp2.org.
A. Giulietti et al. (2002), Parallel Turbo code interleavers: Avoiding collisions in
accesses to storage elements, Electron. Lett., vol. 38, no. 5, pp. 232–234, Feb.
2002.
A. J. Viterbi. (1998). An intuitive justification of the MAP decoder for convolutional codes,
IEEE J. Select. Areas Commun., vol.16, pp. 260-264, February 1998.
A. Raghupathy. (1998). Low power and high speed algorithms and VLSI architectures for
error control coding and adaptive video scaling, Ph.D. dissertation, Univ. of
Maryland, College Park, 1998.
B. Bougard et al. (2003). A scalable 8.7 nJ/bit 75.6 Mb/s parallel concatenated
convolutional (Turbo-) codec, in IEEE ISSCC Dig. Tech. Papers, 2003, pp.
152–153.
C. Berrou, A. Clavieux & P. Thitimajshia. (1993). Near Shannon limit error correcting coding
and decoding: Turbo codes, ICC’93, pp 1064-70.
E. Boutillon, W. Gross & P. Gulak. (2003). VLSI architectures for the MAP algorithm, IEEE
Trans. Commun., Volume 51, Issue 2, Feb. 2003, pp. 175 – 185.
E. Yeo, S. Augsburger, W. Davis & B. Nikolic. (2003). A 500-Mb/s soft-output Viterbi
decoder, IEEE Journal of Solid-State Circuits, Volume 38, Issue 7, July 2003,
pp:1234 – 1241.
172 VLSI

H. Suzuki, Z. Wang & K. K. Parhi. (2000). A K=3, 2Mbps Low Power Turbo Decoder for 3rd
Generation W-CDMA Systems, in Proc. IEEE 1999 Custom Integrated Circuits Conf.
(CICC), 2000, pp 39-42.
J. Hagenauer & P. Hoher. (1989). A Viterbi algorithm with soft decision outputs and its
applications, IEEE GLOBECOM, Dallas, TX, USA, Nov. 1989, pp 47.1.1-7.
J. Kwak & K Lee (2002). Design of dividable interleaver for parallel decoding in
turbo codes. Electronics Letters, vol. 38, issue 22, pp. 1362-64, Oct. 2002.
J. Kwak, S. M. Park, S. S. Yoon & K Lee (2003). Implementation of a parallel Turbo
decoder with dividable interleaver, ISCAS’03, vol. 2, pp. II-65- II68, May
2003.
K. K. Parhi. (1999). VLSI Digital signal Processing Systems, John Wiley & Sons, 1999.
L.Bahl, J.Jelinek, J.Raviv & F.Raviv. (1974). Optimal Decoding of Linear Codes for
minimizing symbol error rate, IEEE Trans.s Inf. Theory, vol. IT-20, pp.284-287,
March 1974.
M. Bickerstaff, L. Davis, C. Thomas, D. Garret & C. Nicol. (2003). A 24 Mb/s radix-4
LogMAP Turbo decoder for 3 GPP-HSDPA mobile wireless, in Proc. IEEE ISSCC
Dig. Tech. Papers, 2003, pp. 150–151.
P. Urard et al. (2004). A generic 350 Mb/s Turbo codec based on a 16-state Turbo decoder, in
Proc. IEEE ISSCC Dig. Tech. Papers, 2004, pp. 424–433.
S. Lee, N. Shanbhag & A. Singer. (2005). A 285-MHz pipelined MAP decoder in 0.18 um
CMOS, IEEE J. Solid-State Circuits, vol. 40, no. 8, Aug. 2005, pp. 1718 – 1725.
T. Miyauchi, K. Yamamoto & T. Yokokawa. (2001). High-performance programmable SISO
decoder VLSI implementation for decoding Turbo codes, in Proc. IEEE Global
Telecommunications Conf., vol. 1, 2001, pp. 305–309.
T.C. Denk & K.K. Parhi. (1998). Exhaustive Scheduling and Retiming of Digital Signal
Processing Systems, IEEE Trans. Circuits and Syst. II, vol. 45, no.7, pp. 821-838, July
1998
W. Gross & P. G. Gulak. (1998). Simplified MAP algorithm suitable for implementation of
Turbo decoders, Electronics Letters, vol. 34, no. 16, pp. 1577-78, Aug. 1998.
Y. Wang, C. Pai & X. Song. (2002). The design of hybrid carry-lookahead/carry-select
adders, IEEE Trans. Circuits and Syst. II, vol.49, no.1, Jan. 2002, pp. 16 -24
Y. Wu, B. D. Woerner & T. K. Blankenship, Data width requirement in SISO decoding with
module normalization, IEEE Trans. Commun., vol. 49, no. 11, pp. 1861–1868, Nov.
2001.
Y. Zhang & K. Parhi (2004). Parallel Turbo decoding, Proceedings of the 2004 International
Symposium on Circuits and Systems (ISCAS04), Volume 2, 23-26 May 2004, pp: II -
509-12 Vol.2.
Z Wang, Y. Tan & Y. Wang. (2003b). Low Hardware Complexity Parallel Turbo Decoder
Architecture, ISCAS’2003, vol. II, pp. 53-56, May 2003.
Z. He, S. Roy & Fortier (2005). High-Speed and Low-Power Design of Parallel
Turbo Decoder Circuits and Systems, IEEE International Symposium on
Circuits and Systems (ISCAS’05), 23-26 May 2005, pp. 6018 – 6021.
Z. Wang & K. Parhi. (2003a). Efficient Interleaver Memory Architectures for Serial Turbo
Decoding, ICASSP’2003, vol. II, pp. 629-32, May 2003.
Z. Wang & K. Parhi. (2003c). High Performance, High Throughput Turbo/SOVA Decoder
Design, IEEE Trans. on Commun., vol. 51, no 4, April 2003, pp. 570-79.
High-Speed VLSI Architectures for Turbo Decoders 173

Z. Wang, H. Suzuki & K. K. Parhi. (1999). VLSI implementation issues of Turbo decoder
design for wireless applications, in Proc. IEEE Workshop on Signal Process. Syst.
(SiPS), 1999, pp. 503-512.
Z. Wang, Z. Chi & K. K. Parhi. (2001). Area-Efficient High Speed Decoding Schemes for
Turbo/MAP Decoders, in Proc. IEEE ICASSP'2001, pp 2633-36, vol. 4, Salt Lake
City, Utah, 2001.
Z. Wang. (2000). Low complexity, high performance Turbo decoder design, Ph.D.
dissertation, University of Minnesota, Aug. 2000.
Z. Wang. (2007). High-Speed Recursion Architectures for MAP-Based Turbo Decoders, in
IEEE Trans. on VLSI Syst., vol. 15, issue 4, pp: 470-74, Apr. 2007.
174 VLSI
Ultra-High Speed LDPC Code Design and Implementation 175

X9

Ultra-High Speed LDPC Code

Design and Implementation
Jin Sha1, Zhongfeng Wang2 and Minglun Gao1
1Nanjing University, 2Broadcom Corporation
1China, 2U.S.A.

1. Introduction
The digital communications are ubiquitous and have provided tremendous benefits to
everyday life. Error Correction Codes (ECC) are widely applied in modern digital
communication systems. Low-Density Parity-Check (LDPC) code, invented by Gallager
(1962) and rediscovered by MacKay (1996), is one of the two most promising near-optimal
error correction codes in practice. Since its rediscovery, significant improvements have been
achieved on the design and analysis of LDPC codes to further enhance the communication
system performance. Due to its outstanding error-correcting performance (Chung 2001),
LDPC code has been widely considered in next generation communication standards such
as IEEE 802.16e, IEEE 802.3an, IEEE 802.11n, and DVB-S2. LDPC code is characterized by
sparse parity check matrix. One key feature associated with LDPC code is the iterative
decoding process, which enables LDPC decoder to achieve outstanding performance with
moderate complexity. However, the iterative process directly leads to large hardware
consumption and low throughput. Thus efficient Very Large Scale Integration (VLSI)
implementation of high data rate LDPC decoder is very challenging and critical in practical
applications.
A satisfying LDPC decoder usually means: good error correction performance, low
hardware complexity and high throughput. There are various existing design methods
which usually encounter the limitations of high routing overhead, large message memory
requirement or long decoding latency. To implement the decoder directly in its inherent
parallel manner may get the highest decoding throughput. But for large codeword lengths
(e.g., larger than 1000 bits), to avoid routing conflict, the complex interconnection may take
up more than half of the chip area. Both serial and partly parallel VLSI architectures are well
studied nowadays. However, none of these approaches are good for very high throughput
(i.e., multi-Gb/s) applications.
In this chapter, we present the construction of a new class of implementation-oriented LDPC
codes, namely shift-LDPC codes to solve these issues integrally. Shift-LDPC codes have
been shown to perform as well as computer generated random codes following some
optimization rules. A specific high-speed decoder architecture targeting for multi-Gb/s
applications namely shift decoder architecture, is developed for this class of codes. In
176 VLSI

contrast with conventional decoder architectures, the shift decoder architecture has three
major merits:
1) Memory efficient. By exploring the special features of the min-sum decoding algorithm,
the proposed architecture stores the message in a compact way which normally leads to
approximately 50% savings on message memory over the conventional design for high rate
codes.
2) Low routing complexity. Through introducing some novel check node information
transfer mechanisms, the complex message passing between variable nodes and check
nodes can be alleviated by the regular communication between check nodes and thus the
complex global interconnection networks can be replaced by local wires. One important fact
is that in the new architecture, check node processing units have few and fixed connections
with variable node processing units.
3) High parallel level of decoding. The architecture can normally exploit more levels of
parallelism in the decoding algorithm than conventional partially parallel decoder
architectures. In addition, the decoding parallel level can be further linearly increased and
the critical path can be significantly reduced by proper pipelining.
The chapter is organized as follows. Section 2 gives an overview of LDPC codes, with the
main attention being paid to the decoding algorithm and decoder architecture design.
Section 3 introduces the code construction of shift-LDPC codes and Section 4 presents the
decoder design with shift decoder architecture and demonstrates the benefits of proposed
techniques. In Section 5, we consider the application of shift architecture to some well
known LDPC codes such as RS-based LDPC codes and QC-LDPC codes. Section 6 concludes
the chapter.

2. Review of LDPC codes

2.1 LDPC decoding algorithms
LDPC codes are a set of linear block codes corresponding to the M×N parity check matrix H,
which has very low density of 1’s. M is the parity check number and N is the codeword
length. With the aid of the bipartite graph called Tanner graph (Tanner 1981), LDPC codes
can be effectively represented. There are two kinds of nodes in Tanner graph, variable nodes
and check nodes, corresponding to the columns and rows of parity check matrix
respectively. Check node fi is connected to variable node cj only when the element hij of H is
a 1. Fig.1 shows the parity check matrix H of an (8, 4) regular LDPC code and its Tanner
graph. A LDPC code is called (c, t) regular if the number of 1’s in every column is a constant
c and the number of 1’s in every row is also a constant t.

f1 f2 f3 f4

c1 c2 c3 c4 c5 c6 c7 c8
(a)
Ultra-High Speed LDPC Code Design and Implementation 177

1 0 1 0 1 0 1 0
1 0 0 1 0 1 0 1 
H
0 1 1 0 0 1 1 0
 
0 1 0 1 1 0 0 1
(b)
Fig. 1. An LDPC code example. (a) Tanner graph. (b) Parity check matrix.

The typical LDPC decoding algorithm is the Sum-Product (or belief propagation) algorithm.
After variable nodes are initialized with the channel information, the decoding messages are
iteratively computed by all the variable nodes and check nodes and exchanged through the
edges between the neighbouring nodes (Kschischang 2004).
The modified min-sum decoding algorithm studied in Guilloud (2003), Chen (2005) and
Zhao (2005) is similar to the Sum-Product algorithm, with a simplification of check node
process. It has some advantages in implementation over the Sum-Product algorithm, such as
less computation complexity and no requirement of knowledge of SNR (signal-to-noise
ratio) for AWGN channels. In the modified min-sum decoding algorithm, the check node
computes the check-to-variable messages Rcv as follows:
(1)
Rcv     sign (Lvc )  min Lvc
nN( c )\v
nN(c )\v

where α is a scaling factor around 0.75, Lvc is the variable-to-check messages, N(c) denotes
the set of variable nodes that participate in c-th check node. The variable node computes the
variable-to-check messages Lvc as the following:
Lvc   Rmv  I v
mM ( v ) \ c
(2)
where M(v)\c denotes the set of check nodes connected to the variable node v excluding the
variable node c, Iv denotes the intrinsic message of variable node v.

2.2 Implementation options

From the decoding algorithm described above, it can be seen that the calculations between
different variable nodes or check nodes are independent at each iteration. Therefore, LDPC
codes are especially suitable for parallel implementation. In addition, the computation in the
decoder is fairly simple, which means low logic consumption. The main obstacle in the
implementation of fully parallel LDPC decoders is the complicated interconnection between
variable node processing units (VNU) and check node processing units (CNU). This
interconnection complexity is due to the random-like locations of ones in the code’s parity-
check matrix and it is a crucial problem for practical fully parallel decoders when the code
block length is large. The routing congestion or interconnection problem will arise and thus
causes high area consumption and low logic utilization in the decoder. For instance, with 4-
bit precision of probability messages, the 52.5 mm2 die size of the (1024, 512) decoder in
Blanksby (2002) has a logic utilization of 50% in its core and the rest of the core area is
occupied by wires. There are some approaches developed to alleviate this routing
178 VLSI

congestion problem. Darabiha (2008) uses bit-serial architectures and message broadcasting
technique to reduce the number of wires in fully parallel decoders. Sharifi Tehrani (2008)
presents a stochastic decoder design where the messages or probabilities are converted to
streams of stochastic bits and complex probability operations can be performed on
stochastic bits using simple bit-serial structures.
The more general methods to alleviate the routing problem is through reducing the
parallelism (by using partially parallel or serial processing) and using storage elements to
store the intermediate messages passed along the edges of the graph (e.g., see Mansour
2002, Yeo 2003, Cocco 2004). Various approaches are investigated in the literature at both
code design and hardware implementation levels. One approach is to design
“implementation-aware” codes (e.g., see Boutillon 2000, Zhang 2002, Mansour 2003a, Liao
2004, Zhang 2004, Sha 2009). In this approach, instead of randomly choosing the locations of
ones in the parity-check matrix at the code design stage, the parity-check matrix of an LDPC
code is decided with constraints allowing a suitable structure for decoder implementation
and providing acceptable decoding performance. In these cases, the problem becomes how
to reduce the complexity incurred by the message storage elements and how to speedup the
decoding process. For example, one important subclass of LDPC codes with regular
structure is quasi-cyclic (QC) LDPC codes which has received the most attentions and has
been selected by many industry standards. Due to their regular structure, QC-LDPC codes
lend themselves conveniently to efficient hardware implementations. Not only the partly
parallel decoder architectures for QC-LDPC codes require simpler logic control (e.g., see
Chen 2004b, Wang 2007) but also the encoders can be efficiently built with shift registers
(e.g., see Li 2006). Construction of QC-LDPC codes with good error performances is
therefore of both theoretical and practical interest. Various construction methods of QC-
LDPC codes have been proposed and the satisfying error correcting performances are
reported (e.g., see Fossorier 2004, Chen 2004a).
On the other hand, LDPC decoders can be implemented with a programmable architecture
or processor, which lend themselves to Software Defined Radio (SDR). SDR offers flexibility
to support codes with different block lengths and rates, however, Seo (2007) shows that the
throughput of SDR-based LDPC decoders is usually low. In addition to digital decoders,
continuous time analog implementations have also been considered for LDPC codes.
Compared to their digital counterparts, analog decoders offer improvements in speed or
power. However, because of the complex and technology-dependent design process, the
analog approach has been only considered for very short error-correcting codes, for example,
the (32, 8) LDPC code in Hemati (2006).
In the following, a new ensemble of LDPC codes, called shift-LDPC codes, will be
introduced to mitigate the above decoder implementation problems.

3. Shift-LDPC Code Construction

3.1 Construction of Shift-LDPC Code
Shift-LDPC code is constructed in a well-regulated way. A regular (N, M) (c, t) shift LDPC
code example is shown in Fig. 1, with N=t×q and M=c×q, where each submatrix has a
dimension of q×q. As displayed, the parity check matrix consists of c×t submatrices. The
structures of the leftmost c submatrices Hi1 (1 ≤ i ≤ c) are random column permutations of
the identity matrix. The other submatrices are decided by these Hi1 submatrices.
Ultra-High Speed LDPC Code Design and Implementation 179

 H11 H12  H1t 

H H 22  H 2t 
H   21
   
 
 H c1 H c2  H ct 
Fig. 2. The construction of shift-LDPC parity check matrix and an example of the (2,3) shift-
LDPC code

The 1’s in the shift LDPC code parity check matrix are arranged as the example in Fig. 1. At
first, the 1’s in the leftmost submatrices are arranged randomly under the constraint of
keeping one “1” in each row and one “1” in each column. Thus the submatrix Hi1 is a
permutation matrix of identity matrix. Next, for each submatrix on the right hand side, the
1’s are cyclic-shifted up by 1 space. The “1” at the top is moved down to the bottom. Finally
we can get a (c, t) regular shift-LDPC code.
It should be noted that shift-LDPC code is a kind of code specially designed for hardware
implementation. In fact, the shift decoder architecture is designed first and then the code
construction is found. This follows the “decoder-first code design” methodology proposed
in Boutillon (2000).

3.2 Code examples and performance

Given a fixed block length, column weight, and row weight, the ensemble of shift-LDPC
codes can have considerable variations in performance. An effective method to find good
LDPC codes is to find the codes with large girth. Girth is referred to the length of the
shortest cycle in a Tanner graph. A large girth generally improves the bit error performance
of the codes. Hence, LDPC codes with large girth are particularly desired. By exploiting the
structured property of shift-LDPC codes, an efficient girth optimization algorithm is
developed. Normally, by using this girth optimization soft program, the cycle-4 can be
eliminated, and the cycle-6 can be avoided for moderate code rate shift-LDPC codes.
Through extensive simulation, we found that the performance of the optimized shift LDPC
codes can be comparable to the computer generated random codes. Fig. 2 shows the
performance comparison between a girth optimized (1008, 504) (3, 6) shift-LDPC code, the
same size rate-1/2 code downloaded from Mackey’s website and a (960,480) irregular
WiMAX code as well as comparison between a (8192, 7168) (4, 32)-regular shift-LDPC code
and a girth optimized QC-LDPC code. It can be seen that the WiMAX code performs better
due to its irregular property, and in other cases the shift-LDPC codes have comparable
performance. The performance of shift-LDPC codes could be further improved by
introducing limited irregularity, e.g., to zero out some of the submatrices, which will cause
tiny hardware overhead.
Shift-LDPC codes are a kind of implementation oriented codes and are constructed after the
decoder design. Thus in this chapter we will focus on the efficient decoder design instead of
the code performance. On the other hand, the shift decoder architecture can be applied to a
large number of known structured LDPC codes such as quasi cyclic LDPC codes and RS
180 VLSI

based LDPC codes. For example, QC-LDPC codes can be converted to shift architecture
compliant code with proper matrix permutation. It will be explained later in the chapter.

Fig. 3. Performance of (1008, 504) (3, 6) shift-LDPC code and (8192, 7168) (4, 32) shift-LDPC
code compared with the same length WiMAX standard code and computer generated
random codes

4. Shift LPDC Decoder Architecture

In this section, we present the shift decoder architecture specifically designed for shift-LDPC
codes.

4.1 Decoding schedule

Firstly, let us discuss about the decoding schedule. For a regular (N, M) (c, t) shift-LDPC
code, we assume q variable node processing units and M check node processing units are
instantiated in the decoder where q is the size of submatrix (q=N/t=M/c). Fig. 3 illustrates
the decoding flow for a simple shift-LDPC code example with q=4. The schedule can be
applied to other cases with different q, c, and t parameters.

Fig. 4. Decoding schedule for a sample shift-LDPC code

Ultra-High Speed LDPC Code Design and Implementation 181

The leftmost q columns are processed first, then the second left-most q columns, and so on.
So q columns are processed concurrently in one clock cycle. The column process and the row
process are interleaved. The whole check node process is divided into t steps. In each clock
cycle, q VNUs get M check-to-variable messages and compute the M variable-to-check
messages, so that M CNUs get one message each, so each CNU can deal with one step of the
check node process. With this decoding schedule, we can finish one iteration in t clock
cycles. It is normally much faster than the traditional partly parallel decoder architectures.
By combining the decoding schedule and the min-sum algorithm, the decoding process can
be expressed as below:

Min-Sum Algorithm and Decoding Schedule

1: Initialization: Lvc  I v , i  0,1, , N  1;
2: repeat
3: for k  0 to k  d c -1 do
4: { process of kq th ~ (k  1)q th columns }
4: computeRcv  fromrow processresultof lastiteration
5: magnitude of Rcv  minimum or2nd -minimum
6: sign of Rcv  
nN ( c )
sign( Lnc )  sign( Lvc )

7: forqcolumns process :
8: Lvc  
mM ( v )
Rmv  I v  Rcv

9: forallcheck nodes(receiveoneLvc  perrow)

10: update minimum; 2nd -minimum; location of minimum;
11: record signs
12: endfor
13: until max iteration times reached or convergence to a codeword
14: Output: decoded bit
Table 1. Decoding schedule designed for shift decoder architecture

4.2 Overall decoder architecture

The overall decoder block diagram is shown in Fig. 4. q variable node processing units and
M check node processing units are instantiated in the decoder. The critical part of this
implementation is the network connecting VNUs and CNUs. The shuffle network A
transmits variable-to-check messages; the shuffle network B transmits check-to-variable
messages. Shuffle network B is the same as shuffle network A except that the data flow is in
the reverse direction.
In LDPC decoding algorithms, CNUs only communicate with VNUs. The row processes of
different check nodes are independent of each other and so do the variable nodes processes.
In shift-LDPC decoder architecture, a novel block called CNU communication network is
added into the decoder. This new block brings some communications between CNUs. By
introducing these originally unwanted communications, the complexity of connections
182 VLSI

between CNUs and VNUs and thus the complexity of shuffle networks can be significantly
reduced. This will be shown next.

Fig. 5. Overall shift decoder block diagram. Massages are iteratively exchanged between
VNUs and CNUs through the shuffle network

For random codes, the shuffle network routing complexity is normally intolerable, while for
shift-LDPC codes we introduced above, the shuffle networks become really simple. Through
the introduction of the CNU communication network, we can ensure that each CNU processes
the variable-to-check messages from and transmit the check-to-variable messages to a fixed
VNU during the entire decoding process. Therefore, the shuffle network connecting CNUs
and VNUs only consists of M*b wires, where we assume each message is quantized as b bits.
(Normally the quantization bits b is chosen as 6.) In contrast, the number of wires required
by original LDPC decoding algorithm is M*t*b. For high rate codes, for example, t can be as
large as 32.
Fig. 5 explains why each CNU always computes with the messages from a fixed VNU
during the entire decoding process. With the simple shuffle network, CNU i is connected
with VNU x. The row process of each check node is separated into t steps with one variable-
to-check message being processed in each step. At the first clock cycle of an iteration, having
received the message from VNU x, CNU i performs the first step of i-th row process. After
that, CNU i passes the i-th row process intermediate result to CNU i+1 through the CNU
communication network. Meanwhile, CNU i receives the (i-1)-th row process intermediate
result from CNU i-1. In the second clock cycle of the iteration, CNU i still receives the
variable-to-check message from VNU x. This message and the row process intermediate
result in CNU i are arranged both corresponding to row i-1. Therefore, CNU i can perform
the second step of (i-1)-th row process. At the same time, CNU i+1 is performing the second
step of i-th row process. For variable node processing, VNU x receives the check-to-variable
Ultra-High Speed LDPC Code Design and Implementation 183

message in sequential from row i to i-1, i-2, …, etc. In the same way, the CNU communication
network also can ensure that VNU x only need to receive its message from CNU i.

Fig. 6. Data flow of row process intermediate result

To show the decoding procedure more clearly, we illustrate the whole process with the
simple (12, 4) (2, 3) shift LDPC code presented in Fig. 1 as an example. The procedure for
one entire iteration is shown in Fig. 6. It can be seen that the CNU communication network is
separated into two independent networks: CNU communication network (intra iteration) and
CNU communication network (inter iterations) which transfer the row processing results
during iteration and between iteration respectively.
The connections between CNUs and VNUs are fixed and simplified. CNU 1 connects with
VNU 4; CNU 2 connects with VNU 2…These connections are based on the 1’s positions in
the leftmost submatrices.
One time iteration consists of three steps or three cycles (the row weight t equals 3). In each
step/clock cycle, each CNU generates a check-to-variable message, passes the message to its
connected VNU and processes a variable-to-check message which is received from its
connected VNU. At the starting of iteration, in the first clock cycle, CNU 1 process the data
of row 1; CNU 2 process the data of row 2. After this one step row process, the row process
intermediate results are shifted between CNUs through the CNU communication network
(intra iteration). The result according to row 1 is passed to CNU 2; the result according to row
2 is passed to CNU 3, etc.
In the second clock cycle, CNU 2 still transmits to and gets message from VNU 2. This time
the variable-to-check message it gets is corresponding to row 1, thus it can deal with one
step of row 1 process. Likely CNU 3 performs one step of row 2 process; CNU 4 performs
one step of row 3 process; etc. The third step is performed in the same way.
Once the iteration is finished, the row process results will be transferred back to appropriate
CNUs for the check-to-variable messages generation in next iteration. The results are
transferred through the CNU communication network (inter iterations). This network is a one-
to-one communication network as shown in Fig. 6.
As a result, the communication between CNUs and VNUs required by the original LDPC
decoding algorithm is decomposed into three kinds of connections: the simplified
connection between CNUs and VNUs; the CNU communication network (intra iteration) and
the CNU communication network (inter iterations).
184 VLSI

Fig. 7. Decoder scheduling of the simple shift-LDPC code in Fig. 1

In the following, we will introduce the architecture of CNU applying the min-sum decoding
algorithm. The (8192, 7168) (4, 32) shift LDPC code is chosen as the design example.

4.3 Architecture of check node processing unit

The check node processing unit executes the check-to-variable message computation and the
magnitude comparison between the variable-to-check message and the intermediate result
of row process. Fig. 7 shows the architecture of CNU. The old register, new register, and
sign register store the row process results of last iteration, the intermediate results of row
process, and the sign bits of the check-to-variable messages respectively. To manage the
message storage more efficiently, only row process results are stored in a compressed way
that only the minimum magnitude, 2nd-minimum magnitude, the location of the minimum
and the sign bits are saved.
The data in new registers and old registers are always being passed between CNUs through
the CNU communication network (intra iteration and inter iterations). The signs in the sign
register are from the same VNU, so they do not need to be passed. In each clock cycle, sign
register shifts in a sign bit of the variable-to-check message and shifts out a sign bit of a
check-to-variable message.
Ultra-High Speed LDPC Code Design and Implementation 185

Fig. 8. Architecture of check node processing unit

Because only one variable-to-check message is processed in one time, the computation in
CNU has very low complexity. The magnitude comparison part compares the magnitude of
input message with the current row process intermediate result to update the magnitudes,
sign and index. Then the updated row process intermediate result is registered and passed
to its next CNU neighbour through the CNU communication network (intra iteration). The
message computation part selects the proper message magnitude according to the index
value, and computes the sign of the message. At the beginning of each iteration, it performs
the message scale calculation (α = 0.75).

4.4 Architecture of variable node processing unit

The architecture of variable node processing unit is the same with other LDPC decoder
designs. It can be designed with only combinational logic. Fig. 8 shows the VNU
architecture. The inputted check-to-variable messages are firstly converted from sign-
magnitude format to two’s complement format. Then the adder tree computes the variable-
to-check messages according to equation (2). Finally the messages are converted back to
sign-magnitude format and transmitted to CNUs.
186 VLSI

Fig. 9. Architecture of the variable node processing unit. StoT module converts from the
sign-magnitude format to two’s complement format. TtoS is for reverse conversion

4.5 Implementation results

To further increase the decoding speed, more parallel levels can be added to the decoder
architecture. For example, to double the parallel level, 2*q VNUs will be instantiated. The
architecture of CNU should be changed to process two messages per clock cycle. Then each
iteration can be finished in t/2 clock cycles. A double parallel level decoder example is
designed for the (8192, 7168) (4, 32) shift LDPC code to investigate the tradeoff between
hardware cost and decoding speed.
In addition, to increase the clock frequency, some pipeline levels can be added to the data
path. The critical path is from old register, through CNU message computation part, to VNU
computation, to CNU row processing part, and finally ends at the new register. Pipeline will
increase the number of clock cycles needed for each iteration. For example, a three stage
pipeline will add two clock cycles to each iteration.

Through-
CNU VNU total Message Total
frequenc put per
Gate count combination combina- combina- area area
y (MHz) iteration
logic tion logic tion logic (mm2) (mm2)
(Gbps)
One parallel
148 953 395K 160 40.96 6.1 10
level design
One parallel
148 1410 512K 317 76.4 ★ 6.1 11.3
+ pipelining
Double
342 953 838K 151 77.3 6.1 14
parallel level
Double
parallel 342 1410 1072K 317 153 6.1 16.7
+ pipelining

Table 2. Synthesis result for (8192, 7168) (4, 32) shift LDPC decoder under 0.18μm
technology

Table 2 shows the synthesis result of four design examples: one-level parallel design, one-
level parallel with pipelining, two-level parallel design, and two-level parallel with
Ultra-High Speed LDPC Code Design and Implementation 187

pipelining. Due to the large number of CNUs and VNUs (submatrix size q = 256), the
bottom-up synthesis strategy is applied. These results are achieved with 0.18μm technology.
It can be seen that, by applying pipelining, the clock frequency can be almost doubled, and
the overhead is 81 registers per CNU. The one-level parallel design with pipelining can
achieve the best trade-off between speed and hardware complexity.
For the message storage, it is implemented with registers. The transferring messages take
(16*2+32)*1024 = 65536 bits. In comparison, the traditional partially parallel architecture
(e.g., Wang (2007)) needs to store 8192*4*6 = 196224 bits in total. It saves 67% of the message
memory needed. The initial intrinsic information takes 8192*6 = 49152 bits. It is the same for
both cases.
To further examine the efficiency of the shift decoder architecture, backend design of the
sample (8192, 7168) (4, 32) shift LDPC decoder is completed. The technology used for
implementation is a 0.18μm CMOS process with 6 metal layers. Fig. 9 shows the floor-plan
and layout of the decoder chip with a die size of 4.1 mm ×4.1mm. The logic density is 70%.
The placement of CNUs is critical to reducing the routing congestion of CNU
communication networks. The CNU array located in the centre of the chip is specially
arranged: the 1024 CNUs are aligned with 31 CNUs per row, so that the communications
between CNUs can be locally routed.

Fig. 10. The floor-plan and layout of the (8192, 7168) (4, 32) shift LDPC decoder
188 VLSI

Partly parallel Ultra sparse

Proposed
Mansor (2006) Brack (2007)
Code length 2048 9600 8192
Code rate 1/2 3/4 7/8
Edges 6144 26400 32768
Quantization 4b 6b 6b
Algorithm Min-Sum Min-Sum Min-Sum
technology 180nm 65nm 180nm
frequency 125 MHz 500MHz 317 MHz
Iterations 10 10 15
Throughput 640Mb/s 1.45Gb/s 5.1 Gb/s
Area (mm2) 14.3 0.504 11.3/16.8
Area [scaled
3.57 1.008 2.82 /4.2
to 90nm]
Hardware
1778 14384 25500
Efficiency

Table 3. Comparisons with other decoder architectures

Table 3 shows the decoder implementation results compared with some other LDPC
decoder architectures. The throughput achieved here is 5.1 Giga bits per second at a
maximum iteration times of 15. One parameter “Hardware Efficiency” as (Throughput *
Iterations / Area) is defined to evaluate the efficiency of each architecture. The area metrics
are scaled to resemble 90nm CMOS results (65nm by a factor of 2 and 180nm by a factor of
1/4). From this table, we can see that the proposed design can achieve more than 70%
improvement in hardware efficiency compared with an advanced existing design while the
higher clock speed benefited from much more advanced CMOS technology is not
considered for this comparison. Otherwise the improvement would be even more
significant. So shift decoder architecture is very efficient for high-speed LDPC decoder
implementation.

5. Apply the shift architecture to known LDPC codes

5.1 RS-based LDPC code
In Djurdjevic (2004), the authors constructed a kind of LDPC codes based on Reed-Solomon
codes and called them RS-based LDPC codes. These codes have good minimum distances,
girth properties, and perform very well with iterative decoding. One code constructed in
this way was chosen as the forward error correction coding scheme in IEEE 802.3an
10GBase-T Ethernet standard (LAN/MAN CSMA/CD Access Method). RS-based LDPC
codes are constructed in an algebraic way and by exploiting the shift-structured properties
hidden in the parity check matrices, the shift decoder architectures presented above can be
applied to this kind of code. In the following, we will explain the shift properties in these
codes by going through the code generation procedure.
We start from the extended (q, 2, q -1) RS code Cb over GF(q) which has a length of q and two
information symbols. By using the generation matrix
Ultra-High Speed LDPC Code Design and Implementation 189

1 1  1 1 0
G
1 
q2
 2  1 , (3)
we can get the extended RS code with q2 codewords in total. Let v be a nonzero codeword in
Cb with weight q, for example v = (1, αq-2, ···, α2, α, 1). Then, the set Cb(0) = {cv : c ∈ GF(q)} of q
codewords forms the subcode of Cb. Set Cb(0) always contains the all zero codeword and q-1
weight q codewords, no matter which weight q codeword v is chosen.
Partition Cb into q additive cosets, Cb(0), Cb(1),…, Cb(q-1) based on the subcode Cb(0). Each coset
Cb(i) is composed of q codewords, W0(i) , W1(i) , … , Wq-1(i) and each codeword has a length of q,
as below:
W (i )   w w0,1  w0,q 1 
 0   0,0 
W (i )   w w1,1  w1,q1 
Cb(i )   1    1,0 
        
 (i )  
Wq 1   wq 1,0 wq 1,1  wq 1,q 1 
  
 1 i  q  2   i  q 3   i   i 1 
 
    i
1  i  q2   i  2 i  
  2 i   i
1  i  3 i 2 
 
       
 q  2   i  q 3   i  q  4   i  1 i  q2 
 
 0   0 i 0 i 0 i
i
 0 
. (4)
Then the cosets are arranged together to get the q2 × q matrix Hrs

 Cb( 0 ) 
 (1) 
C
H rs   b 
  
 ( q 1) 
Cb 
. (5)
Replace each symbol in Hrs with a location vector:
z ( i )  ( z  , z0 , z1, z2 ,  , zq  2 )
, (6)
where the i-th componet zi = 1 and all the other components equal zero. The exponential
value can be denoted as the position of “1” in the location vector replacement. Finally,
choose a dv × dc subarray from Hrs to get a (dv, dc) regular LDPC code. This can also be stated
as that select dv cosets randomly and then select dc columns in them, finally replace each
symbol with a location vector to get a sparse parity check matrix H.
A coset example is given here to show the matrix properties clearly. It is generated in GF(8):
190 VLSI

4 6 7 3 1 5 0 1
0 4 6 7 3 1 5 2 

5 0 4 6 7 3 1 3
 
1 5 0 4 6 7 3 4
3 1 5 0 4 6 7 5
 
7 3 1 5 0 4 6 6
6 7 3 1 5 0 4 7
 
 2 2 2 2 2 2 2 0  . (7)

The examination of the matrix and equation (4) reveals the four following properties:
1) Each column is composed of the q elements of Galois Field GF(q).
2) qth row is a special row. It is composed with only one value αi except the last one “0”.
3) qth column is a special column. It is always “1, α, α2, … αq-2, 0”.
4) Except the special row and special column, the remaining q-1 × q-1 matrix is a special type
of circulant in which each row is the right cyclic-shift of the row above it while the first row
is the right cyclic-shift of the last row.
Careful readers will find that property 4) is exactly the shift property which is introduced
earlier in the chapter. Thus the shift decoder architecture can be applied to this kind of RS-
based LDPC codes. Fig. 10 shows the location vector replacement in RS-based LDPC code
construction and the mapping of the parity check matrix to decoder hardware. The chosen
code sample is a (32, 24) (1, 4) RS-based LDPC code generated in GF(8).

4 6 7 3
0 4 6 7 CNU 0 0 0 0 0 1 0 0 0 00000010 00000001 00010000
5 0 4 6 CNU 1 1 0 0 0 0 0 0 0 00001000 00000010 00000001
1 5 0 4 CNU 2 0 0 0 0 0 1 0 0 10000000 00001000 00000010
3 1 5 0 CNU 3 0 1 0 0 0 0 0 0 00000100 10000000 00001000
CNU 4 0 0 0 1 0 0 0 0 01000000 00000100 10000000
7 3 1 5
CNU 5 0 0 0 0 0 0 0 1 00010000 01000000 00000100
6 7 3 1
CNU 6 0 0 0 0 0 0 1 0 00000001 00010000 01000000
2 2 2 2 CNU 7 0 0 1 0 0 0 0 0 00100000 00100000 00100000
clock cycle 1 clock cycle 2 clock cycle 3 clock cycle 4
Fig. 11. The location vector replacement in RS-based LDPC code construction and the
mapping of the parity check matrix to decoder hardware
Ultra-High Speed LDPC Code Design and Implementation 191

Fig. 12. Block diagram of the shift decoder design for the matrix defined in Fig.10

Fig. 11 shows the block diagram of decoder designed for matrix defined in Fig. 10. It can be
decomposed into two parts: the processing units and the message storage registers. There is
one row result registers (RR) corresponding to each CNU. One RR is composed of current
iteration row result registers (RRC) and last iteration row result registers (RRL). Both
contain minimum value, 2nd-minimum value, minimum value index and sign bit.
Since the offset value in qth row is fixed at “2”, qth row’s message always comes from VNU
2. CNU 7 is only in charge of the qth row’s processing and RR 7 stores processing result of
the qth row, which corresponds to the special row property 2).
With the shift decoder architecture, two decoder examples are designed. The target code is
the (2048, 1723) (6, 32) LDPC code generated from (64, 32, 2) RS code. A 6 × 32 sub-array at
the upper left corner of Hrs is selected.
Table 4 shows the synthesis results under 90nm CMOS technology of two design examples:
one basic design and a four times parallel level design. Compared to the basic design, the
four times parallel level design can achieve an approximately 3x decoding speed with 2x
hardware consumption. Table 3 also compares the shift architecture decoders with state-of-
the-art bit serial fully parallel LDPC decoder (Darabiha (2008)) and partially parallel LDPC
decoder (Chen (2003)). Comparing the throughput to area ratio metric, it can be clearly seen
that the shift architecture-based designs are very much hardware efficient.
192 VLSI

Gate count × Total Clock

Design Module parameters throughput
number gate speed
CNU 148 × 384
basic VNU 1450 × 64 6 bit quan
420K 500MHz 3.8 Gbps
34 cycles 2048 × 6 bits initial 8 iterations
storage
64 × 384 bits message
CNU 515 × 384
4parallel VNU 1450 × 256 6 bit quan
820K 400MHz 10 Gbps
10 cycles 2048 × 6 bits initial 8 iterations
storage
64 × 384 bits message

(2048, 1723) (6,32)

Darabiha 4 bit quan
RS-based LDPC code 2230K 250MHz 16 Gbps
(2008) 8 iterations
fully parallel architecture

Chen 8088 codeword length, 1/2 rate 6 bit quan

742K 212MHz 188 Mbps
(2003) partially parallel architecture 25 iterations

Table 4. Synthesis results of (2048, 1723) RS-LDPC code decoder under shift architecture and
comparisons with other designs

5.2 Quasi-Cyclic LDPC codes

Quasi-cyclic LDPC (QC-LDPC) codes are well known that they can achieve comparable
performance with equivalent random-like LDPC codes. In addition, more importantly, they
are well suited for hardware implementation because of the regularity in their parity check
matrices. In particular the encoder of a QC-LDPC code can be easily built with shift registers
while usually it is hardware complex to encode an LDPC code.
In general, a regular (c, t) quasi-cyclic LDPC code is defined as its parity check matrix H
(M×N) consisting of a c×t array of submatrices. The dimension of each submatrix is q×q (q =
M/c = N/t). Every submatrix Hij is a circulant matrix of identity matrix I. Fig.12 shows an
example of (24, 8) (2, 3) regular QC-LDPC code.

Fig. 13. The construction of QC-LDPC parity check matrix Hqc: an array of circulant
submatrices
Ultra-High Speed LDPC Code Design and Implementation 193

Fig. 14. The transformation from the sample QC-LDPC parity check matrix into a shift like
LDPC code through column permutation

Fig. 13 illustrates the way to transform the QC-LDPC code parity check matrix into a shift
kind matrix whose decoder can be implemented with the shift decoder architecture. Firstly
the q columns are distributed to t block columns of Hqcs1 in a round-robin fashion. Then the
second q columns are permutated in the same way and so on until all columns are
distributed into new matrix Hqcs1. Careful readers will find that a new property is inducted in
the new matrix that in some submatrices there are multiple “1”s in one row. This newly
inducted property requires a special CNU design which processes multiple messages
instead of one at each step. It will cause some additional logic delay in CNU. In Cui (2008a),
the authors proposed an efficient matrix permutation optimization method to minimize the
maximum row weight. Normally, it can be no bigger than 2. In addition, in Cui (2008a) the
authors optimized the column layered decoding algorithm and applied it to shift decoder
architecture. By this way, the iteration number to converge can be reduced and the decoding
throughput can be further increased.
There are other matrix transformation methods to apply the extended shift decoder
architecture on QC-LDPC codes. For example, the quasi cyclic matrix can be transformed
into a row shift construction by row permutations as shown in Fig. 14. Firstly the q rows are
distributed to 4 block rows of Hqcs2 in a round-robin fashion (i.e., rows A-H of Hqc are
distributed to row 1, 5, 9, 13, 2, 6, 10 and 14 of Hqcs2). Then the second q rows are permutated
in the same way and so on until all rows are distributed into new matrix Hqcs2. The quasi
cyclic matrix is converted to a form as:

 A1 A2 A3  At 
 A A 2 A3  At  (8)

1
H qcrs
      
 q /m 
 A1 A 2 q /m A3 q /m  At  q / m 

where α is a q × q permutation matrix representing a single right cyclic shift. m is an integer

such that q can be divided by m.
194 VLSI

Fig. 15. The transformation from the sample QC-LDPC parity check matrix into a shift like
LDPC code through row permutation

After the matrix conversion, the row shift property can be exploited to apply a row shift
decoder architecture just like the presented column shift architecture. By this means, the row
layered decoding algorithm should be applied. Fortunately the row layer algorithm can
increase the convergence speed significantly (Mansour (2003b)). Interested readers can be
referred to Cui (2008b) for more information.

5.3 Array LDPC code

Array LDPC code is a more structured QC-LDPC code which can achieve comparable
performance to random codes. Efficient encoding, minimum distance properties, and
performance of array-code based LDPC codes for binary as well as multilevel modulation
are addressed in Fan (2000). The parity check matrix based on array codes can be
represented as follows:

I I I  I 
I  2   t 1 
 
(9)
H array  I  2 4   2( t 1) 
 
     
 I  c  1
 c 1)
2(
  ( c 1)( t 1) 
Where I is a q × q identity matrix with a prime number q. α is a permutation matrix with a
single cyclic right shift or left shift. For example, a (2209, 2024) (4, 47) array LDPC code can
be constructed with q = 47. We use this code in the decoder explanation.
For array LDPC codes, they have the similar shift property as shift-LDPC codes. Some
modifications in the CNU communication network can be enough for the decoder
implementation. Fig. 15 shows the modified CNU connections. They are grouped into four
Ultra-High Speed LDPC Code Design and Implementation 195

CNU groups. Each group has 47 members. Group 1 contains from CNU-1 to CNU-47, in
charge of the row operation of row 1~47. Group 2 contains from CNU-48 to CNU-94, in
charge of the row operation of row 48~94... There is no communication between groups. The
difference between array LDPC codes and shift LDPC codes is that the shift values are
changing from 0 to 3 instead of keeping at 1. In addition, for this special kind of LDPC
codes, the CNU communication network (intra iteration) is exactly the same with CNU
communication network (inter iteration), thus these two networks can be incorporated into one
simple network and the routing complexity can be further simplified. As a result, except the
CNUs in Group 1, every 47 CNUs in a group form a chain and each CNU only
communicates with its two neighbors. For example, in CNU group 3, CNU-i only gets data
from CNU-(i-2) and delivers date to CNU-(i+2).
A sample decoder is designed based on an ALTERA FPGA EP2C35. 50MHz clock speed and
120Mbps throughput (20 iterations) are achieved with 23k logic elements and 26k memory
bits. Interested readers are referred to Sha (2006) for more information.

Fig. 16. The CNU communication network designed for (2209, 2024) (4, 47) array LDPC code

6. Concluding Remarks
Since MacKay’s (1997) rediscovery of LDPC codes, the LDPC decoder design has
experienced considerable development and enhancement over the last ten years. Various
optimization techniques on both decoding algorithms and decoder designs have been
developed. This chapter provided a brief overview of the problems in existing LDPC
decoder designs and presented a novel shift-structured decoder design approach to tackle
these problems. The codes suited for this kind of decoder architecture are called shift-LDPC
codes. Several decoder samples are discussed to illustrate the effectiveness of the proposed
architecture. In addition, it was shown in the chapter that some popular classes of LDPC
codes such as RS-based LDPC codes and QC-LDPC codes can be implemented with the shift
decoder architecture through simple matrix permutations or architecture extensions. As a
conclusion, the presented shift decoder architecture will be a competent candidate in future
high-speed communication system designs.
196 VLSI

7. References
A. Darabiha, A. C. Carusone, & F. R. Kschischang. (2008). Power Reduction Techniques for
LDPC Decoders, IEEE Journal of Solid-State Circuits, vol. 43, no. 8, pp. 1835-1845
A. J. Blanksby & C .J .Howland. (2002). A 690-mW 1-Gbps 1024-b, rate-1/2 Low-Density
Parity-Check code decoder, IEEE J. Solid-State Circuits, vol. 37, no. 3, pp. 404-412
D. J. C. MacKay & R. M.Neal. (1996). Near Shannon limit performance of low density parity
check codes, Electron. Lett., vol. 32, pp. 1645-1646
E. Boutillon, J. Castura, & F. R. Kschischang, (2000). Decoder-First Code Design, Proceedings
of the 2nd International Symposium on Turbo Codes and Related Topics, Brest, France,
pp. 459-462
F. Guilloud, E. Boutillon & J.L. Danger. (2003). λ-Min Decoding Algorithm of Regular and
Irregular LDPC Codes, Proc. 3nd International Symposium on Turbo Codes and Related
Topics, pp. 451-454
E. Liao, E. Yeo & B. Nikolic. (2004). Low-density parity-check code constructions for
hardware implementation, Proc. IEEE Int. Conf. on Commun. vol. 5, pp. 2573-2577
E. Yeo, B. Nikolic & V. Anantharam. (2003) Iterative decoder architectures, IEEE Commun.
Mag., vol. 41, pp. 132-140
F. R. Kschischang, B. J. Frey & H. A. Loeliger. (2001). Factor graphs and the sum-product
algorithm, IEEE Trans. Inf. Theory, vol. 47, pp. 498-519
G. Liva, S. Song, L. Lan, Y. Zhang, S. Lin, & W. E. Ryan. (2006). Design of LDPC Codes: A
Survey and New Results. J. Comm. Software and Systems, vol. 2, pp. 191
I. Djurdjevic, J. Xu, K. Abdel-Ghaffar, & S. Lin. (2004). Construction of low-density parity-
check codes based on Reed-Solomon codes with two information symbols, IEEE
Commu. Lett., vol. 8, no. 7, pp. 317-319
J. Chen, A. Dholakia, E. Eleftheriou, M. P. C. Fossorier, & X. Hu. (2005). Reduced-complexity
decoding of LDPC codes, IEEE Trans. Commun., vol. 53, pp. 1288-1299
J. L. Fan. (2000). Array codes as low-density parity-check codes, Proc. 2nd Int. Symp. Turbo
Codes and Related Topics Brest, France, pp. 543
J. Sha, M. Gao, Z. Zhang, L. Li, Z. Wang. (2006). An FPGA Implementation of array LDPC
decoder, IEEE Asia Pacific Conference on Circuits and Systems, pp. 1675-1678
J. Sha, Z. Wang, M. Gao, & Li. (2009). Multi-Gb/s LDPC Code Design and Implementation,
IEEE Trans. on VLSI Systems, vol. 17, no. 2, pp. 262-268
LAN/MAN CSMA/CD Access Method, IEEE 802.3 Standard Online available: http://
standards.ieee.org/getieee802/802.3.html
J. Zhao, F. Zarkeshvari, & A. H. Banihashemi. (2005). On implementation of min-sum
algorithm and its modifications for decoding Low-Density Parity-Check (LDPC)
codes, IEEE Trans. Commun., vol. 53, no. 4, pp. 549-554
L. Chen, J. Xu, I. Djurdjevic & S. Lin. (2004a) Near-Shannon limit quasi-cyclic low-density
parity-check codes, IEEE Trans. Commun, vol. 52, pp. 1038
M. Cocco, J. Dielissen, M. Heijligers, A. Hekstra, & J. Huisken. (2004). A scalable architecture
for LDPC decoding, Proc. Design, Automation and Test in Europe, vol. 3, pp. 88-93
M. M. Mansour & N. R. Shanbhag. (2002). Low power VLSI decoder architecture for LDPC
codes, Proc. IEEE Int. Symp. on Low Power Electron. Design, pp. 284-289
M. M. Mansour & N. R. Shanbhag, (2003a). Architecture-Aware Low-Density Parity-Check
Codes, Proc. IEEE ISCAS, pp. 57-60
Ultra-High Speed LDPC Code Design and Implementation 197

M. M. Mansour & N. R. Shanbhag. (2003b). High throughput LDPC decoders, IEEE Trans.
Very Large Scale Integr. (VLSI) Syst., vol. 11, pp. 976-996
M.M. Mansour & N. R. Shanbhag. (2006). A 640-Mb/s 2048-bit programmable LDPC
decoder chip, IEEE J. Solid-State Circuits, vol. 41, no. 3, pp. 684- 698
M. P. C. Fossorier. (2004). Quasi-cyclic low-density parity-check codes from circulant
permutation matrices, IEEE Trans. Inf. Theory, vol. 50, no. 8, pp. 1788–1793
R. G. Gallager. (1962). Low-density parity-check codes, IRE Transactions on Information
Theory, vol. IT-8, pp. 21-28
R. M. Tanner. (1981). A recursive approach to low complexity codes, IEEE Trans. Inf. Theory,
vol. IT-27, pp. 533-547
S. Hemati, A. Banihashemi, & C. Plett. (2006). A 0.18 µm analog min-sum iterative decoder
for a (32,8) low-density parity-check (LDPC) code, IEEE J. Solid-State Circuits, vol.
41, pp. 2531–2540
S. Seo, T. Mudge, Y. Zhu & C. Chakrabarti. (2007). Design and analysis of LDPC decoders
for software defined radio, Proc. IEEE Workshop on Signal Processing Systems,
Shanghai, China, pp.210–215
S. Sharifi Tehrani, S. Mannor & W. J. Gross. (2008). Fully Parallel Stochastic LDPC Decoders
IEEE Trans. Signal Processing, Vol. 56, no. 11, pp. 5692 - 5703
S. Y. Chung, G. D. Forney, T. J. Richardson & R. Urbanke. (2001). On the design of low-
density parity-check codes within 0.0045 dB of the Shannon limit, IEEE Commun.
Lett., vol. 5, pp. 58-60
T. Brack, M. Alles, T. Lehnigk-Emden, F. Kienle, N. Wehn, & L. Fanucci. (2007). Low
Complexity LDPC Code Decoders for Next Generation Standards, Proc. Design,
Automation and Test in Europe, pp. 1-6
T. Zhang & K. Parhi. (2002). A 54Mbps (3,6)-regular FPGA LDPC decoder, Proc. IEEE Sips’,
pp. 127-132
T. Zhang & K. K. Parhi. (2004). Joint (3,k)-regular LDPC code and decoder/encoder design,
IEEE Trans. Signal Process., vol. 52, no. 4, pp.1065–1079
Y. Chen & D. Hocevar. (2003). A FPGA and ASIC implementation of rate 1/2, 8088-b
irregular low density parity check decoder, Proc. IEEE GLOBECOM, San Francisco,
CA, pp. 113–117
Y. Chen & K. K. Parhi. (2004b). Overlapped message passing for quasi-cyclic low density
parity check codes, IEEE Trans. Circuits Syst. I, vol. 51, pp. 1106-1113
Z.–W. Li, L. Chen, L.-Q. Zeng, S. Lin & W.H. Fong. (2006). Efficient encoding of quasi-cyclic
low-density parity-check codes, IEEE Transactions on Communications, Vol. 54 , no. 1,
pp. 71 – 81
Z. Cui, Z. Wang, X. Zhang & Q. Jia. (2008a). Efficient decoder design for high-throughput
LDPC decoding, APCCAS, pp. 1640-1643
Z. Cui, Z. Wang & Y. Liu. (2008b). High-throughput layered LDPC decoding architecture,
IEEE Trans. VLSI Systems, vol. 17, no. 4, pp. 582-587
Z. Wang & Z. Cui. (2007). Low-complexity high-speed decoder design for quasi-cyclic LDPC
codes, IEEE Trans. on VLSI Systems, vol. 15, no. 1, pp. 104-114
198 VLSI
A Methodology for Parabolic Synthesis 199

10
X

A Methodology for Parabolic Synthesis

Erik Hertz and Peter Nilsson
Department of Electrical and Information Technology, Lund University
Sweden

1. Introduction
In relatively recent research of the history of science interpolation theory, in particular of
mathematical astronomy, revealed rudimentary solutions of interpolation problems date
back to early antiquity (Meijering, 2002). Examples of interpolation techniques originally
conceived by ancient Babylonian as well as early-medieval Chinese, Indian, and Arabic
astronomers and mathematicians can be linked to the classical interpolation techniques
developed in Western countries from the 17th until the 19th century. The available historical
material has not yet given a reason to suspect that the earliest known contributors to
classical interpolation theory were influenced in any way by mentioned ancient and
medieval Eastern works. For the classical interpolation theory it is justified to say that there
is no single person who did so much for this field as Newton. Therefore, Newton deserves
the credit for having put classical interpolation theory on a foundation. In the course of the
18th and 19th century Newton’s theories were further studied by many others, including
Stirling, Gauss, Waring, Euler, Lagrange, Bessel, Laplace, and Everett. Whereas the
developments until the end of 19th century had been impressive, the developments in the
past century have been explosive. Another important development from the late 1800s is the
rise of approximation theory. In 1885, Weierstrass justified the use of approximations by
establishing the so-called approximation theorem, which states that every continuous
function on a closed interval can be approximated uniformly to any prescribed accuracy by
a polynomial. In the 20th century two major extensions of classical interpolation theory is
introduced: firstly the concept of the cardinal function, mainly due to E. T. Whittaker, but
also studied before him by Borel and others, and eventually leading to the sampling
theorem for band limited functions as found in the works of J. M. Whittaker, Kotel'nikov,
Shannon, and several others, and secondly the concept of oscillatory interpolation,
researched by many and eventually resulting in Schoenberg's theory of mathematical
splines.

The parabolic synthesis methodology

Unary functions, such as trigonometric functions, logarithms as well as square root and
division functions are extensively used in computer graphics, digital signal processing,
communication systems, robotics, astrophysics, fluid physics, etc. For these high-speed
applications, software solutions are in many cases not sufficient and a hardware
implementation is therefore needed. Implementing a numerical function f(x), by a single
200 VLSI

look-up table (Tang, 1991) is simple and fast which is strait forward for low-precision
computations of f(x), i.e., when x only has a few bits. However, when performing high-
precision computations a single look-up table implementation is impractical due to the huge
table size and the long execution time.
Approximations only using polynomials have the advantage of being ROM-less, but they
can impose large computational complexities and delays (Muller, 2006). By introducing
table based methods to the polynomials methods the computational complexity can be
reduced and the delays can also be decreased to some extent (Muller, 2006).
The CORDIC (COordinate Rotation DIgital Computer) algorithm (Volder, 1959) (Andrata,
1998) has been used for these applications since it is faster than a software approach.
CORDIC is an iterative method and therefore slow which makes the method insufficient for
this kind of applications.
The proposed methodology of parabolic synthesis (Hertz & Nilsson, 2008) develops
functions that perform an approximation of original functions in hardware. The architecture
of the processing part of the methodology is using parallelism to reduce the execution time.
For the development of approximations of functions a parabolic synthesis methodology has
been applied. Only low complexity operations that are simple to implement in hardware are
used

2. Methodology

  x
Fig. 1. Example of normalized function, in this case sin  .
 2 
A Methodology for Parabolic Synthesis 201

The methodology is developed for implementing approximations of unary functions in

hardware. The approximation part is of course the important part of this work but there are
sometimes two other steps that are necessary, a preprocessing normalization and
postprocessing transformation as described by (P.T.P. Tang, 1991) (Muller, 2006). The
computation is therefore divided into three steps, normalizing, approximation and
transforming.

2.1 Normalizing
The purpose with the normalization is to facilitate the hardware implementation by limiting
the numerical range.
The normalization has to satisfy that the values are in the interval 0 ≤ x < 1 on the x-axis and
0 ≤ y < 1 on the y-axis. The coordinates of the starting point shall be (0,0). Furthermore, the
ending point shall have coordinates smaller than (1,1) and the function must be strictly
concave or strictly convex through the interval. An example of such a function, called an
original function forg(x), is shown in Fig. 1.

2.2 Developing the Hardware Architecture

When developing a hardware architecture that approximates an original function, only low
complexity operations are used. Operations such as shifts, additions and multiplications are
efficient to implement in hardware and therefore searched for. The downscaling of the
semiconductor technologies and the development of efficient multiplier architectures has
made the multiplication operation efficient in both size and execution, time when
implemented in hardware. The multiplier is therefore commonly used in this methodology
when developing the hardware.
As in Fourier analysis (Fourier, 1822) the proposed methodology is based on decomposition
of basic functions. The proposed methodology is not, as in Fourier analysis, a decomposition
method in terms of sinusoidal functions but in second order parabolic functions. Second
order parabolic functions are used since they can be implemented using low complexity
operations. The proposed methodology also differs from the Fourier synthesis process since
the proposed methodology is using multiplications in the recombination process and not
additions as in the Fourier case.
The proposed methodology is founded on terms of second ordered parabolic functions
called sub-functions sn(x), that when recombined, as shown in (1), obtains to the original
function forg(x). When developing the approximate function, the accuracy depends on the
number of sub-functions used.

f org (x)  s1 (x)  s2 (x) ... s (x) (1)

The procedure when developing sub-functions is to divide the original function forg(x), with
the first sub-function s1(x). This division generates the first function f1(x), as shown in (2).

f org (x) (2)

f1 (x) 
s1 (x)
202 VLSI

The first sub-function s1(x), will be chosen to be feasible for hardware, according to the
methodology described in (4). In the same manner the following functions fn(x), are
generated, as shown in (3).

f n (x) (3)
f n1 (x) 
sn1 (x)

The purpose with the normalization is to facilitate the hardware implementation by limiting
the numerical range.

2.3 Methodology for developing sub-functions

The methodology for developing sub-functions is founded on decomposition of the original
function forg(x), in terms of second order parabolic functions for the interval 0 ≤ x < 1.0 and
the sub intervals within the interval. The second order parabolic function is chosen as
decomposition function since the structure is reasonable simple to implement in hardware
i.e. only low complexity operations such as additions and multiplications are used.

First sub-function
The first sub-function s1(x), is developed by dividing the original function forg(x), with x as an
approximation.
As shown in Fig. 2 there are two possible results after dividing the original function with x,
one where f(x)>1 and one where f(x)<1.
A Methodology for Parabolic Synthesis 203

Fig. 2. Two possible results after dividing an original function with x.

The first sub-function s1(x), is according to (4). To approximate these functions 1+(c1.(1-x)) is
used. The first sub-function s1(x), is given by a multiplication of x and 1+(c1.(1-x)) which
results is a second order parabolic function according to (4).

s1 (x)  x (1 (c1 (1 x)))  x  (c1 (x  x 2 )) (4)

In (4) the coefficient c1 is determined as the limit from the division of the original function
with x and subtracted with 1, according to (5).

f org (x) (5)

c1  lim 1
x0 x

Second sub-function
The first function f1(x), is calculated according to (2) and the result of this operation is a
function which appearance is similar to a parabolic function, as shown in Fig. 3.

Fig. 3. Example of the first function f1(x) compared with sub-function s2(x).

The second sub-function s2(x), is chosen according to the methodology as a second order
parabolic function, see (6).

s2 (x)  1 (c2 (x  x 2 )) (6)

204 VLSI

In (6) the coefficient c2, is chosen to satisfy that the quotient between the function f1(x), and
the second sub-function s2(x), is equal to 1 when x is equal to 0.5, see (7).

  1  (7)
c2  4   f1    1
  2 

Thereby the second function f2(x), will get a shape of a lying S, as shown in Fig. 4.

Fig. 4. Example of the second function f2(x), shaped like a lying S.

When developing the third sub-function s3(x), the function is to be split into two parabolic
functions where the first function is restricted by function f2(x) to be in the interval
0 ≤ x < 0.5 and the second function is thus restricted to the interval 0.5 ≤ x < 1.0. By splitting
the function we get strictly convex and concave functions in each interval. The intervals can
be chosen differently but that will lead to a more complex hardware, as shown in section 3.

Sub-functions when n > 2

For functions fn(x) when n > 2, the function is characterized by the form of one or more S
shaped functions. When developing the higher order sub-functions, each S shaped function
is divided into two parabolic functions. For each sub interval, a parabolic sub-function is
developed as an approximation of the function fn(x) in the sub interval. To show which sub
A Methodology for Parabolic Synthesis 205

interval the partial functions is valid for, the subscript index is increased with the index m,
which gives the following appearance of the partial function fn,m(x).

 1 (8)
 f n,0 (x), 0  x  n1
 2
 1 2
 f (x),  x  n1
f n (x)   n,1 2 n1 2
 ...

 2 n1  1
 f n,2n1 1 (x), 2 n1
 x 1

In equation (8) it is shown how the function fn(x), is divided into partial functions fn,m(x),
when n > 2.

As shown in (8), the number of partial functions is doubled for each order of n > 1 i.e. the
number of partial functions is 2n-1. From these partial functions, the corresponding sub-
functions are developed. Analogous to the function fn(x), also the sub-function sn+1(x), will
have partial sub-functions sn+1,m(x). In equation (9) it is shown how the sub-function sn(x), is
divided into partial functions when n > 2.

 1 (9)
sn,0 (x), 0  x  n2
 2
 1 2
s (x),  x  n2
sn (x)   n,1 2 n2 2
 ...

 2 n2  1
sn,2n2 1 (x), 2 n2
 x 1

Note that in (9), the partial functions to the sub-functions; x has been changed to xn. The
change to xn is normalization to the corresponding interval, which simplifies the hardware
implementation of the parabolic function. To simplify the normalization of the interval of xn
it is selected as an exponentiation by 2 of x where the integer part is removed. The
normalization of x is therefore done by multiplying x with 2n-2, which in hardware is n-2 left
shifts and the integer part is dropped, which gives xn as a fractional part (frac( )) of x, as
shown in (10).


xn  frac 2 n2  x  (10)

As in the second sub-function s2(x), the second order parabolic function is used as an
approximation of the interval of the function fn-1(x), as shown in (11).

 
sn,m (xn )  1 cn,m  xn  xn2  (11)

Where the coefficients cn,m is computed according to (12).

206 VLSI

  2 (m  1)  1  (12)
cn,m  4   f n1,m   1
  2 n1
After the approximation part the result is transformed into its desired form.

3. Hardware Implementation
For the hardware implementation two’s complement representation (Parhami, 2000) is used.
The implementation is divided into three hardware parts, preprocessing, processing, and
postprocessing as shown in Fig. 5, which was introduced by (P.T.P. Tang, 1991), (Muller,
2006).

Fig. 5. The hardware architecture of the methodology.

3.1 Preprocessing
In this part the incoming operand v is normalized to prepare the input to the processing
part, according to section 2.1.
If the approximation is implemented as a block in a system the preprocessing part can be
taken into consideration in the previous blocks, which implies that the preprocessing part
can be excluded.

3.2 Processing
In the processing part the approximation of the original function is directly computed in
either iterative or parallel hardware architecture.
The three equations (4), (6) and (11) has the same structure which gives that the
approximation can be implemented as an iterative architecture as shown in Fig. 6.
A Methodology for Parabolic Synthesis 207

Fig. 6. The principle of an iterative hardware architecture.

The benefit of the iterative architecture is the small chip area whereas the disadvantage is
longer computation time.
The advantages with parallel hardware architecture are that it gives a short critical path and
fast computation to the prize of a larger chip area. The principle of the parallel hardware
architecture for four sub-functions is shown in Fig. 7.

Fig. 7. The architecture principle for four sub-functions.

208 VLSI

Fig. 8. Squaring algorithm for the partial products xn2.

To increase the throughput even more, pipeline stages can be implemented in the parallel
hardware architecture.
In the sub-functions (4), (6) and (11) x2 and xn2 are reoccurring operations. Since the square
operation xn2, in the parallel hardware architecture is a partial result of x2 a unique squarer
has been developed. In Fig. 8 the algorithm that performs the squaring and delivers partial
product of xn2 is described.
The squaring algorithm for the partial products xn2 can be simplified as shown in Fig. 9.
A Methodology for Parabolic Synthesis 209

Fig. 9. Simplified squaring algorithm for the partial products xn2.

In Fig. 8 and Fig. 9, the squaring algorithm that performs the partial products xn2, shown.
The first partial product p, is the squaring of the least significant bit in x. The second partial
product q, is the squaring of the two least significant bits in x. The partial product r, is the
result of the squaring of the three least significant bits in x and s is the result of the squaring
of x. The squaring operation is performed with unsigned numbers. When analyzing the
squarer in Fig. 8 and Fig. 9, it was found that the resemblance to a bit-serial squarer (Ienne &
Viredaz, 1994) (Pekmestz et al., 2001) is large. By introducing registers in the design of the
bit-serial squarer the partial results of xn2 is easily extracted. The squaring algorithm can
thus be simplified to one addition only when computing each partial product.
From (4), (6) and (11) it is found that only the coefficients values differentiate when
implementing different unary functions. This implies that different unary functions can be
realized in the same hardware in the processing part, just by using different sets of
coefficients.
Since the methodology is calculating an approximation of the original function the error to
the desired precision can be both positive and negative. Especially, if the value of the
approximation is less than the desired precision, the word length can have to be increased
compared with the word length needed to accomplish the desired precision. If the order of
the last used sub-function is n > 1, an improvement of the precision can be done by
optimizing one or more coefficients c2 in (7) or cn,m in (12). The optimization of the
coefficients will minimize the error in the last used sub-function and thereby it can reduce
the word length needed to accomplish the desired accuracy. Computer simulations perform
such coefficient optimization numerically.

3.3 Postprocessing
The postprocessing part transforms the value to the output result z. If the approximation is
implemented as a block in a system the postprocessing part can be taken into consideration
in the following blocks, which implies that the postprocessing part can be excluded.
210 VLSI

4. Implementation of the sine function

An implementation of the function sin(v), using the proposed methodology (Hertz &
Nilsson, 2009) is described in this section as an example.

4.1 Preprocessing

Fig. 10. The function f(v) before normalization and the original function forg(x).

To satisfy that the values of the incoming operand x is in the interval 0 ≤ x < 1 a π/2 is
multiplied with the operand as shown in (13).

 (13)
v x
2

To normalize the f(v)=sin(v) function v is substituted with x which gives the original
function forg(x) (14).

  (14)
f org (x)  sin   x 
2 

In Fig. 10 the f(v) function is shown together with the original function forg(x).
A Methodology for Parabolic Synthesis 211

4.2 Processing
For the processing part, sub-functions are developed according to the proposed
methodology. For the first sub-function s1(x), the coefficient c1 is defined according to (5).
The determined value of the coefficient is shown in (15).

    (15)

s1 (x)  x     1  x  x 2 
 2  


The first function f1(x), is computed as shown in (16).

f org (x) (16)

f1 (x) 
s1 (x)

To develop the second sub-function s2(x), the coefficient c2 is defined according to (7). The
determined value of the coefficient is shown in (17).


s2 (x)  1 0.400858  x  x 2   (17)

The second function f2(x), is computed as shown in (18).

f1 (x) (18)
f 2 (x) 
s2 (x)

To develop the third sub-functions s3(x), the second function f2(x), is divided into its two
partial functions as shown in (8). The third order of sub-functions is thereby divided into
two sub-functions, where s3,0(x3) is restricted to the interval 0 ≤ x < 0.5 and s3,1(x3) is
restricted to the interval 0.5 ≤ x < 1.0 according to (9). A normalization of x to x3 is done to
simplify in the implementation in hardware, which is described in (10).
For each sub-function, the corresponding coefficients c3,0 and c3,1 is determined. These
coefficients are determined according to (12) where higher order sub-functions can be
developed. The determined values of the coefficients are shown in (19).

  
s3,0 (x3 )  1 0.0122452  x3  x32 , 0  x  0.5
(19)

s3,1 (x )  1 0.0105947  x  x ,

3 3
2
3
0.5  x  1

The third function f3(x), is computed as shown in (20).

f 2 (x) (20)
f3 (x) 
s3 (x)

To develop the fourth sub-functions s4(x), the third function f3(x), is divided into its four
partial functions as shown in (8). The fourth order of sub-functions is thereby divided into
four sub-functions, where s4,0(x4) is restricted to the interval 0 ≤ x < 0.25, s4,1(x4) is restricted
212 VLSI

to the interval 0.25 ≤ x < 0.5, s4,2(x4) is restricted to the interval 0.5 ≤ x < 0.75 and s4,3(x4) is
restricted to the interval 0.75 ≤ x < 1.0 according to (9). A normalization of x to x4 is done to
simplify the hardware implementation, which is described in (10).
For each sub-function, the corresponding coefficients c4,0, c4,1, c4,2 and c4,3 is determined.
These coefficients are determined according to (12) which accomplish that higher order of
sub-functions can be developed. The determined values of the coefficients are shown in (21).

  
s4,0 (x4 )  1 0.00223363 x4  x42 , 0  x  0.25
(21)

s (x )  1 0.00192558 x  x , 0.25  x  0.5

4,1 4 4
2
4

s (x )  1 0.00119216  x  x , 0.5  x  0.75

4,2 4 4
2
4

s (x )  1 0.00126488 x  x , 0.75  x  1

4,3 4 4
2
4

No postprocessing is needed since the result out from the processing part has the right size.

4.3 Optimization
If no more sub-functions are to be developed the precision of the approximation can be
further improved by optimization of coefficients c4,0, c4,1, c4,2 and c4,3. As shown in Fig. 12
sub-function s4,3(x) in the interval 0.75 ≤ x < 1.0 has the largest relative error. When
performing an optimization of sub-function s4,3(x) in the interval 0.75 ≤ x < 1.0 it was found
that the word length in the computations could be reduced from 17 bits to 16 bits.

4.4 Architecture
In Fig. 11, architecture of the approximation of the sine function using the proposed
methodology is shown.
The x2 block in Fig. 11 is the special designed multiplier described in Fig. 8 and Fig. 9 that
delivers the partial results q, q3 and q4 used in the following blocks. In the x-q block, x is
subtracted with the partial result q, from the x2 block. The result r from the x-q block is then
used in the two following blocks as shown in Fig. 11. In the x+(c1·r) block is s1(x) performed,
in 1+(c2·r) is s2(x) performed, in 1+(c3·(x3-q3)) is s3(x) performed and in 1+(c4·(x4-q4)) is s4(x)
performed. Note, that in the blocks for sub-function s3(x) and s4(x), the individual index m is
addressing the MUX that selects the coefficients in the block.

4.5 Optimization of Word Length

As shown in (19) and (21) the absolute value of the coefficients decreases in size with
increasing index number of the coefficient. In similarity to the word length of the coefficients
the word length of the (xn-qn) part, shown in Fig. 11, will decrease in size with increasing
index number. The decreased word length will course that the size of the multiplier used in
a sub-function to decreas accordingly to the highest value bit in the coefficients and of the
(xn-qn) part. In resemblance to above the size of the multipliers computing the multiplication
of the sub-functions can be analyzed. This analysis will also result in that some of the
following multipliers accordingly can be decrease in size.
A Methodology for Parabolic Synthesis 213

Fig. 11. The architecture of the implementation of the sine function.

4.6 Precision
In Fig. 12 the resulting precision when using one to four sub-functions is shown. Decibel
scale is used to visualize the precision since the combination of binary numbers and dB
works very well together. In dB scale 2 is equal to 20log10(2) = 20·(0.3) = 6 dB and since 6 dB
corresponds to 1 bit, this will make it simpler to understand the result. As shown in Fig. 12,
the relative error decreases with the number of used sub-functions. With 4 sub-functions we
can see that we have accuracy better that 14 bits that will result in at least a latency of 14
adders in the CORDIC algorithm is used.
214 VLSI

Fig. 12. Estimation of the relative error between the original function and different numbers
of sub-functions.

As shown in Fig. 12, the relative error decreases with the number of sub-functions used.
However, increases the delay with the number of sub-function as shown in Table 1.

Number of sub-functions Delay

1 2 mult + 2 add
2 3 mult + 2 add
3 to 4 4 mult + 2 add
5 to 8 5 mult + 2 add
Table 1. Delay relative the number of sub-functions.

The delay of one multiplier is about two adders.

5. Comparison
The most common methods used when implementing approximation of a unary functions
in hardware are look-up tables, polynomials, table-based methods with polynomials and
CORDIC. Computation by table look-up is attractive since memory is much denser than
random logic in VLSI realizations. However, since the size of the look-up table grows
exponentially with increasing word lengths, both the table size and execution time becomes
totally intolerable. Computation by polynomials is attractive since it is ROM-less. The
disadvantages are that it can impose large computational complexities and delays.
A Methodology for Parabolic Synthesis 215

Computation by table-based methods combined with polynomials is attractive since it

reduces the computational complexity and decreases the delays. Since the size of the look-
up tables grows with the accuracy the execution time also increases with the needed
accuracy. Computation by using CORDIC is attractive since it is using an angular rotation
algorithm that can be implemented with small look-up tables and hardware, which is
limited to simple shifts and additions. The CORDIC algorithm is an iterative method with
high latency and long delays. This makes the method insufficient for applications where
short execution time is essential.
In all methods including the proposed method, it is a trade-off between complexity and
memory storage. By using parallelism in the computation and parabolic synthesis in the
recombination process, the proposed methodology thereby gets a short critical path, which
assures fast computation.

6. Using the Methodology

It has been shown that the methodology of parabolic synthesis can directly compute the sine
function but the methodology is also able to compute other trigonometric functions,
logarithms as well as square root and division. In the following parts algorithms for
elementary functions will be shown.
When describing the implementation of each function the different parts are shown in a
table. The first row in the table shows the function to be implemented and in which interval
the function is implemented. In the second row it is described how to perform the
normalization of the function. The third row shows the original function to be used when
developing the approximation. The last row describes how to perform transformation of the
approximation into desired interval.

6.1 The Sine Function

When developing the algorithm that performs the approximation of the sine function, the
normalization in the preprocessing part is performed as a substitution according to Table 2.
Since the outcome of the approximation has the desired form no postprocessing is needed.

Algorithm Range
f (v)  sin v  
Function 0v
2
2
Preprocessing x v 0  x 1

 
Processing y  sin   x  0  y 1
2 
Postprocessing zy 0  z 1
Table 2. The algorithms for the sine function.

6.2 The Cosine Function

The algorithm that performs the approximation of the cosine function is founded on the
algorithm that performs the approximation of the sine function. To perform the
216 VLSI

approximation of the cosine function x is substituted with 1-x in the preprocessing part of
the approximation for the sine function.

Algorithm Range
f (v)  cos v  
Function 0v
2
2
Preprocessing x  1 v 0  x 1

 
Processing y  sin   x  0  y 1
2 
Postprocessing zy 0  z 1
Table 3. The algorithm for the cosine function.

6.3 The Arcsine Function

When developing the algorithm that performs the approximation of the arcsine function, the
methodology has problems to perform an approximation for angels larger than /4.
Therefore, the range of the approximation has been limited according to the range of the
function in Table 4. To satisfy the requirements of the methodology in the preprocessing
part a substitution according to Table 4 has to be performed. To get the desired outcome the
approximation is multiplied with a factor according to Table 4.

Algorithm Range
1
Function f (v)  arcsin v  0v
2
Preprocessing x  2 v 0  x 1
 x 
Processing y  arcsin  0  y 1
 2 
 
Postprocessing z y 0z
4 4
Table 4. The algorithm for the arcsine function.

6.4 The Arccosine Function

The algorithm that performs the approximation of the arccosine function is founded on the
algorithm performing the approximation of the arcsine function. The difference between the
two approximations is in the transformation in the postprocessing part, as shown in Table 5.
A Methodology for Parabolic Synthesis 217

Algorithm Range
1
Function f (v)  arccos v  0v
2
Preprocessing x  2 v 0  x 1
 x 
Processing y  arcsin  0  y 1
 2 
   
Postprocessing z   1 y  z
4 4 4 2
Table 5. The algorithm for the arccosine function.

6.5 The Tangent Function

When developing the algorithm that performs the approximation of the tangent function the
angle range is from 0 to /4, since the tangent function is not strictly concave or convex for
higher angles. To perform the normalization the preprocessing part is performed as a
substitution according to Table 6. Since the outcome of the approximation has the desired
form no postprocessing is needed.

Algorithm Range
f (v)  tan v  
Function 0v
4
4v
Preprocessing x 0  x 1


Processing y  tan   x  0  y 1
4 
Postprocessing zy 0  z 1
Table 6. The algorithm for the tangent function.

6.6 The Arctangent Function

When developing the algorithm that performs the approximation of the arctangent function
it can only be performed in the range from 0 to 1 where the function is strictly concave or
convex. To get the desired outcome the approximation is in the postprocessing part
multiplied with a factor according to Table 7.

Algorithm Range
Function f (v)  arctan v  0  v 1
Preprocessing xv 0  x 1
4
Processing y  arctan x  0  y 1

 
Postprocessing z y 0z
4 4
Table 7. The algorithm for the arctangent function.
218 VLSI

6.7 The Logarithmic Function

When developing the algorithm that performs the approximation of the logarithm function
with the base two, it is only perform on the mantissa of the floating-point number, since the
exponent part is scaling the mantissa. For the preprocessing part a substitution according to
Table 8 has to be performed to satisfy the normalization criteria’s for the methodology. Since
the outcome of the approximation has the desired form no postprocessing is needed.

Algorithm Range
Function f (v )  log 2 v  1 v  2
Preprocessing x  v 1 0  x 1
Processing y  log 2 1 x  0  y 1
Postprocessing zy 0  z 1
Table 8. The algorithm for the logarithm function.

6.8 The Exponential Function

When developing the algorithm that performs the approximation of the exponential
function with the base two, it is only performed on the fractional part of the logarithm since
the integer part is scaling the fractional part of the logarithm. As shown in Table 9 only a
one needs to be added in the postprocessing part to get the desired outcome.

Algorithm Range
Function f (v)  2 v 0  v 1
Preprocessing xv 0  x 1
Processing y  2 1x
0  y 1
Postprocessing z  1 y 1 z  2
Table 9. The algorithm for the exponential function.

6.9 The Division Function

When developing the algorithm that performs the approximation of the division it is limited
to the range according to Table 10, since the division is not strictly concave or convex
outside this range. The pre- and post-processing part both needs computation when
performing the approximation of the division.

Algorithm Range
1
Function f (v)  0.5  v  1
1 v
Preprocessing x  2  1  v  0  x 1
6
y 3
Processing  x 0  y 1
1  1 
 2
3 y 1 2
Postprocessing z z
6 2 3
Table 10. The algorithm for the division function.
A Methodology for Parabolic Synthesis 219

6.10 The Square Root Function

When developing the algorithm that performs the approximation of the square root function
the range is limited according to Table 11. The pre- and post-processing part both needs
computation when performing the approximation of the square root function.

Algorithm Range
Function f (v)  1 v 1 v  2
Preprocessing x  v 1 0  x 1
2x  2
Processing y 0  y 1
3 2

Postprocessing
z  2  y  3  2 2z 3

Table 11. The algorithm for the square root function.

7. Conclusions
A novel methodology for implementing approximations of unary functions such as
trigonometric functions, logarithmic functions, as well as square root and division functions
etc. in hardware is introduced. The architecture of the processing part automatically gives a
high degree of parallelism. The methodology to develop the approximation algorithm is
founded on parabolic synthesis. This combined with that the methodology is founded on
operations that are simple to implement in hardware such as addition, shifts, multiplication,
contributes to that the implementation in hardware is simple to perform. By using the
parallelism and parabolic synthesis, one of the most important characteristics with the out
coming hardware is the parallelism that gives a short critical path and fast computation. The
structure of the methodology will also assure an area efficient hardware implementation.
The methodology is also suitable for automatic synthesis.

8. References
B. Parhami (2000), Computer Arithmetic, Oxford University Press Inc., ISBN: 0-19-512583-5,
198 Madison Avenue, New York, New York 10016, USA.
E. Hertz, P. Nilsson (2008), A Methodology for Parabolic Synthesis of Unary Functions for
Hardware Implementation, Proc. of the 2nd International Conference on Signals,
Circuits and Systems, SCS08_9.pdf, pp 1-6, ISBN-13: 978-1-4244-2628-7, Hammamet,
Tunisia, Nov. 2008, Inst. of Elec. and Elec. Eng. Computer Society, 445 Hoes Lane -
P.O.Box 1331, Piscataway, NJ 08855-1331, United States.
E. Hertz, P. Nilsson (2009), Parabolic Synthesis Methodology Implemented on the Sine
Function, Proc. of the 2009 IEEE International Symposium on Circuits and Systems, pp
253-256, ISBN: 978-1-4244-3828-0, Taipei, Taiwan, May. 2009.
E. Meijering (2002), A Chronology of Interpolation From Ancient Astronomy to Modern
Signal and Image Processing, Proceedings of the IEEE, vol. 90, no. 3, March 2002, pp.
319-342, ISSN: 00189219, Institute of Electrical and Electronics Engineers Inc.
J. E. Volder (1959), The CORDIC Trigonometric Computing Technique, IRE Transactions on
Electronic Computers, vol. EC-8, no. 3, 1959, pp. 330–334.
220 VLSI

J. Fourier (1822), Théorie Analytique de la Chaleur, Paris, France.

J.-M. Muller (2006), Elementary Functions: Algorithm Implementation, second ed. Birkhauser,
ISBN 0-8176-4372-9, Birkhauser Boston, c/o Springer Science+Business Media Inc.,
233 Spring Street, New York, NY 10013, USA.
K. Z. Pekmestzi, P. Kalivas, and N. Moshopoulos (2001), Long unsigned number systolic
serial multipliers and squarers, IEEE Circuits and Systems II:, vol. 48, no 3, March
2001, pp. 316 -321, ISSN 1057-7130.
P. Ienne, M. A. Viredaz (1994), Bit-serial multipliers and squarers, IEEE Transactions on
Computer, vol. 43, no12, Dec. 1994, pp. 1445 -1450, ISSN 0018-9340.
P.T.P. Tang (1991), Table-lookup algorithms for elementary functions and their error
analysis, Proc. of the 10th IEEE Symposium on Computer Arithmetic, pp. 232 – 236,
ISBN: 0-8186-9151-4, Grenoble, France, June 1991.
R. Andrata (1998), A survey of CORDIC algorithms for FPGA based computers, Proc. of the
1998 ACM/SIGDA Sixth Inter. Symp. on Field Programmable Gate Array (FPGA’98),
pp. 191-200, ISBN: 0-89791-978-5 , Monterey, CA, Feb. 1998, ACM Inc.
Fully Systolic FFT Architectures for Giga-sample Applications 221

11
X

Fully Systolic FFT Architectures

for Giga-sample Applications
D. Reisis
Department of Physics, Electronics Laboratory,
National Kapodistrian University of Athens,
Greece

1. Introduction
This chapter presents a technique for designing architectures, which execute Fast Fourier
Transform (FFT) algorithms involving a large number of complex points and they target
real-time applications (Ersoy, 1997). Hitherto published techniques and FFT architectures
include mainly ASIC designs (Thompson, 1983; Wold and Despain, 1984; He and Torkelson,
1996; Choi et al., 2003; Uzun et al., 2005; Bouguezel et al., 2004, 2006; Jo & Sunwoo, 2005;
Chang & Nguyen, 2006; Yang et al., 2006; Lin et al., 2005; Takala & Punkka, 2006; Wang &
Li, 2007; Reisis & Vlassopoulos, 2006), which vary with respect to the level of parallelism,
the throughput rate, the latency, the hardware cost and the power consumption. The most
common ASIC architectures are the fully unfolded FFT realizations (Rabiner & Gold)
utilizing large memory arrays between their successive stages and also the latency and
memory efficient cascade FFT topologies (Thomson, 1983; He & Torkelson, 1996, 1998). The
cascade solutions though, as well as the high radix techniques lead to complicated designs
in the case of architectures parameterized with respect to the pipelining of computations,
the FFT size and the data length.
This chapter describes a technique to design efficient, very high speed, deeply-pipelined
FFT architectures maximizing throughput and keeping control and memory organizations
simple compared to the cascade and the fully unfolded FFT architectures. Moreover, the
design is proven more efficient comparing to the previously mentioned architectures in
terms of scalability, maximum operating frequency and consequently, in terms of power
consumption, pipeline depth and data and/or twiddle bit-widths. The technique improves
the latency and the memory requirements -particularly for large input data sets- of systolic
FFT architectures by combining three (3) Radix-4 circuits to result in a 64-point FFT engine.
The efficiency of organizing 64-point FFT engines based on Radix-4 FFT engines is shown by
a 4096-complex point design. This architecture requires only two dual memory banks of
4096 words each and on a Xilinx Virtex II FPGA performs at 200 MHz to sustain a
throughput of 4096 points/20.48 us. The design implemented on a high performance 0.13
um, 1P8M CMOS (standard cell) process from UMC achieved a worst-case (0.9V, 125 C)
post-route frequency of 604.5 MHz, while consuming 4.4 Watts. It is interesting to point out
222 VLSI

that the design exceeded the 1 GHz frequency (rate of 1 GSample/sec) for typical conditions
(1.0V, 25C).
Towards designing architectures FFTs with large input data sets we consider the 4096
complex point FFT architecture as a core. The core constitutes the basis of FFT architectures
computing transforms of 16K, 64K and 256K complex points. These architectures
implemented in the 0.13 CMOS process perform at 352, 256 and 188 Mhz worst-case (0.9V,
125 C) post-route frequencies respectively. The 16K and the 64K architectures have a four
point parallel input/output achieving throughputs of 1.4 and 1 GSample/sec respectively.
Further, this chapter will present a technique, which allows the parallelization of the
memory accesses in hardware implementations of the FFT algorithm. This technique enables
each processor to perform a radix-b butterfly by loading the b-tuple data from b memory
banks in parallel, then by operating on the b data and finally, by storing the resulting b-
tuple in b memory banks in parallel. Hence, the speedup and the throughput increase by b.
Techniques parallelizing the FFT accesses are reported in (Johnson, 1992; Ma, 1999; Reisis &
Vlassopoulos, 2006). We describe the technique in (Reisis & Vlassopoulos, 2006), which is
developed for arbitrary radix and straightforward to implement.
The chapter is organized in three technical sections. Section 2 shows how to organize radix-4
computations to result in a radix-43–equivalent to radix-64–computation and describes the
fully systolic 4096-point architecture. Section 3 will describe the 16K, 64K and 256K point
architectures. Section 4 presents the access parallelization technique in the FFT architectures
and section 5 concludes the chapter.

2. A Fully Systolic 4096 complex point FFT Architecture

This section presents an attractive solution combining the advantages of both the cascaded
(Thompson, 1983; He & Torkelson, 1998; OH & Lim, 2005) and the unfolded (Rabiner &
Gold; Thompson, 1983) structures. The organization uses the unfolded structure with radix-
64 (R64) stages. We note that 64 is a power of either 2 or 4 and it is sufficiently large to result
in small number of stages. Realizing the 64 point FFT at high speed with radices greater than
2 and 4 becomes inefficient, because the use of higher radices imposes complicated
structures and leads to lower the architecture’s performance. R4 computations are equally
straightforward to implement as the R2, but R4 designs reduce the number of complex
multipliers to the half of those required by the R2 designs. Also, R4 process 4 points in a 2-
step computation accumulating-multiplying with the multiplying reduced to sign-inversion and
swap operations. This result minimizes the number of possible bottlenecks in the data flow
and allows the maximal pipelining and/or parallelization of the four computing steps.
Consequently, the choice of R4 leads to a more efficient design with respect to the
multipliers, the pipeline depth and the scalability of the operating frequency and/or of the
word length.

2.1 Radix-4³ Algorithm

To produce the R4³ structure we start with the Discrete Fourier Transform (DFT) of a signal
x[n] of length N,
N 1
X[k ]   x[n ]W
n 0
kn
N (1)
Fully Systolic FFT Architectures for Giga-sample Applications 223

2
j
where, WN  e N are the twiddle factors and denote the N-th primitive root of unity
(Oppenheim, 1975). The architecture presented in this section is based on a four-
dimensional index map and on a R4 decomposition of the DFT series. The derivation of the
Radix − 4³ algorithm lies on three (3) steps in the cascade decomposition. In the frame work
of these 3 steps, the linear mapping transforms into a four-dimentional index map [10,4] as
follows:

N N N
n  n1  n2  n3  n4
64 16 4 (2)
k  64k 1  16k 2  4k 3  k 4

Applying equation 2 to the DFT equation yields

N
1
64 3 3 3
N N N
X(64k 1  16k 2  4k 3  k 4 )      x(n
n 1 0 n 2 0 n 3 0 n 4 0
1 
64
n2 
16 4
nk
n 3  n 4 )WN (3)

The composite twiddle factors can be expressed as follows

N N N
[n 1  n 2  n 3  n 4 ][ 64 k 1  16 k 2  4 k 3  k 4 ]
kn 64 16 4
WN  WN 
n 2 k 2  n 3k 3  n 4k 4 n 3k 4 n ( 4k  k ) n ( 16 k 2  4 k 3  k 4 ] n 1k 1
(4)
kn
WN  (  j) W16 W642 3 4  WN1 WN
64

Applying equation 4 into equation 3 and expanding the summation with index n 4 yield

X(64k 1  16k 2  4k 3  k 4 ) 
N
1
64 3 3  
Bk 4  n 1  N n 2  N n 3  
  N
n 3 0  4
64 16 
(5)
n 1 0 n 2  0 
 (  j)n 2 k 2  n 3k 3 W16
n 3k 4 n 2 ( 4k 3  k 4 )
W64 n 1 ( 16k 2  4k 3  k 4 ) n 1k 1
 WN WN
64
 N N 
Where B kN4  n 1  n2  n 3  denotes the first butterfly unit and can be written as
 64 16 
4

 N N   N N   N N N
B kN4  n 1  n2  n 3   x n 1  n2  n 3   (  j )k 4 x  n 1  n2  n3  
 64 16   64 16   64 16 4  (6)
4
k4  N N N  N N 3N 
 ( 1 ) x n 1  n2  n 3    jk 4 x n 1  n2  n3  
 64 16 2   64 16 4 
224 VLSI

Expanding equation 5 with respect to the next summation with index n3 yields
X(64k 1  16k 2  4k 3  k 4 ) 
N
1
64 3 (7)
N
 H
n 1 0 n 2 0
k 3k 4 
N  n1


64 

n 2  j n 2 k 2  W64
n 2 ( 4 k 3  k 4 ) n 1 ( 16 k 2  4 k 3  k 4 ) n 1k 1
WN WN
16 64

k k  N 
where H N3 4  n 1  n 2  is the secondary butterfly structure and can be expressed as
 64 
16
 N   N  k4 k4  N N
HkN3 k 4  n 1  n 2   BkN4  n 1  n 2    jk 3 W16 B N  n1  n2   
 64   64   64 16 
16 4 4 (8)
 N N k4 k4  N 3N 
  1 k3
W8k 4 BkN4  n 1  n 2    jk 3 W8k 4 W16 B N  n1  n2  
 64 8  64 16 
4 4

Finally, expanding the summation of equation 7 with regard to index n2 provides a set
of 64 DFTs of length N/64.

T W
( N / 64 )1
X(64k 1  16k 2  4k 3  k 4 )  
n 1 0
k 2 k 3k 4
N / 64
n ( 16 k 2  4 k 3  k 4 )
( n 1 )WN1
n 1k 1
N / 64
(9)

k k k
where TN2/ 64
3 4
( n 1 ) represents the third butterfly and is expressed according to equation 10

k k 3k 4
T N2 n 1   H kN3k 4 n 1    jk 2 W16k 3 W64k 4 H kN3k 4  n 1  N  
 64 
64 16 16
k k k k4  N
  1k 2 W8 3 W324 H N3 n1   (10)
 32 
16
3( 4 k 3  k 4 ) k k4  3N 
 j k 2 W64 H N3 n1  
 64 
16

Equations 5 to 10 describe a radix-64 based FFT. Further, equations 6, 8 and 10 describe the
internal structure of the radix-64 butterfly based on three radix-4 butterflies and therefore
called Radix - 43 (R43).

2.2 R43 Features

Figure 1 shows the flow diagram of the R43 computations where  ,  and  present
k3 k k k k
multiplications by (  j) W163 , ( 1) W8 4 and j k 3 W8 4 W164 respectively, while the
k3
, ,
k k k k 3( 4 k  k )
 present multiplications by (  j) k 3 W163 W644 , ( 1) k 2 W8 3 W324 and j k 2 W64 3 4 . Note
that the order of the twiddle factors in the R43 differs from the respective order of the R64.
Fully Systolic FFT Architectures for Giga-sample Applications 225

The FFT architecture with R43 stages and input N complex points uses log4N−1 complex
multipliers. With each R4 using 3 complex adders to produce 1 result/cycle the architecture
has 3log4N complex adders. The memory size is (N/3)log4N.
Comparing to the cascaded R22 (He & Torkelson, 1998) and R24 (OH & Lim, 2005) the
proposed unfolded R43 has equal number of multipliers (Table 1). The R43 has less complex
adders (3log4N) than the cascaded (4log4N) but uses larger memory size. The R43 FFT design
achieves the least number of multipliers and adders in the literature, equal to that in (Bidet
et al., 1995), and although it requires larger memory, it uses simple control and it can
achieve high frequency performance as it is shown in the following subsection.

2.3 Architecture
Figure 2 depicts the architecture of the 4096-point FFT. The implementation of the 4096-
point FFT using R43 butterflies includes: first the use of a 2-dimensional index map based on
R64, second the decomposition of the 4096-point series into two sums using the R64

Complex Multipliers Complex Adders Memory Control

R2MDC * 2log4N − 1 4log4N 3N/2-2 Simple
R2SDF * 2log4N − 1 4log4N N−1 Simple
R4SDF * log4N − 1 8log4N N−1 Medium
R4MDC * 3(log4N − 1) 8log4N 5N/2 − 4 Simple
R4SDC * log4N − 1 3log4N 2N − 2 Complex
R22SDF ** log4N − 1 4log4N N−1 Simple
R43 log4N − 1 3log4N (N/3)log4N Simple
*(He & Torkelson, 1998), **(OH & Lim, 2005)
Table 1. FFT architecture hardware complexity comparison

butterfly and finally the replacement of each R64 with a R43. Consequently, the architecture
consists of two R43 engines, two 4096-word dual bank memory modules, one 4096 point
read-only memory storing the W4096 twiddles and one complex multiplier. The control is
local to each module and a global sync signal synchronizes the modules. The throughput is
1 complex-point/clock-cycle.
The overall design allows the optimization of the architecture, either globally, or locally: The
architecture can be improved with respect to the operating frequency, or the area, or both
without alterations in the control, the scheduling of the 4K buffers, the registers and the
memories within the R43 engines. Hence, we can modify the pipelined computations and
exploit the specifics of each technology to maximize the operating frequency.
The following subsections describe the details of the R4 processor used as the basis of the
R43, the R43 architecture, the addressing scheme used for each R43 engine and discuss the
performance of the entire FFT architecture.

2.3.1 The Radix-4 basis of the R43

Figure 3 depicts the Radix-4 engine architecture used in the R43. Each Radix-4 engine
consists of one complex Accumulator and one ”Swap” module. The task of the ”Swap”
module is to switch the real part of each input value with the corresponding imaginary and
226 VLSI

the opposite in order to perform the correct butterfly operations. Figure 4 depicts the
architecture of the Accumulator module. The accumulator consists of 8 registers and 3
add/sub units. The four ”A” registers are used as input registers storing every 4-tuple of
data, on which the R4 butterfly operation will be applied. The four ”B” resisters are used as
an intermediate stage holding the 4-tuple (while the next 4-tuple of input data is shifted into
the ”A” registers) and loading the data in parallel to the add/sub units. The add-sub units
form an adder-tree architecture, thus avoiding the feedback loops that common
accumulators use.

Fig. 1. Radix − 43 Butterfly SFG

The data flow of the Radix-4 engine starts with the data entering the Radix-4 engine as a
word serial input stream at a rate of one complex-point/cycle. During each period of 4
cycles four consecutive input data are shifted in the 4 input registers (Reg A). In the fifth
cycle this 4-tuple is latched in the 4 ”storage” registers (Reg B), while the next input 4-tuple
is starting its input to the ”A” registers. During the sixth cycle the 4-tuple enters in parallel
the adder-tree to produce the first of the four R4 results. During the following 3 cycles, the
Fully Systolic FFT Architectures for Giga-sample Applications 227

adder-tree uses as input the 4-tuple stored in the ”B” registers and it produces the remaining
3 of the four R4 results, one result/cycle, by following the Radix-4 computational flow.
Therefore, the total latency of each accumulator is 5+2k cycles, where k is the latency of each
add/sub unit, k = 5 in the implementation.
A control unit synchronizes the operations of the ’Swap’ module and the accumulator. A 2-
bit counter generates the necessary signals, controlling the add/sub units in order to
perform the correct additions and/or subtractions according to the Radix-4 schema.

Fig. 2. Block diagram of the proposed 4K FFT Architecture

Fig. 3. Overall Engine of the Radix − 4

2.3.2 The R43 Architecture

Figure 5 shows the R43 engine consisting of 3 R4 butterflies, 2 complex multipliers, 2 Dual
Bank memory elements (1 consisting of 2x64 words and 1 of 2x16 words) and 2 Read Only
Memories storing the W64 and W16 twiddles. The complex twiddle multipliers are 4 integer-
multipliers of input 14 × 14 and are pipelined with depth 5. The control of each pipeline-
stage is local and all stages are synchronized through a global sync signal. Realizing the R64
stage by the R43 engine offers a combination of advantages: the simplicity of the circuits as a
result of using as basic butterfly small radix (4) and the regularity of the architecture, which
is built hierarchically with contiguous R43 engines at the higher level and with each R43
formed by three R4 pipelined computations at the basic level. Furthermore, the R43 design is
VSLI area efficient because it utilizes the complex multipliers 75% of the cycles and 100% of
the other resources at each cycle. The FFT achieves the dynamic range of 84 dB (amplitude)
by using the width of 14 bits for each real and imaginary part of each sample. To maintain
the maximum precision (14 bits) for each part, the R43 uses increased precision for internal
228 VLSI

calculations. The bit width of each part is 16 and 20 bits at the output of the first and second
R4 stages respectively. At the output of the final (third) R4 stage, the real and imaginary
parts are truncated to provide an output of 14 bits each. The twiddle factors are 18 bits real

Fig. 4. The Adder Tree of the Complex Accumulator

Fig. 5. Block diagram of the Radix − 43 Butterfly Architecture

and 18 bits imaginary. By applying the above mentioned data format, the implementation of
the 4096 point FFT with fixed point calculations is almost as accurate as a floating point
implementation given the dynamic range of 84 dB. The fixed point FFT’s outputs are within
a margin of +/- (1) compared to the output of a floating point FFT whose values are
normalized (divided by N = 4096) and truncated to 14 bit integers.
Fully Systolic FFT Architectures for Giga-sample Applications 229

2.3.3 Data Scheduling and Addressing in R43

The first stage of the 4K point FFT algorithm is realized by the first R43 processor, which
computes 4096 FFTs with each transform consisting of 64 points. Each transformation 64-
64
tuple contains the elements whose address is of the form 64i + k, where i = 0, . . . , 63 and k is
the tuple index, ranging from 0 to 4096 − 1.
64
In the first R43 processor the data within each 64 -tuple are processed in terms of
quadruples, so that the first R4 processor computes the DFT of the elements with address
16i1 + k1, i1 = 0, . . . , 3 and k1 = 0, . . . , 64 − 1 is again the tuple index. We can use the
4
notation [a11, . . . , a0] for each point’s input address (position within the 4096 points) and
[d 5 d 4 d 3 d 2 d 1 d 0 ] for the address of each bank of the first intermediate memory. The
resulting 64 elements are written in one bank of the first intermediate memory in
consecutive addresses, i.e. [a 11 a 10 a 9 a 8 a 7 a 6 ] → [d 5 d 4 d 3 d 2 d 1 d 0 ] .
The second R4 reads the data from the first intermediate bank of size 64 by using the
address permutation [a 1a0 a 3a 2 a 5a 4 ] → [d 5d 4d 3d 2d1d0 ] . Every 4 consecutive results of 4-
point DFTs produced by the second R4 are stored in one of the two banks of the second
intermediate memories of size 16. When the third R4 starts reading these data the second R4
will store the next 4 DFT results into the second intermediate memory of size 16.
Note that the second R4 produces results so that the third R4 can use all the data stored in
the intermediate memory to complete its 4-point DFT fully pipelined and with 100%
resource utilization. With [d 3 d 2 d 1 d 0 ] denoting the address within the second intermediate
memory, the third R4 reads the data with address permutation [a 1 a 0 a 3 a 2 ] → [d 3 d 2 d 1 d 0 ] .
During the second stage of the 4K FFT algorithm, the second R43 processor computes 4096
64
FFTs with each transform consisting of 64 points, so that each transformation 64-tuple
contains the elements whose address is of the form 64k + i, where i = 0, . . . , 63 and k is the
tuple index, ranging from 0 to 4096 −1. The first R4 produces 64 elements, which are written
64
in one bank of the first intermediate memory in consecutive addresses, i.e.
[a 5 a 4 a 3 a 2 a 1 a 0 ] → [d 5 d 4 d 3 d 2 d 1 d 0 ] . Apart from this difference the addressing and the data
scheduling within the second R43 is exactly the same as in the first R43 described above.

2.3.4 FPGA and VLSI Implementation

The proposed architecture has been implemented in RTL VHDL using fixed point
arithmetic. The project aims at both high capacity and high speed FPGA devices such as the
XILINX Virtex II 6000, as well as a high performance 0.13um standard cell process. For the
Xilinx implementation we have used the Xilinx tool and the automatic floor planning
option. The Xilinx implementation on the 6000 part results in 20% of logic area utilization
and 25% Blocks RAM utilization. The use of optimized Xilinx components (CoreLib
multipliers) reduces the area to 13% of the total resources of the 6000 part.
The TSMC 0.13 library implementation achieved 604 MHz worst case. It includes 96K
standard cells and 64 RAMs (1361 standard cell rows), (×64 configuration in Table 2)
occupying a silicon area of 2630x5129 um square (1.42*10E7 um sq) at 84.2% utilization and
achieves a worst-case (0.9V, 125C) post-route performance of 604.5 MHz and a 4.4 Watts
230 VLSI

power consumption. The use of typical process parameters (1V, 25C) results in exceeding
the 1GHz post-route frequency mark (data rate 1 GSample/sec), making the proposed
architecture the fastest standard-cell 4096 complex point FFT implementation reported in
the literature. In addition, a second 4094 complex point engine has been implemented in the
same standard cell library, this time using deeper RAMs (×16 configuration in Table 2).
Power consumption in this case was substantially reduced to 722.8 mW from the 4.4 W of
the ×64 configuration. Figure 6 depicts the VLSI layout of the Radix − 43 engine, while figure
7 depicts the final layouts of the 4K FFT, of both the ×16 and the ×64 implemented.
The 4K FFT has been designed for an experiment involving a frequency analyzer for a
bandwidth of 200 MHz. The band has been divided into four sub-bands and each sub-band
has been accommodated by a 4K FFT architecture. The FPGAs perform at 102.5 MHz on a
18-layer board which has a compact-PCI interface performing at 51.25 MHz. The task is to
perform FFT and use a ”Threshold” filter to identify the frequencies of high power within
each sub-band. The expected output set includes at most 10 frequencies per sub-band, per
FFT. After the prototype completion the 4K FFT has been delivered as an IP core with
specifications achieving 5 times the performance of the FPGA prototype. The 16K, 64K and
256K have been realized as IP cores for research purposes.

2.4 Architecture’s Performance and Advantages

The proposed architecture demonstrates improved latency compared to other unfolded
architectures because it requires data buffering only between the two R43 stages instead of
the buffering required at all the 6 stages of the unfolded FFT using R4. The existing cascade
FFT architectures require 6 R22 (He & Torkelson, 1996, 1998) and although they use 1/4 of
the memory of the proposed R43 architecture for 4096 points, they are not scalable and they
must be designed for specific performance due to their organization involving “closed
computing loops”.
The achieved performance establishes the proposed architecture as a real-time high
performance FFT realization, the importance of which can be further pointed out by

Fig. 6. VLSI implementation layout of the Radix − 43 engine

Fully Systolic FFT Architectures for Giga-sample Applications 231

comparing its characteristics to other relevant results. Tables 3 and 4 present a comparison
of the proposed FFT architecture with related results. We distinguish related published
results into two categories.

FFT 4096 Ramsx16 FFT 4096 Ramsx64

Clock (period) 1.6 ns 1.2 ns
Fmax in MHz 386.8 604.5
Std Cells (RAMs) 100347(16) 95897(64)
Std Cells (rows) 805 1361
Chip Size (um sq) 5.34E+06 1.42E+07
Util (%) 80.30% 84.20%
X(um) 1681 2630
Y(um) 3165 5129
Power (mW) 722.8 4414.1
Table 2. 4K FFT VLSI Implementation Routing Results

In the first category we compare the hitherto published architectures executing FFT
algorithms up to 128 points to the features of the R−43 playing the role of a complete 64
complex point FFT architecture. This comparison is shown in Table 3. The second category
includes the architectures solving FFTs of size 1024 to 4096 points presented in Table 4. The
comparison includes FFT size, word length, algorithm, FFT architecture, technology process,
voltage, area, power, maximum operating frequency and sustained throughput in both
MSamples/s and Gbits/s. Since the FFT designs vary with respect to FFT size, algorithm
and architecture we have also included the Normalized Area (Bidet et al., 1995), in order to
evaluate the silicon cost. Moreover, we compare the efficiency (performance/cost) of the
proposed design to the related results by using the fraction Sustained
Throughput/Normalized Area.
The proposed FFT architecture achieves the highest sustained throughput compared to all
the other designs. Furthermore, the efficiency expressed as the fraction Sustained
Throughput/Normalized Area of the proposed design is the highest considering both small
input size (Table 3) and large input size FFT architectures (Table 4). Note that, in both
categories the architectures R − 43 and the proposed 4K occupy more area than their
competitors respectively. This is a penalty though in achieving the highest throughput
possible.
Also note that (Lin et al., 2005) performs transformations of only 128 points, and there is no
provision taken so that it will constitute a core for scalable architectures with respect to FFT
size. The 64-point Fourier transform chip, presented in (Maharatna et al, 2004) operates at 20
MHz with latency 3.85 us, comparing to the R43 processor performing a 64 complex point
FFT while operating at a 200 MHz clock frequency with latency 0.32 us.
The architecture described in (Lenart & Owall, 2003) is a 2K complex point FFT processor
which achieves maximum operating frequency of 76MHz and sustains a throughput of 2048
points/26us. The design presented in (Cortes et al., 2006) implements a 2K/4K/8K
multimode FFT and achieves 9 MHz clock frequency, at a computation time of up to 450us.
232 VLSI

Fig. 7. VLSI implementation layout of the 4K FFT processor: a) 16 × 4096 Configuration, b)

64 × 4096 Configuration

(Lin et al., (Maharatna Proposed Design

Characteristics
2005) et al., 2004) (R−43)
FFT size 128p 64p 64p
Word Length, bit 10 16 14
Algorithm R-2, R-8 R-2 R-43
FFT Architecture MRMDF Unrolled Unrolled
Process (um) 0.18 0.25 0.13
Voltage (V) 3.3 1.8 0.9
Area (mm2) 3.52 6.8 5.09
Normalized Area (mm2) 3.52 3.5 9.7
Power (mW) 77.6 41 78.2
Fmax (MHz) 110 20 604.5
Sustained Throughput in 480 16.6 1052
MSamples/s
Sustained Throughput in Gb/s 9.6 0.53 29.5
Throughput(Gb/s) 2.72 0.15 3.04
Norm.Area
Table 3. Comparison Table with FFT designs from 64-128 points
Fully Systolic FFT Architectures for Giga-sample Applications 233

(Lenart &
(Cortes et (Swartzlander, Proposed
Characteristics Owall,
al., 2006) 2007) Design
2003)
FFT size 2K 2K/4K/8K 4K 4K
Word Length, bit 10 16 16 14
Algorithm R − 22 R − 22 R−2 R − 43
FFT Architecture SDF SDF Split Systolic Unrolled
Process (um) 0.35 0.35 0.25 0.13
Voltage (V) N/A 3.3 N/A 0.9
Area (mm2) 6 18.7 N/A 13.48
Normalized Area (mm2) 1.58 4.9 N/A 25.84
Power (mW) N/A 114.65 260 4414.1
Fmax (MHz) 76 9.1 100 604.5
Sustained Throughput in 75.8 18.2 200 1052
MSamples/s
Sustained Throughput in 1.5 0.582 6.4 29.5
Gb/s
Throughput(Gb/s) 0.94 0.11 N/A 1.14
Norm.Area
Table 4. Comparison table with FFT designs from 1K-4K

Finally, a single ASIC chip, systolic FFT processor, developed by the Mayo Foundation
computes 4096-point FFTs sustaining a throughput of 200 Ms/s (Swartzlander, 2007).
Considering FPGA implementations, the corresponding XILINX designs (www.xilinx.com)
achieve equal maximum operating frequency of 200MHz, but occupy considerably larger
chip area than the R43 approach. Also note that, ALTERA designs (www.altera.com) utilize
FFT cores with FFT length varying from 64 points up to 4K points. They demonstrate a
maximum operating frequency of 300 MHz. Among the ALTERA’s FFT designs we compare
the 64 point FFT at 300 MHz to the R43 performance, which realized on the same ALTERA
FPGA (ALTERA STRATIX II EP2S30F484C3) achieves operating frequency of 350 MHz.

3. 16K, 64K and 256K complex point FFT Architectures

The high utility prospects of the efficient 4K point FFT design as presented above becomes
tangible through the utilization of this architecture to develop larger size FFT architectures
to compute 16K, 64K and 256K point transforms. This section describes the exploitation of
the 4K point FFT as a “radix-4096” processor resulting in architectures meeting different
requirements with respect to parallelism, silicon area and throughput. These architectures
can be tailored to a broad area of applications.

3.1 Deriving the 16K point Architecture

The DFT equation for a 16K point FFT takes the form
234 VLSI

16383 44096 1 (11)

X[ k ]  
n 0
nk
x[ n ]W16384  
n 0
x[ n ]W4nk
4096

N
Setting n  n 1  n 2 and k  k 1 4096  k 2 :
4096
 N  N
k  n   n1  n 2 4096k 1  k 2  ⇒ k  n  n 1k 1  n 1k 2  Nn 2 k 1  k2 ⇒
 4096  4096
n 1k 1
W4kn
 4096  W4 W4n4096
1k 2 n 2k 2
W4096

Therefore, the transform becomes

3  4095 
n 1k 1   n 1k 2
X[k ]  ∑W 4

∑x[n]W n 2k 2
4096 W44096

(12)
n 1 0  n 2 0 

According to equation 12, the 4K points FFT can be extended to 16K points. This is
accomplished by first performing four 4K point FFT transforms. Next, the data is multiplied
by the twiddle factors that correspond to a 16K points FFT and finally, a radix-4 stage
completes the 16K point FFT computation.
The architecture is presented in figure 8. There are four 4K FFT blocks operating in parallel.
The architecture has a four complex point input per cycle. The 16K FFT architecture was
implemented in a high performance, 0.13um, 1Poly-8Copper layer standard cell technology
from TSMC. A flat back-end flow was used in which the design was first synthesized to
gates using Synopsys Design Compiler and optimized for a frequency of 300 MHz. The
uniquified netlists was then read into Cadence SoC Encounter where floorplanning, power
planning, clock-tree-synthesis, placement, routing and IPO routing took place. Finally, the
design was brought back into the top-level for final top-level placement, routing and timing
analysis. Table 5 shows the Implementation Routing Results and figure 8 depicts the VLSI
Cells for the 16K FFT. The throughput is 1.4 Gs/sec (39.2 Gbits/sec).

Fig. 8. Block diagram of the 16K FFT - Parallel/Parallel Architecture

Fully Systolic FFT Architectures for Giga-sample Applications 235

Parameter FFT 16K

Core Size 1.6561e + 07um2
Std Cell Rows 1102
Number of Cells 885456
Number of RAMs 64
Statistical Power 3.45 W
Fmax 352 MHz
Table 5. VLSI Implementation Routing Results of the 16K FFT architecture

The 16K architecture implemented on the Xilinx Virtex 5 (-2) achieves operating frequency
of 250MHz, occupies 12264 slices and sustains a throughput of 1Gs/s (28 Gbits/sec).

3.2 64K and 256K complex points FFT Architectures

The DFT equation for a 64K point FFT takes the form

65535 416384 - 1
X[k ]  ∑ nk
x[ n ]W65536  ∑x[n]W nk
4 16384 (13)
n 0 n 0
N
Setting n  n 1  n 2 and k  k 1 16384  k 2 :
16384
N n 1k 1
k  n  16384 n 1k 1  n 1k 2  Nn 2 k 1  k 2  W4kn
16384  W4 W4n16384
1k 2 n2k 2
W16384
16384
Therefore, the transform becomes
3  16383 
n 1k 1  n 2 k 2  n 1k 2
X[k ]  ∑ W4  ∑
 n 0
x[ n ]W16384 W416384

(14)
n 1 0  2 

According to equation 14, the 16K points FFT can be extended to 64K points. This is
accomplished by first performing a 16K point FFT transform. Next, the data is multiplied by
the twiddle factors that correspond to a 64K points FFT and finally, a R4 stage completes the
64K point FFT computation, as shown in figure 10. The 64K FFT has been VLSI (and FPGA)
implemented by using the 16K parallel/parallel computation with a R4 stage with four
parallel inputs and outputs. The architecture has a post routing frequency of 256 Mhz with a
throughput of 1 Gs/sec (28 Gbits/sec). Figure 11 depicts the VLSI Cell for the 64K FFT
design. The 64K has been implemented on the Xilinx Virtex 5 (-2) achieving an operating
frequency of 125MHz and using 13461 slices. The architecture has a four data parallel input
and output and sustains a throughput of 500Ms/s (14 Gbits/sec).
236 VLSI

3.3 256K complex point FFT Architecture

Figure 12 depicts the 256K FFT architecture, which is a straightforward application of the
R43 algorithm. Three consecutive R43 engines are used to accomplish the task.

Fig. 9. VLSI implementation layout of the 16K FFT Parallel/Parallel architecture

Fig. 10. Block diagram of the 64K complex point FFT Architecture
Fully Systolic FFT Architectures for Giga-sample Applications 237

Parameter FFT 64K FFT 256K

Core Size 3.4894e + 08um2 2.7647e + 08um2
Std Cell Rows 1600 4500
Number of Cells 896148 735945
Number of RAMs 192 384
Statistical Power 9.8 W 35.75 W
Fmax 256.5 MHz 188 MHz
Table 6. VLSI Implementation Routing Results of the 64K and 256K FFT designs

Fig. 11. VLSI implementation layout of the 64K FFT

In the case of the very large 256K FFT, tool capacity mandated the use of a hierarchical flow.
Following the same front-end synthesis process (again optimized for 300 MHz), the
optimized netlist was read into Cadence SoC encounter where partitioning was first
performed. This process created six instances of the 256K memory block (for a total of 1.5
MB of on-chip SRAM) which were individually placed and routed. The same process was
performed for the R43 engines and the twiddle ROM block. Note that to complete the entire
FFT computation it requires a 3-frame latency (786792 cycles) and the computation latency
within the three R43 (360 cycles), which is 4.1 msec. The 256K FFT architecture has a post
routing frequency of 188 Mhz. Finally, the design was brought back into the top-level for
238 VLSI

final top-level placement, routing and timing analysis. Figure 13 depicts the VLSI Cell for
the 256K FFT design and Table 6 shows the VLSI Implementation Routing Results of the 64K
and 256K FFT designs.

Fig. 12. Block diagram of the 256K FFT Architecture

4. Parallel Accessing in FFT Architectures

This section presents a technique, which allows the parallelization of the memory accesses in
hardware implementations of the FFT algorithm by enabling a processor during a FFT stage
to access in parallel b memory banks without conflicts. Consequently the processor
performs a radix-b butterfly by loading the b-tuple data from b memory banks in parallel,
then by operating on the b data and finally by storing the resulting b-tuple in b memory
banks in parallel. Hence, the speedup and the throughput increase by b.

4.1 Problem Definition and Related Work

Let N = b p denote the number of points of the Fast Fourier Transform, where b is an
arbitrary base such that b ≥2 and p ≥1 . Using b as a base, the input address 0,…,N-1 (or
index) of each element is represented as [ap-1 ....a0 ] , where 0 ≤ai ≤b - 1 . We will denote each
element only by its address [ap-1 ,...,a0 ] , which represents the element in the input stream
with index a p-1 b p-1 + ... + a 1 b + a 0 . This notation holds for each of the p stages of the FFT.
We will use Pb[ n →m ] to denote a processor performing all radix-b computations, which can
also read (load) n data in parallel and write (store) m data in parallel. We assume a generic
architecture using ψ, 1 ≤ ≤log b N , Pb[ b→b ] processors and 2  b  (  1) memory banks,
so that each bank can store N/b elements. Each processor realizes N/b consecutive stages of
the FFT computation.
Fully Systolic FFT Architectures for Giga-sample Applications 239

Fig. 13. VLSI implementation layout of the 256K FFT

The memory banks are arranged according to the scheme described in figure 14. Since, we
are using a radix-b processor, the DFT decomposition consists of log b N stages. If we
assume  = log b N then we consider a Pb[ b→b ] processor realizing each stage of the
algorithm. The stages are indexed from 0 to log b N - 1 . Between two (2) consecutive stages i
and i+1 there is a memory of size 2  N divided into two (2) sets of size N each. We will
denote the memory between stage i and stage i+1 as memory with index i. The memories
among the FFT stages are indexed as 0 to log b N - 2 . There are memories indexed as -1 and
log b N - 1 serving as input and output memories respectively. The memory of each set i is
divided into b memory banks with indices 0 to b-1.
At each clock cycle, the Pb[ b→b ] processor of stage i loads from its ‘’read’’ memory (memory
i-1), b elements that form a transform b-tuple. These are the elements whose address is of
the form [a p-1 ,..., a i+1 la i-1 ,..., a 0 ] , where the indices [ a 0 ,..., a i-1 , a i+1 ,..., a p-1 ] are the same for
all elements in the b-tuple and l ranges from 0 to b-1. These elements will be loaded in
parallel if they are stored in distinct memory banks. During the next ((i + 1)th) step of the
algorithm, these b elements have the same ai+1 address digit and hence they belong to
240 VLSI

different transformation b-tuples. These elements must be stored in the processor’s write
memory so that they can be read in parallel from the processor at stage i + 1.

Fig.14. Generic arrangement of a Pb[ b→b ] processor and its write memories

The straightforward approach is to distribute the N elements to the b memory banks using
their ith address digit. Then the processor at stage i can load them in parallel. However, if we
try to store the same b elements of each b-tuple to the memory bank at stage (i+1) according
to their (i+1)th digit of their address, we will notice that all the elements in each b-tuple must
be stored in the same memory bank. Since we cannot store more than one element in a
memory bank at each clock cycle this situation constitutes a “conflict”.
To illustrate the situation where a “conflict” occurs, consider a 16 point transform that is
based on P4[ 4→4 ] processors. The address of each element in the input stream is [a 1a 0 ] ,
where a i = 0 ,1,2 ,3. Figure 15 shows the first stage of the example. The first 4-tuple to be
transformed consists of the elements [(0, 0) (0, 1) (0, 2) (0, 3)]. Note that, these elements can
be read in parallel from the set of memory banks of stage i − 1, but must be stored in the
same memory bank of memory i.
The first technique for parallelizing the memory accesses in FFT architectures (Johnson,
1992) describes a hardware architecture designed for in-place computation of the FFT
algorithm. This technique uses r (r being the radix) banks and permutes the output data of
each processor, to be written in the same memory locations within each bank as the input
data have been read. (Ma, 1999) describes a technique for radix-2 based FFT by using queues
at the input of each memory to rearrange the data before they are stored. Johnson’s
technique can be extended to the more general case of a pipelined FFT implementation at
the cost of complex addressing hardware implementation. (Thomas & Yelick, 1999) present
a technique that is used in vector processors. This technique permutes the input data prior
to the radix calculations in order to maximize the efficiency of the algorithm. These
permutations are performed within the vector registers (In-register transpose). This
functionality has been implemented by extending the instruction set of the vector processor
with two instructions that shift the data within a register, so that each permutation is
performed by using a register to register copy and a shift instruction.
Fully Systolic FFT Architectures for Giga-sample Applications 241

All the FFT techniques (Johnson, 1992; Ma, 1999), with b parallel memory accesses, involve
at each stage b banks for reading and b banks for writing the N FFT data with bank size
N/b. The in-place case uses the same b banks (each bank of size N/b) for both read and
write. To compare the hardware complexity of the previously published results, we first
consider designs that are based on queues, such as the one described in (Ma, 1999).

Fig. 15. First stage of a 16-point transform using P4[ 4→4 ] processors

Such designs require O( b 2 )2 w -bits wide registers, where w is the word length, arranged in
b queues with each queue having b − 1 registers. b is the processor radix, or in the general
case, the number of parallel read/write accesses. Each set of b−1 registers is arranged so that
b−1 elements are loaded in parallel while one element is written into the memory bank. The
other b−1 elements are shifted -one element per memory access cycle- into each memory
bank. The technique in (Johnson, 1992) uses fixed hardware to implement first a
permutation on the input data and then this specific hardware can apply a permutation at
each FFT stage. More specifically, (Johnson, 1992)’s approach gives only a subset of the
possible solutions. This subset of solutions considers only in-place implementations (ψ = 1
and memory size N). The circuit in (Johnson, 1992) is optimized for implementing these
solutions with respect to the VLSI area. The address generation is based on a circuit
involving a tree of modulo-b adders with O( b 2 ) exclusive-OR gates.
In this section we describe a technique (Reisis & Vlassopoulos, 2006) that can be used both
for pipelined architectures and in place implementations (memory size equals to N). We
follow a different approach than (Johnson,1992) leading to a different proof and providing a
242 VLSI

general solution, which includes that in (Johnson,1992). Moreover, the presented technique
results in a simple and improved hardware implementation compared to (Johnson,1992).
The presented solution shows improved performance with respect to the latency and the
hardware cost compared to other solutions (Ma, 1999; Suter & Stevens,1998; He &
Torkelson; Thomas & Yelick, 1999) while it provides the same throughput as these. Our
technique is based on memory address permutations and can be realized using look up
tables with each stage table occupying O( b 2 ) bits. The presented approach results in a set of
permutations, which can be applied in the design of pipeline FFT architectures with parallel
memory accesses per stage. A subset of these permutations accommodates the in-place
implementations.

4.2 Parallelizing the Memory Accesses

The following Lemmas 1 and 2 prove the existence of the required permutations and
Theorem 1 proves that these permutations result in a correct FFT algorithm.
Lemma 1: Assume that the N = b p elements with indices [a p-1 ...a 0 ] , 0 ≤ a i ≤ b - 1 , are
distributed in b sets Sl, l = 0, . . . , b − 1, such that: S l = {[a p-1 ...a i+1 la i-1 ...a 0 ]}
Let fj : {0, . . . , b−1}→{0, . . . , b−1}, where 0 ≤ j ≤ b−1, be a set of functions such that
- ∃f j-1 ⇒ f is 1-1 and onto, and
- ∀m ,∀n , where m, n ∈{0 ,..., b - 1},
If m ≠n , then ∀x , fm ( x ) ≠fn ( x)
The sets S'l' , where l' = fai +1 (l ) are defined as follows S'l' = {[a p-1 ...a i+2 fai +1 (l ) l a i-1 ...a 0 ]}
Where l is fixed and a i+1 = 0 ,..., b - 1
Under the above conditions, the following hold:
1) Each S′l′ contains at most N/b items
2) Every two elements of two sets S l 1 , S l 2 whose indices differ only in the ith digit, i.e. q and
r will be distributed in different S′ sets, namely the S′q′ and S′r′ .

Proof of Lemma 1: Each set S’l′ contains the data with indices [a p-1 ...a i+2 f( ai +1 ) (l ) l a i-1 ...a 0 ] ,
where a i+1 = 0 ,..., b - 1 and l is fixed. To prove (1) assume that there exists a set that contains
more than N/b elements. Then, there is at least one element with index
[ap-1...ai+2f(ai+1 ) (l' ) l'' ai-1...a0 ] where l′ is different than l′′. Since such an element does not
belong to the set S′l′ , we conclude that each set cannot contain more than N/b elements.
Similarly, to prove (2), assume that two elements of the sets Sq and Sr, respectively, whose
indices differ in the ith digit are distributed in the same set, S′s′ . Then, for these elements,
[a p-1 ...a i  2 fai  1 (q ) l a i -1 ...a 0 ]  [a p-1 ...a i  2 fai 1 (r ) l a i -1 ...a 0 ]  fai +1 (q ) = fai +1 (r ) . Since the

fj are invertible, fa111 (fa i 1 (q))  fa111 (fa i 1 (r)) ⇒q  r , which contradicts our initial assumption
that l 1 ≠l 2 . The following Lemma shows that these functions exist.
Lemma 2: There exists a set of functions f j : {0 ,..., b - 1} → {0 ,..., b - 1}, j = 0 ,..., b - 1 and b ≥2
such that the conditions of Lemma 1 are satisfied.
Fully Systolic FFT Architectures for Giga-sample Applications 243

Proof of Lemma 2: Let M = {0, . . . , b − 1}. Further, let C(M) denote the set of cyclic
permutations of length b on the set M. These permutations are of the form

0 1 ... (b - 1)
( )
( p ) mod b ( 1 + p ) mod b ... ( b - 1 + p ) mod b
for p = 0, . . . , b − 1. Now, these permutations are invertible and there are exactly b such
permutations. To prove that
fi ( x ) ≠f j ( x ) (15)
for i ≠j and i, j = 0, . . . , b − 1 it suffices to show that ( x + i ) mod b ≠( x + j) mod b , which is
true since i and j could only be in the same equivalence class, if i = k ・ j, where k is an
integer, but this can be true only for k = 1, since i, j ≤ b − 1. Therefore equation 15 holds.

Theorem 1: Let N = b p . Assume that we have a Pb[ b→b ] processor. Then, a radix-b based
FFT having log b N stages can use b memory banks at each stage, such that all the read and
write operations are performed in parallel.

Proof: Let [a p-1 ...a 0 ] , 0 ≤ ai ≤ b − 1, be the address of each element according to its initial
position within the input data set of the algorithm. The ith stage of the algorithm has
arranged the N elements in b memory banks according to their ith digit of their address:

[a p-1 ...a i+1 a i-1 ...a 0 ][a i ] (16)

where ai is the index of the memory bank.
To write the outputs of a radix-b processor into b distinct memory banks in parallel, we use
the results of Lemma 1. First, we assign each memory bank to a set, Sl, so that each
transform b-tuple consists of elements that differ in the ith element of their address. Next, we
select a set of b functions f0, f1, . . . , f(b−1) using Lemma 2.
Since the requirements of Lemma 1 hold, we can distribute the elements according to the
equation:
[a p-1 ...a i+2 a i ...a 0 ][ fai +1 (a i )] (17)
These elements will be distributed in distinct memory banks and can be stored in parallel.
Executing the following stage, we will use the inverse permutation to read each b-tuple in
parallel: During the (i+1)th stage of the FFT algorithm, the (i+1)th processor should read the
data whose address is
[a p-1 ...a i+2 a i ...a 0 ][a i+1 ] (18)
where ai+1 = 0, . . . , b−1 and aj = const, for j ≠i + 1 . The inverse permutation is given by
[a p -1 ...a i  2 a i ...a 0 ][fa1 (a i )] (19)
11

where ai = const and ai+1 = 0, . . . , b − 1, since this permutation yields a b-tuple whose
elements differ only in the (i + 1)th digit. To prove that these elements are stored in b distinct
memory banks assume that this is not the case, i.e. there exist two distinct
elements, q 1  [a p -1 ...a i  2 a i ...a 0 ][ fa1 (a i )] and q 2  [a' p - 1 ...a' i  2 a' i ...a' 0 ][fa'1 (a' i )] that are
i 1 1 1
244 VLSI

in the same b-tuple and at the same time are stored in the same memory bank. Since q1 and
1 1
q2 are on the same b-tuple, therefore aj = a′j for j ≠i + 1 . Further, fa' i  1 (a i )  fa i  1 (a i ) ,
which is true only when a'i+1 = a'i+1 . Therefore, the elements q1 and q2 coincide.
To conclude the proof we show that the forward and inverse permutations at all stages
result in a correct FFT algorithm. We proceed by induction on the number of stages. During
the first stage of the FFT the data are written in the input memory banks (memory −1)
according to [ a p-1 ...a 1 ][a 0 ] . The outputs of the butterfly are written to the first intermediate
memory (memory 0) using the permutation [a p-1 ...a 2 a 0 ][ fa1 (a 0 )] , a 0 = 0 ,..., b - 1 . We assume
that the functions fj are identical for all stages, although this may not be the case. The only
actual restriction is to use a permutation Pi for writing the data at stage i and its inverse
permutation Pi-1 for retrieving the data at stage i+1. Applying the inverse permutation we
obtain [a p-1 ...a 2 a 0 ][ f -1 a1 (a 0 )] , with a0 = constant and a1 = 0, . . . , b−1. The set { f -1 a1 (a 0 )}
consists of b distinct elements, such that their address differs only in the digit a1 and
therefore, they constitute a valid transform b-tuple. Further, assume that in the ith stage the
elements are stored according to equation 17. Using the same reasoning as above, we can see
that the inverse transform yields a valid b-tuple and the proof is complete. Note that the
write operation of the elements in the first set of banks (memory 0) does not require any
permutation on the input data set. The elements are written to the bank corresponding to
their 0th (zero) address digit, in radix b notation. Similarly, during the final step of the
algorithm, the data can be written in the same addresses and banks as those that have been
read from, since no computation follows
The current section has shown the technique and the properties of the permutations
required to the parallel access of each b-tuple at each FFT stage. An engineer can realize the
technique by choosing among the straightforward solutions, e.g. the cyclic permutations. A
ninteresting research topic is to identify those permutations, which can be realized with
minimal interconnection and address generation circuits and thus lower the VLSI cost.

5. Conclusion
The present chapter has shown a technique to design FFT architectures for real-time
applications, which involve input with large number of complex points. The technique bases
on combining three consecutive R4 stages to realize a R64 computation. The resulting R43 as
well as the systolic architectures, which utilize a R43 as a stage for executing 4K, 16K, 64K
and 256K point FFTs, have been shown to provide higher throughput compared to hitherto
published architectures solving the corresponding transformations. Moreover, the R43 and
4K FFT architectures achieve the highest ratio of throughput to normalized area.
Furthermore, the chapter has proven a technique for parallelizing the memory access for
each butterfly radix-b computation so that the throughput can be improved further by a
factor of b.
Fully Systolic FFT Architectures for Giga-sample Applications 245

6. References
A. Cortes, I. Velez, I. Zalbide, A. Irizar, J.F. Sevillano ”An FFT Core for DVB-T/DVB-H
Receivers ICECS’06, page(s):102-105, December 2006.
A. Oppenheim, R. Schafer Digital Signal Processing. Prentice Hall, 1975.
B. G. Jo and M. H. Sunwoo, ”New Continuous-Flow Mixed-Radix (CFMR) FFT Processor
Using Novel In-Place Strategy,” IEEE Trans. Circuits Syst. I, vol.52, no.5, May 2005.
B. Suter and K. S. Stevens ”A Low Power, High Performance approach for Time-Frequency
/ Time-Scale Computations,” Proceedings SPIE98 Conference on Advanced Signal
Processing Algorithms, Architectures and Implementations VIII. Vol. 3461, pp. 86-
90, July 1998.
C. D. Thompson, “Fourier Transforms in VLSI”, IEEE Transactions on Computers, Vol. 32,
1047 - 1057, 1983
D. Harper, “Block , Multistride Vector, and FFT Accesses in Parallel Memory Systems”, IEEE
Trans. on Parallel and Distributed Systems, Vol. 2, No. 1, pp. 43 - 51, January 1991
D. Reisis, N.Vlassopoulos, ”Address Generation Techniques for Conflict Free Parallel
Memory Accessing in FFT Architectures” ICECS, pp.1188-1191, December 2006.
E. Bidet, D. Castelain, C. Joanblanq and P. Stenn “A fast single-chip implementation of 8192
complex point FFT” IEEE Journ. of SSC, 30(3):300-305, Mar. 1995.
E. H. Wold and A. M. Despain ”Pipeline and Parallel FFT Processors for VLSI
Implementations,” IEEE Transactions on Computers, vol. C-33, 1984.
Earl E. Swartzlander, Jr ”Systolic FFT Processors: A Personal Perspective Journal of VLSI
Signal Processing, June 2007.
I. S. Uzun, A. Amira and A. Bouridane, ”FPGA implementations of Fast Fourier Transforms
for real-time signal and image processing,” IEEE Vision, Image and Signal
Processing, 2005.
J. Lee, J. Lee, M. H. Sunwoo, S. Moh and S. Oh ”A DSP Architecture for High-Speed FFT in
OFDM Systems,” ETRI Journal, 2002.
J. Takala and K. Punkka Scalable FFT Processors and Pipelined Butterfly Units Journal of
VLSI Signal Processing 43, 113-123, 2006.
J. Y. OH and M. S. Lim, ”New Radix-2 to the 4th Power Pipeline FFT Processor,” IEICE
Trans. Electron., VOL. E88-C, NO. 8, August 2005.
J.W. Cooley and J.W. Tukey, “An algorithm for the machine computation of complex
Fourier series”, Mathematics of Computation, 1965
K. Babionitakis, K. Manolopoulos, K. Nakos, N. Vlassopoulos, D. Reisis, V. Chouliaras, “A
High Performance VLSI FFT Architecture”, 13th IEEE International Conference on
Electronics, Circuits and Systems, pp. 810-813, December 2006
K. Maharatna, E. Grass, and U. Jagdhold, ”A 64-Point Fourier Transform Chip for High-
Speed Wireless LAN Applications Using OFDM,” IEEE Journal of Solid State
Circuits, VOL. 39, NO. 3, March 2004.
K. Manolopoulos, K. Nakos, D. Reisis, N.Vlassopoulos, V.A. Chouliaras ”High Performance
16K, 64K, 256K complex points VLSI Systolic FFT Architecture” ICECS, pp. 146-149,
December 2007.
L. G. Johnson, “Conflict Free Memory Addressing for Dedicated FFT Hardware”, IEEE
Transactions on Circuits and Systems - II: Analog and Digital Signal Processing,
Vol. 39, No. 5, May 1992
246 VLSI

L. R. Rabiner and B. Gold ”Theory and Application of Digital Signal Processing,” Prentice-
Hall.
L. Yang, K. Zhang, H. Liu, J. Huang and S. Huang ”An Efficient Locally Pipelined FFT
Processor,” IEEE Trans. Circuits Syst. II, vol. 53, no. 7, July 2006.
N. Hu, O. Ersoy, “Fast Computation of Real Discrete Fourier Transform for Any Number of
Data Points”, IEEE Transactions on Circuits and Systems, Vol. 38, No. 11, pp. 1280 -
1292, November 1991
O.K . Ersoy, Fourier-Related Transforms, Fast Algorithms and Applications. Englewood
Cliffs, NJ:Prentice Hall, 1997.
R. Thomas and K. Yelick, “Efficient FFTs on IRAM”, Proceedings of the 1st Workshop on
Media Processors and DSPs, 1999
S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy, ”A New Radix-2/8 FFT Algorithm for
Length−q × 2m DFTs,” IEEE Trans. Circuits Syst. I, vol. 51, no. 9, September 2004.
S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy, ”New Radix-(2×2×2)/(4×4×4) and Radix-
(2×2×2)/(8×8×8) DIF FFT Algorithms for 3-D DFT,” IEEE Trans. Circuits Syst. I,
vol. 53, no. 2, February 2006.
S. Choi, G. Govindu, J. W. Jang, V. K. Prasanna ”Energy-Efficient and Parameterized
Designs of Fast Fourier Transforms on FPGAs,” The 28th International Conference
on Acoustics, Speech, and Signal Processing (ICASSP), April 2003.
S. He and M. Torkelson ”A New Approach to Pipeline FFT Processor,” Proceedings of the
IPPS, 1996.
S. He and M. Torkelson ”Design and Implementation of a 1024-point Pipeline FFT
Processor,” IEEE 1998 Custom Integrated Circuits.
S.S. Wang and C.S. Li An Area-Efficient Design of Variable-Length Fast Fourier Transform
Processor Journal of VLSI Signal Processing, March 2007.
T. Lenart and V. Owall, ”A 2048 Complex Point FFT Processor Using a Novel Data Scaling
Approach,” IEEE ISCAS 2003.
W. H. Chang, T. Nguyen, ”An OFDM-Specified Lossless FFT Architecture,” IEEE Trans.
Circuits Syst. I, vol. 53, no. 6, June 2006.
www.altera.com
www.xilinx.com
Y. Ma, “An Effective Memory Addressing Scheme for FFT Processors”, IEEE Transactions
on Signal Processing, Vol. 47, No. 3, 907 - 911, May 1999
Y. Ma, L.Wanhammar, “A Hardware Efficient Control ofMemory Addressing for High-
Performance FFT Processors”, IEEE Transactions on Signal Processing, Vol. 48, No.
3, 917 - 921, March 2000.
Y.N. Lin, H.Y. Liu, and C.Y.Lee “A 1-GS/s FFT/IFFT Processor for UWB Applications,”
IEEE Journ. of SSC, vol. 40, Issue 8, Aug. 2005.
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 247

12
X

Radio-Frequency (RF) Beamforming

Using Systolic FPGA-based Two
Dimensional (2D) IIR Space-time Filters
Arjuna Madanayake, Member, IEEE, and Leonard T. Bruton, Fellow, IEEE
Multidimensional Signal Processing Group, University of Calgary
Calgary, Alberta, Canada
(hmadanay, [email protected])

1. Introduction
Plane-waves are far-field solutions to (1) the vector wave equation, for the case of
electromagnetic waves, (2) the scalar wave equation, for the case of longitudinal pressure
waves in seismic, acoustic, and ultrasonic systems, as well to (3) linear surface waves, such
as those created by dropping a pebble into still the waters of a pond. Far-field beamforming
refers to the highly-selective directional enhancement of propagating spatio-temporal plane-
waves based on their directions-of-arrival (DOA).

The directional enhancement (beamforming) of electromagnetic plane-waves is of

importance in many areas of electrical engineering, such as in wireless communications,
radar and radio-frequency (RF) imaging (multi-GHz range). Of particular importance are
applications in radio-astronomy and space physics (Van Ardenne,2000), where far-field
beamforming is increasingly employed in aperture arrays, and in wireless mobile voice and
data communication systems. In the case of data communications, beamforming is used for
mitigating the fading effects of multipath propagation (Litva and Lo,1996; Liberti Jr. and
Rappaport,1999; Huseyin Arslan, Zhi-Ning Chen and Maria-Gabriella Di Benedetto,2006) in
satellite-borne remote sensing applications involving synthetic aperture and Doppler radar,
in navigation and location devices based on GPS (Silva, Worrel and Brown), as well as in
various ultra-wideband location technologies (Ghavami, Michael and Kohno).

Traditionally, receiver-side beamforming has been achieved using highly-directional

receiving antennas, typically employing parabolic reflectors and horns, antenna array
configurations with passive phasing networks (such as delay-and-sum networks and
phased-array feeds) and reflect-arrays (Hum, Okoniewski and Davies). Digital beamforming
algorithms are often based on fractional-delay steering algorithms and/or finite impulse
response (FIR) digital filters (Ghavami, Michael et al.; Johnson and Dudgeon,1993; Liberti Jr.
and Rappaport,1999; Staderini,2002; Huseyin Arslan, Zhi-Ning Chen et al.,2006; J. Roderick,
H. Krishnaswamy, K.Newton et al.,2006). The use of digital signal processing (DSP) in far-
248 VLSI

field broadband beamforming for smart antenna array applications is currently receiving
much attention, mainly due to the continuously increasing availability of digital
programmable logic and custom silicon fabrication technologies that are gradually enabling
the typically high levels of real-time computational throughputs necessitated by such DSP-
based broadband smart antenna arrays.

In this contribution, we describe a particular type of recently proposed far-field beamformer

that is based on two-dimensional (2D) space-time digital filters having infinite impulse
responses (IIRs) (Ramamoorthy and Bruton; Agathoklis and Bruton,1983; Bruton and
Bartley,1985). Unlike the more widely-used DSP-based 2D FIR beamformers, the described
2D IIR beamformers have 2D z-domain transfer-functions H ( z1 , z2 )  N ( z1 , z2 ) / D ( z1 , z2 )
havng pole-manifolds, as well as zero-manifolds, in the 2D complex plane  2 . Further, for
beamforming applications, D ( z1 , z2 ) must be non-separable, implying non-trivial design
challenges to avoid multidimensional instability and computability constraints (such
challenges are not encountered in 1D design or if D ( z1 , z2 ) is separable in the 2D case).
However, suitable closed-form algebraic expressions for beamforming discrete-domain 2D
IIR transfer functions are available in the literature that are computable, stable and
realizable using hardware with lower complexity than FIR beamformers having similar
directional selectivity, angular half-power bandwidth, etc. Furthermore, the existence of
such closed-form algebraic transfer functions facilitates the real-time continuous steering of
the direction of the beam and adjustment of its bandwidth, making these filters attractive for
applications in emerging software-defined radio (SDR), microwave imaging and cognitive
radio systems.

This paper is organized as follows: in section 2, we provide a brief review of space-time

plane-waves and their properties in the 2D space-time frequency domain, followed by
section 3, where we provide a comprehensive review of broadband plane-wave digital filter
design. Thereafter, in section 4, we discuss practical hardware implementations starting
from 2D difference equations that lead to signal flow graphs and massively-parallel systolic-
array hardware realizations. Section 5 describes some recent progress we have achieved in
prototyping these new systolic-array circuits using field programmable gate array (FPGA)
technology. Finally, in section 6 we describe recently available technology and electronic
design automation (EDA) tools that may eventually lead to 2D IIR beamformers that operate
in real-time for various broadband microwave beamforming applications.

2. Electromagnetic Plane-waves in Space-time

We consider here either the transverse electric field E y ( x, y, z , ct ) or magnetic field
H x ( x, y, z , ct ) of a propagating electromagnetic plane wave, where ( x, y, z , ct )   4 is the
4D space-time continuous-domain, ( x, y, z )   3 is 3D space, t   is time and
c  3  108 ms 1 is the speed of light in air/vacuum. By analogy with 3D planes, the
equation
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 249

3
1 x   2 y   3 z  ct    , 1,2,3  , and, 
k 1
2
k 1 (1)

is a 4D hyper-plane in the 4D continuous-domain ( x, y, z, ct )   4 . Propagating

electromagnetic plane waves are 4D hyper-plane waves in ( x, y , z , ct ) given by

  3
w( x, y, z, ct )  wPW  1 x   2 y   3 z  ct  ,
    2
k  1   1 (2)
   k 1

and therefore have the property that they are constant-valued in each of the hyper-planes
(1): that is, for each   1 . Equivalently, for each value of  , wPW ( ) is a
corresponding 4D iso-surface in ( x, y , z , ct ) . In Fig. 1, we show the 4D plane wave
wPW 1 x   2 y   3 z  ct  in the 3D spatial domain ( x, y, z )   3 in an iso-plane which, by
simply 3D geometry, is perpendicular distance ct from the origin, as shown. In 3D, we may
therefore visualize the 4D space-time plane wave of equation (2) as an infinite set of such
iso-planes w(  ) , each of which is propagating in ( x, y, z ) over time t with speed c in a
direction normal to the iso-planes. Depending on the 1D spectral properties of the c-scaled
temporal signal wPW (ct ) , the plane wave might be temporally-narrowband or temporally-
broadband.

Note that, for the case of the ideal plane wave, the region of support (ROS) of equation (2) in
( x, y, z , ct )   4 extends, in general, to infinity in at least some directions in  4 . Equation
(2) represents either the electric or magnetic field of the plane wave in 4D space-time. In this
chapter, we are only concerned with the values of the 4D plane wave signal as received on a
straight line in ( x, y, z ) . Therefore, we consider only the special case of the 2D space-time
 
representation for which equation (2) reduces to the form w( x, 0, 0, ct )  wPW  1 x  ct  for



  
signals on the x-axis. With the DOA in 3D space defined by the angles
 o (measured on x - z plane) and  o as shown in Fig. 1, it is easily shown that (2) may be
written in the form

 
w( x, y, z , ct )  wPW    sin  o cos o  x   cos  o cos o  z   sin o  y  ct  (3)
   
  
with  / 2   o , o   / 2, from which it follows that the corresponding 2D space-time
plane wave signal received on the x-axis is given by
250 VLSI

 
w( x, ct )  wPW    sin  o cos o  x  ct  (4)
  
  

The 4D hyper-planar iso-surfaces of constant  in (3) become 2D iso-lines of (4) in

( x, ct )   2 given by
  sin  o cos o  x  ct   (5)

As shown in Fig. 2, the space-time direction of the 2D space-time plane wave is defined by
the normal to these contours and is given by (Gunaratne and Bruton; Khademi)

  tan 1 (sin  o cos o ) (6)

with respect to the ct-axis in the 2D Cartesian space-time domain ( x, ct ), where

sin   sin  o cos o . Note from (6) that the 2D spatio-temporal direction   tan 1 (sin  ) is
constrained by  / 4     / 4 where the extreme values  / 4 occur where
((  o   / 2 ) and/or  o  0 . These extreme directions are known as the ‘end-fire’ angles,
corresponding to plane waves having DOAs in 3D space that are in the direction of the x-
axis. The so-called ‘broadside’ DOAs, with respect to the x-axis, are those directions for
which the 4D space-time signal in (3) is constant everywhere on the x-axis at all instants of
time t, which corresponds to the 2D space-time direction   0 and is equivalent to the
DOAs in 3D space given by (  o  0,  and/or  o   / 2 ).

Fig. 1. The direction of arrival (DOA) of a plane-wave in 3D space, ( o , o ) , and the

appararent spatial DOA seen by a linear array along x-axis,  .
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 251

Fig. 2. Propagating plane-wave in 3D space (a), 2D spatial view on the y  0 plane (b), 2D
spatio-temporal DOA (c), and region of support (ROS) on 2D frequency-domain, aligned
along the spatio-temporal DOA (d).

2.3 On The Region of Support (ROS) of the 2D Fourier Transform of 2D space-time

Plane Waves
Given the 2D Fourier transform pair for 2D space-time plane
waves wPW   x sin   ct   WPW (e
j x
, e jct ) , it may be shown (Bruton and Bartley,1985)
j x
that the Region of Support (ROS) of the spectrum WPW (e , e jct ) in the 2D Cartesian
frequency domain ( x , ct ) is confined to the straight line

 x  ct sin   0 (7)

which passes through the origin and subtends angle  to the ct axis. It lies on the ct axis
for broadside DOAs and on ct   x for end-fire DOAs. Importantly therefore, the ROS of
all 2D space-time electromagnetic plane wave signals, propagating at speed c, cannot lie
outside the 90-degree wide 2D fan-shaped region  x  ct in ( x , ct ) .
252 VLSI

2.4 Spectral-Filtering a Desired 2D Space-time Plane-wave in the Presence of other

Plane Waves and Noise
Typically, the signal received on the x-axis may be represented by M multiple broadband
plane-waves, each having a different orientation  o , k , o , k , k  0,1, 2, ( M  1) , in 2D space-
time, and additive 2D noise. We assume the first plane wave, given by k  0 is the desired
plane wave to be recovered by filtering. Then the received signal may be written in the form

M 1
w( x, ct )  w
k 0
PW , k ( x sin k  ct )  nv ( x, ct ) (8)

Where sin k  sin  o , k cos o , k , and where nv ( x, ct ) represents 2D space-time noise. The
Fourier transform of (8) is therefore given by

M 1
W (e jx , e jct )  W
k 0
PW , k (e jx , e jct )  N v (e jx , e jct ) (9)

2D
where N v ( x , ct )  nv ( x, ct ) . Typically, nv ( x, ct ) corresponds to non-plane-wave
electromagnetic propagating interference or other sources of 2D broadband noise, modelled
as additive white Gaussian noise (AWGN). Therefore, the 2D ROS of the noise spectrum
N v ( x , ct ) is typically uniform throughout the 2D frequency-domain ( x , ct )   2 . The
j x
ROS of W (e , e jct ) therefore consists of the uniform ROS of N v ( x , ct ) and M lines
through the origin, where the orientation of each line is given by the M different angles
 k  tan 1 (sin k ) .

For notational convenience in the rest of the chapter, we will use 1   x as the spatial frequency
variable, and 2  ct as the time-frequency variable corresponding to spacetime ct.

2.5 Beamforming of a Broadband Plane Wave using Space-time Filters

The simplest of 2D space-time plane-wave filter is known as a ‘frequency-planar beam’ filter
(Bruton and Bartley,1985), because it’s passband lies on a line-through the origin and ideally
has a ‘beam’ shaped 2D passband of uniform width. The beam-shaped 2D passband is
oriented to enclose the ROS of the 2D spectrum of the desired plane-wave over its full
temporal bandwidth while attenuating all spectra away from this narrow passband. Such
beam filters can be realized using several methods involving both FIR or IIR digital filters
(Gunaratne and Bruton; Khademi; Bruton and Bartley,1985). FIR filters are inherently stable
but are of high arithmetic complexity due to the relatively higher order of the filter that is
required for a given selectivity, relative to approximately-equivalent IIR filters. However,
the latter are not as straightforward to design and to implement.

The focus of this chapter is on the design and real-time hardware implementation of a first-
order 2D IIR beam digital plane-wave filter.
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 253

2.6 Effects of 2D Space-time Sampling using a Uniform Linear Array (ULA)

The 2D continuous-domain space-time signal w( x, ct ) is sampled in space using a number
of equi-spaced antennas along the x-axis and sampled in time to yield the corresponding 2D
sampled space-time signal w( n1 x, n2 cTCLK ) , n1,2  0,1, 2,3, , and where. In order to
prevent undesirable spatial aliasing, such uniform linear arrays (ULAs) of antennas require
that the uniform distance between antennas satisfy the Nyquist condition (the distance
between antennas is x  cT , where fU  1/ T is the upper temporal frequency of the
input signal beyond which its spectrum lacks support). For N antennas, the 2D spatially-
sampled continuous-time antenna array signals are given by w( n1x, ct ), n1  0,1,..., N  1 .
The 2D spectrum of w( n1x, ct ) replicates on the spatial-frequency 1 axis with periodicity
2 . The N continuous-time signals w(n1x, ct ) are amplified, using a low noise amplifier
having the required temporal bandwidth and noise performance, then low-pass filtered
prior to A/D conversion, resulting in the 2D discrete-domain antenna signal
w(n1x, n2 cTCLK ). The ADC clock frequency FCLK  1/ TCLK where is chosen by selecting
the inter-sample time TCLK  T such that Nyquist sampling theorem is satisfied in both
the spatial and temporal frequency domains. Usually, TCLK  KU T , KU  1, where KU
is the so-called temporal oversampling factor and is sometimes necessary for minimizing the
frequency-warping effects that are introduced in the design of the digital filter transfer
function.

Methods have been proposed and implemented for significantly reducing the required
number of antennas without significantly reducing performance, for wireless
communications and other applications. These methods also lead to much reduced
arithmetic complexity of the filter and are based on allowing a controlled amount of
multidimensional spatial aliasing and thereby spatial under-sampling, as reported in
(Khademi and Bruton; Madanayake, Hum and Bruton).

3. First-order 2D IIR Frequency-Beam Plane-wave Filter

In the above, it is established that the directional enhancement of an ideal desired space-
time plane-wave may be achieved using a 2D space-time filter having a 2D passband that
encompasses the line-shaped ROS of the desired plane-wave signal. Further, undesired
signals, such asother plane-waves and/or noise may be attenuated by ensuring that the ROS
of the 2D stopband correspondas to the ROS of the spectrum of the undesired signals.
Although this approach has been extended to 3D and 4D space-time signals (Kuenzle and
Bruton; Bolle,1994; Bruton,2003; Dansereau,2003; Kuenzle and Bruton,2005; Dansereau and
Bruton,2007), here we focus on the simplest 2D case by describing the design and
implementation of a suitable 2D filter.
254 VLSI

3.1 The Prototype Resistively-terminated 2D Passive Frequency-Beam Network

Consider a 2D first-order continuous-domain inductance-resistance network (Bruton and
Bartley,1985), where s1 and s2 are spatial and temporal Laplace variables (Dudgeon and
Mersereau,1990; Johnson and Dudgeon,1993; Schroeder and Blume,2000), respectively. The
input-output 2D Laplace voltage transfer function of this network is given by

R Y ( s1 , s2 )
T ( s1 , s2 )   (10)
R  L1 s1  L2 s2 W ( s1 , s2 )

where the parameters L1  0, L2  0, and R  0 correspond to a passive spatial inductor,

passive temporal inductor and passive resistance, respectively, with transform inputs and
outputs W ( s1 , s2 ) and Y ( s1 , s2 ) , respectively. We denote the respective transform pairs by
2D 2D
w( x, ct )  W ( s1 , s2 ) and y ( x, ct )  Y ( s1 , s2 ), respectively. The steady-state input-output
frequency-response of (10) is found by setting s1  j1 and s2  j2 , leading to the 2D
frequency response transfer function

R Y ( j1 , j2 )
T ( j1 , j2 )   (11)
R  j  L11  L22  W ( j1 , j2 )
From (5), the network under consideration is 2D resonant on the 2D line-shaped region

1 L1  2 L2  0 (12)

passing through the frequency-origin (Note: In 2D, capacitors are not required to induce
resonance). At all finite frequencies where (12) is satisfied (i.e. throughout the 2D passband) ,
network energy resonates between the two inductance elements and T ( j1 , j2 ) is unity. By
choosing L1  cos  and L2  sin  , 0    90o , we can orient the axis of the 2D passband to
the angle  . A typical response is shown in Fig. 3. The shape of the 2D gain T ( j1 , j2 ) of
the filter may be envisaged in 2D frequency space by noting that L11  L22   describes,
for constant  , a line that is parallel to the 2D passband and along which T ( j1 , j2 ) is
constant and less than unity. Importantly, T ( j1 , j2 ) decreases monotonically with
increasing values of  with the two –3dB lines having gain 0.707 given by    R . We
make the following summary observations (see (Bruton and Bartley,1985) for details):

1. At 2D resonance, that is on the frequency-line in (1 , 2 ) where   0, the transfer

function T ( j1 , j2 )  1 . This defines the 2D passband axis and unity gain on the
centre of the beam-shaped passband.
2. Along all directions orthogonal to the passband axis in (1 , 2 ) , the magnitude
and phase frequency response of the filter correspond to that of a first-order low-
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 255

pass transfer function, monotonically-decreasing with distance from the passband

axis.
3. The uniform –3dB bandwidth of the beam is given by 3 dB  R / L12  L2 2 .

3.2 The Transfer-functions of the First-order Beam Filter in the 2D s- and z- Domains
Although the inverse 2D Laplace transform of equation (10) yields a continuous-domain
partial differential equation for the input-output transfer-function, practical
implementations have so far been in the discrete-domain of 2D finite-difference equations,
implemented in the form of digital circuits. Transformation to the discrete-domain is
zk  1
achieved by applying the normalized 2D bilinear transform (2D BLT) sk  , k  1, 2, to
zk  1
equation (10) leads, after considerable algebraic manipulation (Bruton and Bartley,1985), to
the 2D z-transform transfer function

H ( z1 , z2 ) T  z1 1, z2 1  
1  z 1  z 
1
1
1
2

Y ( z1 , z2 )
(13)
 z 1 z 1 
 1 2  1  b  b z  z  b
10
1
11 2
1
1 z 1
01 2
W ( z1 , z2 )

2D 2D
where W ( z1 , z2 )  w( n1x, n2 cTCLK ) and Y ( z1 , z 2 )  y ( n1x, n2 cTCLK ), respectively, and
where bij   R  (1)i L1  ( 1) j L2  /  R  L1  L2  . The above application of the 2D BLT, which is
a conformal mapping between the 2D Laplace and 2D z-domain, results in a distortion ofn
the high frequency part of the 2D passband, known an bilinear warping, that leads to a
practical limitation of the upper frequency  0.5 | 2 |   of the beam-shaped passband.
This effects of this limitation may be avoided by suitable temporal and/or spatial over-
sampling of the input signal.

For example, here we shall employ a temporal over-sampling factor of 2 for which we show
in Fig. 4 the correpsonding ‘weakly-warped’ magnitude response of the discrete-domain
frequency-response transfer-function over the useful range | 2 | 0.5 .
256 VLSI

Fig. 3. The 2D continuous-domain steady-state Magnitude Frequency Response T ( j1 , j2 )

of the filter in (10) for | k |  , k  1, 2, and R  0.1, L1  cos(30o ), L2  sin(30o ).

Fig. 4. A Beam-shaped Response that is warped by the 2D BLT, shown in the usable range
  1   and  0.5  2  0.5 . The beam shape at frequencies  0.5 | 2 |   are
not used in our application because the beam-shape is significantly off-axis, due to binear
warping. The interested reader is referred to (Madanayake and Bruton; Bruton,2003) for
details.
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 257

3.3 On Implementing the 2D Difference-Equations Using Differential Operators

Taking the inverse 2D z-transform of (10) leads to the 2D space-time input-output direct-
form difference-equation, which we have shown can be implemented in massively-parallel
systolic-array hardware for real-time filtering applications. However, the recently proposed
hybrid-form signal flow graph (Madanayake, Hum et al.; Madanayake,2008), although
relatively complicated to design, offers lower complexity and high-speeds of operation than
direct-form methods. In this paper, we pursue the hybrid-form signal flow graph method.
The 2D z-domain transfer-function that leads to the so-called hybrid-form structure (an
architecture that enjoys the high-speeds of operation of direct-form structures while being as
low in complexity as so-called differential-form structures) can be obtained by methods
available in the literature (Madanayake, Hum et al.; Madanayake and Bruton,2007;
Madanayake,2008). For example, we may write a suitable transfer function in terms of the
1 1 1
spatial differential operator z1D  z1 /(1  z1 ) as

1  z21 Y ( z1 , z2 )
H ( z1 , z2 )  1
 (14)
z W ( z1 , z2 )
1 1
1  z11
1  z21    z21

z1D1

where we require

2 cos  2sin 
 ,   1 (15)
R  cos   sin  R  cos   sin 

R
Note that the passband gain in (14) is scaled by the constant R  L1
, relative to the direct-form
case (Bertschmann, Bartley and Bruton; Madanayake, Hum et al.; Liu and Bruton,1989;
Madanayake and Bruton,2007; Madanayake,2008), and is ignored in the following because it
is not of practical significance. Re-writing the direct-form transfer-function using the spatial-
differential operator results in just two filter coefficients in the denominator of (14) instead
of three, implying a 33% reduction in the number of parallel hardware multipliers required
in circuit realizations, relative to direct-form realizations.

The difference-equation realizations described here lead to practical-bounded-input-

bounded-output (practical-BIBO) stable (Agathoklis and Bruton,1983) performance under
finite precision arithmetic, assuming zero initial conditions (ZICs) for both spatial and
temporal iterations. Methods that guarantee ZICs are considered later in the following.

4. A Real-time High Throughput Implementation using the Hybrid-form

Systolic-Array Processor Architecture
Systolic-array processors are massively-parallel computers having identical, synchronously
clocked, fully-pipelined, high throughput identical processing elements, which are
connected in a linear- or meshed-array configuration
258 VLSI

Fig. 5. Overview of the plane-wave filter implementation consisting of N parallel processing

core modules (PPCMs), which have an internal signal flow graph based on the hybrid-form
transfer-function in equation (14).

(Sid-Ahmed; Kung,1988a; Kung,1988b; Shanbhag,1991; Rader,1996; Zajc, Sernec and

Tasic,2000). Such systolic-array processors are modular, regular, and locally interconnected,
making them well-suited for real-time signal processing using application-specific VLSI
hardware for digital signal processing applications at radio-frequency.

Research on novel systolic-array architectures for 2D/3D IIR frequency-planar digital plane-
wave filters for beamforming applications has lead to field-programmable gate-array
(FPGA) based single-chip multiprocessor implementations capable of real-time operation at
a sustained arithmetic throughput of one-frame-per-clock-cycle (OFPCC), a requirement for
real-time plane-wave filtering at RF using linear- or rectangular-arrays of antenna elements
(Hum, Madanayake and Bruton; Madanayake, Bruton and Comis; Madanayake and Bruton;
Madanayake and Bruton; Madanayake, Hum et al.; Madanayake, Hum and Bruton;
Madanayake,2004; Madanayake,2008; Madanayake and Bruton,2008). The required OFPCC
throughput rate, required for multi-GHz implementations, arises due to the fact that the
signals of interest are of ultra-wide RF bandwidth, which leads to Nyquist sample rates that
are at least twice the full RF bandwidth of the signal.

The beamformers therefore directly sample RF signals from the antennas without down-
conversion (or bandpass sampling), and leads to frame sample rates in the GHz. Such
excessively-high frame sample rates (multiple GHz) make software-based realizations
infeasible using traditional DSP technologies. Our research indicates (Madanayake and
Bruton; Madanayake,2008) that massively-parallel synchronously-clocked, speed-optimized,
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 259

fully-pipelined systolic-array processors are currently the best available solution for the
broadband real-time DSP-based radio-frequency (RF) beamforming applications using
sampled antenna arrays (Arnold Van Ardenne; Ellingson,1999; Liberti Jr. and
Rappaport,1999; Weem, Noratos and Popovic,1999; Frederick, Wang and Itoh,2002; Do-
Hong and Russer,2004; Rodenbeck, Sang-Gyu, Wen-Hua et al.,2005; Madanayake,2008;
Devlin,Spring 2003).

4.1 Overview of the Architecture

The massively-parallel systolic-array architecture consists of an array of identical parallel-
processing core-modules (PPCMs), sometimes called “processing elements” in the literature
(Kung,1988a). Each PPCM is dedicated to processing an antenna element, and signals from
all N elements are amplified using a low-noise amplifier (LNA), low-pass filtered (LPF),
and time-synchronously sampled using N identical analog signal processing chains. The
PPCMs and analog-to-digital converters (ADCs) are clocked using a single-phase master
clock signal, of frequency FCLK  1/ TCLK . The PPCMs are derived using the recently-
proposed hybrid-form signal flow graph having the required z-domain transfer-function
(14) (Hum, Madanayake et al.; Madanayake, Hum et al.).

The PPCMs that comprise the systolic-array processor are fully-parallel, speed-maximized,
fully-pipelined, multi-input-multi-output (MIMO) processors, each consisting of 2 input
ports and 2 output ports. A PPCM at spatial location n1 has its input port A connected to the
ADC at location n1 and input port B connected to the output port C of the PPCM at location
n1  1. Port D provides the computed output signal y (n1x, n2 cTCLK ) for spatial location
n1 .

4.2 Inter-PPCM and Intra-PPCM Pipelines

The PPCMs are pipelined such that signals entering through the input ports
An1 and Bn1 undergo p additional clocked delays, as a result of internal pipelining. These
additional delays can be compensated by delaying the input signal An1 1 by ( n1  1) p clock
cycles, leading to a delay of (n1  1)Tx seconds, where Tx  pTCLK is the pipelining
th
latency of a PPCM. For N PPCMs, the final output signal at the output of the N PPCM
therefore undergoes a pipelining delay of Np clock cycles. When 2D space-time output
signals are required, the output signals from each PPCM must be fed through additional
clocked FIFOs, having depth ( N  1  n1 ) p , so that the signals at all spatial output locations
are uniformly delayed by Np clock cycles. The implemented transfer-function is therefore
modified, in the presence of pipelining, to the linear phase-delayed form H ( z1 , z2 ) z2 Np
which has no effect on the magnitude frequency-response function (because
 j2 Np CLK
e  1. )
260 VLSI

Fig. 6. Hybrid-form PPCM circuit having 2 inputs, 2 outputs, 4 two-input parallel

adder/subtractors, and 2 parallel hardware multipliers.

4.3 Design of the Hybrid-form PPCMs

Having obtained a basic overview of the systolic-array architecture, we now derive the
internal signal flow and internal components of each PPCM. Recall the 2D hybrid-form
transfer-function (14):
1  z21 Y ( z1 , z2 )
H ( z1 , z2 )  1
 (16)
z W ( z1 , z2 )
1 
1 1
1  z21    z21
1  z1

Cross-multiplying terms in (16), we get the 2D z-domain input-output form, given by

 1

1  z W ( z , z )  1   1 z z
1
2 1 2
1
1
1  z21   z21  Y ( z1 , z2 ), (17)
 1 
leading to
 z11 
 W ( z1 , z2 )   1
Y ( z1 , z2 ) 
1  z1
Y ( z1 , z2 )    1  z 1
 2 (18)
1   z21

p
Multiplying both sides by z2 yields the required form
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 261

Fig. 7. Interconnections between PPCMs, shown here in the mixed domain (n1 , z2 )  ,
leads to the massively-parallel systolic-array processor implementation of the beam plane-
wave filter.

 z11 
 W ( z1 , z2 ) z2   z2
p p
1
Y ( z1 , z2 ) 
1  z1
Y ( z1 , z2 ) z2 p   1  z 1
 2 (19)
1   z21

Computing the inverse z1-transform of (19) under spatial ZICs, we obtain the 2D mixed-
domain (n1 , z2 )   form, given by

W (n , z ) z p
  z2 pU (n1 , z2 ) 
Y (n1 , z2 ) z2 p 
1 2 2

1   z21
1  z 
1
2 (20)

where
U (n1 , z2 )  Y (n1  1, z2 ) z2 p  U (n1  1, z2 ) z2 p (21)

and, where z2 p is the z2-transform of the internal pipelining delays at each PPCM. Because
the depth of pipelining is arbitrary, the numerator of (20) can be pipelined at will using
straightforward 1D FIR filter pipelining methods, noting that this 1D FIR section has two
terms W ( n1 , z2 ) and U (n1 , z2 ) which are obtained using digital ports An1 and
Bn1 , respectively. Equations (20) and (21) describe the 2-input-2-output z2-domain transfer-
functions of a PPCM at location 0  n1  N  1. The hybrid-form signal flow-graph is
thereby obtained, and is shown in Fig. 6. The first 3 PPCMs in the systolic-array are shown
in Fig. 7 as an interconnection of processors.
262 VLSI

4.4 A Pipelining Example

In order to familiarize the reader with pipelining concepts, we now provide a simple
example where it is assumed that p  12 internal pipelining stages are sufficient for
achieving the required throughput. The 12 stage pipeline will be distributed as follows: the
multiplier  will consist of 3 level pipelining; the three 2-input adders/subtractors denoted
A1, A2, and A3, are to have 3 levels of pipelining. It is important to ensure that all signal
components that connect to a particular 2-input adder/subtractor undergo equal delays.
This is essential for correct operation, and must be satisfied for all pipelined designs.

The hybrid-form signal-flow graph does not allow pipelining of A4, because only one unit-
delay buffer is available in the first-order feedback loop, which is usually absorbed inside
the parallel logic of multiplier  . Provided all feed-forward paths are fully pipelined, the
critical path delay of the hybrid-form PPCM cannot be reduced beyond
TCPD  TMul  TA / S where TMul and TA / S are the propagation delays of a parallel multiplier
and adder/subtractor circuit, respectively. The maximum speed of operation for a hybrid-
form PPCM is therefore less than FCLK  1/ TCPD unless additional speed-optimization
methods, based on look-ahead optimization, are employed. This method is discussed in the
next section.

Fig. 8. Signal flow graph of a hybrid-form PPCM having 12 cycles of pipeline latency
(arbitrarily chosen for the purpose of demonstration). The 12 clock-cycles of additional
pipelining can be used as required to reduce the critical-path delay (CPD) of the systolic-
array.
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 263

The pipelined version, having p  12 for the hybrid form PPCM, is shown in Fig. 8. We
now describe look-ahead speed-optimization of the internal 1D temporal IIR digital filter
1  z21
section having transfer-function that enables much greater levels of real-time
1   z21
throughput at the cost of additonal circuit complexity.

4.5 Additional Look-Ahead Speed-Maximization

We extend well-known 1D IIR pipelining using “look-ahead” optimization, a method
pioneered by Parhi et al (Parhi; Parhi,1991; Parhi,1999). Look-ahead is a method for
reducing additional delays into a critical feedback loop and is based on pole-zero
cancellation of 1D z-domain transfer-functions.

In section 4.4, we described an example for which 10 additonal delays are distributed in the
forward (that is, FIR, also known as feed-forward) signal paths of the PPCM, such that the
critical path delay of the PPCM (and therefore, of the systolic-array) is reduced to the
latency for a multiply-add-operation, denoted TCPD . The speed-bottleneck for this example
lies within the first-order feed-back IIR filter, which has a simple real-pole at z2    where
it may be shown that   1 for passive filter network prototypes. Because this pole is
within (or on) the unit circle z2  1 the 1D IIR filter section is unconditionally stable
(ignoring effects due to finite precision).

Let us further assume that our objective is to halve the critical path delay using look-ahead
optimization of the IIR section. This can be achieved by increasing the number of internal
delays in the first-order feedback loop to 2 (causing the feedback loop to increas in order):
this may be easilty achieved by multiplying both numerator and denominator of
1  z21
by 1   z 1
leading to the 2 nd order section, given by
1  z21 1   z21  leading to
2
1   z21 1   2 z22
a new critical path delay in the feedback loop TCPD , LA  TCPD / 2, implying an almost 100%
increase in the maximum speed of operation (Parhi; Parhi; Parhi and Messerschmitt; Parhi
and Messerschmitt; Sundarajan and Parhi; Parhi and Messerschmitt,1989; Parhi,1991;
Parhi,1999). This “look-ahead” speed-maximization process may be repeated: for example,

by multiplying the numerator and denominator of

1  z 1   z 
1
2
1
2
by 1   z2 leads
2 2

1   2 z22

to a 4th-order feedback loop having transfer-function

1  z 1   z 1  
1
2
1
2
2 2
z2  which
1  z 4 4
2

allows the multiplier  4 to consist of 3 levels of internal pipelining, while the fourth delay
can be used in the 2-input adder that completes the feedback loop (Madanayake and Bruton;
Madanayake,2008). The additional terms in the numerator that appear due to the
application of look-ahead speed-maximization lead to additional circuit complexity – this is
264 VLSI

the price for the extensive gain in real-time throughput, which is 300% for 4th order feedback
loops. The additional arithnetic circuits that appear in the feed-forward sections can be
easily pipelined by increasing the depth of pipelining p as required. In our example, we
have increase the depth of pipelining up to p  22 which allows 3-level pipelining to all
additional adders/subtractors and multipliers in the PPCM.

5. FPGA Circuit Prototypes

In this section, we provide a proof-of-concept circuit design using a field programmable
gate array (FPGA) device. An example implementation of a hybrid-form systolic-array
processor containing 21 fully pipelined PPCMs is provided. The target FPGA is a Xilinx
Virtex-4 Sx35-10ff668 device installed on a Nallatech BenADDA daughter card, which in
turn is installed on a Nallatech BenONE mainboard. This particular combination is widely
known as the Xilinx XtremeDSP Kit-4.

The logic design flow starts with the Xilinx System Generator (XSG) design tool, which is a
plug-in for Matlab/Simulink. We chose XSG as our FPGA design tool, although
conventional design methods based on hardware description languages such as VHDL or
Verilog may also be attempted. The modular regular nature of the systolic-array, together
with the complicated pipelines and dataflow structure, makes the use of a graphical FPGA
design method such as XSG, easier, compared to text-based design tools. We however note
that XSG, in the end, leades to synthesizable VHDL (or Verilog), which is subsequently
processed by conventional FPGA logic synthesis tools such as the Xilinx Synthesis Tool
(XST) or Synplify Pro.

5.1 Finite Precision Arithmetic and the FPGA Circuit

The arithmetic circuits on the FPGA are obviously based on finite precision hardware blocks
for the multipliers, adders/subtractors, and memory devices. The designation of precisions
(word sizes) is an important design step that requires extensive further research. Our
example is based on experience with many similar circuits, and is largely a result of
experiential learning accumulated over several years of research on similar systolic-arrays.
At this time, a comprehensive design method that can lead to optimal finite precision levels
(in terms of hardware resource consumption, quantization noise statistics, power
consumption, and throughput) is not available, and is an interesting subject for research
activities.

The following example assumes input signals obtaining from 4-bit A/D converters.
Preliminary studies show that 3 bit A/D converters are quite sufficient for ultra-wideband
wireless communications applications. Our choice of 4-bits in our A/D converters results
from 1-bit overdesign, mainly as a margin of safety, in order to ensure good performance
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 265

Fig. 9. Xilinx FPGA circuit for a hybrid-form PPCM having 12 cycles of pipeline latency
(corresponding to the signal-flow graph in Fig. 8).

Fig. 10. First 4 PPCMs of a hybrid-form systolic-array FPGA circuit showing inter-PPCM
interconnections. The FPGA circuit is tested on-chip using stepped hardware co-simulation
using a 2D unit impulse input at PPCM #1, with inputs of PPCM #2, #3, …, #21, set to zero,
leading to the 2D measured impulse response h(n1x, n2 cTCLK ) . A bit-true cycle-accurate
FPGA circuit simulation of the 2D impulse response is available in the Matlab variables
simout,simout1,…,simout20, and the measured on-chip FPGA circuit response are available in
Matlab variables h0,h1,…,h20.
266 VLSI

from a real-world application. The multiplier coefficients are assumed to be 12 bits, with the
binary point at position 10. All other registers, including quantized outputs from multipliers
and adder/subtractor blocks, are fixed at 14-bits, with binary point assumed at position 10.
The design of a PPCM is shown in Fig. 9, followed by the systolic-array, in Fig. 10. The finite
precision values at various locations on the PPCM signal flow graph can be widely
optimized against various requirements, but is not attempted here, because we are only
interested in giving our readers a basic design overview of the hybrid-form systolic-array
processor.

The FPGA circuit was tested, using on-chip hardware-in-the-loop co-simulation, using
Matlab/Simulink, XSG, and FUSE, using the XtremeDSP Kit-4 device, which was installed
on the 5V 32-bit PCI slot of the host PC. Figure 11 shows the measured 2D magnitude
frequency response of an example beam filter having spatial DOA o  25o , and bandwidth
parameter R  0.02, computed for 21 spatial samples, and 256 time samples, of the impulse
response. The “uneven” nature of the measured response is attributed to quantization
effects, and magnitude sensitivity, for which a comprehensive study remains as useful
future research.

Fig. 11. Measured 2D magnitude response,   1  and  0.5  2  0.5 , obtained

from a 21 PPCM FPGA implementation using on-chip hardware co-simulation using the
Xilinx XtremeDSP Kit-4. Quantization effects cause the implementation to have extra ripples
in both pass- and stop-bands, and lead to addition of AGWN. A detailed study of
quantization effects remain for future work.
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 267

5.2 High-speed Implementation Technologies

At present, systolic-array implementations of the proposed 2D IIR beam plane-wave filters
have been limited to proof-of-concept realizations on FPGA circuits that operate at clock
rates of up to 100 MHz. However, real-world electromagnetic applications requires frame
rates in excess of 1 GHz and can be high as 21 GHz for full-band UWB radio systems. FPGA
circuit implementations are often impractical for product applications and are mostly used
as prototypes for eventual implementation using application-specific integrated circuits
(ASICs) using high-speed VLSI platforms such as the state-of-the-art 40nm digital CMOS
process. Porting the available FPGA designs to 40nm CMOS (or similar) VLSI technology
remains an exciting field for future research.

5.3 Field-Programmable Object Arrays (FPOAs) and Asynchronous FPGA Circuits

In general, although FPGAs from vendors such as Xilinx and Altera are limited to speeds
less than  300 MHz for most recursive filter designs, future developments may facilitate the
use of conventional FPGA technology to implement the proposed systolic arrays at 1 GHz or
higher. Furthermore, it should be noted that fab-less semiconductor technologies, such as
MathStar’s (http://www.mathstar.com/) Arrix field programmable object array (FPOA) devices
(Anonymous,2007) and high-speed “picoPIPE” FPGAs capable of 1.5 GHz operation
(Anonymous,2008) from Achronix Semiconductor (http://www.achronix.com/), are emerging as
an alternative to conventional ASIC solutions, and forms a basis for future research.

FPOAs, from MathStar, consists of an array of hard IP blocks such as arithmetic-and-logic

units (ALUs) and multiply-accumulate (MACs) blocks, within a reconfigurable switching
fabric, which are ideal for systolic-array realizations due to their modularity, regularity, and
local interconnectivity. These “objects” are pre-fabricated onto the FPOA structure, and
meet stringent timing standards, enabling deterministic design at 1 GHz which are
independent of the logic being implemented. On the other hand, Speedster FPGAs from
Achronix Semiconductor, employ asynchronous handshaking protocols between
combinational logic blocks, which they describe as the merging of clock and data tokens into
one signal, which in turn, according to Achronix documentation, enables faster operation
compared to conventional synchronous FPGA architectures. The Speedster family boasts 1.5
GHz, and is potential candidate for real-time implementation of the systolic-array
architectures described herein, following future research.

6. Conclusions
The above new systolic implementation of a 2D IIR frequency-beam filter transfer function
has promising engineering applications for the directional enhancement of a propagating
broadband space-time plane-wave received on an array of sensors. A particularly important
case is the use of an array of broadband antennas for the directional enhancement (that is,
beamforming) of ultra-wideband electromagnetic plane-waves.

A massively-parallel systolic-array custom architecture, that is capable of processing one

linear frame per clock cycle (OFPCC) with detailed design and optimization information,
has been described. The architecture is based on the recently proposed hybrid-form 2D
signal flow graph, which has been shown to be optimal in terms of critical path delay (hence
268 VLSI

maximum throughput, because at OFPCC, the clock rate is equal to the frame rate in these
architectures) and low computational complexity.

A design example for the proposed systolic-array processor architecture has been described
using a Xilinx Virtex-4 Sx35 FPGA device, and the Matlab/Simulink based FPGA design
tool called Xilinx System Generator. The example FPGA implementation of the 2D IIR
frequency- beam filter was tested on-chip using the hardware-in-the-loop verification
method called ‘hardware co-simulation’, and the on-chip 2D unit-impulse response was
measured, which in turn led to measured 2D frequency response results that confirm correct
implementation of the hardware.

Although the FPGA-based example is generally too slow for microwave imaging
applications, it serves as a validation of the proposed OFPCC systolic-array processor and
can be used in its current form for slower applications in audio, ultra-sound, and lower
radio frequencies (of up to approximately 100 MHz). Finally, promising new VLSI
implementation platforms are described here , which may eventually enable the proposed
architecture to operate at the required multi-GHz clock frequency to enable real-time ultra-
wideband digital smart antenna array applications.

7. References
Agathoklis, P. and L. T. Bruton (1983). "Practical-BIBO stability of N-dimensional discrete
systems." Proc. Inst. Elec. Eng. 130, Pt. G(6): 236-242.
Anonymous (2007). Arrix FPOA Overview. Available online at http://www.mathstar.com.
Anonymous (2008). Using High-Performance FPGAs for Advanced Radio Signal Processing.
. Available online at http://www.achronix.com.
Arnold Van Ardenne. The Technology Challenges for the Next Generation Radio
Telescopes. Perspectives on Radio Astronomy - Technologies for Large Antenna
Arrays, Netherlands Foundation for Research in Astronomy.
Bertschmann, R. K., N. R. Bartley and L. T. Bruton A 3-D integrator-differentiator double-
loop (IDD) filter for raster-scan video processing. IEEE Intl. Symp. on Circuits and
Systems, ISCAS'95.
Bolle, M. (1994). A Closed-form Design Method for 3-D Recursive Cone Filters IEEE
International Conference on Acoustics, Speech, and Signal Processing, ICASSP.
Bruton, L. T. (2003). "Three-dimensional cone filter banks." IEEE Trans. on Circuits and
Systems I: Fundamental Theory and Applications 50(2): 208-216.
Bruton, L. T. and N. R. Bartley (1985). "Three-dimensional image processing using the
concept of network resonance." IEEE Trans. on Circuits and Systems 32(7): 664-672.
Dansereau, D. (2003). 4D Light Field Processing and its Application to Computer Vision.
Electrical and Computer Engineering. Calgary, University of Calgary. MSc.
Dansereau, D. and L. T. Bruton (2007). "A 4-D Dual-Fan Filter Bank for Depth Filtering in
Light Fields." Signal Processing, IEEE Transactions on 55(2): 542-549.
Devlin, M. (Spring 2003) "How to Make Smart Antenna Arrays." Xcell Journal Online
Do-Hong, T. and P. Russer (2004). Signal Processing for Wideband Smart Antenna Array
Applications. IEEE Microwave Magazine. 5.
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 269

Dudgeon, D. E. and R. M. Mersereau (1990). Multidimensional Digital Signal Processing.

Englewood Cliffs, N.J. 07632, Prentice-Hall.
Ellingson, S. W. (1999). A DSP Engine for a 64-Element Array. Proceedings on Perspectives
for Radio Astronomy-- Technologies for Large Antenna Arrays, Netherlands.
Frederick, J. D., Y. Wang and T. Itoh (2002). "A smart antenna receiver array using a aingle
RF channel and digital beamforming." IEEE Trans. on Microwave Theory and
Techniques 50(12): 3052-3058.
Ghavami, M., L. B. Michael and R. Kohno Ultra wideband signals and systems in
communication engineering, John Wiley and Sons., Inc.
Gunaratne, T. K. and L. T. Bruton "Beamforming of Broadband-bandpass Plane Waves using
Polyphase 2D FIR Trapezoidal Filters." IEEE Trans. on. Circuits and Systems:
Regular Papers 55(3): 838-850.
Hum, S. V., H. L. P. A. Madanayake and L. T. Bruton "UWB Beamforming using 2D Beam
Digital Filters." IEEE Trans. on Antennas and Propagation 57(3): 804-807.
Hum, S. V., M. Okoniewski and R. J. Davies "Modeling and Design of Electronically Tunable
Reflectarrays." IEEE Trans. on Antennas and Propagation 55(8): 2200-2210.
Huseyin Arslan, Zhi-Ning Chen and Maria-Gabriella Di Benedetto (2006). Ultra-wideband
Wireless Communication, Wiley Interscience.
J. Roderick, H. Krishnaswamy, K.Newton, et al. (2006). "Silicon-based ultra-wideband beam-
forming." IEEE Journal of Solid-State Circuits 41(8): 1726-1739.
Johnson, D. H. and D. E. Dudgeon (1993). Array Signal Processing-Concepts and
Techniques. Englewood Cliffs, N.J. 07632, Prentice-Hall.
Khademi, L. Reducing the computational complexity of FIR 2D fan and 3D cone filters [MSc
Thesis]. Electrical and Computer Engineering, University of Calgary.
Khademi, L. and L. T. Bruton On the limitations of narrow 2D fan filters speech processing.
IEEE 2003 Pacific Rim Conference on Communications, Computers, and Signal
Processing (PACRIM'03).
Kuenzle, B. and L. T. Bruton "3-D IIR filtering using decimated DFT-polyphase filter bank
structures." IEEE Trans. on Circuits and Systems I: Regular Papers 53(2): 394-408.
Kuenzle, B. and L. T. Bruton (2005). A novel low-complexity spatio-temporal ultra wide-
angle polyphase cone filter bank applied to sub-pixel motion discrimination. IEEE
Intl. Symp. on Circuits and Systems, ISCAS'05, Kobe, Japan.
Kung, S. Y. (1988a). VLSI Array Processors, Prentice-Hall, Englewood Cliffs, N.J.
Kung, S. Y. (1988b). VLSI Array Processors: Designs and Applications. 1988 IEEE
International Symp. on Circuits and Systems, ISCAS'88.
Liberti Jr., J. C. and T. S. Rappaport (1999). Smart Antennas for Wireless Communications-
IS-95 and Third Generation CDMA Applications. Upper Saddle River, N.J. 07632,
Prentice-Hall.
Litva, J. and T. K.-Y. Lo (1996). Digital Beamforming in Wireless Communications, Artech
House.
Liu, Q. and L. T. Bruton (1989). "Design of 3-D planar and beam recursive digital filters
using spectral transformation." IEEE Trans. Circuits and Systems 36(3): 365-374.
Madanayake, A. (2004). FPGA Architectures for 2D/3D Digital Filters. Electrical and
Computer Engineering. Calgary, University of Calgary. MSc: 205.
270 VLSI

Madanayake, A. (2008). Real-time FPGA Architectures for Space-time Frequency-planar

MDSP. Electrical and Computer Engineering. Calgary, University of Calgary. PhD:
371.
Madanayake, A. and L. Bruton A Review of 2D/3D IIR Plane-wave Real-time Digital Filter
Circuits. IEEE Canadian Conference on Electrical and Computer Engineering,
CCECE'05, Saskatoon, Sasketchawan, Canada.
Madanayake, A., L. Bruton and C. Comis FPGA architectures for real-time 2D/3D FIR/IIR
plane wave filters. IEEE Intl. Symp. on Circuits and Systems, ISCAS'04.
Madanayake, A. and L. T. Bruton "A Speed-optimized systolic-array processor architecture
for spatio-temporal 2D IIR broadband beam filters." IEEE Trans. on. Circuits and
Systems: Regular Papers 55(7): 1953 - 1966.
Madanayake, H. L. P. A. and L. T. Bruton "A Systolic-array Architecture for First-Order 3D
IIR Frequency-planar Filters." IEEE Trans. Circuits and Systems: Regular Papers
55(6): 1546-1559.
Madanayake, H. L. P. A. and L. T. Bruton (2007). "Low-complexity distributed-parallel-
processor for 2D IIR broadband beam plane-wave filters." Canadian Journal of
Electrical and Computer Engineering (CJECE) 32(3): 123-131.
Madanayake, H. L. P. A. and L. T. Bruton (2008). A Real-time Systolic Array Processor
Implementation of Two-dimensional IIR Filters for Radio-frequency Smart Antenna
Applications. IEEE Intl. Symp. on Circuits and Systems (ISCAS'08), Seattle.
Madanayake, H. L. P. A., S. V. Hum and L. T. Bruton "A Systolic Array 2D IIR Broadband
RF Beamformer." IEEE Trans. on Circuits and Systems-II: Express Briefs 55(12):
1244-1248.
Madanayake, H. L. P. A., S. V. Hum and L. T. Bruton UWB Beamforming Using Digital 2D
Frequency-planar Filters. IEEE 2008 Antenna and Propagation Society
Symposium/URSI Symposium, San Diego.
Parhi, K. K. "Finite word effects in pipelined recursive filters." IEEE Trans. on Signal
Processing, IEEE Trans. on Acoustics, Speech, and Signal Processing. 39(6): 1450-
1454.
Parhi, K. K. "Pipelining in algorithms with quantizer loops." IEEE Trans. on Circuits and
Systems 38(7): 745-754.
Parhi, K. K. (1991). "Finite word effects in pipelined recursive filters." IEEE Trans. on Signal
Processing [see also IEEE Trans. on Acoustics, Speech, and Signal Processing] 39(6):
1450-1454.
Parhi, K. K. (1999). VLSI Digital Signal Processing Systems: Design and Implementation,
John Wiley and Sons.
Parhi, K. K. and D. Messerschmitt Look-ahead computation: Improving iteration bound in
linear recursions. IEEE Intl. Conf. on Acoustics, Speech and Signal Processing,
ICASSP'87.
Parhi, K. K. and D. G. Messerschmitt Pipelined VLSI Recursive Filter Architectures using
Scattered Look-Ahead and Decomposition. 1988 IEEE Intl. Conf. on Acoustics,
Speech and Signal Processing, ICASSP88, New York, N.Y., USA.
Parhi, K. K. and D. G. Messerschmitt (1989). "Concurrent architectures for two-dimensional
recursive digital filtering." IEEE Trans. on Circuits and Systems 36(6): 813-829.
Rader, C. M. (1996). VLSI Systolic Arrays for Adaptive Nulling. IEEE Signal Processing
Magazine. 13: 29-49.
Radio-Frequency (RF) Beamforming Using Systolic FPGA-based
Two Dimensional (2D) IIR Space-time Filters 271

Ramamoorthy, P. A. and L. T. Bruton "Design of stable two-dimensional analog and digital

filters with applications in image processing." Int. J. Circuit Theory Appl. 7: 229-
245.
Rodenbeck, C. T., K. Sang-Gyu, T. Wen-Hua, et al. (2005). "Ultra-wideband low-cost phased-
array radars." Microwave Theory and Techniques, IEEE Transactions on 53(12):
3697-3703.
Schroeder, H. and H. Blume (2000). One- and Multidimensional Signal Processing-
Algorithms and Applications in Image Processing, John Wiley and Sons, Ltd.
Shanbhag, N. R. (1991). "An improved systolic architecture for 2-D digital filters." Signal
Processing, IEEE Transactions on [see also Acoustics, Speech, and Signal
Processing, IEEE Transactions on] 39(5): 1195-1202.
Sid-Ahmed, M. A. "A systolic realization for 2-D digital filters." IEEE Trans. on Acoustics,
Speech, and Signal Processing 37(4): 560-565.
Silva, R., R. Worrel and A. Brown Reprogrammable, Digital Beam Steering GPS Receiver
Technology for Enhanced Space Vehicle Operations. Core Technologies for Space
Systems Conference, Colorado Springs, CO.
Staderini, E. M. (2002). "UWB radars in medicine." Aerospace and Electronic Systems
Magazine, IEEE 17(1): 13-18.
Sundarajan, V. and K. K. Parhi Synthesis of folded multidimensional DSP systems. IEEE
Intl. Symp. on Circuits and Systems (ISCAS'98).
Van Ardenne, A. (2000). Concepts of the Square Kilometre Array; toward the new
generation radio telescopes. IEEE 2000 Intl. Symp. on Antennas and Propagation.
Weem, J. P., B. M. Noratos and Z. Popovic (1999). Broadband Array Considerations for SKA.
Proceedings on Perspectives for Radio Astronomy-- Technologies for Large
Antenna Arrays.
Zajc, M., R. Sernec and J. Tasic (2000). Array processors for DSP: implementation
considerations. 10th Mediterranean Electrotechnical Conference, 2000, MELECON
2000.
272 VLSI
A VLSI Architecture for Output Probability Computations of HMM-based Recognition Systems 273

13
X

A VLSI Architecture for Output Probability

Computations of HMM-based
Recognition Systems
Kazuhiro Nakamura, Masatoshi Yamamoto,
Kazuyoshi Takagi and Naofumi Takagi
Nagoya University
Japan

1. Introduction
Mobile embedded systems with natural human interfaces, such as speech recognition, lip
reading, and gesture recognition, are required for the realization of future ubiquitous
computing. Recognition tasks can be implemented either on processors (CPUs and DSPs) or
dedicated hardware (ASICs). Although processor-based approaches offer flexibility, real-
time recognition tasks using state-of-the-art recognition algorithms exceed the performance
level of current embedded processors, and require modern high-performance processors
that consume far more power than dedicated hardware. Dedicated hardware, which is
optimized for low-power, real-time recognition tasks, is more suitable for implementing
natural human interfaces in low power mobile embedded systems. VLSI architectures
optimized for recognition tasks with low power dissipation have been developed.
Yoshizawa et al. investigated a block-wise parallel processing method for output probability
computations of continuous hidden Markof models (HMMs), and proposed a low power,
high-speed VLSI architecture. Output probability computations are the most time-
consuming part of HMM-based recognition systems. Mathew et al. developed low-power
accelerators for the SPHINX 3 speech recognition system, and also developed perception
accelerators for embedded systems. In this chapter, we present a fast and memory efficient
VLSI architecture for output probability computations of continuous HMMs using a new
block-wise parallel processing method. We show block-wise frame parallel processing
(BFPP) for output probability computations and present an appropriate VLSI architecture
for its implementation. Compared with a conventional block-wise state parallel processing
(BSPP) architecture, when there are a sufficient number of HMM states for accurate
recognition, the BFPP architecture requires fewer registers and processing elements (PEs),
and less processing time. The PEs used in the BFPP architecture are identical to those used
in the BSPP architecture. From a VLSI architectural viewpoint, a comparison shows the
efficiency of the BFPP architecture through efficient use of registers for storing input feature
vectors and intermediate results during computation. The remainder of this chapter is
organized as follows: the structure of HMM based recognition systems is described in
274 VLSI

voice HMM parameters words

Feature vectors
etc. (stored in ROM) etc.
(stored in RAM)

exclusive access Bus

Output probability computation circuit

PEs Register arrays

PE
………
Output probabilities of HMMs
Viterbi scorer /search algorithm

Likelihood scores
Fig. 1. Basic structure of HMM-based recognition hardware

Section 2, BFPP and BFPP-based VLSI architecture are introduced in Section 3, the
evaluation of the BFPP architecture is described in Section 4, and conclusions are presented
in Section 5.

2. HMM-based Recognition Systems

2.1 HMM-based Recognition Hardware
Due to their effectiveness and efficiency for user-independent recognition, HMMs are
widely used in applications such as speech recognition, lip-reading, and gesture recognition.
Figure 1 shows the basic structure of HMM-based recognition hardware (Yoshizawa et al.,
2006, Yoshizawa et al., 2004, Yoshizawa et al., 2002, Mathew et al., 2003a). The output
probability computation circuit and Viterbi scorer work together as a recognition engine.
The inputs to the output probability computation circuit are feature vectors of several
dimensions and model parameters of HMMs. These values are stored in RAM and ROM
respectively. The RAM, ROM and output probability computation circuit interconnect via a
single bus, and memory accesses are exclusive. The output probability computation circuit
outputs the results of the output probability computation of HMMs. The Viterbi scorer
outputs likelihood score using the Viterbi algorithm. In HMM-based recognition systems,
the most time-consuming task is output probability computations, and the output
probability computation circuit accelerates these computations. The output probability
computation circuit has several register arrays and processing elements (PEs) for efficient
high-speed parallel processing.

2.2 Output Probability Computation of HMMs

Ret O1, O2, ..., OT be a sequence of P-dimensional input feature vectors (input frames) to
HMMs, where Ot = (ot1, ot2, ..., otP), 1  t  T. T is the number of input feature vectors, and P is
the dimension of the input feature vector. For an input frame Ot, the output probability of
N-state left-to-right continuous HMM at the j-th state is given by
A VLSI Architecture for Output Probability Computations of HMM-based Recognition Systems 275

v=0
v=v+1 Loop D
t=0
t=t+1 Loop C
j=0
Loop B
j=j+1
p=0
p=p+1 Loop A

Partial computation of logbj (Ot)

no
Pp
yes
no
Nj
yes
no
Tt
yes
no
Vv
yes
Fig. 2. Flowchart of output probability computation

P
log b j Ot    j    jp otp   jp  , 1  j  N , 1  t  T ,
2
(1)
p 1

where j, jp and jp are the factors of the Gaussian probability density function (Yoshizawa
et al., 2006).
The output probability computation circuit computes log b j(Ot) based on Eq. (1), where all
HMM parameters j, jp and jp are stored in ROM, and the input frames are stored in RAM.
The values of T, N, P and the number of HMMs V differ for each recognition system. For a
recent isolated word recognition system (Yoshizawa et al., 2006, Yoshizawa et al., 2004), T, N,
P and V are 86, 32, 38, and 800, respectively, and for another word recognition system
(Yoshizawa et al., 2002), T, N, P and V are 89, 12, 16 and 100 respectively. For a continuous
speech recognition system (Mathew et al., 2003a), T, N, P and V are approximately 20, 10, 40,
and 50, respectively. Different applications require different output probability computation
circuit architectures. A flowchart of output probability computations for V HMMs is shown
in Fig. 2. Output probabilities are obtained by T  N  P  V times the partial computation of
log b j(Ot). Partial computation of log b j(Ot) performs four arithmetic operations, a
subtraction (a = otp  jp), an addition (acc = acc + b, where the initial value of acc is j), and
two multiplications (b = a  a  jp) for Eq. (1), and computes log b j(Ot).

3. Fast and memory efficient VLSI architecture

3.1 Block-wise frame parallel processing
Block parallel processing (BPP) for output probability computations was proposed as an
efficient parallel processing method for word HMM-based speech recognition by Yoshizawa
276 VLSI

v=0
v = v + 1 Loop D
t=0
Loop C
t=t+1
p=0
Loop A
p=p+1
otp
PE1 PE2 PEN
logb1 (Ot) logb2 (Ot) . . . . . . . . . . . . . . . . . . . . . logbN (Ot)
N-parallel computation with N PEs

Pp no
yes
no
Tt
yes
no
Vv
yes

Fig. 3. Flowchart of output probability computation using BSPP

et al. (Yoshizawa et al., 2006, Yoshizawa et al., 2004, Yoshizawa et al., 2002). In this method,
the set of input frames is called a block, and HMM parameters are effectively shared between
different input frames in the computation. N-parallel computation is performed by their BPP.
In this chapter, we classify two types of BPP according to data flow of output probability
computations: block-wise frame parallel processing (BFPP) and block-wise state parallel
processing (BSPP). A block can be seen as a set of M (≤ T) input frames, whose elements are
Ot’’s, 1 ≤ t’ ≤ M. M frames in T input frames are processed in block. BFPP performs
arithmetic operations to locally stored input frames, which are O1, O2, ..., OM, and output
probability computations for multiple frames are carried out simultaneously. On the other
hand, a block can also be seen as a M  P matrix whose elements are ot’p, 1 ≤ t’ ≤ M, 1 ≤ p ≤ P.
BSPP performs arithmetic operations to an input sequence, which is o11, ..., o1P, o21, ..., o2P, ...,
oM1, ..., oMP, and output probability computations for multiple states are carried out
simultaneously.
The BPP proposed by Yoshizawa et al. (Yoshizawa et al., 2006, Yoshizawa et al., 2004,
Yoshizawa et al., 2002) is classified as a BSPP. In this chapter, we present BFPP for output
probability computations. M/2-parallel computations are performed by our BFPP.
A flowchart of the output probability computations with the conventional BSPP (Yoshizawa
et al., 2006, Yoshizawa et al., 2004, Yoshizawa et al., 2002) is shown in Fig. 3.
PEi represents the i-th processing element, which computes log b i(Ot) by a subtraction, an
addition, and two multiplications for Eq. (1). Loop B (Fig. 2) is expanded as shown in Fig. 3,
and log b 1(Ot), log b 2(Ot), ..., and log b N(Ot) are computed simultaneously with N PEs, where
otp is fed to the N PEs in Loop A. In addition to the N-state parallel computation, the same
HMM parameters jp’s and jp’s, and j’s, 1 ≤ j ≤ N, 1 ≤ p ≤ P, are used repeatedly during
Loop C in Fig. 3.
A flowchart of the output probability computation with BFPP is shown in Fig. 4. The PEs in
A VLSI Architecture for Output Probability Computations of HMM-based Recognition Systems 277

v’max = 0
v’max = v’max + L Loop D2
t’max = 0
Loop C1
t’ = t’max, t’max = t’max + M, v’ = v’max  L

Loop D1
v’ = v’ + 1
j = 0, v = v’
Loop B
j=j+1
p=0
Loop A
p=p+1
jp , jp
PE1 PE2 PEM/2
logbj (Ot’+1) logbj (Ot’+2) . . . . . . . . . . . . . . . . . . logbj (Ot’+M/2)

logbj (Ot’+M/2+1) logbj (Ot’+M/2+2) . . . . . . . . . . . . . . . . logbj (Ot’+M)

double M/2-parallel computation
with M/2 PEs

Pp no
yes
no
Nj
yes
no
v’max  v’
yes
no
T  t’max
yes
no
V  v’max
yes
Fig. 4. Flowchart of output probability computation using BFPP

Figs. 4 and 3 are identical, but in a different number. Loop C in Fig. 2 is partially expanded in
Fig. 4, and log b j(Ot’+1), log b j(Ot’+2), ..., and log b j(Ot’+M/2) are computed simultaneously with
M/2 PEs in Loop C1, where jp and jp are fed to the M/2 PEs in Loop A. In addition to the
M/2-frame parallel computations, log b j(Ot’+M/2+1), log b j(Ot’+M/2+2), ..., and log b j(Ot’+M) are
also computed with the same M/2 PEs. In this double M/2-parallel computation, the same
HMM parameters jp and jp are used twice, because the parameters are independent of t. In
addition to the M/2-parallel computations, Loop D (Fig. 2) is divided into Loops D1 and D2
(Fig. 4). The same input frames Ot’+1, Ot’+2, ..., and Ot’+M are used repeatedly during Loop D1,
because the input frames are independent of v.
278 VLSI

3.2 A VLSI architecture for output probability computation

Our BFPP VLSI architecture for output probability computations is shown in Fig. 5. The
architecture consists of five register arrays and M/2 PEs. RegO stores M input frames Ot’+1,
Ot’+2, ..., Ot’+M. Reg and Reg store HMM parameters jp, and jp, respectively. Reg stores

ROM RAM
(, , ) (O)

Output probability computation circuit

Reg Reg
Reg Reg
RegO P
Ot’+1 … PE1    
Ot’+2 …
otp M/2 PE2 M/2 M/2
   

…
…

…
PEM/21    
Ot’+M/21 …
Ot’+M/2 …
PEM/2    
M
Ot’+M/2+1 …
… M/2-parallel computation
Ot’+M/2+2
M/2 M/2
…

…
M/2
…

Ot’+M1 …
Ot’+M …

Fig. 5. BFPP VLSI architecture

HMM parameter j and intermediate results. Reg stores computed output probabilities for
a Viterbi scorer. Each PEi consists of two adders and two multipliers, which are used for
  jp ot ' p   jp 2 .
P
computing  j 
p 1

Figure 6 shows the flowchart of output probability computations using the BFPP
architecture. The computation starts by reading M input frames from RAM and storing
them to RegO in Loop C1, which are Ot’+1, Ot’+2, ..., Ot’+M/2, Ot’+M/2+1, Ot’ +M/2+2, ..., Ot’+M. The
HMM parameters of v-th HMM are read from ROM and stored in Reg, Reg and Reg,
which are 11, 11, and 1. The value of all registers in Reg is set to 1. For the first half of
the stored input frames O t’+1 , O t’+2 , ..., and O t’+M/2 , M/2 intermediate results are
simultaneously computed with the stored 11, 11, and 1 by M/2 PEs, where the HMM
parameters are shared by all PEs. At the same time, an HMM parameter j p+1 of v-th HMM
A VLSI Architecture for Output Probability Computations of HMM-based Recognition Systems 279

is read from ROM and stored in Reg. Then, for the other half of the stored input frames
Ot’+M/2+1, Ot’+M/2+2, ..., and Ot’+M, M/2 intermediate results are simultaneously computed with
the same 11, 11, and 1 by M/2 PEs. At the same time, an HMM parameter j p+1 of v-th
HMM is read from ROM and stored in Reg. In this double M/2-parallel computation, the

v’max = 0
v’max = v’max + L Loop D2
t’max = 0
Loop C1
t’ = t’max, t’max = t’max + M, v’ = v’max  L
v = v’+1, j = 1, p = 1
Load Ot to RegO (t=t’+1, t=t’+2, …, t=t’+M, MP cycles)
Load 11, 11 to Reg, Reg, respectively (2 cycles)

Loop D1
v’ = v’ + 1
j = 0, v = v’
Loop B
j=j+1

Load j to Reg (1 cycle)

p=0
Loop A
p=p+1
jp , jp
PE1 PE2 PEM/2
Load j p+1 to Reg logbj (Ot’+1) logbj (Ot’+2) . . . . . . . . . . . . . . . . . . logbj (Ot’+M/2)

Load j p+1 to Reg logbj (Ot’+M/2+1) logbj (Ot’+M/2+2) . . . . . . . . . . . . . . . . logbj (Ot’+M)

double M/2-parallel computation
with M/2 PEs

Pp no
yes
Copy Reg to Reg
no
Nj
yes
no
v’max  v’
yes
no
T  t’max
yes
no
V  v’max
yes
Fig. 6. Flowchart of computations using the BFPP architecture

same HMM parameters 11, 11, and 1 are used twice. In the next double M/2-parallel
computation, the stored HMM parameters  j p+1 and j p+1 are used twice. M output
probabilities log b j(Ot’+1), log b j(Ot’+2), ..., and log b j(Ot’+M) of v-th HMM are obtained by Loop
280 VLSI

A. The obtained results are transfered from Reg to Reg for starting the next output
probability computation, log b j+1(Ot’+1), log b j+1(Ot’+2), ..., log b j+1(Ot’+M) of v-th HMM. The
stored results are fed to the Viterbi scorer. The MN output probabilities of v-th HMM are
obtained by Loop B. MNL output probabilities of HMM v’  1, v’  2, ..., v’  L are obtained
by Loop D1 with the same M input frames Ot’+1, Ot’+2, ..., and frames Ot’+1, Ot’+2, ..., and Ot’+M.

ROM RAM
(, , ) (O)

Output probability computation circuit

N Reg N Reg
… …
… …
P P
…

… ……
… …
Reg

PE1    

PE2    
N
…
…

otp
PEN1    

PEN    

N-parallel computation

Fig. 7. BSPP VLSI architecture

frames Ot’+1, Ot’+2, ..., and Ot’+M. The MNL(T/M) output probabilities of HMM v’  1, v’ 
2, ..., v ’  L are obtained by Loop C1, and finally the MNL(T/M)(V/L) output probabilities
of all HMMs are obtained by Loop D2.

4. Evaluation
We compared the proposed BFPP with BSPP (Fig. 7) VLSI architecture (Yoshizawa et al.,
2006, Yoshizawa et al., 2004, Yoshizawa et al., 2002). The BSPP architecture consists of three
register arrays and N PEs. Reg and Reg store HMM parameters jp and jp, respectively,
and Reg stores HMM parameter j and intermediate results. The PEs in Figs. 7 and 5 are
identical.
A VLSI Architecture for Output Probability Computations of HMM-based Recognition Systems 281

Figure 8 shows the flowchart of the computations of BSPP architecture. The computation
starts by reading all 2NP  N HMM parameters of v-th HMM from ROM and storing them
to Reg, Reg, and Reg in Loop D. For input otp, the intermediate results are computed with
stored HMM parameters by N PEs. N output probabilities log b 1(Ot), log b 2(Ot), ..., log b N(Ot)
of the HMM are obtained by Loop A. The obtained results are fed to a Viterbi scorer. NT

v=0
v = v + 1 Loop D

Load jp and jp of HMM v to Reg and Reg

(j = 1, 2, …, N, p = 1, 2, …, P, 2NP cycles)

Load j to Reg (j = 1, 2, …, N, N cycles)

t=0
t = t + 1 Loop C
p=0
p = p + 1 Loop A

PE1 PE2 PEN

Load otp logb1 (Ot) logb2 (Ot) . . . . . . . . . . . . . . . . . . . . . logbN (Ot)
N-parallel computation with N PEs

Pp no
yes
no
Tt
yes
no
Vv
yes

Fig. 8. Flowchart of computations using the BSPP architecture

Register size (bit)

BFPP (ours) PMxo  2x  x  2Mxf
BSPP NPx  NPx  Nxf
Table 1. Register size

Processing time (cycles)

BFPP (ours) V/L{PM  (1  2P)LN}T/M
BSPP V(2NP  N  PT)
Table 2. Processing time

output probabilities of v-th HMM are obtained by Loop C with the same HMM parameters.
The NTV output probabilities of all HMMs are obtained by Loop D.
Table 1 shows the register size of the BSPP and BFPP architectures, where x, x, xo, and xf
represent the bit length of jp, jp, otp, and the output of PE, respectively. N, P, and M are the
282 VLSI

number of HMM states, the dimension of input feature vector (frame), and the number of
input frames in a block, respectively.
Table 2 shows the processing time for computing output probabilities of V HMMs with the
BFPP and BSPP architectures, where T and L are the number of input frames and the
number of HMMs whose output probabilities are computed with the same input frames
during Loop D1 of Fig. 6, respectively.

Register size (bit) Processing time (cycles) #PEs

BFPP (ours) 15,512 4,477,440 22
BSPP 20,224 4,585,600 32
Table 3. Evaluation of the BSPP and BFPP performance

#PEs (BFPP)
6
7
8
9
15
#PEs=32 22 #PEs=32 (BSPP)

#PEs=43

44

Fig. 9. Evaluation of the BSPP and BFPP performance, and the value of M of the BFPP (N =
32, P = 38, T = 86, V = 800)

Table 3 shows the register size, the processing time, and the number of PEs for computing
output probabilities of 800 HMMs, where we assume that N = 32, P = 38, T = 86, x = 8, x = 8,
xo = 8, xf = 24, and V = 800, the same values used in a recent circuit design for isolated word
recognition (Yoshizawa et al., 2006, Yoshizawa et al., 2004). We also assume that M = 44 and
L = 5 for the BFPP architecture. The PEs used in the BSPP and BFPP architectures are
identical. Compared with the BSPP architecture, the BFPP architecture has fewer registers,
requires less processing time, and has fewer PEs. From the VLSI architecture viewpoint, this
is because the register size of the BFPP architecture is independent of N, and its PEs can
repeatedly use the same input frames. The BFPP architecture has fewer wait cycles for
A VLSI Architecture for Output Probability Computations of HMM-based Recognition Systems 283

reading data from ROM before parallel computations, 586,240 (V/L(PM  LN)T/M),
than the BSPP architecture, which has 1,971,200 (V(2NP  N)).
Fig. 9 shows the processing time and the number of PEs of the BFPP and BSPP architectures,
and the value of M of the BFPP architecture. The processing time and the number of PEs of
the BFPP architecture are less than those of the BSPP architecture when M = 44 (Fig. 9).
From a logic design viewpoint, the register arrays of the BSPP and BFPP architectures are
designed with Flip-Flops or on-chip multi-port memories of different sizes. Data paths are
designed with identical PEs, but in a different number. The control paths of these
architectures are designed, as shown in the flowcharts Figs. 8 and 6. The data path delay is
the same for both the BSPP and BFPP designs, equal to the delay time of one PE. The delay
times of control paths differ between the two, but the control path delay is small compared
with the data path delay.

5. Conclusions
We presented BFPP for output probability computations and presented an appropriate VLSI
architecture for its implementation. BFPP performs arithmetic operations to locally stored
input frames, and output probability computations for multiple frames are carried out
simultaneously. Compared with the conventional BSPP architecture, when the number of
HMM states is large enough for accurate recognition, the BFPP architecture requires fewer
registers and PEs, and less processing time. In terms of the VLSI architecture, a fast and
memory efficient VLSI architecture for output probability computations of HMM-based
recognition systems has been presented. A logic design, a Viterbi scorer for the BFPP
architecture, and a reconfigurable architecture for both the BSPP and BFPP architectures are
our future works.

6. References
B. Mathew, A. Davis & Z. Fang (2003a). Perception Coprocessors for Embedded Systems,
Proc. of Workshop on Embedded Systems for Real-Time Multimedia (ESTIMedia), pp. 109-
116, 2003.
B. Mathew, A. Davis & Z. Fang (2003b). A Low-Power Accelerator for the SPHINX 3 Speech
Recognition System, Proc. of Int'l Conf. on Compilers, Architecture and Synthesis for
Embedded Systems, pp. 210-219, 2003.
S. Yoshizawa, Y. Miyanaga & N. Yoshida (2002). On a High-Speed HMM VLSI Module with
Block Parallel Processing, IEICE Trans. Fundamentals (Japanese Edition), Vol. J85-A,
No. 12, pp. 1440-1450, 2002.
S. Yoshizawa, N. Wada, N. Hayasaka & Y. Miyanaga (2004). Scalable Architecture for Word
HMM-Based Speech Recognition, Proc. of 2004 IEEE Int'l Symposium on Circuits and
Systems (ISCAS'04), pp. 417-420, 2004.
S. Yoshizawa, N. Wada, N. Hayasaka & Y. Miyanaga (2006). Scalable Architecture for Word
HMM-Based Speech Recognition and VLSI Implementation in Complete System,
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS, Vol. 53, No. 1, pp. 70-77,
2006.
284 VLSI

X. Huang, F. Alleva, H. W. Hon, M. Y. Hwang, K. f. Lee & R. Rosenfeld (1992). The SPHINX-
II speech recognition system: an overview, Computer Speech and Language, Vol. 7(2),
pp. 137-148, 1992.
Efficient Built-in Self-Test for Video Coding Cores: A Case Study
on Motion Estimation Computing Array 285

14
X

Efficient Built-in Self-Test for Video

Coding Cores: A Case Study on
Motion Estimation Computing Array
Chun-Lung Hsu, Yu-Sheng Huang and Chen-Kai Chen
Department of Electrical Engineering, National Dong Hwa University
Taiwan, R.O.C

1. Introduction
In more recent years, multimedia technology applications have been becoming more flexible
and powerful with the development of semiconductors, digital signal processing (DSP), and
communication technology. The latest video standard, H.264/AVC/MPEG-4 Part 10
(Advance Video Coding) (Wiegand, 2003), is regarded as the next generation video
compression standard (VCS). For video compression standards, the motion estimation
computing array (MECA) is the most computationally demanding component in a video
encoder/decoder (Kuhn, 1999; Komarek and Pirsch 1989). It is known that about 60-90% of
the total of computation time is consumed in motion estimation. Additionally, the motion
estimation algorithms used also profoundly influences the visual quality of reconstructed
images. More accurate predictions increase the compression ratio and improve peak signal-
to-noise ratio (PSNR) at a given bit-rate. Since the motion estimation algorithm is not
specified in the video coding standards that many algorithms are applied to with different
hardware platform, system frequencies, operating voltage and power dissipation.
On the other hand, due to the rapid advance in semiconductor fabrication technology, a
large number of transistors can be integrated on a single chip. However, integrating large
number of processor on a single chip results in the increase in the logic-per-pin ratio, which
drastically reduces the controllability and observability of the logic on the chip.
Consequently, testing such highly complex and dense circuits become very difficult and
expensive (Lu et al., 2005).
For a commercial chip, the VCS must introduce design-for-testability (DFT), especially in an
MECA. The objective of DFT is to increase the ease with which a device can be tested to
guarantee high system reliability. Many DFT approaches have been developed such as ad
hoc, structured, and built-in self-test (BIST) (Wu et al., 2007; Nagle et al., 1989; McCluskey,
1985; Kung et al., 1995; Touba and McCluskey, 1997). Among these DFT approaches, BIST
has an obvious advantage in that reduces the need for testing of expensive test equipment,
since the circuit/chip and its tester are implemented in the same circuit/chip. In a word, the
BIST can generate test simulations and analyze test responses without outside support,
making tests and diagnoses of digital systems quick and effective.
286 VLSI

Thus, this chapter proposed a minimal performance penalty BIST design with significantly
smaller area overhead. In normal mode, the BIST circuit does not be performed and do not
deliver the test pattern to AD-PEs for testing. Thus, each AD-PE in MECA performs its
normal operation to determine the SAD values. In testing mode, the BIST circuits, comprise
test pattern generator (TPG) and output response analyzer (ORA), are performed to testing
itself. In terms of results, after employing the presents BIST design, the circuit guarantee
100% fault coverage with low test application time at low area overhead. Moreover, the
experimental results prove the effectiveness and value of this work.

2. Motion estimation computing array

The motion estimation process in any VCS comprises a search strategy to determine the
motion vectors (MVs). The motion vector is to describe the transformation from adjacent
frames in the video sequences. On the other hand, the motion estimation also comprises a
process to determine a matching cost metric computation such as the sum of absolute
differences (SAD). In a word, the search strategy in motion estimation aims at selecting a set
of candidates and selects the one with the minimum cost metric. After selecting minimizes
the cost metric, it encodes the prediction residual between the original and motion
compensated blocks. Each residual block is transformed, quantized, and entropy coded.
Additionally, the main objective of motion estimation algorithms is to exploit the temporal
redundancy of a segmented video sequences. For example, the full-search block-matching
algorithm has been developed that introduced the best results in terms of finding motion
vectors. Such algorithms are implemented in two stages, namely the calculation of the sum
of absolute differences (SAD) for each displacement vector, followed by method for finding
the smallest SAD values. This is summarized by Eq. (1) and (2).

N -1 N -1
SAD (i, j ) =   C (k , l )
k 0 l 0
- R (i  k , j  l ) (1)

SAD MIN  min(SAD(i, j )) (2)

where C (k , l ) and R (i  k , j  l ) represent the current frame and search region's macroblock,
respectively (Pirsch et al., 1995). In the other hand, block-matching in MECA is performed
by a sequential exploration of the search region when the computations are performed in
parallel (see Fig. 1). In Fig.1 , the MECA is the parallel architecture for computing the SAD
value and its corresponding AD-PE structure are shown in Fig.1 (a) and (b), respectively.
The purpose of AD-PEs is to store the value of C (k , l ) and R (i  k , j  l ) and receive the value
of corresponding to the current position of the reference in search region. In other word, the
AD-PEs perform the processing of the subtraction and the absolute value computation.
After that, AD-PE adds the results with the partial result coming from the upper AD-PE (see
Fig. 1 (b)). The partial results are added on columns and a linear array of adders performs
the horizontal summation of the row sums, and compute SAD(i, j ) . For each position (i, j ) of
the reference block, the M-PE checks if the matching cost metric computation, SAD(i, j ) , is
smaller the previous smaller SAD value that updates the smaller SAD value to the register
which stores the previous smaller SAD value.
Efficient Built-in Self-Test for Video Coding Cores: A Case Study
on Motion Estimation Computing Array 287

16

AD-PE
AD-PE AD-PE
AD-PE
… AD-PE
AD-PE AD-PE
AD-PE

AD-PE
AD-PE AD-PE
AD-PE
… AD-PE
AD-PE AD-PE
AD-PE

…
…

…
16

AD-PE
AD-PE AD-PE
AD-PE
… AD-PE
AD-PE AD-PE
AD-PE

AD-PE
AD-PE AD-PE
AD-PE
… AD-PE
AD-PE AD-PE
AD-PE

SADMIN SAD(i,j)
M-PE
M-PE Adder
Adder Adder
Adder … Adder
Adder Adder
Adder

(a)

C R From upper
AD-PE

AD
AD

CR

Adder
Adder

SAD(i , j )

(b)
Fig. 1. The generic architecture of motion estimation and structure of AD-PE

3. The BIST design and its test strategy

Figure 2 shows the proposed BIST design, which consists of test pattern generator and
output response analyzer. The main objective of the proposed BIST design is to effectively
test the MECA by itself. Moreover, using the BIST design can drastically reduce the cost of
288 VLSI

expensive test equipment and testing time. The targeted fault, test pattern generator and
output response analyzer are explicitly described as follows.

Frame data TPG

TPG

System Clock

Circuit Under Test（CUT）

AD-PE
AD-PE AD-PE
AD-PE AD-PE
AD-PE AD-PE
AD-PE
Reset
Test_mode
AD-PE Select AD-PE
AD-PE AD-PE
AD-PE AD-PE
AD-PE AD-PE
AD-PE

AD-PE
AD-PE AD-PE
AD-PE AD-PE
AD-PE AD-PE
AD-PE

++ ++ ++ ++

ORA
ORA
SAD Value

Fig. 2. The proposed BIST design for MECA architecture

3.1 Fault model, AD-PE design and test strategy

In this chapter, the proposed BIST design aimed at the single stuck-at fault (SSAF). In
general, considerable physical faults in a circuit/chip are be discussed individually, such as
stuck-at fault, bridge fault, open fault, delay fault, crosstalk, etc.. However, the stuck-at fault
is classical and most widely fault, which covers 80-90% of possible manufacturing defect in
CMOS circuits. Thus, the targeted faults of this work are the single stuck-at faults. On the
other hand, the proposed BIST design consists of two major objectives: 1) the targeted fault
coverage is first achieved by using efficient BIST design. Meanwhile, the test pattern
generators can be implemented with linear-feedback shift register (LFSR). 2) only a few test
patterns are required to test the entire MECA architecture.

In addition, to achieve higher controllability and observability for single AD-PE, each AD-
PE uses the ripple-carry adders (RCAs) to produce the processing of the absolute difference
value computation and addition units, as shown in Fig.3. In Fig. 3, each multiplexers (mux)
are designed for testing requirement and are performed normal and test mode by using the
signal sel. The functions of the absolute difference value computation and addition are
designed by using ripple-carry adder. Besides, the frame data delivered the pixel value to
RCAs to produce the partial SAD value when the normal mode is performed. Then, the test
patterns from test pattern generator are delivered to testing the RCAs while the test mode is
selected.
Efficient Built-in Self-Test for Video Coding Cores: A Case Study
on Motion Estimation Computing Array 289

Frame
Frame TPG
TPG

Frame
Frame data
data Upper
Upper AD-PE
AD-PE

sel mux sel mux

mux

++
|X－Y|

sel
dmux
dmux

To ORA

sel mux mux sel

++

sel
dmux
dmux
AD-PE

To
To next
next AD-PE
AD-PE To
To ORA
ORA

Fig. 3. The internal architecture of AD-PE

In this chapter, the proposed BIST design includes three modes: 1) Normal mode: In this
mode, each AD-PE performs its normal function, which is to determine the minimum SAD
value. 2) Test mode: The test pattern generator delivers the test patterns to each AD-PE for
testing. Then, the testing output results are compressed by the output response analyzer for
testing analysis. 3) Analyze mode: In analyze mode, output response analyzer used the
compressed signature to determine each AD-PE in MECA architecture that are fault-free or
not.

3.2 Test pattern generator

According to the RCA structure, it was comprised by many one bit full-adder (FA). in other
word, each AD-PE consists of many one bit full-adder. Consequently, according to the
characteristic of one bit full-adder, the FA has three primary inputs. Thus, the test patterns
of FA1 has 23 different inputs that can input to one bit full-adder for testing requirement. In
a word, one bit full-adder can be obtained 100% fault coverage for single stuck-at faults by
using test patterns (000, 001, 010, 011, 100, 101, 110, 111 ). Based on the above-mentioned
concept, the Table 1 shows the proposed test patterns for each one bit full-adder in AD-PE
which can achieve 100% single stuck-at fault coverage. On the other hand, in order to realize
a simple and low-cost test pattern generator, this chapter exploits the LFSR to realize the test
pattern generator of BIST design. Figure 4 shows the proposed test pattern generator which
can deliver the test patterns to each RCA by using LFSRs. In order to effectively realize,
these test patterns can be simplified as segment patterns and generated by a LFSR. The
connection between the proposed test patterns generators (8 bits and 12 bits) and RCAs (8
290 VLSI

bits and 12 bits) are shown in Fig. 4 and 5, respectively. For example, the test patterns for 8
bits RCA are shown in Table I, which can achieve 100% single stuck-at fault coverage while
the LFSRs deliver the test patterns to RCA for testing. Furthermore, the test pattern
generator of 8 bits RCA comprises four LFSR to generate the test pattern for testing
requirement. Firstly, LFSR0 generated and delivered the test patterns to three inputs of FA0
in 8 bits RCA that the sequences of test pattern listed as follow: 111→011→001→100→010→
101→110. Then, the next stage carry is generated by itself. Secondly, the inputs of FA1 in 8
bits RCA are inputted the test pattern sequences, 00→10→01→10→11→11→01, by LFRS1.
Similarly, the LFSR2 and LFSR3 deliver the test pattern sequences, 10→11→11→01→00→10
→01 and 11→01→00→10→01→10→11, to the inputs of FA2 and FA3 in 8 bits RCA,
respectively. Besides, these inputs, an and bn, need to match up different carry as shown in
Table 1. in other words, the propagations of test pattern are generated by LFSRs as shown in
Fig. 6.

TEST Test patterns

TP1 1 1 1 0 0 1 0 1 1 0 0 1 0 1 1 0 0
TP2 0 1 1 1 0 1 1 0 1 1 0 1 1 0 1 1 0
TP3 0 0 1 0 1 1 1 0 0 0 1 1 1 0 0 0 1
TP4 1 0 0 1 0 0 1 1 0 1 0 0 1 1 0 1 0
TP5 0 1 0 1 1 0 0 0 1 1 1 0 0 0 1 1 1
TP6 1 0 1 1 1 1 0 1 0 1 1 1 0 1 0 1 1
TP7 1 1 0 0 1 0 1 1 1 0 1 0 1 1 1 0 1
Table 1. 17-bit test pattern for 8-bit RCA

LFSR0
LFSR0 LFSR1
LFSR1 LFSR2
LFSR2 LFSR3
LFSR3

TPG

a0 b0 a1 b1 a2 b2 a3 b3 a4 b4 a5 b5 a6 b6 a7 b7
Cin Cout

FA0
FA0 FA1
FA1 FA2
FA2 FA3
FA3 FA4
FA4 FA5
FA5 FA6
FA6 FA7
FA7

RCA

Sum0 Sum1 Sum2 Sum3 Sum4 Sum5 Sum6 Sum7

Fig. 4. The connection between the proposed test pattern generator and 8 bits RCA
Efficient Built-in Self-Test for Video Coding Cores: A Case Study
on Motion Estimation Computing Array 291

LFSR0
LFSR0 LFSR1
LFSR1 LFSR2
LFSR2 LFSR3
LFSR3

TPG

a0 b0 a1 b1 a2 b2 a3 b3 a4 b4 a5 b5 a6 b6 a7 b7 a8 b8 a9 b9 a10 b10 a11 b11

Cin Cout

FA0
FA0 FA1
FA1 FA2
FA2 FA3
FA3 FA4
FA4 FA5
FA5 FA6
FA6 FA7
FA7 FA8
FA8 FA9
FA9 FA10
FA10 FA11
FA11

RCA

Sum0 Sum1 Sum2 Sum3 Sum4 Sum5 Sum6 Sum7 Sum8 Sum9 Sum10 Sum11
Fig. 5. The proposed test pattern generator for 12 bits RCA

111
111 100
100

110
110 011
011 101
101 110
110

101
101 001
001 111
111 001
001

010
010 100
100 011
011 010
010

(a) LFSR 0 (b) LFSR 1

010
010 011
011

101
101 111
111 111
111 101
101

110
110 011
011 110
110 100
100

100
100 001
001 001
001 010
010

(c) LFSR 2 (d) LFSR 3

Fig. 6. The test patterns propagation of 8 bits RCA

3.3 Output response analyzer

Multi-Input Shift Register (MISR) is widely used as the signature analyzer to compact the
output response of the circuit under test (CUT). Thus, this chapter exploits the MISR to
analyze the testing result. Figure 7 shows the output response analyzer (ORA) that the ORA
of the two adders can be implemented by MISR and adding some extra logics. In test mode,
the MISR compresses the output responses and the final result of the MISR to serve as the
292 VLSI

signature which is used to check and determine whether the CUT is fault-free or not. In a
word, testing with signature analyzer, MISR, has the merits of simplicity and low cost
hardware because MIRS does not need to store the entire responses of test patterns. The
MISR of ORA has two kinds of input data which are from AD-PE output data through some
XOR gates and the test patterns (see Fig. 7). On the other hand, the results are checked by
MISR and propagated the final result to the last model.

From AD-PE output data

From TPG generate test patterns

D0 Q0 D1 Q1 D2 Q2 D6 Q6 D7 Q7

CLK

Reset

Fig. 7. The output response analyzer for 8 bits RCA

3.4 Test strategy

Based on the discussion of each sub-circuit (AD-PE, TPG and ORA), the overall working
flow of the proposed BIST design is summarized in Fig. 8, and is briefly described as follow.

Step 1: Selecting the operation modes

The operation mode, normal and test mode, are determined by test controller. In normal
mode, the CUT circuit performs its normal operation to determine the SAD values, whereas
the MECA architecture is tested by itself.

Step 2: Initial setting of BIST design

In test mode, the test patterns of BIST design are generated by TPG which was made by four
LFSRs. Then, the TPG delivered the test patterns to each AD-PE for testing.

Step 3: Analyzing the test result of AD-PE

The testing results of each AD-PE analyzed by the ORA until all AD-PE are tested and
analyzed completely.
Efficient Built-in Self-Test for Video Coding Cores: A Case Study
on Motion Estimation Computing Array 293

Start

Test Controller

No
Testing ?

Yes

• Initial TPG values

• Select AD-PE tested
• ORA on

No Yes
Have fault on AD-PE ?

Take down
Next AD-PE
the fail AD-PE

No
Test completely ?

Yes

Test done Normal mode

End

Fig. 8. The overall working flow of the proposed BIST design

4. Results discussion
The proposed BIST design was realized using Verilog HDL and synthesized Design
Compiler of Synopsys. The performance comparisons aim at area overhead, fault coverage
and test patterns discussion which are also presented here to verify the good performance of
the proposed BIST design.
The design is carried out top-down at the gate-level in the system of Quartus II by means of
waveform and the design finally passes both the unit test and the integrated test. Figure 9
shows a functionality of the waveforms for AD-PE (as you can see in section II). Figure 10
shows the functionality of test pattern generator. For test mode, initial test patterns are
scanned in, an AD-PE is tested, and test responses are scanned out. All related control
signals are generated from controller. And, the fault coverage of AD-PE reaches 100% with 7
test patterns.
294 VLSI

1 2 3 4 5 6 7 8 9 10

CLK

Pixel_x 35 255 128 49 4 107 86 240

Pixel_y 28 145 92 74 213 107 87 250

y(2’s) 228 111 164 182 43 149 169 6

AD 7 110 36 25 209 0 1 10

Upward AD-PE 0 45 17 29 175 84 204 155 113

AD-PE 0 0 127 65 200 293 204 156 123 52

Fig. 9. The logic simulation of AD-PE

1 2 3 4 5 6 7 8 9 10

CLK

Reset

LFSR0_Set 7 0 0

LFSR1_Set 1 0 0

LFSR2_Set 3 0 0

LFSR3_Set 6 0 0

TPG_out_a 0 109 183 36 218 147 254 73 109 0

TPG_out_b 0 73 109 183 36 218 147 254 73 0

Fig. 10. The waveform for test pattern generator

Comparing with previous work (Li et al., 2004), numbers of test patterns, the pin overheads,
test time and fault coverage are listed in Table 2.
By using the proposed BIST design and fault models, the single stuck-at fault coverage of
each AD-PE can achieve 100%. This perfectly proves the validity of the test patterns. And,
the area overhead of the MECA architecture including BIST is 0.2%, which is tolerable in
industry
Efficient Built-in Self-Test for Video Coding Cores: A Case Study
on Motion Estimation Computing Array 295

Proposed (Li et al., 2004)

No. of test pattern 8 30
Total test time 35us 13us
Pins overhead 25 1904
Fault coverage 100% 96%
Table 2. Comparisons

5. Conclusions
This chapter describes a BIST design for MECA architecture in video coding systems. In test
mode, test patterns are generated by TPG and scanned into each AD-PE of MECA
architecture to testing, and test responses are scanned out. All of control signals are
generated by the controller. And, experimental results show that the area overhead of the
BIST architecture for motion estimation architecture is less than 1%. The fault coverage of
each AD-PE can achieve 100%, and it perfectly proves the validity of the test patterns.
Moreover, BIST structure can easily be designed and applied to the MECA architecture.
That means the simplification in the BIST design of AD-PE in MECA architecture is
reasonable.

6. References

Abramovici, M.; Breuer, M. A. & Friedman, A. D. (1990). Digital Systems Testing and
Testable Design. Boston, MA: Computer Science Press, 1990.
Gallagher, P.; Chickermane, V.; Gregor, S. & Pierre, T. S. (2001). A Building Block BIST
Methodology for SOC Designs: A Case Study, Proceedings of International Test
Conference, PP. 111-120, Oct. 2001.
Komarek, T. & Pirsch, P. (1989). Array Architectures for Block Matching Algorithms, IEEE
Transactions on Circuits and Systems, Vol. 36, No. 2, PP. 1301–1308, Oct. 1989.
Kuhn, P. (1999). Algorithms, Complexity Analysis and VLSI Architectures for MPEG-4
Motion Estimation. New York, NY: Kluwer Academic Publishers, 1999.
Kung, C. P.; Huang, C. J. & Lin, C. S. (1995). Fast Fault Simulation for Bist Applications,”
Proceedings of the Fourth Asian Test Symposium, PP. 93–99, 1995.
Li, D.; HU, M. & Mohamed, O. (2004). Built-In Self-Test Design of Motion Estimation
Computing Array, Proceedings of IEEE Northeast Workshop on Circuits and Systems
(NEWCAS’04), June 2004.
Lu, S. K.; Shih, J. S. & Huang, S. C. (2005). Design-for-Testability and Fault-Tolerant
Techniques for FFT Processors,” IEEE Trans. VLSI Systems, vol. 13, no. 6, pp. 732-
741, June 2005.
McCluskey, E. J. (1985). Built-In Self-Test Technique, IEEE Design and Test of Computers, Vol.
2, No. 2, PP. 29–36, Apr. 1985.
Nagle, H. T.; Roy, S. C.; Hawkins, C. F.; Macnamer, M. G. & Fritzemeier, R. R. (1989). Design
for Testability and Built-in Self Test : A review, IEEE Transactions on Industrial
Electronics, Vol. 36, No. 2, PP. 129–140, May. 1989.
Pirsch, P.; Demassieux, N. & Gehrke, W. (1995). Vlsi architecture for video compression-a
survey, Proceedings of the IEEE, No. 2, P. 220, Feb. 1995.
296 VLSI

Touba, N. A. & McCluskey, E. J. (1997). Pseudo-Random Pattern Testing of Bridging Faults,

Proceeding sof International Conference on Computer Design , PP. 54–60, 1997.
Wiegand, T. (2003) Draft ITU-T Recommendation and Final Draft International Standard of
Joint Video Specification (ITU-T Rec. H.264/ ISO/ IEC 14496-10 AVC), Joint Video
Team (JVT) of ISO/IEC MPEG and ITU-T VCEG, Mar. 2003.
Wu, T. H.; Tsai, Y. L.; & Chang, S. J. (2007). An Efficient Design-for-Testability Scheme for
Motion Estimation in H.264/AVC, Proceedings of International Symposium VLSI
Design Automation and Test, PP.25-27, Apr. 2007.
SOC Design for Speech-to-Speech Translation 297

15
X

SOC Design for Speech-to-Speech Translation

Shun-Chieh Lin1, Jia-Ching Wang2, Jhing-Fa Wang3,
Fan-Min Li4 and Jer-Hao Hsu5
1Industrial Technology Research Institute
2National Central University
3, 4, 5National Cheng Kung University

1. Introduction
In today’s globalised world, information exchange between different languages is
indispensable. Accordingly, speech-to-speech translation researches [1]-[6] grew to a
leading-edge technology enabling multilingual human-to-human and human-to- machine
interaction. In previous work, two different architectures are adopted by many researchers
for speech-to-speech translation research [4],[15] – a conventional sequential architecture
and a fully integrated architecture. The sequential architecture is composed of a speech
recognition system followed by a linguistic (or non-linguistic) text-to-text translation system
and a text-to-speech system [1],[2],[5],[16-20]. The integrated architecture combines speech
feature models and translation models in a manner similar to that used for speech
recognition. This integration brings about efficient translation by searching for an optimal
word sequence of target language through the integrated network such as finite-state
transducer [4],[21] and the others [22],[23].
According to these two architectures, there are two system implementations – client-server-
based systems and stand-alone handheld systems. Nevertheless, a critical shortcoming of
client-server based speech translation systems is that they should be built on the server
computer. In other words, it is not available anytime or anywhere. An obvious solution
would be to build portable stand-alone speech-to-speech translation handheld devices. To
mention just a few: Isotani et al. [7] used a Pocket PC PDA with 206 MHz StrongARM/64
MB RAM to build a speech-to-speech translation system for the use in various situations
while traveling. Waibel et al. [8] adopted a Pocket PC PDA with 400 MHz StrongARM/64
MB RAM to construct a speech-to-speech translation system for medical interviews.
Watanabe et al. [9] developed a mobile device running on 400 MHz Pentium II-class
processor/192 MB RAM that helps speech-to-speech translation in various situations during
their travel abroad. However, these works show that real-time speech-to-speech translation
with a resource-limited device is still a problem.
Table 1 shows a comparison among related speech-to-speech translation systems, including
JANUS [1], Verbmobil [2], EUTRANS [4] etc. Clearly, the client-server based architecture
can work in real time, but is not portable; while the stand-alone system is portable, but lacks
the real-time performance. Therefore, a VLSI solution is presented by realizing the entire
298 VLSI

speech-to-speech translation algorithm within a single chip. This SOC chip only requires a
few peripheral components for complete operation, and is characterized by small size, low
cost, real-time operation, and high reliability. The construction of this chip is accomplished
in two main phases: the software simulation phase and the SOC design phase.
In the software simulation phase, the simulation is based on the multiple-translation
spotting (MTS) method, a kind of integration of speech analysis and language translation
[23]. The proposed multiple-translation spotting approach is directly from speech to speech
without language models like other automatic speech recognition (ASR) approaches. With
identifying speech features, translation primarily stays in the speech modality and does not
go through a textual modality. The proposed MTS method not only retrieves the optimal
multiple-translation spotting template, but also extracts the appropriate target patterns.
With the extracted patterns, the target speech can be generated by a concatenation-based
waveform segment synthesis method.
In the SOC design phase, besides a cost efficient programmable core used for system control
and non-computation-intensive tasks, three specific hardware cores were designed to
perform cepstrum extraction, template retrieval, target pattern extraction. Moreover, the
A/D converter (ADC) and D/A converter (DAC) are also designed.
The rest of this paper is organised as follows. Section 2 gives the software simulation of the
proposed speech-to-speech translation system. Section 3 discusses the SOC architecture for
the MTS-based speech-to-speech translation system. Finally, a short conclusion is provided
in Section 4.

System Vocabulary Size Response time Specification

JANUS [1] 3,000~5,000  2 times real-time PC-based platform (server-
client)

Verbmobil [2] 10,000 4 times real-time PC-based platform (server-

client)

EUTRANS [4] 1,701(English)  3 times real-time PC-based platform (server-

2,459(Italian) client)

ATR-MATRIX [5] 13,000 0.1 sec for TDMT PC-based platform (server-
client)

Isotani et al. [7] 20,000(English) Slower than VR5500 Pocket PC PDA (206 MHz
50,000(Japan) processor StrongARM/64 MB RAM)

Speechalator [8] – 2~3 sec Pocket PC PDA (400 MHz

StrongARM/64 MB RAM)

Watanabe et al. [9] 10000 (English) 2~3 sec Mobile PC (400 MHz Pentium
50000 (Japan) II-class processor/192 MB
RAM)
Table 1. A comparison among related speech-to-speech translation systems
SOC Design for Speech-to-Speech Translation 299

Table 2. Examples for TS and MTS.1

2. The Proposed Speech-to-Speech Translation System

2.1 System Overview
Multiple-translation spotting (MTS) is proposed as an improvement on the traditional
translation spotting (TS) method [11], [12] and uses multiple-translation spotting templates
derived from multiple pairs of translations for spotting all hypothesised target-language
patterns. Table 2 lists some examples of TS and MTS. The source-language input “I want a
double room” can only provide five target-language patterns <我,要,一,間,房> in traditional
TS method but the proposed MTS provides all target-language patterns <我,要,一,間,雙人,房
>. All target-language patterns <明天,有,一,間,單人,房,嗎> are also extracted by MTS while
inputting the sentence “Is there a single room for tomorrow.” For each multiple-translation
spotting template, hypothesised target patterns are generated in the MTS process.
Figure 1 shows an example of the MTS process for the direction of English-to-Chinese while
inputting a speech. The multiple-translation spotting template of this example possesses
eleven English patterns in a feature template. Based on the one-stage algorithm [13], when a
speaker inputs an English speech “Is a single room still available for tonight”, the proposed
system can obtain the identified results – “is”, ”a”, “single room”, “still”, “available”, “for”,
and “tonight.” Following the identified results, the translations of “is嗎”, “a一間,”
“single room單人房”, and so on are extracted. A Mandarin Chinese speech pattern set that
includes seven hypothesised speech patterns, “嗎”, “一間”, “單人房”, “還”, “有”, and “今晚”
is generated by the constructed waveform translations in the multiple-translation spotting
template and represented as darker waveforms in Fig. 1.
For speech generation, after determining the optimal target speech sequence, these
waveforms are rearranged with adequate overlapping portions to generate speech with the

1
The gray tablets list the translation spotting results.
300 VLSI

waveform similarity overlap and add (WSOLA) algorithm [24]. WSOLA introduces a
tolerance on the desired time-warping function to ensure signal continuity at waveform
segment joins. With a proper windows length and a timing tolerance, WSOLA usually
produces high quality time-scaled speech [25],[26]. Therefore, the system can generate high
phonetic/prosodic quality in the translated speech output. The advantages of this method
are the small computational cost during the generation process and the high intelligibility of
the generated speech. The following subsections further discuss the details of the kernel
spotting algorithm within the speech translation.

tonight
a single room
laundry service

Multiple-
room service

Translation
Spotting
Template
tomorrow

Speech
Feature features
Template
for
available

Identified
Waveform results
Translation
still
breakfast

Speech
Pattern
Is

Speech
feature input
Is a single room still available for tonight
An hypothesized Mandarin Chinese speech pattern set:

還有 嗎 一 間 單人房今晚
Fig. 1. An example of MTS process for the direction of English-to-Mandarin Chinese.

2.2 Multiple-Translation Spotting

To formulate the concept of MTS between input speech ( X 1L ) and the v-th multiple-
translation spotting template ( rv ), we use the following notations: l represents the frame
index within X 1L , 1  l  L , j represents the translation pair s vj , t vj index within rv , 1  j  J ,
and k represents the frame index within the j-th source pattern s vj and its mapped target
pattern denoted by t vj , 1  k  K j . Then for each input frame, the accumulated distortion
d A (l , k , j ) is defined by:
SOC Design for Speech-to-Speech Translation 301

d A (l , k , j )  d (l , k , j )  min d A (l  1, m, j )  , (1)
k  2 m  k

for 2  k  K j , 1  j  J , where d (l , k , j ) is the local distortion between the l-th frame of

X 1L and the k-th frame of the source pattern s vj . The recursion in (1) is carried out for the
internal frames (i.e., k  2 ) of each source pattern. At the pattern boundary, i.e., when k  1 ,
the recursion can be calculated as:
A 
1 m  J
A m A 
d (l ,1, j )  d (l ,1, j )  min min d (l  1, K , m) , d (l  1,1, j ) . (2)

And the best path is determined by


d Gv  min d A L , K j , j  .
1 j  J
 (3)

After ranking all the templates, the hypothesized spotting template is decided from the Top
N candidates with minimum distortion by
J
(4)
vˆ  arg max   vj
1 v  N j 1

According to the decided v̂ -th template from (4), the target patterns  t vˆ  J can be obtained
j j 1

by  τ vˆ  J . With the determining target speech patterns, these waveforms are rearranged
j j 1

with adequate overlapping portions to generate speech with the waveform similarity
overlap and add (WSOLA) algorithm.

2.3 Software Simulation

The translation experiments were performed on both a PC-based platform and iPAQ PDAs.
On the PC-based platform, the software simulations were done using Windows CE 3.0 on a
Pentium IV 1.8 GHz, 1 GB RAM, Windows XP PC. On the COMPAQ iPAQ PDAs, the
system was implemented on a 400 MHz Intel® XScale processor with 128 MB RAM. This
work built a collection of English sentences and their Chinese translations that frequently
appear in phrasebooks for foreign tourists. The phrasebooks are referred to [27] and [28]. Six
sub-domains are used for collection including accommodation, restaurant, traffic, shopping,
tourism, and asking for directions. Each sub-domain contains 174~251 translations.
Experiments were conducted between Mandarin Chinese and English. To evaluate the
system performance, the 1,260 utterances of the training set used for constructing multiple-
translation spotting templates were collected from one speaker (Sp1) and the 105 utterances
of the test set were also collected from the same speaker (Sp1) for closed testing and two
bilingual male speakers (Sp2 and Sp3) for open testing. First speaker's feature models
(spotting templates) were used to proceed tests on the remaining two speakers. All speeches
had an 8 kHz sampling rate with a precision of 16 bits on both the PCs and the PDAs. The
speech feature analysis was performed using 10th-order linear prediction cepstral
coefficients (LPCCs) using 32 ms frames with 8 ms overlap. The total memory requirement
of the whole system at the startup was about 22.9 MB. The required work memory is about 1
MB, depending on the length of a speech input.
While hesitations or insertions are occurred, the proposed MTS approach can spot critical
patterns as keyword spotting or partial matching and a target output can be properly
generated by the spotted patterns with the decided spotting template. However, the
proposed approach would be sensitive to disturbances including speaking rate, noise,
speaker properties, and so on. For the effect of speaker properties for the proposed system,
302 VLSI

For retrieving Top 5 templates, Table 3 shows that the spotting accuracy of Sp2 and Sp3
drops by 10 to 15 percent. A given spotting template is called a match when it obtains the
same intention of the input speech. In addition, from the research presented in [31], the
speaking rate had a significant effect on recognition accuracy and further adaptation
methods of duration models for spotting templates are needed. Based on the proposed
approach, when template or vocabulary size increases, the increasing spotting templates
will lead to more speech feature vectors and hence more similarities will occur in speech
spotting measurement, thus causing false spotting results and lowering spotting accuracy.
By collecting more speech databases, the system can apply speaker-dependent or speaker-
independent HMM to MTS for more robust speech translation. Speech translation
performance also is degraded by noise. Related works to minimize the effects of the noise on
the system performance are presented in [29],[30] and would be applied to the proposed
system in the future.
To judge the generated translations from the matched templates, a subjective sentence error
rate (SSER) in [3] is used to evaluate and classify the target generation results into three
categories by three bilingual evaluators. Referring to the SSER evaluation method, good (G),
understandable (U), bad (B) levels of translation quality are scaled to 1.0, 0.5, and 0.0,
respectively. The understandable translation rate is calculated by
1.0  no. of G level  0.5  no. of U level (5)
rate   100%
no. of tests
The results revealed that our proposed approach roughly achieves 89% and 92%
understandable translation rate from (5) for the Mandarin Chinese-to-English and the
English-to-Mandarin Chinese translations, respectively.

Table 3. Average spotting accuracy in multiple speaker testing.

3. SOC Design for the MTS-based Speech-to-Speech Translation System

The VLSI architecture for the speech-to-speech translation SOC was developed using the
programmable application-specific technique. Besides a cost efficient programmable core
[14], which is used for system control and other non computation-intensive tasks, three
specific hardware cores are also designed for feature extraction, template retrieval, pattern
extraction. Moreover, the A/D converter and D/A converter are also designed.
Figure 2 shows the overall block diagram of the VLSI architecture for the speech-to-speech
translation SOC. This architecture mainly consists of a cepstrum extraction core, a template
retrieval core, a pattern extraction core, a programmable core, and ADC/DAC. In the
SOC Design for Speech-to-Speech Translation 303

following paragraph, we will discuss the three specific processing hardware cores and the
ADC/DAC for our speech-to-speech translation system.

Cepstrum  Speech
Template 
Template LPC Extraction Feature
Retrieval Core Extraction 
Retrieval Core Core
Core RAM
ADC/
DAC

Low Speed Bus

High Speed Bus

Pattern 
Pattern Programmable
Programmable  High Low
Extraction 
Extraction Core Core Converter
Core

Translation
Synthesis Translation
Template
Pattern ROM Rule ROM
ROM

Fig. 2. Block diagram of the SOC architecture for the proposed speech-to-speech translation
system.

3.1 The Design of the Cepstrum Extraction Core

Feature extraction is one of the most important issues in the field of speech recognition. In
many speech recognition systems, the cepstral coefficients are treated as the feature
parameters, derived by the Fourier Transform or from the LPC coefficients. We selecte the
latter way as it yields high recognition rate and lower computational load.

3.1.1 Autocorrelation Analysis

The sampling rate of the input speech for the cepstrum extraction core is 8 KHz. The pre-
emphasis filter for the digitized speech is defined by
H ( z )  1  hz 1 , 0.9  h  1.0 . (6)
Because we adopt fixed-point implementation, h is chosen as 15/16. The pre-emphasized
speech is then blocked into N samples, with adjacent frames being separated by M samples.
In other words, the adjacent frame begins M samples later than the previous frame and
overlaps with it by N-M samples. In our design, M=192, the frame size N=256, and therefore
the frame rate equals 31.25 frames per second. The next step is to window each frame so that
the signal discontinuities at the beginning and the end of each frame are minimized. The
Hamming window is used and has the form of
 2i 0  i  N 1 (7)
 0.54  0.46 cos( ),
w(i )   N 1
 0 , otherwise

For each frame of the windowed speech signal, autocorrelation analysis is performed by the
following formula:
304 VLSI

N 1
(8)
R( k )   x(i  k ) x(i ),0  k  P ,
i k

where P is the order of the LPC analysis.

The detailed architecture for the autocorrelation analysis of overlapping frames from the
digitized input speech can be seen in [32]. The architecture is based on a calculation
procedure which is depicted in Fig. 3.

R(0) = x(0)x(0)+x(1)x(1) +x(2)x(2)+x(3)x(3) +......+x(10)x(10)+......+ x(n) x(n) + x(n+1)x(n+1) +......+ x(N-1)x(N-1)

R(1) = x(0)x(1) +x(1)x(2)+x(2)x(3) +......+ x(9)x(10)+......+ x(n-1) x(n) + x(n) x(n+1) +......+ x(N-2)x(N-1)
R(2) = x(0)x(2)+x(1)x(3) +......+ x(8)x(10)+......+ x(n-2) x(n) + x(n-1)x(n+1) +......+ x(N-3)x(N-1)

R(10)= x(0)x(10)+......+ x(n-10)x(n) + x(n-9)x(n+1) +......+ x(N-9)x(N-1)

f n (k) x(n1k)x(n1)
Fig. 3. An example illustrating the autocorrelation calculation procedure.

3.1.2 Linear Predictive Analysis

Of the various linear predictive analysis algorithms, we adopt the autocorrelation method
because it leads to more stable results than the covariance method and has lower
computational load than the lattice method [33]. For the autocorrelation method, the matrix
equation for solving the LPC coefficients is in the form of
P
R(i)  a R( i  k ), 1 i  P,  k
(9)
k 1

where R(i) is the autocorrelation coefficient, P is the order of the LPC analysis, chosen as 10
in this paper.
The most efficient method known for solving this particular system of equations is the
Levinson-Durbin recursion, which can be formulated as follows [33]:
E ( 0 )  R(0) (10)
for 1  i  P
i 1
(11)

temp  R(i )  a (i 1) R(i  j )
j 1
j

K i  temp / E ( i 1) (12)

end for
a i( i )  K i (13)
a (i )
j a ( i 1)
j  Ki a ( i 1)
i j , 1  j  i 1. (14)
E (i)
 (1  K ) E , i
2 ( i1)
(15)
The above equations are solved recursively for i = 1, 2,  , P and the final solution is given
as
a j  a (j P ) , 1 j  P . (16)
Substituting (12) into (15), we obtain that
E (i )  E ( i 1)  temp  K ( i ) . (17)
The architecture based on pipeline fashion for the Levinson-Durbin recursion is described in
SOC Design for Speech-to-Speech Translation 305

Fig. 4. There are a small number of divisions needed for the LPC computation. However, it is
not economical to prepare extra hardware for them. Instead, the division operations can be
performed by prune-and-search [34] on the hardware without the use of an individual divider.

Register File
R(0)~R(10), quotient
Constants E, K,Temp, Set up
a1~a10

mux mux

Multiplier

mux mux

Adder

Fig. 4. The architecture for the LPC analysis.

Register File Register File

Constants
Cepstrum LPC

mux mux

Multiplier

mux mux

Adder

Fig. 5. The architecture for the conversion from LPC to cepstrum.

306 VLSI

3.1.3 Conversion from LPC to Cepstrum

The cepstrum is computed from the linear prediction model. The recursive equations of the
LPC based cepstrum are as follows [35]:
C1   a1 , (18a)
n 1
C n  a n   (1  m n)a m C n m , 1  n  P, (18b)
m 1

where Cn is the nth cepstral coefficient, an is the nth LPC coefficient, and P is the cepstrum
order, which is set equal to the LPC order in this paper.
The memory requirement for the constant (1  m n) can be reduced as follows. There are n-1
constants for any n, so the total number of the various constants is 1+2+3+…+P-1=P(P-1)/2.
By taking into consideration the pattern of the constant array, let km,n=(1-m/n), then we can
show that
kn-1-i,n=1- ki,n , (19)
for i=1, 2 ,…, n  1 / 2 , and if n is even then k n / 2,n =1/2.
Equation (19) implies that ki,n is equal to the complement of k n1i,n , therefore, the size of the
constant array can be reduced to half. This algorithm is implemented by a store-and-
accumulate technique and the architecture is described in Fig. 5. We need two storage
elements to hold the LPC coefficients and the previous order cepstral coefficients. In
addition, an accumulator is necessary for computing and totaling the main expression, (1-
m/n)anCn-m to yield the cepstral coefficients defined in (18b).
The dedicated architecture for the autocorrelation analysis, the LPC analysis, and the
conversion from LPC to LSP are individually designed. However, since the three procedures
are performed sequentially, a resource-sharing technique,is performed so that the cepstrum
extraction core will need only one multiplier and one adder.

3.2 The Design of the Template Retrieval Core

To exemplify our design concept, Figs. 6(a) and 6(b) display two dependence graphs (DGs) of
the template retrieval. These dependence graphs are constructed according to (1) and (2), and
can be regarded as the template retrieval space. The DG shown in Fig. 6(a) describes the case
with 12 frames in the input speech ( X 112 ). There are three source patterns ( s1v1 , s 2v1 , s3v1 ) in this
template ( rv1 ), and the frame number of them is 7, 8, and 6, respectively. The DG shown in Fig.
6(b) has 8 frames in the input speech ( X 8 ) and 2 patterns in the template ( rv 2 ). The frame
1

number of the two patterns ( s1v 2 , s 2v 2 ) in this template is 6 and 7, respectively.

The different frame numbers within different input speeches cause the variation along the
horizontal axis in the template retrieval space. The different pattern numbers within different
templates and the different frame numbers within different patterns are responsible for the
variation along the vertical axis in the template retrieval space. The architecture design for
coping with the variation of the structure of the template retrieval space is described as follows.
First, horizontal projection is performed in the original DGs (Fig. 6(a)) and the result is shown
in. Fig. 6(c). This projection reduces the two-dimensional (2-D) template retrieval space into a
one-dimensional (1-D) one and eliminates the variation resulting from different frame
numbers within different input speeches. Each node in Fig. 6(c) requires a register to store the
computational results for the distortion accumulation of the next frame. The nodes within one
SOC Design for Speech-to-Speech Translation 307

column are divided into different blocks according to the patterns they belong to. Thus the DG
in Fig. 6(c) consists of three blocks corresponding to three different patterns.
To further reduce the size of the DG, a vertical projection is performed in Fig. 6(c). This
projection transforms the three-block DG into a one-block DG (see Fig. 6(e)). Because the frame
numbers for the blocks are different, the largest frame number is taken as the frame number
for the new one-block DG. To deal with the discrepancy of having different frame numbers
within different patterns, we added some multiplexers. In addition, two extra registers are
added in front of each node to replace the two eliminated blocks. Figure 6(e) is then modified
as Fig. 6(g) to have a regular wire connection.
Figures 6(b), 6(d), 6(f) and 6(h) provide another example of the use of horizontal and vertical
projections. One can find that the frame number and the register number for a certain node
between the DGs in Figs. 6(g) and 6(h) are different. To combine the two DGs into a single one,
the largest frame number and the largest register number between them are chosen, see Fig.
6(i). The added multiplexers in front of each node provide the selectivity among the one-block
DGs.
rv1

v1
s 3

H o r iz o n t a l
s 2v 1 P r o je c t io n

s 1v 1

x 11 2

( a ) :  2 ‐ D   O r i g i n a l   T e m p l a t e  R e t r i e v a l  S p a c e ( c ) : 1 ‐ D  T e m p la t e  R e t r ie v a l S p a c e

rv 2

s 2v 2
H o r iz o n t a l
P r o je c tio n

s 1v 2

x 18

( b ) :   2 ‐ D  O r i g i n a l   T e m p l a t e   R e t r i e v a l   S p a c e ( d ) :  1 ‐ D  T e m p la t e  R e t r ie v a l S p a c e

R
M
in

e
M

ux
Mn
i

V e r t ic a l
M o d if ic a t io n
P r o je c t io n
(c )
R
e
M

ux
n
M
i

g
M M M M M M M M
u u u u u u u u
x x x x x x x x

DG 
(e ) : O n e ‐ B o c k  D G ( g ) :  M o d i f i e d  O n e ‐ B o c k  D G
M e r g in g
R
M
in

e
M

ux
Mn
i

V e r t ic a l
M o d if ic a t io n
P r o je c t io n ( i ) :   F i n a l   M e r g e d  D G
(d )

(f ) : O n e ‐ B o c k  D G ( h ) : M o d if ie d  O n e ‐ B o c k  D G

Fig. 6. Architecture inference of the template retrieval based on DG.

308 VLSI

3.2.1 The Distortion Unit

Each node in the template retrieval DG associates with a local distortion. The local distortion
between a frame of the template and a frame of the input speech is defined by
P
, (20)
Local distortion  R U 
i 0
i i

where Ri denotes the i-th cepstral coefficient in the R-th frame of the template, U i represents
the i-th cepstral coefficient in the U-th frame of the input speech, and P is the cepstrum order.
The cepstral coefficients of the input speech and the templates are stored in RAM0 and
ROM1, respectively. The distortion unit accesses the cepstral coefficients from the two
memories, and then accumulates the distortion in register R.

3.2.2 The Processing Element

Each node in the DGs shown in Fig. 6(i) is implemented as a processing element. The
architecture design of the PE is displayed in Fig. 7. After receiving all the accumulated
distortions from the registers, the PE selects the minimum accumulated distortion, and then
adds it to the local distortion d (l , k , j ) . The result is the new accumulated distortion d A (l , k , j )
associated with the current node. In addition to the accumulation process, the PE also
generates the decision information that indicates the source node in a path transition. The
use of the decision information is discussed together with the pattern extraction core later.
Based on the data path of the template retrieval core, a corresponding template retrieval
controller is also developed. In addition to producing control signals, this controller also
functions as a memory address generator for accessing the template speech features. The
template retrieval core not only accumulates the minimum local distortion but also writes
the decision information into the memory. The decision information for a node indicates its
source node in a path transition. The best path can be obtained from the decision
information, tracing it backward for extracting the hypothesised speech patterns. The design
of the pattern extraction core is given in the next subsection.

d(l, k, j) d (l, k, j)

dA (l 1, k, j) dA (l 1, k, j)
0 0
dA (l 1, k 1, j)
+ dA (l, k, j) + dA (l, k, j)
Mux

Mux

1
dA (l 1, k  2, j) 2 1

mindA (l 1, Km , m)

1mJ
Comparator decision information Comparator decision information
dk , k 1,2,K 1 d0 , e

For the internal Frames in a node For the first frame in a node
Fig. 7. Architecture of the processing element.

3.3 The Design of the Pattern Extraction Core

To extract the best pattern sequence, this work presents a block-node pattern extraction
method. Figure 8 exemplifies the relationship between nodes and blocks. The block
boundaries are the same as the pattern boundaries. This template retrieval space contains 24
SOC Design for Speech-to-Speech Translation 309

blocks, numbered from 0 to 23. Blocks in the same row belong to the same pattern, and have
the same number of nodes (frames). For example, blocks 2, 5, 8, …, 23 belong to the third
pattern within this template. The pattern extraction method used in this work can be
regarded as a two-level address decoding process. At the first level, the blocks are the
addressing units. Each block consists of frame nodes, which become addressing units at the
second level. If b denotes the block index and k denotes the local node index in a block, each
node can also be referred to by the pair (b, k). For example, node 64 can be referred to as node
(21, 1).
The decision word is generated for each block, and is denoted by
D  {d 0 , d1 , d 2 , ..., d K 1 , e} , (21)
where K is the largest number among all the block node ones, e is the pattern decision
information from the first node of a block, and dk, 0  k  K  1 , is the decision information
from the k-th node of a block. We use e to indicate the source pattern for an external
(between-patterns) transition, and dk to indicate the source node for an internal (within-
patterns) transition. Assume that Kj denotes the number of nodes in the j-th pattern.
Pattern extraction starts from the end node, and recursively update the local node index and
the block index to construct a best path in the template retrieval space. The block index b is
updated by
b  J , for internal transition, (22)
b 
b  J  new _ e  old _ e, for external transition,
where J is the number of patterns in this template, new_e is the current value of the decision
information e, and old_e is the previous value of the decision information e.
As for the local node index k, it is updated by
k , for internal transition with d k  0,
k  1, for internal transition with d  1, (23)

k 
k

 k  2, for internal transition with d k

 2,
 K e  1, for external transition.
Let us denote the best path end in the block by end_block, and its pattern by end_pattern. The
flow chart of the algorithm is depicted in Fig. 9.
The block-node pattern extraction method is implemented as a finite state machine (FSM).
Because template retrieval and pattern extraction are required for all templates, this design
is based on a two-memory scheme. While the decision word is being written into RAM1, the
back tracing is being performed in RAM2; and vice versa. This reduces the overall
computation time of template retrieval and pattern extraction for all templates.
310 VLSI

8 17 26 35 44 53 62 71 (2,1) (5,1) (8,1) (11,1) (14,1) (17,1) (20,1) (23,1)

Third
Pattern 7 16 25 34 43 52 61 70 (2,0) (5,0) (8,0) (11,0) (14,0) (17,0) (20,0) (23,0)

b=2 b=5 b=8 b=11 b=14 b=17 b=20 b=23

6 15 24 33 42 51 60 69 (1,3) (4,3) (7,3) (10,3) (13,3) (16,3) (19,3) (22,3)

5 14 23 32 41 50 59 68 (1,2) (4,2) (7,2) (10,2) (13,2) (16,2) (19,2) (22,2)

Second
Pattern 4 13 22 31 40 49 58 67 (1,1) (4,1) (7,1) (10,1) (13,1) (16,1) (19,1) (22,1)

3 12 21 30 39 48 57 66 (1,0) (4,0) (7,0) (10,0) (13,0) (16,0) (19,0) (22,0)

b=1 b=4 b=7 b=10 b=13 b=16 b=19 b=22

2 11 20 29 38 47 56 65 (0,2) (3,2) (6,2) (9,2) (12,2) (15,2) (18,2) (21,2)

First 55
1 10 19 28 37 46 64 (0,1) (3,1) (6,1) (9,1) (12,1) (15,1) (18,1) (21,1)
Pattern
0 9 18 27 36 45 54 63 (0,0) (3,0) (6,0) (9,0) (12,0) (15,0 (18,0) (21,0)

b=0 b=3 b=6 b=9 b=12 b=15 b=18 b=21

Fig. 8. Example illustrating the block-node addressing method.

Pattern extraction
from the end node (b, k)

Current
block is in the first
End Yes column of the block-
node addressing
method?

No

Current node is a first

node of a block?
node (b, k)

Yes

k=Ke-1;
External Transition? Yes
b=b-J+new_e-old_e;

No
No

k=k, if dk=0;
k=k-1, if dk=1;
k=k-2, if dk=2;
b=b-J;

Fig. 9. Flow chart of the block-node pattern extraction.

SOC Design for Speech-to-Speech Translation 311

3.4 The Design of the ADC/DAC

The last specific hardware core is ADC/DAC. In this part, we mainly present an efficient
architecture to improve the conventional fully differential successive approximation ADC.
The proposed architecture includes two optimal design solutions. First, in order to avoid the
mismatch between positive and negative reference voltages, a single reference voltage (SRV)
method that removes the negative reference voltage is developed by the distinct switched
capacitor controlling algorithm. Figure 10 shows the implementation of the SRV algorithm
in circuit. Second, a small area scheme for successive approximation register (SAR) is
introduced. Comparing to the traditional successive approximation ADC, our proposed
architecture is contributive to improve the area saving in SAR and the accuracy of ADC.

Fig. 10. The architecture of fully differential successive approximation ADC using single
reference voltage.

To reduce the area for successive approximation register, this work presents a simplified
non-redundant successive approximation register (SSAR). The basic architecture of the
SSAR is a multiple input n-bit shift register, shown in Fig. 11. This multiple input register
consists of a general D flip-flop and a multiplexer with two inputs. The function of SSAR is
shown in Fig. 11. Suppose the “1” is the input token in Fig. 12. Whenever the SSAR is
triggered, each register of SSAR must go through three modes: (1) when the token has not
passed yet, the values of the registers are “0”; (1) when the token is staying at a certain
register, its register value is changed to “1”, and receives the result of comparator; (3) when
the token has passed through, the value determined by the result of comparator is held until
the whole conversion is done. The function can be implemented as below.
Therefore, when the value of k-th register of the SSAR is still “0”, the Qk+1 selected by the
multiplexer is connected to Dk. If the k-th register receives the token, the Qk is changed to “1”,
312 VLSI

and the CMP would be selected. The result of comparator Zk would be held in the k-th
register until the whole conversion is done. The system can be implemented by the blocked
CLKk. The combination of all above implementations makes it possible to improve the
accuracy and save the chip area.
MSB

LSB
Qp
Mux D2 Q2 Mux D1 Q1 Mux D0 Q0 Dp Qp
CMP CMP CMP
2th 1th 0th
Q2 Q1 Q0
CLK2 CLK1 CLK0 CLK

CLK
Q1b Q0b Qpb

Fig. 11. The architecture of the simplified non-redundant SAR.

1st cycle 1 0 0 0

2nd cycle Z2 1 0 0

3 rd cycle Z2 Z1 1 0

4th cycle Z2 Z1 Z 0 1
Fig. 12. The function of simplified non-redundant SAR.

4. Summary
Speech-to-speech machine translation is a prospective application of speech and language
technology. This work presents an MTS based speech-to-speech translation system between
Mandarin Chinese and English. The proposed MTS approach achieves about a 90%
understandable translation rate on average. For a portable and real-time speech-to-speech
translation system, this work also proposes the SOC realization. The architecture design is
based on the semi-ASIC technique, which incorporates a cost efficient programmable core
along with specific hardware accelerators: an LPC extraction core, a template retrieval core,
and a pattern extraction core. Besides, the ADC/DAC are also included. A SRV-based fully
differential successive approximation ADC with reduced SSAR is designed. These hardware
cores construct a complete speech-to-speech translation SOC. The proposed SOC chip is the
first one dedicated for speech-to-speech translation.
SOC Design for Speech-to-Speech Translation 313

5. References
[1]A. Lavie et al., “JANUS III: speech-to-speech translation in multiple languages,” in Proc.
IEEE Int. Conf. Acoustics, Speech and Signal Processing, Apr. 1997, pp. 99–102.
[2]W. Wahlster, Verbmobil: Foundations of Speech-to-Speech Translation. Berlin Heidelberg,
New York: Springer-Verlag, 2000.
[3]H. Ney, S. Nießen, F. J. Och, H. Sawaf, C. Tillmann, and S. Vogel, “Algorithms for
statistical translation of spoken language,” IEEE Trans. Speech and Audio Processing,
vol. 8, pp. 24-36, Jan. 2000.
[4]F. Casacuberta et al., ”Speech-to-speech translation based on finite-state transducers,” in
Proc. IEEE Int. Conf. Acoustics, Speech and Signal Processing, May 2001, pp. 613–616.
[5]F. Sugaya, T. Takezawa, A. Yokoo, and S. Yamamoto, “End-to-end evaluation in ATR-
MATRIX: speech translation system between English and Japanese,” in Proc. 6th
Eur. Conf. Speech Communication and Technology, Sep. 1999, pp. 2431–2434.
[6]P. C. Ching and H. H. Chi, “ISIS: a trilingual conversational system with learning
capabilities and combined interaction and delegation dialogs,” in Proc. National
Conf. Man-Machine Speech Communications, Nov. 2001, pp. 119–124.
[7]R. Isotani, K. Yamabana, S. Ando, K. Hanazawa, S. Ishikawa, T. Emori, H. Hattori, A.
Okumura, and T. Watanabe, “An automatic speech translation system on PDAs for
travel conversation,” in Proc. IEEE Int. Conf. Multimodal Interfaces, Oct. 2002, pp.
211–216.
[8]A. Waibel, A. Badran, A. W. Black, R. Frederking, D. Gates, A. Lavie, L. Levin, K. Lenzo,
L. M. Tomokiyo, J. Reichert, T. Schultz, D. Wallace, M. Woszczyna, and J. Zhang,
“Speechalator: two-way speech-to-speech translation on a consumer PDA”, in Proc.
European Conf. Speech Communication and Technology, Sep. 2003, pp. 369–372.
[9]T. Watanabe, A. Okumura, S. Sakai, K. Yamabana, S. Doi, and K. Hanazawa, “An
automatic interpretation system for travel conversation,” in Proc. Int. Conf. Spoken
Language Processing, Sep. 2000, pp. IV-444–IV-447.
[10]J. F. Wang, B. Z. Houg, and S. C. Lin, “A study for Chinese text to Taiwanese speech
system,” in Proc. Int. Conf. Research on Computational Linguistics, Aug. 1999, pp. 37–
53.
[11]M. Simard, “Translation spotting for translation memories,” in Proc. HLT-NAACL
Workshop on Building and Using Parallel Texts: Data Driven Machine Translation and
Beyond, May 2003, pp. 65–72.
[12]J. Véronis and P. Langlais, “Evaluation of parallel text alignment systems – the ARCADE
project,” in Parallel Text Processing, Dordrecht: Kluwer Academic, 2000, pp. 369–388.
[13]L. Rabiner and B. H. Juang, Fundamentals of Speech Recognition. Prentice-Hall, Inc., 1993.
[14]J. F. Wang, A. N. Suen, and C. K. Chieh, “A programmable application specific
architecture for real-time speech recognition,” in Proc. of VLSI Design/CAD
Symposium, Aug. 1995, pp. 261–264.
[15]Y. Zhang, “Survey of current speech translation research,” presented at Multilingual
Speech-to-Speech Translation Seminar, Carnegie Mellon University, Pittsburgh,
PA, 2003.
[16]Y. S. Lee and S. Roukos, “IBM Spoken Language Translation System Evaluation,” in
Proc. INTERSPEECH2004 Workshop on Spoken Language Translation: Evaluation
Campaign on Spoken Language Translation, Oct. 2004, pp.39–46.
314 VLSI

[17]L. Gu and Y. Q. Gao, “On Feature Selection in Maximum Entropy Approach to Statistical
Concept-based Speech-to-Speech Translation,” in Proc. INTERSPEECH2004
Workshop on Spoken Language Translation: Evaluation Campaign on Spoken Language
Translation, Oct. 2004, pp.115–121.
[18]S. Nakamura, K. Markov, T. Jitsuhiro, J. S. Zhang, H. Yamamoto and G. Kikui, “Multi-
Lingual Speech Recognition System for Speech-To-Speech Translation,” in Proc.
INTERSPEECH2004 Workshop on Spoken Language Translation: Evaluation Campaign
on Spoken Language Translation, Oct. 2004, pp.146–154.
[19]K. Matsui, Y. Wakita, T. Konuma, K. Mizutani, M. Endo, and M. Murata, “An
experimental multilingual speech translation system,” in Proc. ICMI-PUI, 2001, pp.
1–4.
[20]S. Rossato, H. Blanchon, and L. Besacier, “Speech-to-speech translation system
evaluation: Results for French for the Nespole! project first showcase,” in Proc.
ICSLP, 2002, pp. 1905–1908.
[21]F. Casacuberta, E. Vidal, and J. M. Vilar, “Architectures for speech-to-speech translation
using finite-state models,” in Proc. ACL Workshop on Speech-to-Speech Translation:
Algorithms and Systems, 2002, pp. 39–44.
[22]H. Ney, “Speech translation: Coupling of recognition and translation,” in Proc. ICASSP,
1999, pp. 517–520.
[23]J. F. Wang and S. C. Lin, and H.W. Yang, “Multiple-Translation Spotting for Mandarin-
Taiwanese Speech-to-Speech Translation,” Int. Journal of Computational Linguistics
and Chinese Language Processing, vol.9, no.2, 2004, pp. 13-28.
[24]W. Verhelst and M. Roelands, “An overlap-add technique based on waveform similarity
(WSOLA) for high quality time-scale modification of speech,” in Proc. ICASSP,
1993, pp. 554–557.
[25]M. Demol, K. Struyve, W. Verhelst, H. Paulussen, P. Desmet, and P. Verhoeve, “Efficient
non-uniform time-scaling of speech with WSOLA for CALL applications,”
presented at InSTIL/ICALL Symp. Computer Assisted Learning, Venice, Italy,
2004.
[26]H. G. Ilk and S. Tugac, “Channel and source considerations of a bit-rate reduction
technique for a possible wireless communications system's performance
enhancement,” IEEE Trans. Wireless Communications, vol. 4, no. 1, pp. 93–99, Jan.
2005.
[27]E. T. Cornelius, English 900. Pace Group International Inc., 1999.
[28]B. E. Bagnell and M. Lee, New Globe English Course on Travel. New Globe Publishing Co.,
LTD, 1990.
[29]J. F. Wang and S. H. Chen, “Speech Enhancement Using Perception Wavelet Packet
Decomposition and Teager Energy Operator”, Journal of VLSI Signal Processing,
Vol. 36 I: 2-3, pp. 125–139, Feb. 2004..
[30]J. F. Wang, C. H. Yang, and K. H. Chang, “Design of a Subspace Tracking Based Speech
Enhancement System”, in Proc. IEEE TENCON 2004, vol. 1, pp. 147–150..
[31]J. T. Chien and C. H. Huang, “Bayesian Learning of Speech Duration Models,” IEEE
Trans. Speech and Audio Processing, vol. 11 I:6, pp. 558–567, Nov. 2003.
[32]Jhing-Fa Wang, Jia-Ching Wang, Han-Chiang Chen, Tai-Lung Chen, Chin-Chan Chang,
and Ming-Chi Shih, “Chip Design of Portable Speech Memopad Suitable for
SOC Design for Speech-to-Speech Translation 315

Persons with Visual Disabilities,” IEEE Transactions on Speech and Audio Processing,
vol. 10, no. 8, pp. 644-658, November 2002.
[33]J. Makhoul, “Linear prediction: a tutorial review,” Speech Analysis, IEEE Press, New
York, 1979.
[34]L. Y. Liu, J. F. Wang, J. Y. Lee, M. H. Sheu, and Y. L. Jeang, "An ASIC design for linear
predictive coding of speech signals," in Proc. Euro ASIC' 92, 1992, pp.288-291.
[35]S. Furui, “Cepstral analysis technique for automatic speaker verification,” IEEE
Transactions on Acoustics, Speech, and Signal Processing, vol. ASSP-29, no. 2, pp. 254-
272, 1981.
316 VLSI
A Novel De Bruijn Based Mesh Topology for Networks-on-Chip 317

16
X

A Novel De Bruijn Based Mesh

Topology for Networks-on-Chip
Reza Sabbaghi-Nadooshan1, Mehdi Modarressi2,3
and Hamid Sarbazi-Azad2,3
1Islamic Azad University Central Tehran Branch, Tehran, Iran
2Sharif University of Technology, Tehran, Iran
3IPM School of computer science, Tehran, Iran

1. Introduction
The mesh topology is the most dominant topology for today’s regular tile-based NoCs. It is
well known that mesh topology is very simple. It has low cost and consumes low power.
During the past few years, much effort has been made toward understanding the
relationship between power consumption and performance for mesh based topologies
(Srivasan et al., 2004). Despite the advantages of meshes for on-chip communication, some
packets may suffer from long latencies due to lack of short paths between remotely located
nodes. A number of previous works try to tackle this shortcoming by adding some
application-specific links between distant nodes in the mesh (Ogras & Marculescu, 2005)
and bypassing some intermediate nodes by inserting express channels (Dally, 1991), or
using some other topologies with lower diameter (sabbaghi et al., 2008).
The fact that the de Bruijn network has a logarithmic diameter and a cost equal to the linear
array topology motivated us to evaluate it as an underlying topology for on-chip networks.
De Bruijn topology is a well-known network structure which was initially proposed by de
Bruijn (de Bruijn, 1946) as an efficient topology for parallel processing. Samathan (Samathan
& Pradhan, 1989) showed that de Bruijn networks are suitable for VLSI implementation, and
several other researchers have studied topological properties, routing algorithms, efficient
VLSI layout and other important aspects of the de Bruijn networks (Park & Agrawal, 1995;
Ganesan & Pradhan, 2003).
In this chapter, we propose a two-dimensional de Bruijn based mesh topology (2D DBM for
short) for NoCs. We will compare equivalent mesh and 2D DBM architectures using the two
most important factors, network latency and power consumption. A routing scheme for the
2D DBM network has been developed and the performance and power consumption of the
two networks under similar working conditions have been evaluated using simulation
experiments. Simulation results show that the proposed network can outperform its
equivalent popular mesh topology in terms of network performance and energy dissipation.
318 VLSI

2. The 2D DBM Topology

2.1 The Structure
The basic idea about our work is based on digraphs, and in this section, we present the
information compiled from studies conducted on the de Bruijn digraph (Liu & Lee, 1993 ;
Mao & Yang , 2000).
The de Bruijn topology has many applications in communication networks and parallel
processing (Samanathan & Pradhan, 1989). A de Bruijn graph has kn nodes. Each node
u  (un1 ,..., u0 ) has an edge to node v  (vn1 ,..., v0 ) if and only if vi  ui 1 1  i  n  1 . Node v
has a unidirectional direct link to node u if and only if
u  v  k  r (mod k n ) , 0  r  k  1 (1)

The in-degree and out-degree of a node is equal to k. Therefore, the degree of each node is
equal to 2k. The diameter of a de Bruijin graph is equal to n which is optimal. The de Bruijn
also has a simple routing algorithm. Case k=2 is the most popular de Brujin network which
is also used in this study. Due to the logarithmic (optimal) diameter and a simple routing
algorithm, it can be expected that the traffic on channels in the network will be less than
other networks, resulting in a better performance (Ganesan & Pradhan, 2003).
Examples of de Bruijn networks are illustrated in Fig. 1. Several researchers have studied the
topological properties (Liu & Lee, 1993 ; Mao & Yang , 2000) and efficient VLSI layout
(Samanathan & Pradhan, 1989; Chen et al., 1993) of the de Bruijn networks. Moreover, the
scalability problem of de Bruijn networks is addressed in (Liu & Lee, 1993). In de Brujin
network data is circulated from node to node until it reaches its destination. Each node has
two outgoing (incoming) connections to (from) other nodes via shuffle (rotate left by one bit)
and shuffle-exchange (rotate left by one bit and complement LSB) operations to neighboring
nodes (Louri & Sung, 1995). Owing to the fact that these connections are unidirectional, the
degree of the network is the same as the one-dimensional mesh networks (or linear array
network). The diameter of a de Bruijn network with size N, that is, the distance between
nodes 0 and N-1, is equal to log (N).

(a)

(b)
Fig. 1. The de Bruijn network with (a) 8 nodes and (b) 16 nodes
A Novel De Bruijn Based Mesh Topology for Networks-on-Chip 319

2.2 The 2D DBM Topology

In a 2D DBM network, the nodes in each row and each column form a de Bruijn network.
Each node has two outgoing edges along which data packets can be sent to other nodes, and
two incoming links receiving data packets from other nodes in each dimension. Thus, node
(u, v), u  (un1 ,..., u0 ) , v  (vn1 ,..., v0 ) , has edges to node (u’, v) if and only if u 'i  ui 1 for
1  i  n  1 , and node (u, v’), if and only if v'i  vi 1 for 1  i  n  1 . Node (u’, v’) has a direct
link to node (u, v’) if and only if

u  2u ' rmod 2n , r  0,1 (2)

and to node (u’, v) if and only if

v  2v' rmod 2n , r  0,1 (3)

The 8×8 2D DBM is shown in Fig. 2.

0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7

1,0 1,1 1,2 1,3 1,4 1,5 1,6 1,7

2,0 2,1 2,2 2,3 2,4 2,5 2,6 2,7

3,0 3,1 3,2 3,3 3,4 3,5 3,6 3,7

4,0 4,1 4,2 4,3 4,4 4,5 4,6 4,7

5,0 5,1 5,2 5,3 5,4 5,5 5,6 5,7

6,0 6,1 6,2 6,3 6,4 6,5 6,6 6,7

7, 0 7,1 7, 2 7,3 7, 4 7,5 7, 6 7, 7

Fig. 2. A 2D DBM with 64 nodes composed from eight 8-node de Bruijn networks (as shown
in Fig.1. a) along each dimension

The 2D DBM networks have some interesting topological properties that motivate us to
consider them as a suitable candidate for on-chip network architectures. The most important
property of 2D DBM networks is that while the number of links in a 2D DBM and an equal-
sized mesh are exactly the same, the network diameter of this network is less than that of the
320 VLSI

mesh. More precisely, the diameter of a 2D DBM and a mesh are 2log N0.5 and 2(N0.5-1),
respectively, where N represents the network size.
Although, establishing the new links removes the link between some adjacent nodes (for
example 1 to 0, 2 to 1 and 3 to 4 connections in Fig. 1) and increases their distance by one
hop. In this network, however, the distance between many nodes is decreased by one or
multiple hops, compared to a mesh, and this can lead to a considerable reduction in the
average inter-node distance in the network.
The 2D DBM links are unidirectional and at most 8 unidirectional links are used per node.
This is equal to the number of links connected to a node in a mesh (which is 4 bidirectional
links). Since the node degree of a topology has an important contribution in (and usually
acts as the dominant factor of) the network cost, the proposed topology can achieve lower
average distance than a 2D mesh while it has almost the same cost. However, we will
discuss the area overhead due to longer links in 2D DBM in next sections.

2.3 Routing Algorithm

During past years, a number of routing algorithms have been developed for the original de
Bruijn network. Ganesan (Ganesan & Pradhan, 2003) proposed a routing algorithm which
routes the packets toward the destination by changing one bit at a time, starting from the
most significant bit of an n-bit address in a network of size 2n. At the ith step of this
algorithm, the n-ith bit of the destination address is compared to the MSB of the current
address. If they are equal, the message is routed over the shuffle channel, to keep the bit
unchanged and rotate the address. Otherwise, the message is routed over the shuffle-
exchange channel to make the two bits identical and then rotate the address (self-loops are
avoided). This algorithm involves a maximum of n steps. In order to be deadlock-free, this
algorithm requires n virtual channels and the message uses the ith channel at the n-ith step.
Since in this virtual channel selection scenario routing is performed in a descending order
regarding channel numbers, the dependency graph of virtual channels is acyclic and the
routing is deadlock-free (Dally & Seitz, 1987).
Ganesan (Ganesan & Pradhan, 2003) splits the networks into two trees: T1 and T2. A
message is routed between T1 to T2, and then in T2, and then the message is routed between
T2 to T1 and then in T1. Therefore, this routing algorithm has four different steps, which
may also decrease, depending on the source and destination nodes. The algorithm has two
phases and initiates with phase 0 (using virtual channel 0). When the message goes through
T2 to T1, the phase number increases (virtual channel 1 is used). Ganesan (Ganesan &
Pradhan, 2003) proved that this method is deadlock-free. The two trees, T1 and T2, are
depicted in Fig. 3 for N=8.

Fig. 3. Trees T1 and T2 for N=8

A Novel De Bruijn Based Mesh Topology for Networks-on-Chip 321

Park (Park & Agrawal, 1995) has deformed the de Bruijn as two graphs, increasing and
decreasing. In the ith stages, the MSB of the current node is compared with the n-ith bit of the
destination node, and if they are the same, shuffle cycle is used; otherwise, shuffle-exchange
is used. If the path switches from increasing graph to the decreasing graph, the virtual
channel number increments by 1. It is proved that the mentioned algorithm for the de Brujin
network is deadlock-free (Park & Agrawal, 1995). Park (Park & Agrawal, 1995) has shown
that for a de Brujin network with N nodes (N=2n) this kind of routing requires

Number of V.C.  n   n  1 (4)

 
 2 

virtual channels to ensure deadlock freedom.

So, for N=8, it requires 2 virtual channels. For N=16 nodes, (n=4), based on equation (4), the
number of virtual channels needed for a deadlock-free routing is equal to 3. Virtual channel
0 is more crowded than the other virtual channels. The routing algorithm uses virtual
channels unevenly and very few packets use all the virtual channels.
We revised the routing algorithm to balance the use of virtual channels. In our proposed
solution, at each node, the header flit has a degree of flexibility in selecting virtual channels,
as used in (Kiasari et al., 2005). For example, for N=16, some of the source nodes may use
only 1 virtual channel. Some nodes use 2 virtual channels and a few source nodes use 3
virtual channels. For the last group (those use 3 virtual channels), virtual channel 0 is
selected in the source node. For the second group, we start with virtual channel 1 and for the
first group we start from virtual channel 2. If virtual channel 2 is occupied, then we can use
virtual channel 1 and finally virtual channel 0 is chosen. Using this method, virtual channels
are used uniformly.
In this chapter, for routing in the 2D DBM, we use the revised method of Park (Park &
Agrawal, 1995) in each dimension and we choose the paths and nodes in a way that the
routing is minimal. Like XY routing in mesh networks, the deterministic routing first applies
the routing mechanism in rows in order to deliver the packet to the column at which the
destination is located. Afterwards, the message is routed to the destination by applying the
same routing algorithm in the columns. Obviously, adding the second dimension in this
routing scheme does not generate a cycle and the whole routing in the 2D network is
deadlock-free provided that the routing in each dimension is deadlock-free (Duato et al.,
2005).

3. Simulation Results
3.1 Simulator
To simulate the proposed NoC topology, we have used an interconnection network
simulator that is developed based on POPNET simulator (Popnet, 2007) with Orion power
library embedded in it. Orion is a library which models the power consumption of the
interconnection networks (Wang et al., 2002). Providing detailed power characteristics of the
network elements, Orion enables the designers to make rapid power performance tradeoffs
at the architecture level (Wang et al., 2002). As mentioned in (Wang et al., 2002), the total
energy each flit consumes at a specified node and its outgoing link is given by
322 VLSI

E flit  Ewrt  E arb  E read  E xb  E link (5)

It consists of five components: 1) the power that is consumed for writing into buffers, 2) the
power of arbiter, 3) the power that is consumed in reading from buffers, 4) the power of the
internal crossbar, and 5) the power that is consumed in the links.
The POPNET simulator is only introduced for a two-dimensional mesh topology (Popnet,
2007). We have customized the simulator to support other topologies and other routing
algorithms such as shuffle-exchange (Sabbaghi et al., 2008), and de Bruijn topologies. The
power is obtained and reported for each layer and each component of the network.
We set the networks link width to 32 bits. Each link has the same bandwidth and one flit
transmission is allowed on a link. The power is calculated based on a NoC with 90 nm
technology whose routers operate at 250 MHz. Based on the core size information presented
in (Mullins et al., 2006), we set the width of the IP cores to 2 mm, and the length of each wire
is set based on the number of cores it passes. The simulation results are obtained for 88
and 16×16 mesh NoCs with XY routing algorithm, and 88 and 16×16 de Bruijn NoCs using
the routing algorithms described in the previous section. The message length is assumed to
be 32 and 64 flits and 2 and 3 virtual channels per physical channel are used. Messages are
generated according to a Poisson distribution with rate λ. The traffic pattern can be Uniform,
Matrix-transpose, and Hotspot (Duato et al., 2005).

3.2 Comparison Results

In this section, we evaluate the 2D DBM and compare it with the mesh, the most common
topology for NoCs. Based on the necklace properties in de Bruijn layout (Chen et al., 1993),
we have considered a more efficient layout for each row and column of the 2D DBM as
shown in Fig. 4. With this new layout the total wire length used in the network is decreased.
For example, for an 8×8 2D DBM about 25% reduction in total wire length is obtained and
this can be increased to a 50% reduction for a 16×16 2D DBM.

Fig. 4. A better node placement of de Bruijn network of 8 nodes

As mentioned before, we have also revised the routing algorithm in (Park & Agrawal, 1995) to
have balanced use of virtual channels. Fig. 5 compares the performance of original routing
algorithm and the new routing algorithm in the 88 2D DBM using 2 virtual channels with
messages of 32 flits. As can be seen in the figure, the new algorithm exhibits better
performance in terms of average message latency.
Figures 6-9 compare power consumption and the performance of simple 2D mesh and 2D
DBM NoCs under various traffic patterns, network sizes and message lengths. In Fig. 6 and
Fig. 7, the average message latency is displayed as a function of message generation rate at
each node for the 8×8 and 16×16 networks under deterministic routing. As can be seen in the
figures, the 2D DBM NoC achieves a reduction in message latency with respect to the popular
2D mesh network for the full range of network load under various traffic patterns (especially
in uniform traffic). Note that for matrix-transpose traffic load, it is assumed that 30% of
messages generated at a node are of matrix-transpose type (i.e. node (x,y) sends the message to
A Novel De Bruijn Based Mesh Topology for Networks-on-Chip 323

node (y,x)) and the rest of 70% messages are sent to other nodes uniformly. For hotspot traffic
load a hotspot rate of 16% is assumed (i.e. each node sends 16% of its messages to the hotspot
node (node (4,4) in 8×8 network and node (8,8) for 16×16 network) and the rest of 84% of
messages are sent to other nodes uniformly). Note that increasing the network size causes
earlier saturation in a simple 2D mesh.
8-8 v=2 m =32
w ith v.c. unif ormity
400
Averagedelay(cycles)

w ithout v.c. unif ormity

300

200

100
0.001 0.002 0.003 0.004 0.005
M e s s age ge ne r ation r ate ()

Fig. 5. The effect of balanced use of virtual channels

500 8-8 v=2 m =32

Delay(cycles )
Average

300 bruijn-u
mesh-u
bruijn-mat
mesh-mat
bruijn-hot
mesh-hot
100
0.001 0.002 0.003 0.004 0.005
Message generation rate ( )

a)
8-8 v=2 m =64
800
Delay(cycles )

600
Average

bruijn-u
mesh-u
400 bruijn-mat
mesh-mat
bruijn-hot
mesh-hot
200
0.0005 0.001 0.0015 0.002 0.0025
Message generation rate ( )

b)
Fig. 6. The average message latency in the 8×8 simple 2D mesh and 2D DBM for different
traffics patterns with message size of (a) 32 flits and (b) 64 flits
324 VLSI

500 16-16 v=3 m =32

Delay(cycles )
Average
300
bruijn-u
mesh-u
bruijn-mat
mesh-mat
bruijn-hot
mesh-hot
100
0.001 0.002 0.003 0.004 0.005
Message generation rate ( )

a)
16-16 v=3 m =64
800
Delay(cycles )

600
Average

bruijn-u
mesh-u
400 bruijn-mat
mesh-mat
bruijn-hot
mesh-hot
200
0.0005 0.001 0.0015 0.002 0.0025
Message generation rate ()

b)
Fig. 7. The average message latency in the 16×16 simple 2D mesh and 16×16 network of 2D
DBM for different traffics patterns with message size of (a) 32 flits and (b) 64 flits

According to the simulation results reported above, the 2D DBM has a better performance
compared to the equivalent simple 2D mesh NoC. The reason is that the average distance a
message travels in the network in a 2D DBM network is lower than that of a simple 2D
mesh. The node degree of the 2D DBM and simple 2D mesh networks (hence the structure
and area of the routers) are the same. However, unlike the simple 2D mesh topology, the 2D
DBM links do not always connect the adjacent nodes and therefore, some links may be
longer than the links in an equivalent mesh. This can lead to an increase in the network area
and also create problems in link placement. The latter can be alleviated by using efficient
VLSI layouts (Samanathan & Pradhan, 1989; Chen et al., 1993) proposed for de Bruijn
networks, as we used.
Fig. 8 demonstrates power consumption of the simple 2D mesh and 2D DBM under
deterministic routing scheme with uniform traffic. It is again the 2D DBM that shows a
better behavior before reaching to the saturation point. Fig. 9 reports similar results for
hotspot and matrix-transpose traffic patterns in the two networks.
A Novel De Bruijn Based Mesh Topology for Networks-on-Chip 325

110

Power (nj / cycles)

90

70
mesh-64f
mesh-32f
50 bruijn-64f
bruijn-32f
30
0.001 0.002 0.003 0.004 0.005 0.006
Message generation rate ( )

a)
550

500
Power (nj / cycles)

450

400

350

300 mesh-64f
mesh-32f
250 bruijn-64f
200
bruijn-32f
150
0.001 0.002 0.003 0.004 0.005 0.006
Message generation rate ( )

b)
Fig. 8. Power consumption of the simple 2D mesh and 2D DBM with uniform traffic pattern
and message size of 32 and 64 flits for (a) 8×8 network and (b) 16×16 network

110
Power (nj / cycles)

90

70 bruijn-32u
bruijn-hot
bruijn-mat
50 mesh-32u
mesh-hot
mesh-mat
30
0.001 0.002 0.003 0.004 0.005 0.006
Message generation rate ( )

a)
326 VLSI

550

500

Power (nj / cycles)

450

400

350
bruijn-u
300 bruijn-hot
bruijn-mat
250 mesh-u
200
mesh-hot
mesh-mat
150
0.001 0.002 0.003 0.004 0.005 0.006
Message generation rate ( )

b)
Fig. 9. Power consumption of simple 2D mesh and 2D DBM for different traffic patterns and
message size 32 flits for (a) 8×8 and (b) 16×16 networks

The results indicate that the power of 2D DBM network is less for light to medium traffic
loads. The main source of this reduction is the long wires which bypass some nodes and
hence, save the power which is consumed in intermediate routers in an equivalent mesh
topology.
Although for low traffic loads the 2D DBM network provides a better power consumption
compared to the simple 2D mesh network, it begins to behave differently near heavy traffic
regions.
It is notable that a usual advice on using any networked system is not to take the network
working near saturation region (Duato et al., 2005). Having considered this and also the fact
that most of the networks rarely enter such traffic regions, we can conclude that the 2D
DBM network can outperform its equivalent mesh network when power consumption is
considered.
The area estimation is done based on the hybrid synthesis-analytical area models presented
in (Mullins et al. , 2006; Kim et al., 2006; Kim et al. 2008). In these papers, the area of the
router building blocks is calculated in 90nm standard cell ASIC technology and then
analytically combined to estimate the router total area. Table 1 outlines the parameters. The
analytical area models for NoC and its components are displayed in Table 2. The area of a
router is estimated based on the area of the input buffers, network interface queues, and
crossbar switch, since the router area is dominated by these components.
The area overhead due to the additional inter-router wires is analyzed by calculating the
number of channels in a mesh-based NoC. An n×n mesh has 2×n×(n-1) channels. The 2D
DBM has the same number of channels as mesh but with longer wires. In the analysis, the
lengths of packetization and depacketization queues are considered as large as 64 flits.
In Table 3, the area overhead of 2D DBM NoC is calculated for 8×8 and 16×16 network sizes
in a 32-bit wide system. The results show that, in an 8×8 mesh, the total area of the 2mm
links and the routers are 0.0633 mm2 and 0.1089 mm2, respectively. Based on these area
estimations, the area of the network part of the 2D DBM network shows a 44% increase
compared to a simple 2D mesh with equal size. Considering 2mm×2mm processing
elements, the increase in the entire chip area is less than 3.5%. Obviously, by increasing the
A Novel De Bruijn Based Mesh Topology for Networks-on-Chip 327

buffer sizes, the network node/configuration switch area increases, leading to much
reduction in the area overhead of the proposed architecture.

Parameter Symbol
Flit Size F
Buffer Depth B
No. of Virtual channels V
Buffer area (0.00002 mm2/bit (Kim et al., 2008)) Barea
Wire pitch (0.00024 mm (ITRS, 2007) Wpitch
No. of Ports P
Network Size N (= n×n)
Packetization queue capacity PQ
Depacketization queue capacity DQ
Channel Area (0.00099 mm2/bit/mm (Mullins et al. , 2006) Warea
Channel Length (2mm ) L
No. Of Channels Nchannel
Table 1. Parameters

Symbol Model
Crossbar RCXarea W2pitch×P×P×F2
Buffer (per RBFarea Barea×F×V×B
port)
Router Rarea RCXarea+P×RBFarea
Network NAarea PQ× Barea +DQ ×Barea
Adaptor
Channel CHarea F×Warea×L×Nchannel
NoC Area NoCarea n2× (Rarea+ NAarea)+ CHarea
Table 2. Area analytical model

Network Link Area Router Increase percent to increase percent in

Area mesh the entire chip
88 mesh .06338 .1089 0 0
88 2D DBM .1086 .1089 44.38 3.46
16×16 mesh .06338 .1217 0 0
16×16 2D DBM .1626 .1217 103.58 9.57
Table 3. 2D DBM area overhead

4. Conclusion
The simple 2D mesh topology has been widely used in a variety of applications especially
for NoC design due to its simplicity and efficiency. However, the de Bruijn network has not
been studied yet as the underlying topology for 2D tiled NoCs. In this chapter, we
introduced the two-dimensional de Bruijn Mesh (2D DBM) network which has the same cost
as the popular mesh, but has a logarithmic diameter. We then conducted a comparative
simulation study to assess the network latency and power consumption of the two
328 VLSI

networks. Results showed that the 2D DBM topology improves on the network latency
especially for heavy traffic loads. The power consumption in the 2D DBM network was also
less than that of the equivalent simple 2D mesh NoC.
Finding a VLSI layout for the 2D and 3D DBM networks based on the design considerations
in deep sub-micron technology, especially in three dimensional design, can be a challenging
future research in this line.

5. References
http://www.princeton.edu/~lshang/popnet.html, August 2007.
Chen, C.; Agrawal, P. & Burke, JR. (1993). dBcube : A New class of Hierarchical
Multiprocessor Interconnection Networks with Area Efficient Layout, IEEE
Transaction on Parallel and Distributed Systems, Vol. 4, No. 12, pp. 1332-1344.
Dally, WJ. & Seitz, C. (1987). Deadlock-free Message Routing in Multiprocessor
Interconnection Networks, IEEE Trans. on Computers, Vol. 36, No. 5, pp. 547-553.
Dally, WJ. (1991). Express Cubes: Improving the Performance of K-ary N-cube
Interconnection Networks, IEEE Trans. on Computers, Vol. 40, No. 9, pp. 1016-1023.
De Bruijn, NG. (1946). A Combinatorial Problem,” Koninklijke Nederlands Akademie van
Wetenschappen Proceedings, 49-2, pp.758–764.
Duato, J. (1995). A Necessary and Sufficient Condition for Deadlock-free Adaptive Routing
in Wormhole Networks, IEEE Transactions on Parallel and Distributed Systems, Vol. 6,
No. 10, pp. 1055–1067.
Duato, J.; Yalamanchili, S. & Ni, L. (2005). Interconnection Networks: An Engineering Approach,
Morgan Kaufmann Publishers.
Ganesan, E. & Pradhan, DK. (2003). Wormhole Routing in de Bruijn Networks and Hyper-
de Bruijn Networks, IEEE International Symposium on Circuits and Systems (ISCAS),
pp. 870-873.
ITRS. (2007). International technology roadmap for semiconductors. Tech. rep., International
Technology Roadmap for Semiconductors.
Kiasari, AE.; Sarbazi-Azad, H. & Rezazad, M. (2005). Performance Comparison of Adaptive
Routing Algorithms in the Star Interconnection Network, Proceedings of the 8th
International Conference on High Performance Computing in Asia-Pacific Region
(HPCAsia), pp. 257-264.
Kim, M.; Kim, D. & Sobelman, E. (2006). NoC link analysis under power and performance
constraints, IEEE International Symposium on Circuits and Systems (ISCAS), Greece.
Kim, MM.; Davis, JD.; Oskin, M & Austin, T. (2008). Polymorphic on-Chip Networks,
International Symposium on Computer Architecture(ISCA), pp. 101 -112.
Liu, GP. & Lee, KY. (1993). Optimal Routing Algorithms for Generalized de Bruijn Digraph,
International Conference on Parallel Processing, pp. 167-174.
Louri, A. & Sung, H. (1995). An Efficient 3D Optical Implementation of Binary de Bruijn
Networks with Applications to Massively Parallel Computing, Second Workshop on
Massively Parallel Processing Using Optical Interconnections, pp.152-159.
Mao, J. & Yang, C. (2000). Shortest Path Routing and Fault-tolerant Routing on de Bruijn
Networks, Networks, vol.35, pp.207-215.
A Novel De Bruijn Based Mesh Topology for Networks-on-Chip 329

Mullins, R.; West, A. & Moore, S. (2006). The Design and Implementation of a Low-Latency
On-Chip Network, Asia and South Pacific Design Automation Conference(ASP-DAC),
pp. 164-169.
Ogras, UY. & Marculescu, R. (2005). Application-Specific Network-on-Chip Architecture
Customization via Long-Range Link Insertion, IEEE/ACM Intl. Conf. on Computer
Aided Design, San Jose, pp. 246-253.
Park, H.; Agrawal, DP. (1995). A Novel Deadlock-free Routing Technique for a class of de
Bruijn based Networks, IPPS, pp. 524-531.
Sabbaghi-Nadooshan, R.; Modarressi, M. & Sarbazi-Azad, H. (2008). A Novel high
Performance low power Based Mesh Topology for NoCs, PMEO-2008, 7th
International Workshop on Performance Modeling, Evaluation, and Optimization, pp. 1-7.
Samanathan, MR.; Pradhan, DK. (1989). The de Bruijn Multiprocessor Network: a Versatile
Parallel Processing and Sorting Network for VLSI, IEEE Trans. On Computers, vol.
38, pp.567-581.
Srivasan, K.; Chata, KS. & Konjevad, G. (2004). Linear Programming Based Techniques for
Synthesis of Networks-on-chip Architectures, IEEE International conference on
Computer Design, pp. 422-429.
Wang, H.; Zhu, X.; Peh, L. & Malik, S. (2002). Orion: A Power-Performance Simulator for
Interconnection Networks, 35th International Symposium on Microarchitecture
(MICRO) , Turkey, pp. 294-305.
330 VLSI
On the Efficient Design & Synthesis of Differential Clock Distribution Networks 331

17
X

On the Efficient Design & Synthesis of

Differential Clock Distribution Networks
Houman Zarrabi1, Zeljko Zilic2, Yvon Savaria3 and A. J. Al-Khalili1
1 Department of Electrical and Computer Engineering, Concordia University
2 Department of Electrical and Computer Engineering, McGill University
3 Department of Electrical Engineering, École Polytechnique de Montréal

Canada

1. Introduction
Almost all high-performance VLSI systems in today technologies are synchronous. These
systems use a clock signal to control the flow of data throughout the chip. This greatly
facilitates the design process of systems because it provides a global framework that allows
many different components to operate simultaneously while sharing data. The only price for
using synchronous type of systems is the additional overhead required to generate and
distribute the clock signal.
Nearly all on-chip Clock Distributions Networks (CDNs) contain a series of buffers and
interconnects that repeatedly power-up the clock signal from the clock source to the clock
sinks. Conventionally, CDNs consisted of only a single stage buffer driving wires to the
clock loads. This is still the case for clock distribution in very small scale systems; yet
contemporary complex systems use multiple buffer stages. A typical clock tree distribution
network in modern complex systems is shown in Figure 1. This design is based on the
reported CDNs in (O’Mahony et al, 2003; Restle et al, 1998; Vasseghi et al, 1996).

1.1 Hierarchy in CDNs

The clock signal is generated with a Phase Lock Loop (PLL). A PLL is a control system that
generates a signal having a fixed relation to the phase of its reference signal. A PLL circuit
responds to both the frequency and the phase of its input signal and automatically
raises/lowers the frequency of the controlled oscillator until it matches the reference
(Wikipedia, 2009). The core clock signal is then amplified through the global buffer and
distributed through a hierarchical network and buffers. The system CDN is generally
defined to span from the PLL to the clock pins. The pin is the input to a buffer that locally
amplifies and distributes the clock signal to clocked storage elements within a macro, the
small blocks that make up a system. There can be any number of buffer levels between the
PLL and the clock pin. In modern VLSI systems, there are up to four buffer levels. The last
buffer level before the clock pin is generally called a sector buffer. This stage drives the
interconnect leading to the macros and the local buffers at the pins. A synchronous VLSI
332 VLSI

system has thousands of loads to be driven by clock signal. In CDNs, the loads are grouped
together creating a (sub-) block. This trend results in a hierarchy in the design of CDNs
including three different levels/categories of clock distribution namely as global, regional and
local as shown in Figure 1. At each level of hierarchy there are buffers associated with that
level to regenerate and to improve the clock signal at that level.
The global clock distribution connects the global clock buffer to the inputs of the sector
buffers. This level of the distribution has usually the longest path in CDN because it relays
the clock signal from the central point on the die to the sector buffers located throughout the
die. The issues in designing the global tree is mostly related to signal integrity which is meant
to maintain a fast edge rate over long wires while not introducing a large amount of timing
uncertainty. Skew and jitter accumulate as the clock signal propagates through the clock
network and both tend to accumulate proportional to the latency of the path. Because most
of the latency occurs in the global clock distribution, this is also a primary source of skew
and jitter (Restle et al, 2001). From a design point of view, achieving low timing uncertainty
is the most critical challenge at this level.
The regional clock level is defined to be the distribution of clock signals from the sector
buffers to the clock pins. This level is the middle ground between global and local clock
distribution; it does not span as much area as the global level and it does not drive as much
load or consume nearly as much power as the local level.
The local level is the part of the CDN that delivers the clock pin to the load of the system to
be synchronized. This network drives the final loads and hence consumes the most power.
As a design challenge, the power at the local level is about one order of magnitude larger
than the power in the global and regional levels combined (Restle et al, 2001).

Fig. 1. A typical hierarchical CDN for a high-performance synchronous VLSI system

1.2 CDNs figures of merit

The main figures of merit for a CDN are the components of timing uncertainty, as well as,
power consumption. All of these performance metrics have significant impacts on the
design, evaluation and verification of synchronous system performance and reliability.
As mentioned previously, the advantage of a synchronous system is to regulate the flow of
data throughout the system. However, this synchronizing approach depends on the ability
to accurately relay a clock signal to millions of individual clocked loads. Any timing error
introduced by the clock distribution has the potential of causing a functional error leading to
On the Efficient Design & Synthesis of Differential Clock Distribution Networks 333

system malfunctioning. Therefore, the timing uncertainty of the clock signal must be
estimated and taken into account in the first design stages. The two categories of timing
uncertainties in a clock distribution are skew and jitter.
Clock skew refers to the absolute time difference in clock signal’s arrival time between two
points in a CDN. Clock skew is generally caused by mismatches in either device or
interconnect within the clock distribution or by temperature or voltage variations around
the chip. There are two components for clock skew: the skew caused due to the static noise
(such as imbalanced routing) which is deterministic and the one caused by the system device
and environmental variations which is random. An ideal clock distribution would have zero
skew, which is usually unachievable.
Jitter is another source of dynamic timing uncertainties at a single clock load. The key
measure of jitter for a synchronous system is the period or cycle-to-cycle jitter, which is the
difference between the nominal cycle time and the actual cycle time. The first cycle, the
period is the same as the clock signal period and the second cycle, the clock period becomes
longer/shorter. The total clock jitter is the sum of the jitter from the clock source and from
the clock distribution. Power supply noise may cause jitter in both the clock source and the
distribution (Herzel et al, 1999).
Clock network also involves long interconnects which implies having lots of parasitics
associated with the network contributing to the power consumption of the clock signal.
Having the highest switching activity of the circuit in a chip is another fact of consuming a
large amount of power of the system. This power consumption can be as high as 50% of the
total power consumption of the chip according to (Zhang et al, 2000). The components of
power consumption of CDN are: static, dynamic and leakage power. The power
consumption due to the leakage current, in CDNs, is relatively small. In the same way,
keeping the proper rise/fall times, minimizes the static power consumption. Thus the main
portion of the power consumption is due to the dynamic power consumption. This is
estimated as:

P=f CL Vdd Vswing

in which f, CL, Vdd and Vswing respectively represent frequency of the clock network, total load
capacitances, supply-voltage and voltage-swing of clock signal. For the case of full swing (in
which the clock signal swing reaches the voltage-supply level) Vswing is the same as Vdd.
Accordingly, methods to reduce the power consumption are:
a. Reduce total load capacitances (CL)
b. Reduce voltage-supply (VDD)
c. Reduce clock signal swing (Vswing)
The intrinsic load capacitance relies on the process technology and there is no handy way to
improve it. Yet, from the design aspects by breaking down interconnects by repeater
insertion the total interconnect load is reduced. Worth mentioning that in coupled lines, the
total load is greater than that of single-node lines, thus compensating design methods
should be taken into consideration for power-saving improvement. Typically, power
reduction is achieved by means of supply and/or swing voltage scaling in CDNs.
334 VLSI

2. Differential Clock Distribution Networks (DCDNs)

In this section, based on the general overview given on CDNs, we will introduce the
concepts and motivations toward the design of Differential CDNs (DCDNs). For this, we
initially address the preliminaries needed for the design of DCDNs. These theories include
differential signaling and differential signal integrity.

Fig. 2. Voltage-mode differential signaling

2.1 Preliminaries

2.1.1 Differential signaling

A digital signal can be transmitted differentially over the medium by utilizing two
conductors. One of which is used for transmitting the signal and the other is used for the
complement of the signal. Figure 2 shows a differential voltage-mode signaling system. To
transmit logic ’1’, the upper voltage source drives V1 and the lower voltage source drives V0.
For logic ‘0’ transmission, the voltages are reversed.
As is shown in Figure 2, the following voltages are defined in a differential system: V1 is the
signal on the first line with respect to common return path, V0 is the signal on the second
line with respect to common return path, Vdiff is the differential signal which is the voltage
difference of the two signal pair, and, Vcomm is the common voltage signal which is in
common between both of signal pair. Differential signal Vdiff carries the information and at
the receiver the information is extracted from this voltage difference. In addition to the
differential voltage there is a common-mode signal. This signal is used to give an initial
biasing to the differential signal pair. In ideal conditions, the common-mode signal is
constant and it does not carry any information. In this case:

Vdiff =V1-V0
Differential signaling requires more routing and wires and pins than its single-ended
counterpart system. In return for this increase, differential signaling offers the following
advantages over single-ended signaling:
a. A differential system, serves its own reference. The receiver at the far end of the
system compares the two signal pair to detect the value of the transmitted
information. Transmitters are less critical in terms of noise issues, since the receiver
is comparing two pair of signals together rather than comparing to a fixed
reference. This results in canceling any noises in common to the signals.
b. The voltage difference for the two signal pair between logic’1’ and ‘0’ is:

ΔV=2(V1-V0)
On the Efficient Design & Synthesis of Differential Clock Distribution Networks 335

which is twice as much as is defined for a single-ended signaling system. This

shows that the noise margin of the differential system is twice as much as the single-ended
signaling system. This doubling effect of signal swing improves the speed of the
signaling system. It affects the transition times (rise/fall time) which is done in half
of the transition time of single-ended signaling system.

rdx ldx
c g dx
cc dx lm dx
rdx
c g dx ldx
Fig. 3. A segment of a coupled interconnect

2.1.2 Differential signal integrity

In order to employ differential signaling, the coupled interconnects model is utilized and
applied to the system. This type of interconnects not only have the intrinsic signal integrity
issues, but also, they are involved with their mutual signal integrity aspects. In Figure 3, a
segment of a coupled interconnect is shown.
The mutual parasitic elements are due to the adjacent line. These are mutual capacitance Cc
and mutual inductance lm in addition to the intrinsic parasitic elements r, Cg and l which
indicate intrinsic resistance, capacitance and inductance of each line. The effective
capacitance Ceff associated with each line, depending on the direction/mode of the signaling
(in-phase or out-of-phase usually called even and odd mode respectively) can be calculated
from the following equations (Hall et al, 2000):

Ceff (odd) = ηCc+Cg

Ceff (even) =Cg
And for effective inductance we have:

leff (odd) =l-lm

leff (even) =l+lm
As the above equations indicate, for the case of differential signaling (or out-of-phase
signaling), the effective capacitance is increased by the factor of η due to coupling
capacitances and the effective inductance is decreased due to the effect of mutual
inductance. In (Kahng et al, 2000) it was shown that η has the value of {0, 2 and 3}
depending on the mode of signaling and slew rates of the coupled signals. The typical value
for η, for typical sharp input signals designs, is taken as 2.
336 VLSI

2.1.3 Differential Buffers

The configuration of differential buffers is based on current steering devices, in which the
output logic can be set by steering the current in the circuit. These devices are also
considered as Current Mode Logic (CML) circuits. CML circuits are known to outperform
the conventional CMOS circuits in Giga Hertz (GHz) operation frequency. A basic
differential buffer is given in Figure 4. The current source in differential buffer is the tail
current Iss. When the common-mode voltage Vcomm is applied to the differential buffer, due
to the symmetry of the differential buffer, the current is split equally between the two wings
(Iss/2). Increasing one of the input voltages which implies the decrease in the other one, will
result in increase in current of one branch and decrease in current of the other branch. Note
that the total possible current to steer is Iss and when one input voltage rises, the other one
decreases by the same amount. When the input differential voltage ΔV=Vin-V’in has passed
a specific threshold, in other words when one of the transistors derives all the possible
current from one branch the other transistors turn off, hence the output voltage reaches Vdd
whereas the first branch drops to Vdd-RIss. Several differential loads also have been
introduced in the literature (Dally et al, 1998). These loads may use resistor, current mirror
and cross-coupled transistors. The differential load is characterized by its differential and
common-mode impedances, known as rΔ and rc respectively. The differential impedance
determines the change in the differential current IΔ when the voltages on the two inputs of
the terminal are varied in opposite directions. The common-mode impedance implies the
average current changes when both input voltages are varied in the same direction.
Depending on the type of application, the design may chose from these design options.
Table I demonstrate the rΔ and rc for each load.

Fig. 4. A basic differential buffer

Load rc rΔ
Resistor R R
Current-mirror 1/gm -1/λI
Cross-coupled 1/gm 1/gm
Table 1. Impedance of differential loads
On the Efficient Design & Synthesis of Differential Clock Distribution Networks 337

2.2 Differential Clock Distribution Networks (DCDNs)

As discussed previously, differential signaling offers higher immunity against external
perturbations. Due to the complexity increase and the need for error-free operation in
contemporary systems, the idea of integrating differential signaling and clock distribution is
seemingly becoming a viable solution for modern and for future IC designs.
Historically the idea of DCDN was to be utilized for off-chip clock distribution and for PCB-
level synchronization. This technique was utilized to reduce and suppress the Electro-
Magnetic Interference (EMI) of the neighboring circuits and systems waves. Due to the
superiority of DCDN, recently there has been a couple of works on on-chip DCDN as well,
such as (Sekar, 2002; Anderson et al, 2002). The idea of utilizing on-chip DCDN has not been
widely used in the literature. In (Anderson et al, 2002) a DCDN is used in global level of the
hierarchical CDN for Itanium Microprocessor. They reported that the use of DCDN has
given the advantage of 10% less skew variation. In (Sekar, 2002) it is reported that DCDN
has 25%-42% less sensitivity to power supply noises and 6% less sensitivity to
manufacturing variations when they utilized H-Tree DCDN.
A general model of a DCDN is given in Figure 5. The DCDN is composed of a differential
signal pair shown in two different patterns. The clock tree generally is a binary tree. The
differential signal is dispersed along the clock network. Throughout the clock network at
branching points the differential clock signals are regenerated by differential buffers to
improve the signal integrity of the clock network. Finally at the last stage, they are all
converted to single-ended signals for compatibility with the rest of the system functionality,
which normally use single-ended signals. For the regenerative buffers a simple differential
buffer introduced in the previous part can be utilized. The only design issue related to the
buffer is the choice of differential loads. Based on the process technology, or design criteria,
this item can be chosen from the design library. For final stage converters, usually the choice
of current mirror load is the superior choice. As Table 1 demonstrates, current mirror loads
have high differential output impedance which results in fast change in the output that is
used to drive the output of the clock network.
Differential clocking eliminates the induced crosstalk due to aggression of clock signals.
Clock signal is spread all over the chip area. It also has full switching activity. Also device
sizes tend to shrink as technology advances. These facts show that as technology advances
the clock signal aggression can be quite harmful for all system components all over the chip
area. Distribution of clock with differential signals eliminates this problem to a certain
extent, as both positive and negative signal values are applied and the noise would be
cancelled. Furthermore, as given in (Anderson et al, 2002), DCDN offers less skew variations
in the presence of external noises; it has less sensitivity in presence of supply and process
variations (Sekar 2005).
The aforementioned points are of the most important criteria/solutions for reliable system
design. Due to technology advances and increase in system complexity, the design with low
or no parameter variation in ideal case, has become the most concerning issue. Timing error
results directly in system malfunctioning. Thus designing a reliable and noise tolerant, clock
distribution may help significantly for a reliable system design. As introduced in the
literature, DCDN has these potentials; thus this design methodology can be a solution for
future robust system design.
Plus the pros and cons of DCDN, there are some design/synthesis challenges associated
with the efficient design of DCDNs. Some of most challenges may be summarized as:
338 VLSI

 Differential signaling is involved with higher parasitic, due the existence of

coupled lines. In this case the total power consumption is commonly increased.
 Coupled lines are commonly routed and synthesized using symmetrical path
models (which is not the general case).
 Routing complex DCDNs may take too much computation time. Using existing
routing methods is not time efficient.
Proposing solutions to address the above challenges can efficiently help the design and
synthesis of DCDNs, needed for modern complex VLSI technologies. These solutions are
given in the following sections.

Fig. 5. A general structure for DCDN

3. Efficient design of DCDNs

3.1 Dynamic Threshold (DT) MOS for low-voltage DCDNs
For the design and synthesis of CDNs, buffers are inserted to improve the performance of
CDNs in order to reduce the overhead capacitances. In this part, Dynamic Threshold (DT)
(sometimes referred to as Variable Threshold) transistors (Assaderaghi et al, 1997) are
utilized in conventional differential buffer structures. These transistors outperform the
conventional transistors in low voltage applications which are suitable for advanced low
voltage technologies. The use of DT transistors helps improve the buffer performance. DT
transistors switch faster since their threshold voltage decrease dynamically when the input
is applied to their gate terminal due to body effect. Such buffers are depicted in Figure 6
(Zarrabi et al, 2006). Figure 6(a) presents a Low-Swing:Low-Swing differential buffer. DT
transistors help improve the speed of these buffers when low swing inputs are applied to
the buffer. The use of cross-coupled differential load with high differential impedance helps
to have a fast transition of the inputs to the outputs. Figure 6(b) represents a Low-Swing:
Full-Swing level converter. This buffer is used at the sinks to restore clock signals to their
single-node full amplitude. The current-source pull-ups help to have asymmetrical fast
transformation of differential to single-ended signals. The structure is based on Chappell
amplifier which offers good common-mode noise rejection (Chappell et al, 1998).
On the Efficient Design & Synthesis of Differential Clock Distribution Networks 339

(a) (b)
Fig. 6. (a) Low-Swing: Low-Swing (b) Low-Swing: Full-Swing DT differential buffers

3.2 Differential low-power buffers

Recalling from Section 2.1.1, two voltage components are associated with differential
signaling: common-mode and differential. Common-mode also refers to DC voltage biasing
and is the voltage used for initial biasing of the differential buffer.
In order for receivers (differential to single-ended converters) to operate efficiently and have
a full/proper output swing, the common-mode voltage or DC biasing of the differential
buffer should be low enough to turn the input transistors off. In the literature, the method
used in order to overcome this issue is to increase the voltage swing as much as possible to
be able to decrease the common-mode voltage to the sufficient supply level (usually used
differential voltage of 50% of Vdd) (Anderson et al, 2002; Sekar, 2005). This method results in
high power consumption in DCDN. Recalling from Section 2.1:

Vdiff_low=Vdd-RIss
The above equation implies that in order to increase the differential voltage swing, the tail
current need to be increased. This technique largely affects the power consumption in
DCDN. Note that, it is not possible to touch the load (R) as it directly affects the bandwidth
of the clock network. Therefore, in previous works, in order to reach sufficient output
swing, the differential voltage swing is increased to reduce common-mode voltage.
Correspondingly, a circuit technique is proposed to address this design problem.
The proposed technique for differential receiver is given in Figure 7 (Zarrabi, 2006). The
buffer configuration is based on Chappell amplifier as introduced in the previous section.
Attached to the buffer are the level-shifting circuits. The buffer functionality is as follows:
The dashed parts in Figure 7 are the level shifters (also referred to as source followers)
(Razavi, 2001). When the input is applied to the gate terminals of the level shifters, the
outputs are dropped and follow their inputs. In other words, the voltage gain equals one (no
voltage amplification), and the following relations are applicable (Broderson, 2005):

I=(β/2)(VIN-VOUT-VT)2
VOUT-=IRs
VOUT =(Rsβ/2)(VIN-VOUT-VT)2
VIN =VOUT+ VT+[(VIN-2)/(Rsβ)]0.5
340 VLSI

Fig. 7. Differential receiver with level shifter

The last result shows that VOUT can be derived by solving the final equation iteratively.
However, by making the first order approximation that RS is large enough (especially in
current sources) to make the third term equal to zero, we can conclude:

VOUT= VIN-VT
This shows that the output of the source follower circuits copies the input of the gates with a
shift of a transistor threshold which is a technology dependent factor. The transistor ratios
for buffers sizing are the same as the ones given in (Chappell et al, 1998). However, the total
size of the buffer is scaled to minimize the skew.
The above configuration for the differential receivers helps lower the common-mode (DC
bias) of the internal input transistors of the receiver. Utilizing this design technique, it is
possible to further reduce the differential voltage swing while maintaining a sufficient
output swing at the final nodes.
In order to perform differential voltage scaling in DCDN, previously a new design for level
converter was given. For the case of intermediate buffers, in order to be able to vary the
differential voltage while maintaining the linearity of the buffer, the differential load should
be reconfigured in a way to establish this design goal. In this part, a new configuration for
differential load is proposed which enables us to have linearity in the buffer. Figure 8 shows
the proposed buffer configuration.

Fig. 8. Differential buffer with composite load

The dashed part demonstrates the proposed composite configuration of the differential load.
Such composition enables the circuit to combine both the characteristics of the diode
connected device and triode transistor together to have a linear operating load in various
On the Efficient Design & Synthesis of Differential Clock Distribution Networks 341

voltage ranges (Dally et al, 1998). The proposed buffer based on composite differential load
is a technology portable design and can be used in any available design process whereas the
use of resistance is limited to current and future advanced technologies. This portable
design method comes at the price of increase in area and parasitic elements. The transistor
ratios (for buffers sizing) are 1 to 3 which refer to the ratio of pull up to pull down
transistors (L=2Lmin to reduce the channel length modulation effect). The total size of the
buffer is scaled to reach the objective frequency of operation.

4. Efficient synthesis of DCDNs

4.1 Zero skew DCDN routing
As seen in the overview section, DCDNs are commonly routed assuming symmetrical CDN
path models. This however is not the general case. Here we will propose a method for zero
skew routing of DCDNs applicable to general (asymmetric) path models. In the literature,
especially in (Cong et al, 1996) a comprehensive study on efficient clock routing is studied.
Tsay’s method (Tsay, 1991) is one of the methods introduced for zero skew routing of clock
trees. In order to route differential clock trees with zero skew characteristic, the existing
methods are modified to satisfy design objectives. To achieve this aim here in this part, a
line equivalent delay model is utilized. The model is applied to Tsay’s method (Tsay, 1991)
for zero skew routing of differential clock trees (DCDNs).

4.1.1 Utilizing Tsay’s method

In this method, zero skew is achieved by locating tapping points throughout the clock tree.
Tapping points are the branching points at which sub-trees are chosen to maintain equal
delay as shown in Figure 9.

Fig. 9. Tapping point extraction through merging decoupled sub-tree(s)

As was seen in Section 2.1.2, the effective capacitance associated with each segment of a
coupled line, considering both intrinsic and mutual effects is:

Ceff (odd) = ηCc+Cg

The effective capacitance is applied to both signal lines independent of each other; in this
way, we call these lines as decoupled lines and we name this model as decoupled line model.
This model is employed for the purpose of clock routing. In this way, the decoupled RC-П
delay model is used to model interconnects. The methodology of tapping extraction is as
342 VLSI

follows. Figure 9 shows a schematic of a decoupled clock tree branch in which each line of
the branch is a decoupled distributed RC model connected to its sub-tree child, for which
the distributed line propagation delay is given by tint=0.37RintCeff. Each sub-tree is modeled
by a total capacitance Csubtree and total propagation delay tsubtree as shown in Figure 9.
Considering tapping location x, to satisfy the equality of the two branch delays, the
following equation is realized:

tint1+0.74Rint1Ceffsubtree1+t1= tint2+0.74Rint2Ceffsubtree2+t2 *
In the second part of the equality, since the interconnect resistance combined with sub-tree
capacitance creates a Lumped loop, it has the lumped propagation delay of 0.74RintCsubtree.
Rewriting interconnect parasitics by per unit length parameters, we have:

Rint1=r0xl, Cint1=c0xl
Rint2=r0(1- x)l, Cint1=c0(1- x)l
in which r0 and c0 are the resistance and capacitance per unit length of the wire, l is total
interconnection length between the two sub-trees and tapping location x. Solving Equation
* with respect to x results into:

x=[1.35(tint2-tint1)+r0l(Ceffsubtree2+0.5c0l)]/[r0l(Ceffsubtree1+c0l+Ceffsubtree2)]
In case of (x ≤ 0 or x ≥ 1), elongation would be needed. Elongation is the process of adding
extra wire length to the sub-tree which has less effective capacitance, in order to equalize the
delay of both sub-trees. The length of elongation to maintain zero skew is given by:

L’=[-20r0Ceffsubtree2+2(100r02Ceffsubtree22+270r0c0(tint2-tint1))0.5]/[20r0c0]
This methodology is applied for zero skew routing in DCDNs. The results are given in
Section 5.1 will validate the efficiency of this methodology.

4.2 Parallel synthesis of DCDNs

CDN synthesis is one of the primary time-consuming steps, performed in the synthesis flow
of VLSI systems. Especially with the growth of complex SoCs in current advanced
technologies, this part has become more complicated and less computational cost-effective.
Many efforts have been put into parallel computer aided design, all with the goal of
reducing the computation time. In literature, methodologies have been proposed for parallel
synthesis of CDN such as the ones proposed in (Banerjee et al, 1992; Banerjee, 1994). These
methods however, focus mainly on the single-ended clock tree structure. In this section, the
goal is to leverage distinctive features of parallel computation to reduce computational time
required to synthesize DCDNs (Zarrabi et al, 2007). The methodology utilizes and extends
the technique proposed in Section 4.1, to synthesize zero skew DCDNs in parallel. This is a
flexible methodology, applicable to symmetric/asymmetric and hybrid (differential and/or
single-ended) clock tree structures.
On the Efficient Design & Synthesis of Differential Clock Distribution Networks 343

P0 P1 P0 P1 P
Clock source

Clock-root
Clock-sink
of the
partition

P3 P2 P3 P2

Fig. 10. Parallel DCDN distribution: a) partitioning the die area into sub-regions, b) locating
the clock-root of each region, c) finding the source of the clock network

The methodology for parallel synthesis of zero skew DCDNs is as follows. Initially the total
chip area is partitioned into sub-regions (partitioning phase). Later, synthesis of zero skew
differential clock distribution networks is performed on each of the partitioned regions
(local clock distribution phase). In the final stage, the global differential clock network is
routed for each of the previously-extracted clock-roots of the sub-regions (global clock
distribution phase). The obtained source of the clock network can end up anywhere in the
whole chip area (Manhattan surface), regardless of the initial partitioning. The proposed
scenario is illustrated in Figure 10. The proposed method may be implemented using C++
language and the Message Passing Interface (MPI) platform (MPI). A pseudo-code
describing the method is given in Figure 11.
A possible negative side effect of parallel synthesis is the increase in the total wire-length in
the clock network. This could be interpreted as the impact of multi-stage distribution of the
clock network which results in initial local zero-skew clock networks and a final global clock
network routed on top of regional clock networks. In general, this parallel processing
approach results in a clock-tree different from the one routed in a single step, due to die area
partitioning; thus, the characteristics of the new clock tree such as total wire-length and
skew may be slightly different. This proposed methodology is flexible, as it allows having a
hybrid (differential and/or single-ended) distribution of the clock network. The global CDN
could be differential, while the local (lower levels) CDNs could be single-ended to alleviate
routing complexity. It is possible to enhance the global/local distribution algorithm with
refined interconnect models. This methodology is also applicable to all
symmetric/asymmetric clock-trees.

Parallel Zero Skew Differential Clock Distribution

(Clock-sinks, Number of Processing-nodes)
{
1. Partition chip area according to the number of processing nodes.
2. Apply ‘local’ zero skew (differential) clock distribution to the partitioned
areas and send the clock-tree root(s) to the root processing node.
3. Receive the processed clock-tree root(s) from processing nodes, and, apply
‘global’ zero skew (differential) clock distribution.
4. Return the obtained final clock-tree root as the source of the (differential)
clock distribution network.
}
Fig. 11. Pseudo-code for parallel synthesis of zero skew differential clock distribution
344 VLSI

5. Results
In this section, the quantitative results related to the given design and synthesis methods for
DCDNs are given.

5.1 Zero skew routing

The zero skew routing method, inspired by Tsay’s algorithm and given in Section 4.1, was
applied to IBM benchmarks r1…r5 (Tsay, 1991). Modified Tsay’s method (for differential
signal integrity) was implemented using C++ for clock routing, and PERL language for
netlist manipulation was utilized for design and simulations. HSPICE simulation results for
the proposed method are tabulated in Table 2. The delay and skew results presented
throughout this work represent the average and absolute difference of clock signal phase
delay at sink nodes respectively.
Two methodologies were used for routing differential lines: Single-Spaced (SS) and Double-
Spaced (DS) routing. In single-spaced routing scheme, the mutual coupling effects are
stronger, therefore differential characteristics of the pair is more dominant. DS offers smaller
mutual coupling, consequently this reduces delay while degrading noise immunity. This is
due to the fact that the stronger the coupling effect is, the stronger common-mode noise
rejection becomes.
Table 2 demonstrates that, on average, clock trees generated with the proposed model show
97% skew reduction compared to those obtained using Elmore model, i.e. neglecting
coupling effects. This improvement is achieved because coupling effects of differential lines
are more accurately considered in the algorithm leading to tapping point selection. As
technology advances the coupling effects increase and we are no longer able to neglect these
effects in system modeling. In this case, neglecting the coupling effects, results in the
misplacement of tapping points and reduces the effectiveness of the considered zero skew
DCDNs. Simulation results also show smaller delay and skew for the DS scheme due to
reduced coupling, however this design strategy as we will see degrades robustness in
presence of external noise.

Single Spaced (SS) Single Spaced (SS) Double Spaced (DS)

Bench
Elmore Proposed Proposed Elmore Proposed Proposed
Mark
Skew (ps) Delay (ns) Skew (ps) Delay (ns) Skew (ps) Delay (ns)
r1 115 1.1 5.9 1.1 2.4 0.9
r2 199 3.8 15 3.7 8.8 3.3
r3 341 5.1 14 5.1 8.0 4.6
r4 759 17.5 36 17.1 20 14.5
r5 1825 34 51 34 39 28
Table 2. Skew and delay of DCDN for r1-r5 benchmarks in 180nm technology
On the Efficient Design & Synthesis of Differential Clock Distribution Networks 345

5.2 Applying DT buffers

Buffers were inserted to improve the performance of benchmark clock networks. The buffer
insertion procedure is as follows. Low-swing buffers were inserted at branching points and
level converting buffers were inserted at sinks. Low-swing buffers are sized to reduce
propagation delay throughout the clock network. To accomplish this, a base size for buffer
according to its delay-power characteristic diagram was chosen to drive a unit length
interconnect segment. Further, the buffers were uniformly sized according to the longest
interconnect. Full-swing level converters (differential to single-ended) were composed of
minimum size transistors to reduce the power consumption and skew reduction; their sizes
were scaled up relative to their load capacitance.

Low-Swing Low-Swing Full-Swing

Bench Conventional Proposed Conventional
Mark
Skew (ps) Delay (ns) Skew (ps) Delay (ns) Skew (ps) Delay (ns)
r1 695 9.6 545 7.0 71 1.1
r2 844 10 679 7.7 163 1.3
r3 808 10 667 7.8 127 1.3
r4 1388 11.9 981 8.8 379 1.6
r5 1566 13.1 1135 9.6 532 1.9
Table 3. Buffered DCDNs (Full-Swing Vdd=1.8V, Low-Swing Vdd=0.5V)

Table 3 shows the skew and delay difference for similarly sized, buffered DCDN based on
the conventional (Dally et al, 1998) and the proposed buffers. It shows 25% delay and skew
improvement on average compared to conventional buffers in low-swing differential
clocking scheme. Results show that the delays are reduced significantly while skews are
degraded as compared with un-buffered DCDNs. It is believed that skews in buffered clock
networks can be reduced significantly by enhancing the process of buffer insertion. For
instance, differential buffers delay model should be considered when tapping points are
selected in the zero skew DCDN design algorithm.

Skew Variations for Different CDNs in presence of Crosstalk

12
Low Swing
Full Swing
10
% Skew Variations

0
Single Differential (SS) Differential (DS)
Clocking Schemes

Fig. 12. Skew variations due to crosstalk

With regards to the skew sensitivity of the proposed DT DCDNs, two types of external
aggressors, resulting into random skew are investigated: power-supply variations and
crosstalk. Comparisons were made between similarly designed single-node CDN, single-
346 VLSI

spaced DCDNs and double-spaced DCDNs (Figure 12). Benchmark r3 is used for
simulations due to its average characteristic in terms of size and simulation time. For the
low-swing scheme, the power supplies were VddH=1.8V & VddL=0.5V, whereas for full-swing
scheme a single supply voltage (VddH=1.8V) was used. In those experiments, supply-
voltages were varied by ±10%.
Simulation results show that for both clocking schemes, the single-spaced DCDN is the most
robust design method in the presence of power-supply variations when compared to other
CDNs. Skew variations increase when low-swing clocking is used. Double-spaced DCDN
has less robustness to supply variations. DCDN is seen to have up to 25% less skew
variations in low-swing clocking scheme and up to 9% less skew variations in full-swing
clocking scheme than single-node CDN, in presence of power-supply variations.
Another source of perturbation that causes delay uncertainty in CDNs is crosstalk. For
experiments, a full-swing aggressor is applied to one of the two big-child of the clock tree.
The same low-swing and full-swing clocking schemes were considered. Simulation results
show that the single-spaced DCDN shows 6% less skew variations when combined with low
swing clocking scheme and 9% when combined with full swing clocking scheme as
compared to single-node CDN subject to crosstalk.

5.3 Applying low-power buffers

In this part, the effect of employing low-power buffers introduced in Section 3.2 is studied.
For comparison and performance evaluation of clock networks, a set of 400 MHz CDNs
were designed for benchmark r3 due to its average size suitable for simulations. All reported
designs were minimally sized to meet the target operating frequency (400MHz) based on
180nm technology parameters. In DCDNs, the differential signal swing was scaled by
adjusting the tail current source of intermediate differential buffers. The lowest potential
reached by either part of the differential signal is Vdd-RIss where Vdd is the supply voltage, Iss
is the tail current and R is the equivalent resistance of the transistor loads. Note that, the
load resistance determines the bandwidth of the clock network; hence the only possible
variable to tune is the tail current to scale the differential voltage swing. In the following, the
effect of differential voltage scaling on power consumption and clock skew variations in the
presence of power supply variations is explored.

Skew Variations for Different Low Swing CDNs Skew Variations for Different Full Swing CDNs
40
Single Single
20 Differential (SS) Differential (SS)
Differential (DS) Differential (DS)
10
0
% Skew Variations
% Skew Variations

-20 5

-40

0
-60

-80
1.62-0.45 1.79-0.475 1.8-0.5 1.89-0.525 1.98-0.55 1.62 1.71 1.8 1.89 1.98
Voltage Supply VddH -VddL (V) Voltage Supply (V)

Fig. 13. Skew due to supply-voltage variations in low/full -swing schemes

On the Efficient Design & Synthesis of Differential Clock Distribution Networks 347

5.3.1 Effect on power consumption

The tail current affects both the short circuit and dynamic power dissipation. Figure 14
demonstrates the effect of differential voltage scaling on the power consumption of DCDN
and its single-node CDN counterpart both optimized for the same benchmark; which is
obtained by scaling the tail current source of the differential buffer.
Figure 14 shows the power consumption obtained with three different clocking schemes:
differential CDNs with composite-load buffers, differential CDNs with only grounded-gate
load transistor buffers and the conventional single-node clocking scheme. The reason for
exploring the network with only grounded-gate load transistor buffers is to demonstrate the
effectiveness of composite-load buffers in terms of reduced sensitivity when large signal
swings are considered. In this case, the size of the load transistors is increased (by
approximately 20%) to get the same amplitude for a given tail current. The CDN used as
reference is a conventional balanced single-node clocking tree, where transistors have the
minimal size necessary to give full output swing while operating at the target frequency
(400MHz).

4.5 DCDN Composite Load

DCDN GND Load
4 Single
Power Consumption (W)

3.5

2.5

1.5

10 15 20 25 30 35 40 45 50
Differential Pair Swing Scaling (%Vdd)

Fig. 14. Power consumption for voltage scaled DCDNs vs. a single-node CDN (r3).

Figure 14 shows that as the differential swing increases beyond 25% of supply-voltage
(450mV, Vdd=1.8V), the power consumption increases drastically. This emphasizes the
significant impact of large differential voltage swings on the power consumption of the
clock network. For differential voltage swings below 450mV, the power consumption is not
reduced much if the differential swing scaling is further reduced. A lower bound of 10% of
Vdd was imposed to the differential swing to ensure a sufficient noise margin. Another
consideration is that even with a differential swing as low as 10% of Vdd (180mV), the power
consumption of the differential clock network remains almost 30% higher than that of
single-node clock distribution network. Thus, trying to match the power dissipation of a
single-node network by decreasing the swing of differential networks does not appear to be
a viable option. A final observation from Figure 14 is that the DCDN with grounded-gate
loads (GND) consumes less power over a limited region where the differential swing is
large. However, as we will see in the following, this slight reduction in dissipated power
comes at a large price in clock skew variability.
348 VLSI

5.3.2 Supply-voltage scaling

In the previous section, we observed that the reduction of the differential swing has a strong
impact on the power consumption. For differential swings smaller than 25% of Vdd, the
power consumption becomes less than half of that observed when the clock network
operates with the voltage swing of 50% of Vdd. Second, as Figure 15 suggests, due to the
fluctuations of common-mode voltage in the clock network, the voltage swing of 25% of Vdd
(450mV) shows the least skew when the clock network is subject to supply variations. We
also observe that DCDN with composite load is more robust. This is due to the linear
characteristic of composite load buffers, as was seen in Section 3.2.

350

300 DCDN Composite Load

Differential GND Load

250
Max Skew Variations (%)

200

150

100

50

0
10 15 20 25 30 35 40 45 50
Differential Pair Swing Scaling (%Vdd)

Fig. 15. Peak to peak skew variations in differential voltage scaled DCDNs.

Taking these considerations into account, we consider a design point for which the
differential swing is 25% of Vdd (450mV) and we reduce the supply voltage to the point
where we reach the same power consumption as that observed for the single-node and
differential clock networks. HSPICE simulations demonstrate that for a supply voltage of
1.4V and differential swing 450mV, we obtain the same power consumption for the
differential clock network. Yet, as can be observed from Figure 16, the variation of clock
skew is still less than that of the comparable single-node CDN. Another interesting point
that was observed during supply voltage scaling in DCDN is a negligible signal latency
difference. This can be justified since as the tail current is lowered to achieve lower
differential swings; the necessary differential voltage needed for the differential buffer to
switch is also decreased. This enables the differential buffer to operate/switch faster than in
the case where greater supply voltage with greater differential swing is used. Also as
observed from Figure 16 and as discussed previously, the DCDN based on only grounded-
gate loads is less resilient.
On the Efficient Design & Synthesis of Differential Clock Distribution Networks 349

Skew Variations for Different Clocking Schemes @ 400 MHz Frequency

Single
Differential (1.8V) Composite
Differential (1.4V) Composite
10
Differential (1.4V) GND Load

% Skew Variations
Less Power
Consumption
5

Less Varations

-10% -5% 0 5% 10%

Voltage Supply Variations (V)

Fig. 16. HSPICE simulation show less variations in DCDN compared to single-node CDN
for equal nominal power consumption.

5.4 Parallel zero skew routing

Benchmarks from (Tsay, 1991) with as many as 3101 clock sinks and total area as large as 1.4
cm * 1.4 cm were chosen. The reported results are based on 180nm CMOS technology file,
for which interconnect parameters are: Cg= 8e-17(F/um), Cc= 8e-17(F/um) and
Rint=0.022(Ω/um). The processing time, speed up and resulting simulated skews when
synthesizing zero skew DCDNs using 1 (sequential), 2 and 4 processing nodes are reported
in Table 4. These results demonstrate a nearly-linear speed-up. It is expected that for very large
benchmarks, the speed-up grows linearly when the number of sinks is sufficiently large
compared to the number of processing nodes. Thus, the processing time overhead due to
performing that synthesis in parallel was almost negligible. Figure 17 confirms these
observations for 2 and 4 processing nodes. As the number of clock sink nodes increases, the
speed-up (dashed lines) converges to the maximum expected speed-up (rigid lines).

Benchmark Skew (ps) Computation Time (s) Speed-Up

(# of sinks) 1 PN 2 PN 4 PN 1 PN 2 PN 4 PN 1 PN 2 PN 4 PN
r3
14 12 14 0.70 0.48 0.21 1.0 1.45 3.24
(862)
r4
36 35 36 1.60 0.90 0.46 1.0 1.76 3.46
(1903)
r5
51 49 52 2.74 1.45 0.73 1.0 1.89 3.74
(3101)
Table 4. Run-time and speed-up results of benchmarks

In general, the parallel processing approach results in a clock-tree different from the one
routed in a single step, due to die area partitioning; thus, the characteristics of the new clock
tree such as total wire-length and skew may be slightly different. This proposed
methodology is flexible, as it allows having a hybrid (differential and/or single-ended)
distribution of the clock network. The global CDN could be differential, while the local
(lower levels) CDNs could be single-ended to alleviate routing complexity. It is possible to
enhance the global/local distribution algorithm with refined interconnect models. This
methodology is also applicable to all symmetric/asymmetric clock-trees.
350 VLSI

3.5

Speed-up
2.5

1.5

1
1000 1500 2000 2500 3000 3500
Number of clock-sinks

Fig. 17. Speed-up approaches its maximum, as the size of clock network increases, for 2 and
4 processing node synthesis cases.

6. Conclusions
In this chapter, some techniques for efficient design and synthesis of on-chip Differential
Clock Distribution Networks (DCDNs) were given.
Initially design techniques were proposed that improve the performance of differential
buffers which result into the performance improvement of DCDNs. This was achieved by
means of introducing configurations for differential buffers based on Dynamic Threshold
(DT) transistors. It was shown that for low supply-voltages, they outperform the
conventional buffers with 25% delay reduction. Also, in order to overcome the high power
consumption of DCDNs, a circuit configuration was proposed by which it is possible to
reduce the differential voltage swings (down to 10% of Vdd) which reduces the power
consumption significantly (30% more than single-node CDN). Furthermore, by scaling the
supply voltage of the system from 1.8V to 1.4V, we reach a design point where the DCDN
consumes the same power as its single-node CDN counterpart but has less variation (in
terms of skew). This however comes at the expense of delay and reduced voltage swing.
Various synthesis techniques were introduced that improve the DCDNs routing to achieve
low (and possibly zero) skew. For this, a line equivalent delay model was suggested by
which it is possible to route DCDNs with low (zero) skew. On average, 97% skew reduction
was obtained utilizing this model compared to the classic Elmore delay model. A
methodology for parallel distribution (routing) of zero skew DCDNs was also proposed.
The method is applicable to all symmetric/asymmetric clock networks with ability for
hybrid implementation (differential and/or single-ended). The proposed method alleviates
the problem of high computational cost of such CDNs in complex VLSI systems. Utilizing
this method, nearly-linear speed-up is achieved for zero skew DCDNs.
In the hierarchy of CDNs in modern high-performance complex systems, DCDNs can be
effectively fit in the global level of CDNs; yet they can be used as the sole solution to the
clock distribution of the system, when noise is the main design issue.
On the Efficient Design & Synthesis of Differential Clock Distribution Networks 351

7. References
Anderson, F. E.; Wells, J. S. & Berta, E. Z. (2002). The core clock system on the next
generation Itanium microprocessor, in ISSCC Digest of Technical Papers, pp. 146-7.
Assaderaghi, F.; Sinitsky, D.; Parke, S.A.; Bokor, J.; Ko, P.K. & Hu, Chenming. (1997).
Dynamic threshold-voltage MOSFET (DTMOS) for ultra-low voltage VLSI”, IEEE
Transactions on Electron Devices, Volume 44, Issue 3, pp.414 – 422.
Banerjee, Prithviraj. & Xing, Zhaoyun. (1992). A parallel algorithm for zero skew clock tree
routing, International Symposium on Physical Design, pp. 118 – 123.
Banerjee, Prithviraj. (1994). Parallel Algorithms for VLSI Computer-Aided Design, PTR
Prentice Hall, Englewood Cliffs, New Jersey 07632.
Broderson, Bob. (2005). Analog Integrated Circuits, online material, available:
http://bwrc.eecs.berkeley.edu/People/Faculty/rb/.
Chappell, B.A.; Chappell, T.I.; Schuster, S.E.; Segmuller, H.M.; Allan, J.W.; Franch, R.L. &
Restle, P.J. (1988). Fast CMOS ECL receivers with 100-mV worst-case sensitivity,
IEEE JSSC, Volume 23, Issue 1, pp:59 – 67.
Cong, J.; He, L.; Koh, C. K. & Madden, P. (1996). Performance Optimization of VLSI
Interconnect Layout, Integration, the VLSI Journal, vol. 21, pp. 1-94.
Dally, William J. & Poulton, John. (1998). Digital Systems Engineering, Cambridge
University Press.
Hall, S.H.; Hall G.W. & McCall, J.A. (2000). High-Speed Digital system Design, A Handbook
of Interconnect theory and Design Practices. John Wiley & Sons INC.
Herzel, F.; Razavi, B. (1999). A study of oscillator jitter due to supply and substrate noise, IEEE J.
Circuits and Systems, Volume 46, pp. 56 - 62.
Kahng, A.B.; Muddu, S.; Sarto, E. (2000). On switch factor based analysis of coupled RC
interconnects, Design Automation Conference, pp. 79 – 84.
MPI, Message Passing Interface, online: http://www.mpi-forum.org/.
O’Mahony, Frank P. (2003). 10GHz Global Clock Distribution Using Coupled Standing-
wave Oscillators, PhD Dissertation, Stanford University.
Razavi, Behzad. (2001). Design of Analog CMOS Integrated Circuits. Mc Graw Hill.
Restle, P. J. & A. Deutsch (1998). Designing the best clock distribution network, in
Symposium VLSI Circuits Digest of Technical Papers.
Restle, P.J.; McNamara, T.G.; Webber, D.A.; Camporese, P.J.; Eng, K.F.; Jenkins, K.A.; Allen,
D.H.; Rohn, M.J.; Quaranta, M.P.; Boerstler, D.W.; Alpert, C.J.; Carter, C.A.; Bailey,
R.N.; Petrovick, J.G.; Krauter, B.L. & McCredie, B.D (2001). A clock distribution
network for microprocessors, IEEE J. Solid-State Circuits, vol. 36, no.5, pp. 792-799.
Sekar, D.C. (2005). Clock trees: differential or single ended? , International Symposium on
Quality of Electronic Design, pp.548 – 553.
Tsay, R. S. (1991). Exact zero skew, in Proc. IEEE Int. Conf. Computer-Aided Design, pp.
336–339, Nov.
Vasseghi, N.; Yeager, K.; Sarto, E. & Seddighnezhad, M. (1996). 200-MHz superscalar RISC
microprocessor, IEEE J. Solid-State Circuits, vol. 31, no. 11, pp. 1675-1685.
Wikipedia, online: http://en.wikipedia.org/wiki/Phase-locked_loop
Zarrabi, Houman. (2006). On the design and synthesis of differential clock distribution
network, MASc Dissertation, Concordia University.
352 VLSI

Zarrabi, Houman; Saaied, Haydar; Al-Khalili, A. J. & Savaria, Yvon. (2006). Zero Skew
Differential Clock Distribution Network, International Symposium on Circuit And
Systems (ISCAS), Greece, Island of Kos.
Zarrabi, Houman; Zilic, Zeljko; Al-Khalili, A. J. & Savaria, Yvon. (2007). A methodology for
parallel synthesis of zero skew differential clock distribution networks, Joint
Conference of MWSCAS/NEWCAS, Montreal, Canada.
Zhang, H.; Varghese George & Rabaey, J. M. (2000). Low-swing on-chip signaling
techniques: effectiveness and robustness, IEEE Trans. on VLSI Syst., Volume 8,
Issue 3, pp. 264 – 272.
Robust Design and Test of Analog/Mixed-Signal Circuits in Deeply Scaled CMOS Technologies 353

18
X

Robust Design and Test

of Analog/Mixed-Signal Circuits in
Deeply Scaled CMOS Technologies
Guo Yu and Peng Li
Texas A&M University
College Station, TX, USA

1. Introduction
The proliferation of communication and consumer electronic systems leads to the large
demands of high-performance & robust analog/mixed-signal circuits. On the other hand,
although deeply scaled CMOS technologies enable greater degrees of semiconductor
integration and lower manufacturing cost, the advancements of technologies also introduce
several new challenges for VLSI circuit design. The increasing parametric variations and
their impacts on circuit performances are becoming key issues which make already complex
circuits even more sophisticated in design practice (Nassif, 2001). Give these barriers,
designing robust mixed-signal circuits in deeply scaled CMOS technologies become a real
challenge for circuit designers.

In this chapter, we propose to solve the problems by first introducing efficient and accurate
modeling techniques for large analog/mixed-signal circuit designs with consideration of
process variations. Powerful statistical dimension reduction techniques are utilized to make
performance modeling of large circuits possible. Then these novel circuit models are used to
achieve efficient parametric system performance analysis. In such ways the designers can
have the full knowledge of real circuit performances under process variations, which makes
it feasible to do robust system topology design and circuit optimization for large mixed-
signal systems. The efficient modeling framework is also extended for circuit test purposes,
which leads to robust Built-in Self-Test (BIST) circuit design and optimization.

We demonstrate the effectiveness of the proposed ideas with two popular types of mixed-
signal circuit examples, Sigma-Delta A/D converters and Phase Looked Loops (PLLs).

For Sigma-Delta ADCs, we present a novel parameterized lookup table (LUT) technique for
capturing performances of building blocks in the systems, and use these LUTs to perform
topology trade-off analysis and system optimization. Modeling of circuit level
nonlinearities, adaptive LUT generation and robust system design are explained in detail
with comprehensive experimental results (Yu & Li, 2006 & 2007a).
354 VLSI

For PLLs, we discuss building parametric Verilog-A models for charge-pump PLLs and use
these models for high-level performance analysis. In order to handle large number of
parametric variables, dimension reduction technique is applied to reduce simulation
complexities. We apply the obtained system simulation framework to evaluate the
efficiencies of parametric failure detecting of different BIST circuits and perform
optimization based on the experimental results (Yu & Li, 2007b).

2. Robust Sigma-Delta ADCs Design

Sigma-Delta ADCs have been widely used in data conversion applications due to the good
resolution. However, oversampling and complex circuit behaviors render the transistor-
level analysis of these designs prohibitively time consuming (Norsworthy & et al., 1997).
The inefficiency of the standard simulation approach also rules out the possibility of
analyzing the impacts of a multitude of environmental and process variations critical in
modern VLSI technologies. We propose a lookup table (LUT) based modelling technique to
facilitate much more efficient performance analysis of Sigma-Delta ADCs. Various
transistor-level circuit nonidealities are systematically characterized at the building block
level and the whole system is simulated much more efficiently using these building block
models. Our approach can provide up to four orders of magnitude runtime speedup over
SPICE-like simulators, hence significantly shortening the CPU time required for evaluating
system performances such as SNDR (signal-to-noise-and-distortion-ratio). The proposed
modeling technique is further extended to enable scalable performance variation analysis of
complex Sigma-Delta ADC designs. Such approach allows us to perform trade-off analysis
of various topologies considering not only nominal performances but also their variabilities.

2.1 Background of Sigma-Delta ADC design

We briefly discuss the background and difficulties for Sigma-Delta ADC design in this
section. Various important circuit nonidealities, which are difficult to model accurately
using analytical equations, are also discussed. The two basic components of Sigma-Delta
ADCs are modulators and digital filters as shown in Fig. 1. The analog input of the ADC is
sampled by a very high frequency clock in the modulator, then the sampled signal is passed
through the loop filter to perform noise-shaping. The output of the loop-filter is quantized
by an internal A/D converter, producing a bit-stream at the same speed as the sampling
clock. The low-pass digital filter in the decimator then removes the out-of-band noise, and
the down-sampler converts the high speed bit-stream to high resolution digital codes.

H ( z)

Fig. 1. Block diagram of a Sigma-Delta ADC

Robust Design and Test of Analog/Mixed-Signal Circuits in Deeply Scaled CMOS Technologies 355

The difference between the ideal digital output of the quantizer and the actual analog signal
is called quantization noise. The goal of Sigma-Delta technique is to eliminate this unwanted
quantization noise as much as possible. By oversampling the input signal, the modulator
moves the majority of the quantization noise out of the signal bandwidth. The principle of
noise-shaping can be analyzed using transfer functions, which can be obtained using a
linear model in the frequency domain. The quantization noise E(z) is modelled as additive
noise to the quantizer, the output of the quantizer Y(z) can be written as
H ( z) 1 (1)
Y ( z)   X ( z)   E ( z)
1  d  H ( z) 1  d  H ( z)
where d is the feedback gain of the DAC, X(z) is the input signal, H(z) is the transfer
function of the loop filter. By configuring H(z) and d, we can have different noise shaping
functions so that the signal-to-noise ratio in the output can be optimized.

Sigma-Delta ADC performances are greatly impacted by various circuit-level nonidealities,

such as the finite DC gain, slew of the operational amplifies, charge injections of the
switches, mismatch of the internal quantizers and D/A converters. The effects are difficult
to analyze accurately by hand calculation or simple analytical models, neither are their
impacts on system performances. There exist several high-level simulators for Sigma-Delta
ADC design, like MIDAS (Babii et al., 1997) and SWITCAP (Fang & Suyama, 1992). These
techniques are only suitable for architecture-level exploration and determination of building
block specifications in the early design phase, but no suitable for the consideration of circuit-
level non-idealities. Often the time, transistor-level simulation is the only choice for current
performance evaluation, which may take a few weeks for a single transient simulation.
Analyzing the impact of process variations is an even greater challenge because a large
number of long transient simulations are needed to derive the performance statistics under
process variations. In the following sections, we address these issues by adopting the lookup
table based models, which can handle the circuit-level details and process variation induced
performance statistics accurately and efficiently.

2.2 Performance modeling with Lookup table (LUT) technique

The need of efficient simulation techniques capable of including transistor-level details is
particularly pressing for assessing the impact of process variations, where the analysis
complexity tends to explode in a large parameter space. We start with modelling of
switched-capacitor type Sigma-Delta ADCs, which are clocked by some global sampling
signals. The major components in the converters are integrators, quantizers and feedback
DACs.

Since the switched-capacitor integrators in the discrete-time Sigma-Delta modulators are

clocked by the sampling clock as shown in Fig. 2, it is possible to use lookup tables to
represent the nonlinear state transfer function of the system (Bishop et al., 1990, Brauns et
al., 1990). The output of integrator can be expressed as a function of input signals and their
previous states as in (2)
y[k  1]  F ( y[k ], x[k  1], d [k  1]) (2)
where y[k+1] is the current output of the integrator, y[k] is the integrator output in the
previous clock cycle, x[k+1] and d[k+1] are the current input signal and the digital feedback
356 VLSI

output of the DAC, respectively. This property of Sigma-Delta ADCs makes it possible to
predict the new integrator output using the previous state and the new input. The previous
state of the integrator, the digital feedback and the new analog input are discretized at a set
of discrete voltage levels that are used as the indices to the lookup table models.

y[k+1]=F(y[k],x[k+1],d[k+1])
Fig. 2. Integrator behaviours under clocking

As illustrated in equation (2) and Fig. 2, the output of an integrator is a function of the input
signals and the initial state of the integrator, which are discretized to generate the lookup
table entries. The number of discretization levels depends on the accuracy requirement of
the simulation. The internal circuit node voltage swings can be estimated by the system
architecture. For low-voltage Sigma-Delta ADC designs, the internal voltages can change
from 0 to supply voltage Vdd. To cover the whole range of voltage swing, we discretize the
inputs and outputs of the integrators uniformly at N levels from 0 to Vdd, where N is in the
range of 10. The extraction setup for an integrator with a multi-bit DAC implemented in
thermometer code is shown in Fig.3. A large inductor L together with a voltage source Vs is
used to set the initial value of the integrator output. The input of the integrator is also set by
a voltage source Vi. The digital output of the quantizer controls the amount of charge to be
fed back. An m-bit DAC implemented in thermometer code has 2m - 1 threshold voltages.
The digital codes from 0 to 2m - 1 can be represented by counting the number of voltage
sources that are connected to the integrator inputs from a set of voltages sources Vd1, Vd2, …,
Vd(2m-1),the voltages of which are set to be either digital “1” or digital “0”.

Fig. 3. LUT generation setup for integrators

Robust Design and Test of Analog/Mixed-Signal Circuits in Deeply Scaled CMOS Technologies 357

The nonlinearities of quantizers can be captured using lookup tables as well. The quantizer
acts as a comparator, the input threshold voltage varies depending on the direction in which
the input voltage changes. To capture the hysteresis effect accurately, we use the transistor-
level simulation to find the input threshold voltages at which the digital output switches
from 0 to 1 ( Voff  ) and from 1 to 0 ( Voff  ), respectively. The quantizer is then modeled as

1 (Vin [ k  1]  Voff  )
 (3)
d [ k  1]   d [k ] (Voff   Vin [k  1]  Voff  )

0 (Vin [k  1]  Voff  )
where d[k+1] is the new output of the quantizer, d[k] is the output in the previous clock
cycle. Multi-bit quantizers can be modeled in a similar way since they are built from several
1-bit quantizers.

Sigma-Delta ADCs with continous-time modulators can also be modeled using the
proposed technique with minor modificiation. Continuous-time Sigma-Delta ADCs are
different from discrete-time counterparts since the integrators are not clocked by the
sampling clock, and the input and the output of integrator changes throughout a clock
period. In order to make the lookup-based modeling possible, we discretize each clock cycle
into M time intervals with a step size dT=T/M. If dT is small enough, then in each small
time period the behaviours of continous-time modulators can be approximated using the
presented technique, detailed implementation for continous-time Sigma-Delta ADCs can be
found in (Yu & Li, 2007a).

2.3 Parametric LUT based modeling

Process variations in the fabrication stage can cause significant performance shift for analog
and mixed-signal circuits. So handling of process variations in the early design stage is
critical for robust analog/mixed-signal design. In order capture the influence of process
variations, we extract parameterized LUT models and perform fast statistical simulation to
evaluate the performance distributions of complex Sigma-Delta ADCs. Using this efficient
modeling technique, we have the ability to find the most suitable system topologies and
design parameters for ADC designs.

Since the number of process variables is large, it is not possible to exhaust all the possible
performances under process variations. Here we use parameterized LUT-based models to
capture the impacts of circuit parametric variations that include both environmental and
process variations. In this case, a nonlinear regression model (macromodel) is extracted for
each table entry. In general, a macromodel correlating the input variables and their
responses can be stated as follows:
given n sets observed responses {y1,y2,…,yn} and n sets of m input variables [x1,x2,…,xm], we
can determine a function to relate input x and response y as (Low & Director, 1989)
358 VLSI

y1  h( x11 , x12 , , x1m )

y2  h( x21 , x22 , , x2 m ) (4)

yn  h( xn1 , xn 2 , , xnm )
where
yi i-th response
h function relating y and x
xi i-th set of process variables
m number of process variables
n number of experimental runs

The task of constructing each macromodel is achieved by applying the response surface
modeling (RSM) technique where empirical polynomial regression models relating the
inputs and their outputs are extracted by performing nonlinear least square fitting over a
chosen set of input and output data (Box et al., 2005). To systematically control the model
accuracy and cost, design of experiment (DOE) technique is applied to properly choose a
smallest set of data points while satisfying a given modeling regulation. For our circuit
modeling task, the input parameters are the parametric circuit variations and the output is
an entry in the lookup tables. Then, a nonlinear function such as a quadratic function
relating each entry in the tables with the circuit parametric variations can be determined via
regression
^ ^ m ^ m m ^
y   0   i xi   ij xi x j (5)
i 1 i 1 j 1
where
xi i-th process variable set
Y approximated response
 estimated model fitting coefficient
m number of process variables

Equation (5) can be rewritten in a more compact matrix form as

X  Y (6)

The fitting coefficient vector  can be calculated using the least-square fitting of
experimental data as
  ( X T X )1 X T Y (7)

A major problem in solving equations (5 – 7) is the number of experimental data, so a

second-order central composite plan consisting of a cube design sub-plan and a star design
sub-plan is employed (Box & et al., 2005). The cube design plan is a two-level fractional
factorial plan that can be used to estimate first-order effects (e.g., xi) and interaction effects
(e.g., xi_xj), but it is not possible to estimate pure quadratic terms (e.g., xi2). The star design
plan is used as a supplementary training set to provide pure quadratic terms in equation (5).
Robust Design and Test of Analog/Mixed-Signal Circuits in Deeply Scaled CMOS Technologies 359

In our implementation, the cube design plan is selected in order to estimate all the first-
order and cross-factor second-order coefficient of the input variables.

The ranges of all parametric variations are usually obtained from the process
characterization. This information is used to setup the model extraction procedure. In the
cube design plan, each factor takes on two values -1 and +1 to represent the minimum and
the maximum values of the parametric variation. Each factor in the star plan takes on three
levels -a, 0, a, where 0 represents the nominal condition and the level range |a| < 1. As
illustrated in Fig. 4, for each point (i,j) in the lookup table, n simulations runs are conducted
using fractional factorial plan to provide the required data to generate the regression model
in equation (5). As long as the lookup tables for specified process variation distributions are
generated, we can perform fast system-level simulation to evaluate the performances under
process variations, and in turn to achieve optimization as to be discussed in the following
section.

^ ^ m ^ m m ^
y   0    i xi   ij xi x j
i 1 i 1 j 1

Fig. 4. Response surface modelling of parameterized LUTs

2.4 System optimization using parametric LUTs

The application of the proposed modeling techniques are demonstrated with three discrete-
time Sigma-Delta ADC designs with different topologies including 2nd-order with 1-bit
quantizer (SDM 1), 2nd-order with 2-bit quantizer (SDM 2) and 3rd-order with 1-bit
quantizer (SDM 3). All these ADCs are implemented in 130nm CMOS technology with a
single 1.5V power supply. The sampling clock and oversampling ratio are chosen to be
1MHz and 128, respectively.

Using our parameterized LUT-based infrastructure, we are able to not only predict the
nominal case design performances but also their sensitivities to parametric variations.
Hence, our technique provides an efficient way for statistical circuit simulation as well as
performance-robustness trade-off analysis. For statistical analysis, a Resolution V 2(8-2)
fractional factorial design plan that includes 64 runs for the cube design plan and 17 runs for
the star design plan is employed by SDM 1. For SDM 2 and SDM 3, a Resolution VI 2(6-1)
360 VLSI

fractional factorial design plan with 45 runs is employed, resulting in 32 runs for the cube
design plan and 13 runs for the star design plan.

In Table 1, the proposed LUT-based simulator is compared with the transistor-level simulator
(Spectre) in terms of model extraction time, simulation time, and predicted nominal SNDR and
THD values. Once the LUT models are extracted, the LUT-based simulator can be efficiently
employed to perform statistical performance analysis, which is infeasible for the transistor-
level simulator. For the 2nd-order Sigma-Delta ADC with 1-bit quantizer, it only takes 20
minutes to conduct 1,000 LUT-based transient simulations each including 64k clock cycles. For
the same analysis, transistor-level simulation with conventional simulators is expected to take
4,500 hours to complete. In terms of accuracy, the SNDRs and THDs predicted by Spectre and
the LUT simulator are also presented in Table 1. The error of SNDR of our LUT-based
simulator is within 1dB, which demonstrates the accuracy of the proposed technique.

LUT-based simulation Spectre simulation

Design Sin. Par. gen. Runtime SNDR THD Runtime SNDR THD
gen.
SDM 1 7 min 9.5 hr 2s 73.8dB -63.1dB 4.5 hr 74.1dB -62.6dB
SDM 2 20 min 15 hr 4s 90.2dB -77.4dB 9.5 hr 90.0dB -76.8dB
SDM 3 8 min 11 hr 2s 83.3dB -67.5dB 5.5 hr 83.5dB -66.9dB
Table 1. Runtime and accuracy of the proposed LUT simulation (© [2007] IEEE, from Yu &
Li, 2007a)

With the powerful LUT-based simulator, we can perform system evaluation very efficiently so
the optimization of system topologies and detailed design become possible. First we use the
optimization of 2nd-order Sigma-Delta ADC with multi-bit quantizer as an example by
investigating the impacts of DAC capacitance mismatch. The capacitor mismatch level
decreases as the capacitance increases, so it is of interest to investigate the trade-offs of system
noise performances and area (Pelgrom & et al., 1989). Statistical simulations are performed to
analyze the influence of the mismatch of the two internal DACs by sweeping the values of the
three charging capacitors in each DAC. The variation of capacitances is modeled using a
Gaussian distribution with 3  1% . The distributions of SNDR due to the capacitance
mismatch in the two DACs are shown in Fig. 5, respectively.
40 40

35 35

SNDR without SNDR without

30 30 mismatching
mismatching

25 25
Number
Number

20 20

15 15

10 10

5 5

0 0
55 60 65 70 75 80 85 90 85 86 87 88 89 90
SNDR(dB) SNDR(dB)

(a) DAC connected to first-stage (b) DAC connected to second-stage

Fig. 5. SDNR distributions for DACs connected to different stages in SDM 2 (© [2007] IEEE,
from Yu & Li, 2007a)
Robust Design and Test of Analog/Mixed-Signal Circuits in Deeply Scaled CMOS Technologies 361

We can see from the two figures that the mismatch of the DAC connected to the first stage
integrator (left figure) has much more influence to the system performance than that of the
other DAC (right figure). This can be explained by the fact that the first stage DAC is
connected directly to the system input, so the feedback error because of the DAC mismatch
will be magnified by the second stage integrator. The result of this analysis indicates that
more attention should be paid to the first stage DAC in the design process.

Another optimization example is to evaluate the charging capacitor mismatches in SDM 3.

The mismatch of each capacitor is modelled using Gaussian distribution. We evaluate the
system performance distributions with variation of capacitances set to   1% and   5% ,
as illustrated in Fig. 6.

50 50
SNDR in
45 SNDR in 45 nominal case
nominal case
40 40

35 35

30 30
Number

25 Number 25

20 20

15 15

10 10

5 5

0 0
82.9 83 83.1 83.2 83.3 83.4 83.5 83.6 83.7 83.8 80 80.5 81 81.5 82 82.5 83 83.5 84 84.5
SNDR(dB) SNDR(dB)

(a)   1% (b)   5%
Fig. 6. SNDR distributions for mismatches of charging and sampling capacitors in SDM 3 (©
[2007] IEEE, from Yu & Li, 2007a)

We can observe from Fig. 6 that performance distribution deviations increase from 1dB to
4.5dB for capacitor variation   5% , and the impact of the mismatch of charging and
sampling capacitors is not as critical as that of the multi-bit DAC even with   5% . It is
also possible to perform more complete system analysis and optimization using the
proposed parametric LUT-based simulation method depending on the target of the design
(Yu & Li, 2007a).

3. Robust PLL Design

As an essential building block, PLLs are widely used in today's communication and digital
systems for purposes such as frequency synthesis, low-jitter clock generation, data recovery
and so on. Although the input and output signals of PLLs are in the digital domain, most
PLLs implementations consist of both digital and analog components, which make them
prone to process variation influences. In this section we propose an efficient parameter-
reduction modelling technique to capture process variations and further achieve low-cost
system performance evaluation using hierarchical system simulation. The proposed method
not only can be used for robust PLL design under process variation, but also paves the road
for effective built-in self-test circuit design as to be discussed in the next section.
362 VLSI

3.1 Background of PLL design

As illustrated in Fig. 7, a typical charge-pump PLL system consists of a frequency detector, a
charge pump, a loop filter, a voltage-controlled oscillator (VCO) and a frequency divider.
The frequency of the output clock signal Fout is N times of that of reference clock signal
Fref, where N can be an integer number or fractional number. The PLL design options
include VCO topologies and component sizes, filter characterizations, charge current in the
charge pump and so on. The metrics of PLL systems usually include acquisition/lock-in
time, output jitter, system power, total area, etc. The goal of PLL design and optimization is
to find the best overall system performances by searching in the design variable spaces.

Fig. 7. Block diagram of charge-pump PLL

Due to the mixed-signal nature, the design and optimization of PLL system is quite complex
and costly. For example, a long transient simulation (in the order of hours or days) is needed
to obtain the lock-in time behavior of PLL, which is one of the most important performance
metrics for a PLL. So the brute-force optimization by searching in the design space with
transistor-level simulation is infeasible for PLL systems.

The difficulties of system performance analysis can be addressed by adopting a bottom-up

modelling and simulation strategy. The performances of analog building blocks can be
evaluated and optimized without too much cost. When the behaviours of analog building
blocks are extracted, these building blocks can be mapped to Verilog-A models for fast
system level evaluation (Zou & et al., 2006). By using this approach we can avoid the
scalability issue associated with time consuming transistor-level simulations.

When process variations are considered, the situation becomes more sophisticated. The
large number of process variables and the correlations between different building blocks
introduce more uncertainties for PLL performance under process variations. In order to
utilize the hierarchical simulation method while taking into consideration of statistical
performance distributions, we propose an efficient macromodeling method to handle this
difficulty. The key aspect of our macromodeling techniques is the extraction of
parameterized behavioral models that can truthfully map the device-level variabilities to
variabilities at the system level, so that the influence of fabrication stage variations can be
propagated to the PLL system performances.

Parameterization can be done for each building block model as follows. First, multiple
behavioural model extractions are conducted at multiple parameter corners, possibly
following a particular design-of-experiments (DOE) plan (Box & et al., 2005). Then, a
Robust Design and Test of Analog/Mixed-Signal Circuits in Deeply Scaled CMOS Technologies 363

parameterized behavioral model is constructed by performing nonlinear regression over the

models extracted at different corners. This detailed parametric modeling step is
advantageous since it systematically maps the device-level parametric variations to each of
the behavioral models. However, difficulties arise when the number of parametric
variations is large, which leads to a prohibitively high parametric model extraction cost. We
address this challenge by applying design-specic parameter dimension reduction
techniques as described in the following section.

3.2 Hierarchical modeling for PLLs

In this section we first describe the nominal behavioral model extraction for each PLL
building block, then we discuss how a parameterized model can be constructed in the next
section.

The voltage controlled oscillator is the core component of a PLL. The two mainstream types
of VCOs are LC-tank oscillators and ring oscillators. In a typical VCO model, the dynamic
(response to input change) and static (V-Freq relation) characteristics of the voltage to
frequency transfer are modeled separately first and then combined to form the complete
model. The static VCO characteristic can be written as Fout=f(V’con), where Fout is the
output signal frequency, V’con is the delayed control voltage, and f(.) is a nonlinear
mapping relating the voltage with the frequency. To generate the analytical model, the
mapping function f(.) can be further represented by an n-th order polynomial function.
' ' 2 ' n
Fout  a0  a1Vcon  a2Vcon    anVcon (8)
where a0, a1, …, an are coefficients of the polynomial. To generate the above polynomial,
multiple VCO steady-state simulations are conducted at different control voltage levels and
a nonlinear regression is performed using the collected simulation data.

Suppose the control voltage is Vcon, the dynamic behavior of the VCO is modeled by
adding a delay element that produces a delayed version of the control voltage (V’con). The
delay element can be expressed using a linear transfer function H(s) (e.g. a second-order RC
network consisting of two R's and two C's). H(s) can be determined via transistor-level
simulation as follows: a step control voltage is applied to the VCO and the time it takes for
the VCO to reach the steady-state output frequency, or the step-input delay of the VCO, is
recorded. H(s) is then synthesized that gives the same step-input delay. The dynamic effect
is usually notable in LC VCOs due to the high-Q LC tank while in ring oscillators this effect
may be neglected.

The charge pump is mainly built with switching current sources. As illustrated in Fig. 8, the
control signals of the two switches M1 and M2 come from the outputs of the phase and
frequency detector. The currents through M1 and M2 can be turned on-and-off to provide
desired charge-up or charge-down currents. The existing charge pump macromodels are
very simplistic. Usually, both the charge-up and charge-down currents are modeled as
constant values. A constant mismatch between the two currents may also be considered
(Zou & et al., 2006). However, this simple approach is not sufficient to model the behavior of
charge pump accurately. In real implementation, the current sources are implemented using
transistors so that the actual output currents will vary according to the voltages across these
364 VLSI

MOSFETs. Therefore, the dependency of charge-up and charge-down currents on Vcon

must be considered.

Fig. 8. Modeling of charge pump (© [2007] IEEE, from Yu & Li, 2007b)

In our charge pump model, for each output current, the current vs. Vcon characteristics is
divided into two regions. When the output voltage Vcon is close to the supply voltage, then
switch M1 will be biased in triode region. The charge-up current Iup in the triode region can
be written as
W
I up   p Cox [(Vgs  Vthp )Vds  0.5Vds 2 ] (9)
L
Vds  Vdd  Von  Vcon
where Vdd is the supply voltage, Von is the on-voltage across the switch, Vgs is the gate-
source voltage,  p is the mobility, Cox is the oxide capacitance, W is width and L is the
length of M1. We can see from Equation (9) that the charge-up current is dependent on the
output voltage Vcon. We use a polynomial to explicitly model such voltage dependency
2 3
I up  b0  bV
1 con  b2Vcon  b3Vcon
(10)
where bi are the polynomial coefficients. Similarly, the charge-down current has a strong
Vcon dependency when Vcon is low. This voltage dependency is modeled in a similar
fashion. When M1 and M2 operate in saturation region, they act as part of the current
mirrors. In this case, constant output current values are assumed while the possible
mismatches between the two are considered in our Verilog-A models.

The phase detector and the frequency divider are digital circuits so that they are more
amenable to behavioral modeling. The two key parameters of the phase detector and the
frequency divider are the output signal delay and the transition time, which are easy to
extract from transistor-level simulation. The loop filters are usually comprised of passive RC
elements, which can be directly modeled in Verilog-A simulation.

3.3 Efficient parameter-reduction modeling for PLLs

The key parametric variations for transistors may include variations in mobility, gate oxide,
threshold voltage, effective length and so on (Nassif, 2001). The consideration of all possible
sources of variations in transistors and interconnects can easily lead to explosion of the
Robust Design and Test of Analog/Mixed-Signal Circuits in Deeply Scaled CMOS Technologies 365

parameter space, rendering the parametric modeling infeasible. Although the widely used
principle component analysis (Reinsel & Velu, 1998) can be adopted to perform parameter
dimension reduction, its effectiveness may be rather limited since the parameter reduction is
achieved by only considering the statistics of controlling parameters while neglecting the
important correspondence between these parameters and the circuit performances of
interest. As such, the extent to which the parameter reduction can be achieved is not
sufficient for our analog macromodeling problems. To address this difficulty, a more
powerful design-specific dimension reduction technique, which is based on reduced rank
regression (RRR), is developed. This new technique considers the crucial structural
information imposed by the design and has been shown to be quite effective for parametric
interconnecting modeling problems (Feng & et al., 2006).

Suppose we have a set of n process variations, X, and a set of N performances, Y. The

objective is to identify a smaller set of new variables Z, based on X, which are statistically
significant to the performances of interest, Y. Without loss of generality, let us assume Y
nonlinearly depends on X through a quadratic model
 X 
Y  f ( X )  C1 C2   
(11)
X  X 
where C1 and C2 are the first and second order coefficients, X  X represents the quadratic
terms of X. The combination of the linear and quadratic terms of X are then defined as a new
X   X T  X  X   . Now the quadratic model in equation (11) can be cast
T T
predictor vector 
 
into a linear model as Y  AX    . To identify the redundancy in X to facilitate parameter
reduction, we seek a reduced rank regression model in the form
Y  AR BR 
X  (12)
where AR and BR have a reduced rank of R (R < n), and BR has only R columns. We denote
the covariance matrix of 
X as Cov ( X )     , and covariance matrix between 
X and Y as
XX

Cov (Y , 
X )  Y X . It can be shown that an optimal reduced rank model (in the sense of mean
square error) is given as (Reinsel & Velu, 1998)
AR  U , BR  U T Y X  X1X (13)
where U contains R normalized eigenvectors corresponding to the R largest eigenvalues of
the matrix: D     1   . It is important to note that a successful construction of the
YX XX XY

above reduced rank model indicates that only a smaller set of R new parameters Z  BR  X
are critical to Y in a statistical sense, hence facilitating the desired parameter reduction.

It should be noted that the reduced rank regression is only employed as a means for
parameter reduction so as to reduce the complexity of the subsequent parameterized
macromodeling step. Hence, Y in the above equations does not have to be the true
performances of interest and can be just some circuit responses that are highly correlated to
the performances. This flexibility can be exploited to more efficiently collect   though
YX

Monte-Carlo sampling if Y are easier to obtain than the true performances in simulation.
366 VLSI

The complete parameterized PLL macromodel extraction flow is shown in Fig. 9. Every
building block is modeled using Verilog-A for efficient system-level simulation. Each model
parameter for building blocks is expressed as a polynomial in the underlying device-level
variations, such as
P  f (Vth1 , Leff 1 , Tox1 ,...., Vthn , Leffn , Toxn ) (14)
where Vthi, Leffi, Toxi, etc represent the parameters of i-th transistor, f(.) is the nonlinear
polynomial function to connect process variations to system performances. f(.) is very
difficult to obtain if the number of parameters is large. Hence, RRR-based parameter
reduction is applied, which leads to a set of R new parameters Z that are the most important
variations for the given circuit performances of interest. If R is small, then a new
parameterized model in terms of Z can be easily obtained through conventional nonlinear
regression for coefficients of equation (11)
C  f ( Z1 , Z 2 , Z R ) (15)

C  f ( Z1 , Z 2 , Z R )

Fig. 9. Flow of PLL macromodeling

In addition to reducing the cost of parameterized macromodeling, parameter dimension

reduction will also lead to more efficient statistical simulation of the complete PLL. This is
because instead of analyzing the design performance variations over the original high-
dimensional parameter space, statistical simulation can now be performed more efficiently
in a much lower dimensional space that carries the essential information of the design
variability, which is a very good property to achieve PLL system optimization. Detailed
optimization method for PLL using the hierarchical macromodels can be found in (Yu & Li,
2008).
Robust Design and Test of Analog/Mixed-Signal Circuits in Deeply Scaled CMOS Technologies 367

With the hierarchical models and parameter reduction technique, we can also perform built-
in self-test circuit design and optimization since lengthy transient simulations can be
relieved by the proposed method. We will discuss this part in the next section.

4. Built-in Self-test Scheme Design for PLLs

Testing of PLLs is very challenging and of great interest since: a) Usually only simple
functional tests such as phase lock test is feasible in production test. However, it may not be
sufficient for guaranteeing all the specifications; b) The operation of PLLs is intrinsically
complex, e.g. simple frequency domain analysis is not applicable for PLL circuits due to the
digitalized input/output signals and the closed-loop dynamics; c) Internal analog signals
are difcult to access outside of the chip and system specifications such as jitter, frequency
range and lock-time are very complex and costly to measure with digital testers. In this
section, we discuss the implementations and optimization of built-in self-test circuits to
capture PLL performance failures using the efficient modelling and simulation framework
in Section 3.

4.1 BIST circuit design

Built-in self-test (BIST) has emerged as a very promising test methodology for integrated
circuits although its application to mixed-signal ICs is more challenging compared to the
digital counterparts. Sunter and Roy propose a BIST scheme to measure key analog
specifications including jitter, open loop gain, lock range and lock time in PLL systems
(Sunter & Roy, 1999), while most of other proposed BIST schemes mainly focus on
catastrophic faults. Kim and Soma present an all-digital BIST scheme using charge pump as
stimulus to charge VCO up and down to detect catastrophic faults (Kim & Soma, 2001). In
(Hsu & et al., 2005) Hsu et al propose a different BIST scheme for catastrophic fault detection
in PLLs by introducing the phase errors to the inputs to the phase frequency detector. Other
BIST schemes proposed like (Azais & et al., 2003) are also targeted at catastrophic faults.

While detection of catastrophic (hard) faults remains as a key consideration in test,

parametric faults caused by the growing process variations in modern nanometer VLSI
technologies are receiving significant concerns. In most cases, chips with parametric faults
may be still functional but cannot achieve the desired performance specifications, and hence
should be screened out. These failing chips are free of catastrophic faults so that specific
BIST schemes targeting at parametric failures must be developed.

Given that process variations and resultant parametric failures will continue to rise in sub-
100-nm technologies, a design-phase PLL BIST development methodology is strongly
desired. Such methodology should facilitate systematic evaluation of parametric variations
of complex PLL system specifications and their relations to specific BIST measurements so
as to enable optimal BIST scheme development. To this end, however, three major
challenges must be addressed: a) Suitable modeling techniques must be developed in order
to enable feasible whole system PLL analysis while considering realistic device-level process
variations and mismatch; b) Device-level parametric variations and mismatch that
contribute to parametric failures form a high-dimensional parameter space and the resulting
368 VLSI

issue of curse of dimensionality brings significant modelling and analysis difficulty; and c)
Effective optimization strategies are desired in order to develop optimal BIST schemes.

With the techniques presented in section 3, the first two difficulties have been addressed,
and now we put more efforts on the optimization of BIST circuits. The most widely used
approach in BIST design is to utilize the existing digital blocks, such as to use the frequency
divider as counter and read out its state in order to detect chip failures (Sunter & Roy, 1999),
(Kim & Soma, 2001), (Hsu & et al., 2005), (Azais & et al., 2003). It is expected that the
frequency divider/counter output will change significantly if there exists a catastrophic
fault. However, parametric failures may produce smaller variations in the readout values.
Hence, they are more difficult to detect and deserve more careful treatments.

We consider three BIST schemes shown in Fig. 10 as potential candidates. Similar in spirit to
the existing BIST schemes, the main idea of the proposed BIST schemes is to control the
charge pump in a way such that the output frequency of the PLL will be altered and the
state of the frequency divider is read out at certain time instance for failure detection.
Device-level variations and mismatch will perturb the operation of the PLL and can push
the system performances out of the specification window. The same parametric variations
may be reflected in the variations in the readout values of the frequency divider. Parametric
failures can be detected if the states of the frequency divider are strongly correlated with the
design performances.

Fref
0
1 Vcon /
Delay1 Frequency Charge Loop Counter
VCO
detector pump filter
Delay2 0
1

Frequency
divider
Delay nT
counter
BIST start/read counter t
signal readout
scheme 1

Fref Vcon /
Frequency Charge Loop Counter
VCO
detector pump filter
0
1

Frequency
divider
BIST Delay nT
counter start/read counter t
scheme 2 signal readout

Fref Vcon /
Frequency Charge Loop Counter
VCO
detector pump filter
0
1

Frequency
divider
Delay nT
BIST counter start/read counter t
scheme 3 signal readout

Fig. 10. Three BIST scheme candidates (© [2007] IEEE, from Yu & Li, 2007b)
Robust Design and Test of Analog/Mixed-Signal Circuits in Deeply Scaled CMOS Technologies 369

Test schemes are described as follows:

a) BIST scheme 1
The first BIST scheme is similar to the one adopted in (Hsu & et al., 2005). In the normal
operation mode, the reference input and the output of the frequency divider are applied to
the frequency detector to form the closed loop configuration. In the test mode, the output of
the frequency divider is disconnected from the input of the frequency detector. The
reference input and or its delayed versions are fed through the muxes to the frequency
divider forming a open loop configuration. The first delay element has a larger delay value
than the second one. To charge up the VCO, the reference input and its delayed version
through delay 2 are applied to the frequency detector. To charge down the VCO, the
delayed versions of the reference input by both delay 1 and delay 2 are selected. The delay
values of the two delay elements determine the phase error introduced at the frequency
detector inputs. Hence, they also dictate the coverage of the VCO tuning range in this BIST
setup. Under typical design values, delay values in the order of ten's of the reference clock
signal period are required, which may cost significant silicon area to implement.

For all these three schemes, the counter read-out signals, which control the start and end
points for a single test run, are generated by passing the reference clock signal Fref through a
series of D flip-flops. As such, the contents of the frequency divider within a defined time
interval are read out.

b) BIST scheme 2
To solve the silicon overhead problem of scheme 1, we propose the second BIST scheme
which employs an inverter to introduce the phase difference. Since this configuration
introduces a constant phase delay of dT at the inputs of frequency divider, the charge pump
experiences the following sequence of operation: charge up  stop  charge up  stop
until the control voltage of the VCO reaches the fully voltage swing.

c) BIST scheme 3
The third BIST scheme is configured as follows: first the PLL is put in a standard closed-loop
configuration and then a standard phase lock test is performed. Once the PLL is locked, the
feedback signal frequency is changed from Fout to 2Fout by using the mux to select the output
of the second last D flip-flop in the divider.

A brief summary of the three BIST schemes: for scheme 1, the area cost is high while the test
time is short, for scheme 2, the area cost is low and the test time is also low, for scheme 3, the
area cost is low, and the test time is medium. The most important aspect for BIST schemes,
however, is the test accuracy. We use the macromodels developed in section 3 to perform
test accuracy evaluation and optimization for these BIST scheme candidates.

4.2 Test scheme evaluation and optimization

The most straightforward way to evaluate a BIST scheme is to perform Monte-Carlo
simulations and examine the ability of the specified BIST measurements to predict the
pass/fail status of the design. Since a set of few hundred Monte-Carlo PLL simulations may
370 VLSI

take tens of hours to complete, more runtime efficient approaches are needed, especially for
the optimization purpose.

The optimization of a given BIST scheme with n digital outputs is illustrated in the second
half of Fig. 11. Since a BIST scheme may be evaluated many times under different setups
(e.g., the time interval within which the states of the frequency divider are read out) within
the optimization loop, the correlation between the BIST schemes and the PLL performance
must be efficiently conducted. This goal is achieved by identifying the critical sources of
variations Zv as in equation (15).

Build
  regression model for Generate a large number of
BIST measurement T with measurements using the
process parameter Z v Ti= f(Zv) nonlinear regression model

  Check the classification   Correlating with T’s

accuracy of SVM model performances using SVM

Achieved Y  Final accuracy

verification with End
good accuracy?
more simulations
N
Adjust BIST setups

Fig. 11. BIST scheme optimization (© [2007] IEEE, from Yu & Li, 2007b)

Noticing that Zv only contains a reduced set of variations, a nonlinear empirical model
relating each measurement Ti of the given BIST scheme and Zv can be rather efficiently
generated. This is achieved by conducting a few Verilog-A based PLL simulations at
different Zv samples and performing nonlinear regression: Ti= f(Zv). Note that this step does
not incur a high simulation cost since regression models are only built over a low-dimension
parameter space represented by Zv. Using these easily obtained regression models, a large
set of samples for each Ti can be efficiently generated.

To capture the potential nonlinear correspondence between the design performances and
the measurements, Support Vector Machine (SVM) is adopted as an accurate classifier
(Vapnik, 1998). Support Vector Machine is a powerful method to build highly nonlinear
multivariate regression/classification models. In SVM regression, we consider a set of
training data {(x1, y1), (x2, y2), …, (xn, yn)}, where xi is the input vector and yi is the
corresponding output. The input X is mapped into a high dimensional feature space using
nonlinear transformation, then a best fitting function is constructed in this feature space as
y  f ( x)     ( x)  b (16)
where  is the nonlinear transformation, b is the bias term, and  represents the model
parameters to be decided. Based on this nonlinear function f(.), we can classify the chip as
faulty or not with the BIST circuit outputs.
Robust Design and Test of Analog/Mixed-Signal Circuits in Deeply Scaled CMOS Technologies 371

We demonstrate the effectiveness of the proposed method using a PLL design implemented
in 90nm CMOS technology. The frequency versus control voltage curve of the VCO is
extracted using the model of equation (8). First we simulate the VCO for a few clock cycles
and gather the time-domain output response as Y. Then RRR is applied to get a reduced
parameter set Z to represent the important device-level parameters. A parametric model of
the VCO in terms of Z is then built. To model the statistical characteristics of the VCO
accurately, a 6-th order polynomial fitting is used to fit the output frequency vs. control
voltage curve.

The Verilog-A models for other building blocks are extracted in a similar fashion.
Specifically, the charge pump model is generated using a 3-rd order polynomial in the
output voltage for each charge-up/down current. There are a total of 17 Verilog-A model
parameters extracted for the complete PLL design.

We evaluate the effectiveness of three BIST schemes. SVM models are generated as in Fig. 11
to predict the pass/fail status of the chips based on the corresponding BIST outputs. 400
Monte-Carlo simulation samples are generated by conducting PLL system simulation using
Verilog-A macromodels. These data are used to generate the SVM model. To evaluate the
effectiveness of the each scheme more reliably, another 100 Monte-Carlo simulations are
carried out and used as the test data for checking the accuracy of the SVM model. The
pass/fail predictions achieved through the three SVM models are compared against the
simulated chip performances, as shown in Fig. 12. Here, the predictions made through the
simulation are labeled as “direct measurement”, and +1 indicates a chip being classified as
“fail”, while -1 indicates the opposite.

1.5

1
Pass/Fail

0.5

Direct measurement
0
BIST scheme 1
BIST scheme 2
−0.5 BIST scheme 3

−1
0 10 20 30 40 50 60 70 80 90 100
Chip Index
Fig. 12. Pass/fail predictions of three BIST schemes (© [2007] IEEE, from Yu & Li, 2007b)

From Fig. 12 we can see that BIST scheme 1 only has only 1 misclassification. The
performance of BIST scheme 2 is verified to be poor as it can only detect two faulty chips
which may have largest variations. BIST scheme 3 can detect more failures than scheme 2
but is still not as good as scheme 1.
372 VLSI

We further look into the trade-offs between the accuracy and the number of digital outputs
for optimization. This is important since fewer test codes will correspond to a shorter test
time, if a similar accuracy can be achieved. This trade-off analysis is conducted for every
scheme in Fig. 12. It can be observed that for a small number of digital outputs, the accuracy
of scheme 2 is actually higher than that of scheme 3. It can be also seen that the accuracy of
scheme 3 becomes quickly saturated as the number of outputs increases. Under all the cases,
scheme 1 is always the optimal choice.

0.2

0.15
Error

0.1 BIST scheme 1

BIST scheme 2
BIST scheme 3
0.05

0
3 4 5 6 7 8 9
Number of test codes
Fig. 13. Accuracy vs. test structures for three BIST schemes (© [2007] IEEE, from Yu & Li,
2007b)

There are other optimizations can be done to improve BIST schemes using the efficient
macromodeling technique in Section 3 and Section 4 (Yu & Li, 2007b). It is of great benefit to
do such kind of analysis and optimization by taking consideration of statistical system
performances in the early design stage so that we can avoid the costly design iterations due
to the influence of process and environmental variations.

5. Conclusion
In this chapter we have discussed the influences of process variations to analog/mixed-
signal circuits in deeply scaled CMOS technologies. The performances of two types of
popular mixed-signal systems, i.e. Sigma-Delta ADCs and Phase-locked Loops are evaluated
under process variations. Parameterized lookup table technique and reduced-rank
regression with hierarchical macromodeling method were proposed to fulfil the
optimization of these two systems, respectively. We also extended the obtained fast system
performance evaluation framework to compare the efficiencies of parametric failure
detecting of different BIST circuits and perform test circuit optimization.

6. Acknowledgement
This work was funded in part by the FCRP Focus Center for Circuit & System Solutions
(C2S2), under contract 2003-CT-888.
Robust Design and Test of Analog/Mixed-Signal Circuits in Deeply Scaled CMOS Technologies 373

7. References
Azais, F. & et al. (2003). An all-digital DFT scheme for testing catastrophic faults in PLLs.
IEEE Design & Test of Computers, Vol. 20, No. 1, pp. 60 - 67, Jan. 2003
Babii, S. & et al. (1997). MIDAS User Manual, Stanford University, Stanford, CA
Bishop, R. & et al. (1990). Table-based modeling of delta-sigma modulators. IEEE Trans. On
Circuits and Systems, Vol. 37, No. 3, pp. 447-451, Mar. 1990
Box, G.; Hunter, D. & Hunter, W. (2005). Statistics for Experiments: Design, Innovation, and
Discovery, John Wiley & Son, 978-0471718130, Hoboken, NJ
Brauns, G. & et al. (1990). Table-based modeling of delta-sigma modulators using ZSIM.
IEEE Trans. On Computer-aided Design, Vol. 9, No. 2, pp. 142-150, Feb. 1990
Fang, S. & Suyama, K. (1992). User's Manual for SWITCAP2, Columbia University, New
York, NY
Feng, Z.; Yu, G. & Li, P. (2007). Reducing the Complexity of VLSI Performance Variation
Modeling Via Parameter Dimension Reduction. Proceedings of International
Symposium on Quality Electronic Design, pp. 737-742, 978-0769527957, Mar. 2007,
IEEE press, San Jose, CA
Hsu, C. ; Lai, Y. & Wang, S. (2005). Built-in self-test for phase-locked loops. IEEE Trans. On
Instrument and Measurement, Vol. 54, No. 3, pp. 996-1002, Jun. 2005
Kim, S. & Soma, M. (2001). An all-digital built-in self-test for high-speed phase-locked loops.
IEEE Trans. On Circuits and Systems II, Vol. 48, No. 2, pp. 141-150, Feb. 2001
Low, K. & Director, S. (1989). An efficient methodology for building macromodels of IC
fabrication processes. IEEE Trans. On Computer-aided Design, Vol. 8, No. 12, pp.
1299-1313, Dec. 1989
Nassif, S. (2001). Modeling and Analysis of Manufacturing Variations. Proceedings of Custom
Integrated Circuits Conference, pp. 223-228, 978-0780365917, May 2001, IEEE press,
San Diego, CA
Norsworthy, S.; Schreier, R. & Temes G. (1997). Delta-Sigma Data Converters: Theory, Design,
and Simulation. IEEE Press, 978-0780310452, Piscataway, NJ
Pelgrom, M.; Duinmaijer, A. & Welbers, A. (1989). Matching properties of MOS transistors.
IEEE Journal of Solid-State Circuits, Vol. 24, No. 5, pp. 1433 - 1440, Oct. 1989
Reinsel, G. & Velu, R. (1998). Multivariate Reduced-Rank Regression, Theory and Applications.
Springer, 978-0387986012, New York, NY
Sunter, S. & Roy A. (1999). BIST for phase-locked loops in digital applications. Proceedings of
International Test Conference, pp. 532-540, 978-0780357531, Sep. 1999, IEEE Press,
Atlantic City, NJ
Vapnik, V. (1998). Statistical Learning Theory. Wiley-Interscience, 978-0471030034, New York,
NY
Yu, G. & Li, P. (2006). Lookup Table Based Simulation and Statistical Modeling of Sigma-
Delta ADCs. Proceedings of Design Automation Conference, pp. 1035-1040, 978-
1595933816, San Francisco, CA, July 2006, IEEE Press
Yu, G. & Li, P. (2007a). Efcient Lookup Table Based Modeling for Robust Design of Σ∆
ADCs. IEEE Trans. on Circuits and Systems – I, Vol.54, No.7, Sep. 2007, pp.1513-1528,
Yu, G. & Li, P. (2007b). A Methodology for Systematic Built-in Self-Test of Phase-locked
Loops Targeting at Parametric Failures. Proceedings of International Test Conference,
pp. 1-10, 978-1424411276, Oct. 2007, IEEE Press, Santa Clara, CA
374 VLSI

Yu, G. & Li, P. (2008). Yield-aware hierarchical optimization of large analog integrated
circuits. Proceedings of International Conference on Computer-Aided Design, pp. 79-84,
978-1424428199, Nov. 2008, IEEE Press, San Jose, CA
Zou, J.; Mueller, D.; Graeb, H. & Schlichtmann, U. (2006). A CPPLL hierarchical
optimization methodology considering jitter, power and locking time. Proceedings of
Design Automation Conference, pp. 19-24, 978-1595933816, San Francisco, CA, July
2006, IEEE Press
Nanoelectronic Design Based on a CNT Nano-Architecture 375

0
19

Nanoelectronic Design Based on

a CNT Nano-Architecture
Bao Liu
Electrical and Computer Engineering Department
The University of Texas at San Antonio
San Antonio, TX, 78249-0669
Email: [email protected]

Abstract — Carbon nanotubes (CNTs) and carbon nanotube field effect transistors (CNFETs) have
demonstrated extraordinary properties and are widely expected to be the building blocks of next
generation VLSI circuits. This chapter presents (1) the first purely CNT and CNFET based nano-
architecture, (2) an adaptive configuration methodology for nanoelectronic design based on the
CNT nano-architecture, and (3) robust differential asynchronous circuits as a promising nano-circuit
paradigm.

1. Introduction
Silicon based CMOS technology scaling has driven the semiconductor industry towards cost
minimization and performance improvement in the past five decades, and is rapidly ap-
proaching its end (30). On the other hand, nanotechnology has achieved significant progress
in recent years, fabricating a variety of nanometer scale devices, e.g., molecular diodes (44)
and carbon nanotube field effect transistors (CNFETs) (46). This provides new opportunities
for VLSI circuits to achieve continuing cost minimization and performance improvement in a
post-silicon-based-CMOS-technology era.
However, we must overcome a number of significant challenges for practical nanoelectronic
systems, including achieving some of the most critical nanoelectronic design metrics as follow.
1. Manufacturability. As minimum layout feature size becomes smaller than lithography
light wavelength, traditional lithography based manufacturing process can no longer
achieve satisfiable resolution, and leads to significant process variations. Resolution
enhancement and other design for manufacturability techniques become less applica-
ble as scaling continues. Alternatively, nanoelectronic systems are expected to be based
on bottom-up self-assembly based manufacturing processes, e.g., molecular beam epi-
taxy (MBE). Such bottom-up self-assembly manufacturing processes provide regular
structures, e.g., perfectly aligned carbon nanotubes (23). Consequently, nanoelectronic
systems need to rely on reconfigurability to achieve functionality and reliability (51).
2. Reliability. Technology scaling has led to increasingly significant process and system
runtime variations, including critical dimension variation, dopant fluctuation, electro-
magnetic emission, alpha particle radiation and cosmos ray strikes. Such variations can-
not be avoided by manufacturing process improvement, and is inherent at nanometer
376 VLSI

CNT

Nano Interface

CNT

Nano Interface

Nano Interface
Nano Interface

Fig. 1. The proposed CNT crossbar nano-architecture: layers of orthogonal carbon nanotubes
form a dense array of RDG-CNFETs and programmable interconnects with voltage-controlled
nano-addressing circuits on the boundaries.

scale due to the uncertainty principle of quantum physics. Robust design techniques,
including redundant, adaptive, and resilient design techniques at multiple (architec-
ture, circuit, layout) levels, are needed to achieve a reliable nanoelectronic system (5).
3. Performance. Nanoscale devices have achieved ultra-high performance in the absence
of load, however, nanoelectronic system performance bottleneck lies in global intercon-
nects. Rent’s rule states that the maximum interconnect length scales with the circuit
size in a power law (24), while signal propagation delay across unit length interconnect
increases as technology scales (30). As a result, interconnect design will be critical to
nanoelectronic system performance.
4. Power consumption. As technology scaling leads to increased device density and de-
sign performance, power consumption is also expected to be critical in nanoelectronic
design.
This chapter presents several recent technical advancements towards manufacturable, reli-
able, high performance and low power nanoelectronic systems.
1. The first purely CNT and CNFET based nano-architecture, which is constructed by lay-
ers of orthogonal CNTs with via-forming and gate-forming molecules sandwiched in
between, forming a dense array of reconfigurable double gate carbon nanotube field
effect transistors (RDG-CNFETs) and programmable interconnects. Such a CNT array
is addressed by novel voltage-controlled nano-addressing circuits on the boundaries,
which do not require precise layout design and achieve yield in aggressive scaling
and adaptivity to process variations. Simulation based on CNFET and molecular de-
vice compact models demonstrates superior logic density, reliability, performance and
power consumption for nano-circuits implemented in this CNT crossbar based nano-
architecture compared with the existing, e.g., molecular diode and MOSFET based
nano-architectures.
Nanoelectronic Design Based on a CNT Nano-Architecture 377

2. A complete set of linear complexity methods for adaptive configuration of nanoelec-

tronic systems based on a CNT crossbar based nano-architecture (Fig. 1) (26), including
(1) adaptive nano-addressing, (2) RDG-CNFET gate matching, and (3) catastrophic de-
fect mapping methods. Compared with the previous nano-architecture defect mapping
and adaptive configuration proposals, these methods are complete, specific, determin-
istic, of low runtime complexity. These methods demonstrate the promising prospect of
achieving nanoelectronic systems of correct functionality, performance, and reliability
based on the CNT crossbar nano-architecture.
3. Robust Differential Asynchronous (RDA) circuits as a promising paradigm for reliable
(noise immune and delay insensitive) high performance and low power nano-circuits
based on the CNT crossbar nano-architecture. Theoretical analysis and SPICE simu-
lation based on 22nm CMOS Predictive Technology Models show that RDA circuits
achieve much enhanced reliability in logic correctness in the presence of a single bit
soft error or common multiple bit soft errors, and timing correctness in the presence of
parametric variations given the physical proximity of the circuit components.
The rest of this chapter is organized as follows. Section 2 reviews the existing nanoelectronic
devices, nano-architectures and nano-addressing circuits. Section 3 presents the proposed
CNT crossbar based nano-architecture including a novel RDG-CNFET device, a multi-layer
CNT crossbar structure, and a voltage-controlled nano-addressing circuit. Section 4 presents
adaptive configuration methods for nanoelectronic systems based on the CNT crossbar nano-
architecture. Section 5 presents robust differential asynchronous circuits as a promising nano-
circuit paradigm. Section 6 presents simulation results which evaluate the CNT crossbar nano-
architecture and robust differential asynchronous nano-circuits. Section 7 concludes this pa-
per with a list of nanotechnologies which enable and improve the proposed CNT crossbar
based nano-architecture.

2. Background
2.1 Existing Nanoscale Devices
Carbon nanotube is one of the most promising candidates for interconnect technology at
nanometer scale, due to its extraordinary properties in electrical current carrying capabil-
ity, thermal conductivity, and mechanical strength. A carbon nanotube is a one-atom-thick
graphene sheet rolled up in a cylinder of a nanometer-order diameter, which is semicon-
ductive or metallic depending on its chirality. The cylinder form eliminates boundaries and
boundary-induced scattering, yielding electron mean free path on the order of micrometers
compared with few tens of nanometers in copper interconnects (32). This gives extraordinary
current carrying capacity, achieving a current density on the order of 109 A/cm2 (56). How-
ever, large resistance exists at CNT-metal contacts, reducing the performance advantage of
CNTs over copper interconnects (38).
Among various nanotechnology devices, carbon nanotube field effect transistors are the most
promising candidates to replace the current CMOS field effect transistors as the building
blocks of nanoelectronic systems. Three kinds of carbon nanotube based field effect transis-
tors (CNFETs) have been manufactured: (1) A Schottky barrier based carbon nanotube field
effect transistor (SB-CNFET) consists of a metal-nanotube-metal junction, and works on the
principle of direct tunneling through the Schottky barrier formed by direct contact of metal
and semiconducting nanotube. The barrier width is modulated by the gate voltage. This de-
vice has the most mature manufacturing technique up to today, while two problems limit its
378 VLSI

future: (a) The metal-nanotube contact severely limits current. (b) The ambipolar conduction
makes this devices cannot be applied to conventional circuit design methods. (2) A MOSFET-
like CNFET is made by doping a continuous nanotube on both sides of the gate, thus forming
the source/drain regions. This is a unipolar device of high on-current. (3) A band-to-band
tunneling carbon nanotube field effect transistor (T-CNFET) is made by doping the source
and the drain regions into p+ and n+ respectively. This device has low on-current and ultra
low off current, making it potential for ultra low power applications. It also has the potential
to achieve ultra fast signal switching with < 60mV/decade subthreshold slope (46).
Molecular electronic devices are based on two families of molecules: the catenanes which
consist of two or more interlocked rings, and the rotaxanes which consist of one or more
rings encircling a dumbbell-shaped component. These molecules can be switched between
states of different conductivities in a redox (reduction/oxidation) process by applying cur-
rents through them, providing reconfigurability for nanoscale devices (44).
A variety of reconfigurable nanoscale devices have been proposed. Resonant tunneling diodes
based on redox active molecules are configurable on/off (44). Nanowire field effect transistors
with redox active molecules at gates are of high/low conductance (17). Spin-RAM devices are
of high/low conductivity based on the parallel/anti-parallel magnetization configuration of
the device which is configured by the polarity of the source voltage (40). A double gate Schot-
tky barrier CNFET is configurable to be a p-type FET, an n-type FET, or off, by the electrical
potential of the back gate (25). A double gate field effect transistor with the back gate driven
by a three state RTD memory cell is configurable to be a transistor or an interconnect, reducing
reconfiguration cost of a gate array (4).

2.2 Existing Nanoelectronic Architectures

At least three categories of nanoelectronic architectures have been proposed. An early nano-
electronic architecture NanoFabrics was based on molecular resonant tunneling diodes (RTDs)
and negative differential resistors (NDRs) (20). The insightful authors have observed that
passive device (diode/resistor) based circuits lack signal gain to recover from signal attenua-
tion, while combining with CMOS circuits compromises scaling advantages. They proposed
latches based on negative differential resistors (NDRs), which, unfortunately, have become
obsolete since the publication.
The majority of the existing nanoelectronic architectures are based on a hybrid nano-CMOS
technology, with CMOS circuits complementing nano-circuits. In FPNI (50) (CMOL (54)),
a nanowire crossbar is placed on top of CMOS logic gates (inverters). The nanowires pro-
vide programmable interconnects (and wired-OR logic), while the CMOS gates(inverters)
provide logic implementation (signal inversion and gain). Such architectures achieve compro-
mised scaling advantage in term of device density. DeHon (11; 13) proposed to combine pro-
grammable nanoscale diode logic arrays with fixed simple CMOS circuitry, e.g., of precharge
and evaluation transistors as in domino logic for signal gain. Sequential elements need also to
be implemented as CMOS circuits. However, the optimal size of a combinational logic block
is typically small (e.g., of 30-50 gates), which results in significant CMOS circuitry overhead
in such architectures. An exception is memory design, where CMOS technology provides pe-
ripheral circuitry such as address decoders and read sensors with moderate overhead, while
nanotechnology provides scaling advantage in memory cells (17; 47; 63).
The third category of existing nanoelectronic architectures rely on DNA-guided self-assembly
to form 2-D scuffles for nanotubes (42; 43) or 3-D DNA-rods (14). Such technologies target
application in the far future.
Nanoelectronic Design Based on a CNT Nano-Architecture 379

Data Lines / Nanoscale Wires

Address Lines / Macroscale Wires

Fig. 2. Layout of undifferentiated nanoscale wires (data lines) addressed by microscale wires
(address lines). Lithography defines high- and low-k dielectric regions, which gives field effect
transistors and direct conduction, respectively.

2.3 Existing Nano-Addressing Circuits

A nano-addressing circuit selectively addresses a nanoscale wire in an array, and enables data
communication between a nano-system and the outside world. The existing nano-addressing
circuits are based on binary decoders, with an array of (microscale) address lines running
across the (nanoscale) data lines, forming transistors at each crossing (e.g., Fig. 2). Each data
line is selected by a unique binary address, given each data line has a unique gate configura-
tion. However, such precise layout design is highly unlikely to achieve at a sublithographic
nanometer scale (without significantly compromised yield).
In details, the existing nano-addressing circuits are in four categories as follow.
1. Randomized contact decoder (59) includes gold particles which are deposited at ran-
dom as contacts between nanoscale and microscale wires. Testing and feedback provide
a one-to-one mapping between a nanoscale wire and an address.
2. Undifferentiated nanoscale wires are addressable by microscale wires with (e.g., lithog-
raphy defined) different gate configurations (which requires nanoscale wire spacing in
the same order of lithography resolution) (22) (Fig. 2).
3. Alternatively, different gate configurations are realized in the nanoscale wires, by grow-
ing lightly-doped and heavily-doped carbon nanotubes of different length alternatively,
while the microscale wires are undifferentiated. A microscale wire crossing a lightly-
doped nanotube segment forms a gate, while a heavily-doped nanotube segment is
always conductive for all possible signals in the microscale wire. In such a case, precise
control of the lengths of the lightly- and heavily-doped nanotube segments would be
critical (12; 21).
4. In radial addressing, multi-walled carbon nanotubes are grown with lightly- and
heavily-doped shells, an etching process removes the heavily-doped outer shells at pre-
cise locations, and defines the gate configurations at each crossing of nanoscale and
microscale wires (48).
380 VLSI

Because process variations are inevitably significant at nanometer scale, these existing nano-
addressing structures achieve limited yield, e.g., there is certain probability that two nanoscale
wires have identical or similar gate configuration due to process variation. Furthermore,
nanoscale wires are mostly partially selected, e.g., they may not achieve the ideal conductivity
upon selected, due to process variations such as misalignment, dopant variation, etc.

2.4 Existing Nano-Architecture Defect-Mapping and Adaptive Configuration Methods

Existing nano-architecture defect mapping techniques are as follow. (1) On a Teramac recon-
figurable computing platform, signals are propagated along each row or each column in a
crossbar structure, a defect is located at the intersection of a defective row and a defective col-
umn, based on the assumption that a single defect is present (10). (2) In the NanoFabrics nano-
architecture, the roughly estimated number of defects for a subset of computing resources are
collected by counter or none-some-many circuits, a simple graph based algorithm or a Bayes’
rule based probabilistic computation procedure gives defect occurrence probability estimates.
E.g., highly likely defects are detected in the probability assignment phase, which accumulates
defect probability in different test configurations, while less likely defects are located in the
defect location phase, which incrementally clears certain spots as non-defects during test of
different configurations (33). (3) A Build-In Self-Test (BIST) method in the NanoFabrics nano-
architecture brings much increased complexity with limited applicability (in finding available
defect-free neighboring nanoBlocks to implement test circuitry) (8; 58).
After a defect map is achieved presumably, logic circutis can be constructed avoiding or utiliz-
ing the defects. For example, a nanoPLA block can be synthesized in the presence of defective
crosspoints (36), a CNT nano-circuit layout can be synthesized in the presence of misaligned
and mispositioned CNTs (41), metallic CNTs (61), and CNTs of variational density (62).

2.5 Existing CNT Nano-Circuit Design

A very limited number of primitive combinational logic circuits have been fabricated based
on CNFETs, including an inverter and two NOR gates in NMOS logic based on SB-CNFETs
(3), and a five-inverter ring oscillator based on MOSFET-like CNFETs (9). While nano-circuits
based on ambipolar SB-CNFETs need different topologies (3; 46; 53), nano-circuits based on
unipolar MOSFET-like CNFETs can be identical to CMOS circuits (9).

3. CNT Crossbar based Nano-Architecture

As we have seen, most existing nanoelectronic architectures are based on diode/resistor logic
and CMOS/nano-technologies (11; 13; 20; 50; 54; 63), which only achieve limited manufactura-
bility, reliability, and performance. Carbon nanotubes (CNTs) and carbon nanotube field effect
transistors (CNFETs) are the most promising candidates as the the building blocks of nanoelec-
tronic systems due to their extraordinary properties. CNTs possess excellent electrical current
carrying capability, thermal conductivity, and mechanical strength. CNFETs are potential to
achieve high on-current, ultra-low off-current, and ultra-fast switching (< 60mV/decade sub-
threshold slope). CNT crossbar structure (Fig. 1) is one of the most promising candidates
for nanoelectronic design platform. Recently, UIUC researchers have achieved fabrication of
dense perfectly aligned CNT arrays (23). Such a CNT crossbar structure forms the basis of
nanoscale memories (17; 47; 63).
However, no nanoelectronic architecture has been proposed which is solely based on CNTs
and CNFETs. The reasons include lack of (1) a reconfigurable CNT based device which could
provide functionality and reliability, (2) a self-assembly process which forms complex CNT
Nanoelectronic Design Based on a CNT Nano-Architecture 381

Front Gate

CNT
Bistable
Molecules

CNT Source n+ p +or i n+ CNT Drain

Dielectric
Redox Active
Molecules
CNT

Back Gate
Fig. 3. A n-type MOSFET-like reconfigurable double gate carbon nanotube field effect tran-
sistor (RDG-CNFET).

structures, and (3) an achievable mechanism which precisely addresses an individual CNT in
an array.
In this section, we investigate the first purely CNT and CNFET based nano-architecture, which
is based on a novel RDG-CNFET device, includes a CNT crossbar structure on multiple layers,
and a novel voltage-controlled nano-addressing circuit.

3.1 RDG-CNFET Device Structure

As the building block of a purely CNT and CNFET based nano-architecture, a reconfig-
urable double-gate CNFET (RDG-CNFET) is constructed by sandwiching electrically bistable
molecules in a double gate CNFET. The double gate CNFET is constructed by three overlap-
ping orthogonal carbon nanotubes. The top and the bottom carbon nanotubes form the front
gate and the back gate, while doping the carbon nanotube in the middle layer forms the source
and the drain of a n- or p-type MOSFET-like CNFET (46). Electrically bistable molecules are
coated around the front gate and sandwiched between the front gate and the source/drain re-
gions. Dielectric and redox active molecules are coated around the back gate and sandwiched
between the back gate and the source/drain regions (Fig. 3).
The redox active molecules at the back gate are electrically reconfigurable to hold/release
charge in a redox process, which controls the CNFET threshold voltage and conductance, or,
turns the CNFET on or off. An example of such configuration is reported in (17), wherein
a ±10V voltage applied to cobalt phthalocyanine (CoPc) molecules triggers a redox process,
and results in a NW-FET conductance change of nearly 104 times. Such reconfiguration of
CoPc molecules is repeatable for more than 100 times.
The bistable molecules sandwiched between the front gate and the source/drain regions are
electrically reconfigurable to be conductive or insular, making the device a via or a FET. An ex-
ample of such electrically bistable molecules is reported in (44), wherein oxidative degradation
reduces resonant tunneling current of the V-shaped amphiphilic [2]-rotaxane 54+ molecules
by nearly a factor of 100. Alternatively, the anti-fuse technologies in the existing reconfig-
urable architectures provide one-time configurability. For example, the QuickLogic ViaLink
technology include a layer of amorphous silicon sandwiched between two layers of metal. A
10V programming voltage provides a resistance difference between GΩ and 80Ω (6).
382 VLSI

S D

BG
Fig. 4. Compact model of a n-type MOSFET-like reconfigurable double gate carbon nanotube
field effect transistor (RDG-CNFET).

Bistable Molecules

L4

Bistable Molecules

L2
Dielectric

Redox Active Molecules

L3

L1
Fig. 5. Carbon nanotube (CNT) layers in the proposed nanoelectronic architecture.

3.2 RDG-CNFET Device Behavior

Such a RDG-CNFET device is described in a compact model as is shown in Fig. 4, and is
reconfigurable to the following components, making it an ideal nanoelectronic architecture
building block.
1. Via, when the front gate bistable molecules are configured to be conductive. The over-
lapping of the front gate and the source/drain regions form conductive contacts. As a
result, the front gate, the source, and the drain are short circuited. The device is config-
ured as a via between the carbon nanotubes on the top and in the middle.
2. Short, when the front gate bistable molecules are configured to be insular, and the back
gate redox active molecules are configured to hold positive(negative) charge in a n-
type(p-type) CNFET. The CNFET is on for any front gate voltage.
3. MOSFET-like CNFET, when the front gate bistable molecules are configured to be insu-
lar, and the back gate redox active molecules are configured to hold negative(positive)
charge in a n-type(p-type) MOSFET-like CNFET. The CNFET threshold voltage is ad-
justable by the doping concentration in the channel (p or n doping for a n- or p-type CN-
FET), such that when the back gate redox active molecules are configured to hold neg-
ative(positive) charge in a n-type(p-type) MOSFET-like CNFET, the CNFET achieves
both performance and leakage control.
4. Open, when the MOSFET-like CNFET is turned off. This is achieved at the architecture
level as follows.
Nanoelectronic Design Based on a CNT Nano-Architecture 383

3.3 CNT Crossbar Structure

At a larger scale, a nanoelectronic architecture is constructed by growing layers of orthog-
onal carbon nanotubes, with via-forming (electrically bistable) and gate-forming (dielectric
and redox active) molecules sandwiched at each crossing (Fig. 1). The carbon nanotubes
are either (1) semiconductive CNTs which are doped to have low resistivity and are recon-
figurable to opens by gate isolation, or (2) metallic CNTs which upon identification can be
utilized as global interconnects if not avoided or removed (1; 64). The (1) via-forming (electri-
cally bistable) and (2) gate-forming (dielectric and redox active) molecules can be first coated
around a carbon nanotube (e.g., as in (17)), then undergo an etching process with the top layer
of carbon nanotubes as masks (e.g., as in (49)). The remaining molecules are sandwiched be-
tween two orthogonal carbon nanotubes on adjacent layers. A top-down (e.g., lithography)
process defines the areas for each type of molecules to assemble on each layer, as well as the
p-wells and n-wells. P-type and n-type of MOSFET-like CNFETs are formed by (e.g., potas-
sium or electrostatic (46)) doping of the carbon nanotubes selectively. E.g., a p-well or n-well
of dimensions in the order of 22nm include about 10 rows of CNFETs.
Configuration of such a CNT crossbar based nanoelectronic architecture gives a nanoscale
VLSI implementation including MOSFET-like CNFETs and interconnects with opens, shorts
and vias (Fig. 6), which can be 2-D (compatible to traditional VLSI systems) or 3-D.
In a 2-D VLSI implementation, MOSFET-like CNFETs are formed on the bottom three layers
of carbon nanotubes, with the first layer (L1) from bottom of carbon nanotubes provides the
back gates, the second layer (L2) provides the source and the drain regions, and the third
layer (L3) provides the front gates of the MOSFET-like CNFETs. Dielectric and redox active
(back gate) molecules are sandwiched between the L1 and L2 layer carbon nanotubes, and
electrically bistable (front gate) molecules are sandwiched between the L2 and L3 layer car-
bon nanotubes. A multi-layer reconfigurable interconnect structure with programmable vias
and opens is achieved with via-forming and gate-forming molecules sandwiched between
interconnects which are formed above the first (L1) layer (Fig. 5).
3-D VLSI circuits are under active research in recent years due to their potential of achieving
reduced wirelength, reduced power consumption and improved performance. However, sili-
con based VLSI circuits are essentially 2-D, because MOSFETs are surface devices on the bulk
of silicon, 3-D MOSFET circuits can only be achieved by bonding chips. It is therefore criti-
cal to achieve (1) bonding technology which provides acceptable mechanical strength, (2) via
technology which provides low resistive interconnects between chips, and (3) heat dissipation
in a multiple chip system for silicon based 3-D circuits. On the contrary, CNFET and CNFET
based nano-architectures provide excellent platforms for 3-D VLSI circuits, because (1) CNTs
and CNFETs are not confined to certain surface and can be manufactured in 3-D space, (2)
CNTs possess excellent current carrying, mechanical and heat dissipation properties which
are critical to 3-D VLSI circuits.
In a 3-D VLSI implementation, the RDG-CNFETs do not need to be confined on the bottom
layers, with the upper layers dedicated to interconnects. Instead, transistors and interconnects
are free to be located on each layer of carbon nanotubes. Gate forming (dielectric and redox
active) molecules and via-forming (electrically bistable) molecules are distributed between
adjacent CNT layers. Combination of the types of molecules surrounding a CNT segment
gives three components.
1. Gate-forming molecules both on top and on bottom of a CNT segment give a device
which is reconfigurable to either open or short,
384 VLSI

Vdd a b c a b c

output
Fig. 6. An RDG-CNFET based Boolean logic a(b + c) implementation.

Vdd
input

clk clk output

Fig. 7. An RDG-CNFET based latch implementation.

2. Gate-forming and via-forming molecules on top and on bottom of a CNT segment give
the RDG-CNFET, which is reconfigurable to via, short, MOSFET-like CNFET, and open,
3. Via-forming molecules both on top and on bottom of a CNT segment give a device
which is reconfigurable to be stacked via, simple via, or double gate FET.
We have the following observations.

Observation 1. Via-forming (electrically bistable) molecules must be present between any two adja-
cent layers.

Observation 2. Gate-forming (redox active) molecules must be present next to each layer for gate
isolation.

Observation 3. Gate-forming (redox active) and via-forming (electrically bistable) molecules need to
be evenly distributed on each layer for performance.

3.4 Circuit Paradigms and Analysis

This CNT crossbar based nano-architecture provides regularity and manufacturability for
high logic density implementations of all CMOS logics, including the standard CMOS logic
(e.g., in Fig. 6), domino logic, pass-transistor logic, etc., for combinational circuits, as well as
latches (e.g., in Fig. 7), flip-flops, memory input address decoder and output sensing circuits.
Such high logic density is achieved via direct connection of CNFETs through their
source/drain regions (e.g., as in an latest Intel microprocessor implementation (15)), without
going through additional (e.g., metal) interconnects. CNT-metal contacts are known to bring
the most significant resistivity in CNT technology (38). Avoiding such CNT-metal contacts
contributes to performance and reliability improvements. Furthermore, reduced interconnect
length also leads to reduced interconnect capacitance, and improved circuit performance.
This CNT crossbar based nano-architecture also provides a high reconfigurability by allow-
ing an arbitrary ratio of logic gates and interconnect switches (a RDG-CNFET device can be
Nanoelectronic Design Based on a CNT Nano-Architecture 385

Data Lines / Nanoscale Wires

Vdda1

Address Line 1
Vssa1

Vdda2

Address Line 2
Vssa2
Fig. 8. Schematic of the proposed voltage-controlled nano-addressing circuit.

configured as either a logic gate or an interconnect switch). A pre-determined ratio of logic de-
vices and interconnect switches (e.g., in standard cell designs and FPGA architectures where
cells and routing channels are separated) constrains design optimization and may lead to
inefficient device or interconnect utilization. Allowing an arbitrary ratio of logic gates and
interconnect switches (e.g., as in sea-of-gate designs) provides increased degree of freedom
for design optimization (4).
The CNT crossbar based nano-architecture is also the first to include multiple routing layers.
Multiple routing layers (as in the current technologies) are necessary for VLSI designs, as
Rent’s rule suggests that the I/O number of a circuit module follows a power law with the gate
number in the module (24). A small routing layer number could lead to infeasible physical
design or significant interconnect detouring, resulting in degraded performance and device
utilization.

3.5 Voltage Controlled Nano-Addressing Structure

The final piece of the CNT crossbar nanoelectronic architecture is the nano-addressing circuits
on the boundary of the carbon nanotube crossbar structure.
Designing a nano-addressing circuit is a challenging task, because (1) the nanoscale lay-
out cannot be manufactured precisely unless it is of a regular structure, and (2) the nano-
addressing circuit cannot be based on reconfigurability since it provides reconfigurability to
the rest of the nanoelectronic system.
A novel voltage-controlled nano-addressing circuit (Fig. 8) is constructed by running two ad-
dress lines (of either microscale or nanoscale wires) on top of the data lines (of nanoscale wires
in an array which are to be addressed). The address lines and the data lines are orthogonal. At
each crossing of an address line and a data line, a field effect transistor is formed by doping
the data line into the source and the drain regions while the address line provides the gate
of the transistor, with a thin layer of dielectric sandwiched between the gate and the tran-
sistor channel. Such field effect transistors have been successfully fabricated based on either
nanowires or carbon nanotubes (17; 37; 46).

3.6 Voltage-Controlled Nano-Addressing Principle

The address line provides the gate voltage for the transistors. Each address line is connected to
two external voltages at the ends (Vdda1 and Vssa1 for address line 1, Vdda2 and Vssa2 for address
line 2). The position of a nanoscale wire in the array gives the gate voltage for the transistor
386 VLSI

on the nanoscale wire alone the address line. For example, a i-th nanoscale wire (starting from
Vss ) in an array of n equally spaced nanoscale wires has a transistor gate voltage

i n−i
Vg (i, n) = V + Vss (1)
n dd n
in an address line connecting to two external voltage sources Vdd and Vss . Here we assume
uniform address lines of negligible external resistance (from the first or the last nanoscale wire
to the nearest external voltage source).
A transistor is on if its gate voltage exceeds the threshold voltage Vg > Vth . A nanoscale
wire is conductive if both transistors on it are on. Because the two address lines provide an
increasing series and a decreasing series of gate voltages respectively, only nanoscale wires at
specific positions in the array are conductive. For example, for Vdda1 = Vdda2 and Vssa1 = Vssa2 ,
the nanoscale wire in the middle of the array gets conductive.
In general, to select the i-th data line from the left in an array of n nanoscale wires, the external
voltages need to be such that all the transistors on the right hand side of the i-th data line in
the first address line are off, and all the transistors on the left hand side of the i-th data line in
the second address line are off:
i+1 i+1
Vga1 (i + 1, n) = (1 − )Vdda1 + Vssa1 < Vth
n n
i−1 i−1
Vga2 (i − 1, n) = Vdda2 + (1 − )Vssa2 < Vth (2)
n n

3.7 Voltage Controlled Nano-Addressing Analysis

Compared with the existing nano-addressing circuits, the proposed voltage-controlled nano-
addressing circuit leads to significant manufacturing yield improvement due to the following
reasons.
The existing nano-addressing circuits are based on binary decoders and require every
nanoscale wire have a unique physical structure to differentiate itself, which is highly unlikely in
a nanotechnology manufacturing process - lithography cannot achieve nanoscale resolution,
while bottom-up self-assembly based nanotechnology manufacturing processes provide only
regular structures. Even at microscale, such a structure is subject to prevalent catastrophic
defects and significant parametric variations, which result in low yield.
On the contrary, the proposed circuit consists of only uniform components in a regular struc-
ture. Every nanoscale wire has a uniform physical structure and is differentiated by their electrical
parameters, e.g., the node voltages. This scheme avoids any precise layout design and signifi-
cantly improves yield and enables aggressive scaling of the addressing circuit with the rest of
the nanoelectronic system.
Furthermore, let us compare voltage-controlled nano-addressing with the existing binary de-
coder based nano-addressing mechanisms in terms of addressing accuracy and resolution.
These two key quantitative metrics for nano-addressing circuits are defined as follow since
such definition is not available in previous publications to the best of the author’s knowledge.

Definition 1. Addressing inaccuracy of a nano-addressing circuit is the offset between the target
data line i and the data line j of maximum current.

AI = |i − j| (3)
Nanoelectronic Design Based on a CNT Nano-Architecture 387

In voltage-controlled nano-addressing, addressing inaccuracy is given by inaccurate address-

ing voltages from the voltage dividers. Such addressing inaccuracy can be further minimized
by adjusting the external voltages to adapt to manufacturing process and system runtime
parametric variations. As a result, a mis-addressing is localized, i.e., the data line j of maxi-
mum current is not far from the target data line i.
In traditional binary decoder based nano-addressing, an n-bit binary address has n neigh-
boring binary addresses of Hamming distance 1. A 1-bit error could lead to n different mis-
addressings. This leads to non-localized mis-addressing and a more significant addressing
inaccuracy.

Definition 2. Addressing resolution of a nano-addressing circuit is the minimum ratio between the
on current Ion (i ) of a target data line i and the off current Io f f ( j) of a non-target data line j (under all
conditions, e.g., different inputs and parametric variations).

Ion (i )
AR = Min{ } (4)
Io f f ( j)

In traditional binary decoder based nano-addressing, the achievable addressing resolution de-
pends on the conductance difference between the target data line and other non-target data
lines. There are n non-target data lines with Hamming distance 1 for a n-bit target address,
which have similar if not identical conductances. The presence of parametric variations fur-
ther reduces addressing resolution.
In voltage-controlled nano-addressing, addressing resolution is largely given by the address-
ing voltage difference between two adjacent data lines. Applying high voltages leads to a
number of reliability issues, such as electromigration and gate dioxide breakdown. Carbon
nanotubes are highly resistive to electromigration, while new material is needed to enhance
reliability for gate dioxide breakdown.
Alternatively, for given gate voltage difference, transistor current difference can be improved
by improving the inverse subthreshold slope. However, MOSFETs and MOSFET-like CNFETs
are limited to an inverse subthreshold slope S (which is the minimum gate voltage variation
needed to bring a 10× source-drain current increase) of 2.3 kT q ≈ 60mV/decade at 300K (46).
This requires development of novel devices for larger inverse subthreshold slopes.

4. Adaptive Configuration of Nanoelectronic Systems Based on the CNT Crossbar

Nano-Architecture
In this section, we examine a list of nanoelectronic design adaptive configuration methods
which cancel the effects of catastrophic defects and parametric variations in the proposed
CNT crossbar nano-architecture.

4.1 Adaptive Nano-Addressing

A variety of parametric variations are expected to be prevalent and significant in nanoelec-
tronic systems. Their effects on the voltage-controlled nano-addressing circuit (Fig. 8) are as
follow.
1. Global address line resistance variations, e.g., due to uniform width, height, and/or
resistivity variations of the address lines, have no effect on the voltage divider hence
the addressing scheme.
2. Address line misalignment (shifting) has no effect on the conductances of the data lines.
388 VLSI

Di Dj

Vh
V l

Rh Rl
R h + ∆R R l − ∆R

Fig. 9. Addressing two CNTs Di and D j with a resistance of ∆R in between.

3. Global data line misalignment (i.e., shifting of all data lines), variations of external volt-
age sources, and variations of external wire/contact resistance (between the resistive
voltage divider and the external voltage sources) lead to potential addressing inaccu-
racy (CNT target offset).
4. Individual data line misalignment (shifting) could decrease the difference between the
gate voltages of two adjacent transistors, leading to degraded addressing resolution
(on/off CNT current ratio between two adjacent CNTs).
5. Process variations of the transistors, including width, length, dopant concentration, and
oxide thickness variations, lead to transistor conductivity uncertainty and degraded
addressing resolution.
The nano-addressing scheme needs to achieve a higher enough addressing resolution which
endures the above-mentioned parametric variation effects (e.g., by applying high external
addressing voltages, and/or novel CNFETs of < 60mV/decade subthreshold slope).
After achieving satisfiable addressing resolution, we need to minimize any addressing inac-
curacy and address the correct CNT data line (Problem 1).

Problem 1 (Adaptive Nano-Addressing). Given a voltage-controlled nano-addressing circuit, ad-

dress the i-th CNT data line in the presence of parametric variations.

Let us first derive the external voltage offset needed for a data address offset. Suppose for an
address line, the external voltages Vh and Vl address the i-th CNT data line Di . The resistance
between CNT data line Di and the high (low) external address voltage Vh (Vl ) is Rh (Rl ) (Fig.
9).1 We have
Rl Rh
V + V = Von (5)
Rh + Rl h Rh + Rl l
where Von is the voltage needed to address a CNT data line of peak current. Shifting the
external voltages to Vh + ∆V and Vl + ∆V addresses another CNT data line D j . The resistance
between CNT data line D j and the high (low) external address voltage is Rh + ∆R (Rl − ∆R).
We have
Rl − ∆R R + ∆R
(V + ∆V ) + h (V + ∆V ) = Von (6)
Rh + Rl h Rh + Rl l
1 For the first address line, the high external voltage Vh = Vl1 is on the left, the low external voltage
Vl = Vr1 is on the right. For the second address line, the high external voltage Vh = Vr2 is on the right,
the low external voltage Vl = Vl2 is on the left.
Nanoelectronic Design Based on a CNT Nano-Architecture 389

As a result,
∆R
∆V = (V − Vl ) (7)
Rh + Rl h
Observation 4. The external voltage offset ∆V is proportional to the resistance offset ∆R between two
CNT data lines, and is proportional to the physical offset ∆L between the two CNT data lines, if the
resistive voltage dividers are uniform (e.g., the CNT data lines are equally spaced and the address lines
have uniform resistivity).

Based on Observation 1, Method 1 gives an adaptive nanoelectronic addressing method,

which finds the external voltage shifts needed to address the left most and the right most
CNT data lines first. Any other external voltage shift needed to address a specific CNT data
line is then computed based on a linear interpolation. To address the left most or the right
most CNT data line, we apply a gradually increasing/decreasing external voltage offset ∆V
at an address line, keep all the transistors at the other address line on, and measure the con-
ductance of the array of CNT data lines. The maximum and the minimum ∆V’s (e.g., ∆Vmink
and ∆Vmaxk , k = 1 or 2) with non-zero CNT data line conductances address the left most and
the right most CNT data lines, respectively.

Algorithm 1: Adaptive Voltage Controlled

Nano Addressing

Input: An array of n CNT data lines, address i

Output: Addressing i-th data line

1. Turn on all transistors at address line 2 (Vl2 =

Vr2 > Vth )
2. Find ∆Vl1 which addresses first data line (bi-
nary search)
3. Find ∆Vr1 which addresses n-th data line (bi-
nary search)
4. Turn on all transistors at address line 1 (Vl1 =
Vr1 > Vth )
5. Find ∆Vl2 which addresses first data line (bi-
nary search)
6. Find ∆Vr2 which addresses n-th data line (bi-
nary search)
7. Shift Vl1 and Vr1 by nn−i ∆Vl1 + ni ∆Vr1
8. Shift Vl2 and Vr2 by nn−i ∆Vl2 + ni ∆Vr2

Observation 5. The addressing accuracy given by Method 1 depends only on the uniformity of the
resistive voltage divider, and the time domain variations of the external voltage differences Vl1 − Vr1
and Vl2 − Vr2 . Any time-invariant (e.g., manufacturing process) variations of the external voltages
(Vl1 , Vr1 , Vl2 , and Vr2 ) or the external address line resistances (from the outer most data lines to the
external voltage sources) do not affect the achievable addressing accuracy.

4.2 RDG-CNFET Gate Matching

Another process variation is the misalignment of the front gate CNT and the back gate CNT of
a reconfigurable double-gate CNFET (RDG-CNFET). This is because that the front gate CNT
and the back gate CNT are on different (i − 1 and i + 1) layers, while CNT arrays on different
layers do not have and are not expected to have a precise alignment mechanism.
390 VLSI

Front Gate Front Gate

CNT CNT
Bistable Bistable
Molecules Molecules

n+ p +or i n+ p +or i n+
Dielectric
Redox Active Redox Active Redox Active
Molecules Molecules Molecules
CNT CNT CNT

Back Gate Back Gate Back Gate

Fig. 10. CNT misalignment in dense CNT arrays. The closest CNT pair forms the front gate
and the back gate of a RDG-CNFET. The neighboring CNTs have cross-coupling effect which
needs to be simulated/tested or avoided by shielding.

Fortunately, we observe that precise alignment between a front gate CNT and a back gate
CNT is not necessarily required as long as the CNT arrays are dense, e.g., with the spacing
between CNTs close to the CNT diameters. In such a case, a double gate field effect transistor
is formed even in the presence of CNT misalignment (Fig. 10). A CNFET channel is formed
by doping the source/drain regions with the front gate CNT on the upper layer as mask. The
resultant CNFET channel aligns with the front gate CNT. A misaligned back gate injects a
weaker electrical field in the CNFET channel from a longer distance. A neighboring back gate
may also injects a weak electrical field in the channel. This is either tolerated (which needs to
be verified by simulation or testing) or avoided (by reserving the neighboring back gates for
shielding).
The question is then how to find the closest CNT pair on different layers which form the front
gate and the back gate of a RDG-CNFET (such that we can address them and configure the
RDG-CNFET).

Problem 2 (RDG-CNFET Gate Matching). Given a CNT i on layer l, locate the closest CNT j on
layer l + 2 (or l − 2) where CNTs i and j form the front gate and the back gate of a RDG-CNFET.

Method 2 solves Problem 2 and finds the closest CNT pairs which form the front gate and the
back gate of a RDG-CNFET.
Nanoelectronic Design Based on a CNT Nano-Architecture 391

Algorithm 2: RDG-CNFET Gate Matching

Input: CNT i on layer l which is a gate of CNFET

T
Output: Closest CNT j to CNT i on layer l + 2
(l − 2) which is the other gate of CNFET T

1. Apply a turn-off gate voltage to CNT i

2. For each CNT j on layer l + 2 (l − 2),
3. Apply a turn-off gate voltage to CNT j
4. Measure the conductance of CNFET T
5. Find CNT j for the smallest CNFET conduc-
tance

Once a matching gate is identified, the CNFET can be characterized (by achieving its I-V
curves). A parasitic CNFET can also be identified by finding the second closest CNT (with the
second smallest CNFET conductance in the algorithm), which is either tolerated or avoided in
a nanoelectronic design.

4.3 Catastrophic Defects and Mapping Techniques

In this subsection, we examine catastrophic defects for CNTs, programmable vias and CN-
FETs, and their corresponding detection and location methods.

4.3.1 Metallic, Open and Crossover CNTs

CNTs are metallic or semiconductive depending on their chirality. One third of CNTs are
metallic if they are grown isotropically. Metallic CNTs can be removed by either chemical
etching (64) or electrical breakdown (1). However, such techniques bring large process vari-
ation effects (34). Mitra et al. propose use of CNT bundles for each nanoelectronic signal to
reduce metallic CNT effect (34). We observe that metallic CNTs need not necessarily to be
removed and CNT bundles are not needed for each nanoelectronic signal as long as metal-
lic CNTs can be detected and located. Upon detection and location, metallic CNTs can be
configured to form global interconnects if not avoided. Their low resistivity helps to reduce
signal propagation delay in global interconnects which are critical to nanoelectronic system
performance.
Open CNTs are expected to be prevalent in a CNT array, as open CNT occurrence is propor-
tional to the length of the CNT. A CNT with a single open can be largely included in a correct
nanoelectronic design, upon detection and location of the single defect. A CNT with two (or
more) opens is not fully utilizable. The segment between the two (extreme) opens are not
accessible by any nano-addressing circuit, and components attached to that segment are not
configurable. Upon detection and location of the extreme opens, the end segments of an open
CNT can be included in a nano-circuit. Or, we can simply avoid open CNTs.
CNTs which are supposedly-parallel may cross over each other, resulting in different ad-
dresses for a CNT on two sides of a crossbar, and unexpected resistive contacts between CNTs.
If not corrected by etching (34), such crossover CNTs can be taken as multi-thread cables and
included in a correct nano-circuit. It is necessary to solve the following problem for nanoelec-
tronic system configuration on a CNT crossbar nano-architecture.

Problem 3. Detect and locate metallic, open and crossover CNTs in an CNT array, which are addressed
on both ends by nano-addressing circuits.
392 VLSI

Such metallic, open, and crossover CNTs can be captured in a n × n resistance matrix RCNT ,
where each entry RCNT (i, j) gives the resistance of CNT between the i-th CNT end and the j-th
CNT end on the opposite sides of an array of n CNTs (if i = j, RCNT (i, j) gives the resistance
of a crossover CNT, otherwise, RCNT (i, i ) gives the i-th CNT’s resistance).
Method 3 solves Problem 3 by giving such a n × n resistance matrix RCNT . With this CNT re-
sistance matrix RCNT , we avoid open CNTs, and consider only semiconductive CNTs, metal-
lic CNTs, and crossover CNT bundles (as multi-thread cables) for the rest of the calibration
(Methods 4 and 5 and 2).

Algorithm 3: Metallic, Open, Crossover CNT

Detection and Location

Input: Array of n CNTs with nano-addressing

circuits on both ends (Fig. 1)
Output: Resistance map RCNT for metallic, open,
crossover CNTs

1. Configure all CNFETs as shorts

2. For each i
3. For each j
4. Address the i-th CNT on one end of CNT
5. Address the j-th CNT on the other end of
CNT
6. Measure resistance RCNT (i, j)
7. If i = j and RCNT (i, j) ≈ ∞
8. Open CNT (i, j)
9. If i = j and RCNT (i, j)  ∞
10. Crossover CNT (i, j)
11. If RCNT (i, j) ≈ Rmetallic
12. Metallic CNT (i, j)
13. If RCNT (i, j) ≈ Rsemiconductive
14. Semiconductive CNT (i, j)

4.3.2 Opens and Shorts in Programmable Vias

A CNT junction with electrically bistable molecules is a programmable via, which is suppos-
edly reconfigured as a conductive via or open. A catastrophic defect at such a junction can
be either (1) permanent open, or (2) permanent short. It is necessary to solve the following
problem for nanoelectronic system configuration on a CNT crossbar nano-architecture.

Problem 4. Detect and locate permanently open or short vias in a CNT crossbar nano-architecture.

Method 4 solves Problem 4 by giving two m × n resistance maps R Pmin and R Pmax , where
each entry R Pmin (i, j) or R Pmax (i, j) gives the resistance of a L-shaped path which includes
the i-th CNT segment on the top(bottom) of the CNT crossbar, the j-th CNT segment on the
left(right) of the CNT crossbar, and a programmable via which is configured as conductive
or open, respectively. Given non-open CNTs, these resistance matrices give a defect map for
permanently open or short vias.
Nanoelectronic Design Based on a CNT Nano-Architecture 393

Algorithm 4: Permanently Open or Short Via

Detection and Location

Input: Two layers of m × n CNT crossbar with

nano-addressing interface on four sides
Output: Resistance maps R Pmin and R Pmax for
permanently open or short vias

1. For each non-open CNT i

2. For each non-open CNT j
3. Address i-th CNT from top(down) of
crossbar
4. Address j-th CNT from left(right) of
crossbar
5. Program via V (i, j) to conductive
6. Measure path resistance R Pmin (i, j)
7. Program via V (i, j) to insular
8. Measure path resistance R Pmax (i, j)
9. If R Pmin (i, j) = R Pmax (i, j) ≈ ∞
10. Permanently open via V (i, j)
11. If R Pmin (i, j) = R Pmax (i, j) ≈ RCNT (i, i )
or RCNT ( j, j)
12. Permanently short via V (i, j)

4.3.3 Opens and Shorts in CNFETs

A CNT junction with dielectric and redox active molecules is supposedly reconfigured as a
FET. A catastrophic defect could lead to (1) short between source and drain (e.g., due to chan-
nel punchthrough, no intrinsic channel area, redox active molecules cannot release charge), (2)
short between gate and source or drain (e.g., due to dielectric breakthrough), or (3) constant
open gate (e.g., redox active molecules cannot hold charge). It is necessary to solve the follow-
ing problem for nanoelectronic system configuration on a CNT crossbar nano-architecture.

Problem 5. Detect and locate permanently open or short CNFETs in a CNT crossbar nano-
architecture.

Shorts between CNFET gate and source or drain can be detected in a method which is sim-
ilar to Method 4 but without via programming. Method 5 finds permanent opens or shorts
between the source and the drain of a CNFET by giving a m × n resistance matrix RCNFET .
Upon detection and location, these catastrophic defects (metallic, open and crossover CNTs,
permanently open or short vias and CNFETs) can be included in a correct nano-circuit. Nano-
circuit physical design needs to be adaptive to the presence of these catastrophic defects, and
will be different from die to die, based on the catastrophic defect maps (RCNT , R Pmin , R Pmax ,
and RCNFET ) for each die.
394 VLSI

Algorithm 5: Permanently Open or Short CN-

FET Detection and Location

Input: CNFETs in crossbar with nano-addressing

interface, CNT resistance matrix RCNT
Output: Resistance map RCNFET for perma-
nently open or short CNFETs

1. Configure all CNFETs as shorts

2. For each non-open CNT i
3. For each non-open CNT j
4. Address the i-th CNT on both ends
5. Configure CNFET (i, j) as open
6. Measure resistance RCNFET (i, j)
7. If RCNFET (i, j) ≈ RCNT (i, i )  ∞
8. Short between CNFET (i, j) source-
drain
9. If RCNFET (i, j) ≈ RCNT (i, i ) ≈ ∞
10. Open between CNFET (i, j) source-
drain
11. Configure CNFET (i, j) as short

4.4 Parametric Variation and Adaptive Design

Other than catastrophic defects, process variations are also critical to nanoelectronic system
performance and reliability. Compared with catastrophic defects, process variations are more
prevalent, and they are more difficult to detect since their effects are accumulated in affecting
the underlying circuit. Adaptive or resilient nano-circuit design techniques are expected to
achieve functionality and reliability in the presence of such process variations, besides online
calibration and adaptive configuration as follows.
In adaptive configuration, each module of the circuit is configured with its test circuit. The test
circuit can be as simple as additional interconnects which connect the inputs and the outputs
of the module to some of the primary inputs and the primary outputs, respectively. In such
cases, function and performance calibration is performed externally. Alternatively, self-test
can be performed given the complexity of the test circuit. If the current configuration passes
online function and performance verification, the auxiliary test circuit will be removed, and
the current configuration of the module is committed. Otherwise, the same circuit module
needs to be realized using other hardware resources on the reconfigurable platform.

5. Reliable, High Performance and Low Power Nano-Circuits

5.1 Nano-Circuit Design Challenges and Promising Techniques
As we have seen, the CNT crossbar nano-architecture provides regularity and manufactura-
bility for high logic density implementations of all CMOS combinational logic families, in-
cluding static logic, domino logic, pass-transistor logic, as well as latches, flip-flops, memory
input address decoder and output sensing circuits such as differential sense amplifiers (Fig. 6
and Fig. 7).
However, nano-circuits in a CNT crossbar nano-architecture face a number of unique chal-
lenges. Nano-circuits must achieve reliability in the presence of prevalent defects and signif-
Nanoelectronic Design Based on a CNT Nano-Architecture 395

icant parametric variations, must achieve performance with highly resistive CNT intercon-
nects, etc. We discuss nano-circuit design in a CNT crossbar nano-architecture in this section.
Nanoscale computing systems are expected to be subject to prevalent defects and significant
process and environmental variations inevitably as a result of the uncertainty principle of
quantum physics. E.g., the conductance of a CNT or a CNFET is very sensitive to chirality,
diameter, etc. (38). Besides adaptive configuration, nanoscale computing systems need new
computing models or circuit paradigms for reliability enhancement, performance improve-
ment and power consumption reduction.
As technology scales, nanoelectronic computing systems are expected to be based on single
electron devices (the average number of electrons in a transistor channel is approaching one
for the current technologies). In quantum mechanics, the occurrence probability of an electron
is the wavefunction given by the Schrödinger equation. How to extract a deterministic com-
putation result from stochastic events such as electron occurrences is one of the fundamental
problems that we are facing in designing nanometer scale computing systems. Traditional
computation based on large devices can be modeled as redundancy and threshold based logic
(which includes majority logic). In redundancy and threshold based logic, the error rate is
given by a binomial distribution (the probability of observing m events in an environment of
expecting an average of n independent events). As a result, a minimum signal-to-noise ratio
is required with performance and power consumption implications. Finding a more efficient
reliable computing model is of essential interest in nanoelectronic design.
Besides stochastic signal occurrence, signal propagation delay variability is another category
of uncertainty in stochastic nanoscale systems. Nanoelectronic design needs to be adaptive to
or resilient in the presence of signal propagation delay variations. Existing techniques (e.g.,
the Razor technology (2; 18) wherein a shadow flip-flop captures a delayed data signal for tim-
ing verification and correction) achieves only limited performance adaptivity, e.g., the circuity
is adaptive to performance variations only within a given range. Asynchronous circuits have
unlimited adaptivity to performance variations, and are ideal for high performance (enabling
performance scaling in the presence of significant performance variations) and low power (be-
ing event-driven and clockless) nanoelectronic design (e.g., multi-core chips are expected to be
increasingly self-timed, global-asynchronous-locally-synchronous, or totally asynchronous).
However, existing asynchronous design techniques suffer in reliability in the presence of soft
errors (e.g., glitches, coupling noises, radiation or cosmos ray strike induced random noises),
which has limited their applications for decades.
A number of robust design techniques at multiple levels help to enhance reliability and reduce
error rate of a nanoelectronic computing system. At the circuit level, differential signaling and
complementary logic reduces parametric variation effects by exploiting spatial and temporal
correlations (e.g., by correlating m and n for reduced error rate in a binomial distribution) (15;
31; 57). At a higher level, we believe that Error Detection/Correction Code (35) is the key to
the stochastic signal occurrence problem in nanoscale systems (e.g., for lower required signal-
to-noise ratios, which lead to high performance and low power), and needs to be applied more
extensively at a variety of design hierarchy levels. Error correction coding has been applied
widely in today’s memories and wireless communication systems. Proposals also exist for
applying (AN or residue) error detection/correction coding in arithmetic circuits (7).
396 VLSI

ack.t 1
C
ack.t n
inputs f f ack.f1
C
ack.fn
valid(i) C f d.t d.t
rs.f rs.t rs.f
valid
f d.f d.f
FF
valid(i) C valid(d)
rs.t valid req
ck.t ck.f
ack.t
valid(d)
inputs f f
ack.f

Fig. 11. A static logic based RDA (robust differential asynchronous) circuit.

5.2 Robust Asynchronous Circuits

A promising CNT nano-circuit paradigm is robust asynchronous circuits, which are formed
by applying Error Detection/Correction Code to asynchronous circuit design for enhanced
reliability in the presence of soft errors, while achieving delay insensitive nano-circuits (28).
For a 1-bit data signal, a robust differential asynchronous (RDA) communication channel in-
cludes three rails, two differential data signals d.t and d. f , and a request signal req. The data
and request signals are valid only if (d.t, d. f , req) = (1, 0, 1) or (0, 1, 1), while (d.t, d. f , req) =
(0, 0, 0) when the data and request signals return to zero. A validity check circuit detects the
arrival of valid data and triggers the receiver flip-flop. The Hamming distance between any
two legal codes for the data and request signals is at least two, instead of one in the dual-rail
asynchronous signaling (d.t, d. f ) = (0, 1), (1, 0) or (0, 0). Differential acknowledgment signals
ack.t and ack. f are also included.2
Fig. 11 gives a static logic based RDA circuit. 3 Two Muller C elements take the input validity
signals valid(i ) and valid(i ) (as well as the rs.t and rs. f signals coming from the downstream
acknowledgment signals) and generate two differential validity signals valid and valid, which
trigger the combinational logic computation. The Muller C elements have the maximum sig-
nal propagation delay for a rising(falling) output for all combinational logic circuits with the
same number of inputs, due to the presence of the longest path of transistors to the power
supply or the ground. Also the valid and valid signals derive from and arrive later than the
data signals. As a result, the valid and valid signals always arrive later than the differential
combinational logic outputs f and f¯.
The later arriving validity signals filter out the internal glitches which come from combina-
tional logic computation before the differential outputs f and f¯ settle to their final values.
Two NMOS transistors clamp the differential outputs f and the f¯ to the ground until the valid
signal rises and the valid signal falls for noise immunity. The differential acknowledgment

2 In general, multiple-bit data can be encoded in a variety of error detection codes (35). A single parity
bit for n-bit data provides a Hamming distance of two which is immune to any single bit error, while
an error detection code of a Hamming distance larger than k is immune to any k-bit error.
3 Alternative implementations (in dynamic circuits, e.g., dual-rail domino, DCVSL, etc.) achieve dif-
ferent cost and reliability tradeoffs, and are potentially preferrable depending on the manufacturing
technology and the environment, e.g., parametric variabilities, soft error rates, etc.
Nanoelectronic Design Based on a CNT Nano-Architecture 397

signals ack.t and ack. f derive from and always appear later than the differential data signals
f and f¯.
At the sender’s end of the interconnect, for sequential elements, flip-flops are preferred over
latches for reliability. A flip-flop is only vulnerable to noise when capturing the signal, while a
latch is vulnerable to noise whenever it is transparent. The flip-flops send out differential data
signals d.t and d. f (which come from the differential combinational logic outputs f and f¯) as
well as a request req signal (which comes from the acknowledgment signal ack.t). At the re-
ceiver end of the interconnect, a group of XOR and AND(NAND) gates verify the differential
data and request signals, and generate two differential validity signals valid(d) and valid(d).
Any single bit soft error or common multiple bit soft errors injected to the interconnects or at
the validity signals will halt the circuit.
Each sender flip-flop is triggered by two differential acknowledgment signals ack.t and ack. f
as the differential clock signals ck.t = 1 and ck. f = 0, and is reset by two differential reset
signals when rs.t = 1 and rs. f = 0. The differential reset signals come from the downstream
differential acknowledgment signals ack.ti and ack. f i via the Muller C elements. They also
generate the valid and valid signals which trigger the combinational logic computation.
In the presence of multiple fanouts, multiple sets of differential acknowledgment signals will
be sent back to the upstream stage. With the Muller C elements holding the input validity sig-
nals, the early arriving acknowledgment signals hold until the latest acknowledgment signal
arrive from the fanouts. At that time the Muller C elements close the inputs to the combina-
tional logic block and reset the flip-flop at the upstream stage, which brings all differential
data and request signals d.t, d. f and req as well as the acknowledgment signals ack.t and ack. f
back to the ground, completing an asynchronous communication cycle.

5.3 Logic and Timing Correctness of RDA Circuits

An RDA circuit achieves logic and timing correctness in the presence of a single bit soft error
or common multiple bit soft errors given the physical proximity of the circuit components.

Definition 3 (Single Bit Soft Error). A single bit soft error is a glitch or toggling caused by a single
event upset as a result of an alpha particle or neutron strike from radioactive material or cosmos rays.

Definition 4 (Common Multiple Bit Soft Error). A common multiple bit soft error is glitches or
togglings of the same magnitude and polarity caused by common noises such as capacitive or inductive
interconnect coupling, or spatially correlated transient parametric (e.g., supply voltage, temperature)
variations (19; 39; 60), which have near identical effects on components at close physical proximity.

Theorem 1 (Logic Correctness). An robust differential asynchronous circuit achieves logic correct-
ness at the event of a single bit soft error or common multiple bit soft errors.

Proof. An RDA circuit achieves logic correctness in the following cases.

1. A single bit soft error or a common multiple bit soft error at the input data signals leads
to invalid data. The valid(i )(valid(i )) signal will not rise(fall).
2. A single bit soft error or a common multiple bit soft error at the valid(i ) and valid(i )
signals, at the Muller C elements computing the valid and valid signals, or at the valid
and valid signals, leads to an early false or a late valid(valid) signal. In this case, the
differential structure for the valid/valid signals prevents any logic error. A false valid
signal turns off the keeper circuit, but can rise neither f nor f¯, because the valid signal
is still high. A false valid signal rises either f or f¯, while the keeper circuit still clamps
398 VLSI

both f and f¯ to the ground. Only when both the valid and valid signals arrive, the
differential combinational logic computation is enabled.
3. A single bit soft error or a common multiple bit soft error at the differential combina-
tional logic block or the differential data signals f and f¯ will not raise the ack.t signal
nor lower the ack. f signal.4
4. A single bit soft error or a common multiple bit soft error at the differential acknowl-
edgment signals ack.t and ack. f does not trigger the flip-flop.
5. A single bit soft error or a common multiple bit soft error at the differential reset signals
RS and RS does not reset the flip-flop.
6. A single bit soft error or a common multiple bit soft error at the differential data signals
d.t and d. f and the request req signal leads to invalid data and does not generate a
validity signal.
In summary, in order to make an RDA circuit to fail, the glitches must follow certain specific
patterns, e.g., to reverse a “01” to a “10”, which is highly unlikely to take place.



Theorem 2 (Timing Correctness). An robust differential asynchronous circuit achieves timing cor-
rectness for any delay variation given the physical proximity of the circuit components.

Proof. Prevalent parametric (process, temperature, supply voltage) variations in nanoelec-

tronic circuits lead to significant delay variations for the components in the circuit. Because
such delay variations are spatially correlated (19; 39; 60), given the physical proximity of the
circuit components, their delay variations are tightly correlated. Consequently, an RDA circuit
achieves timing correctness in the following cases.
1. The input data signals d.t and d. f always arrive earlier than the differential validity
signals valid(d) and valid(d), which derive from the differential data and the request
signals.
2. The valid(valid) signal derives from the input validity signals valid(i ) and valid(i )
through the Muller C elements.
The rising(falling) delay of an Muller C element is the maximum of any combinational
logic block of the same number of inputs with the longest serial transistor path to the
power supply and the ground. Given the tight correlation of parametric variations for
the transistors and the interconnects in the circuit, the valid and valid signals arrive no
early than the final combinational logic computation results f and f¯. The valid and valid
signals enable the differential outputs f and f¯ via the tri-state output structure and the
keeper circuit, thus filtering out the glitches in combinational logic computation.
3. The differential acknowledgment signals ack.t and ack. f derive from and arrive no early
than the differential data signals f and f¯. They are further delayed (e.g., via buffers)
such that at the rising edge of the flip-flop clock signal, the input data signals have
settled to their final values, and the flip-flop captures the correct data f and f¯. Conse-
quently, no setup time constraint is required. The differential data signals f and f¯ hold
4 A glitch does not appear before valid data given the clamp NMOS transistors and fast data rise time
in a static RDA circuit. A dynamic RDA circuit needs to enhance robustness for a soft error strike at
a combinational logic output, e.g., by including weak keeper PMOS transistors, or delaying precharge
PMOS transistors.
Nanoelectronic Design Based on a CNT Nano-Architecture 399

until the differential acknowledgment signals ack.t and ack. f reach the upstream stage,
reset the upstream flip-flop, and lower the valid(i ) and valid signals, which take much
longer time than the hold time of the flip-flop. Consequently, no hold time constraint is
required.
4. The differential acknowledgment signals ack.t and ack. f arrive after the combinational
logic computation completes and the differential data signals f and f¯ settle to their final
values.
5. After all downstream stages send back acknowledgment signals, the flip-flop is reset,
bringing the differential data d.t and d. f and the request req signals to the ground. The
downstream stage acknowledgment signals are also brought back to the ground as a
result. This completes a four-phase asynchronous communication cycle.
As a result, the proposed robust differential asynchronous circuit is delay insensitive, i.e.,
achieves correct timing (signal arrival time sequence) in the presence of delay variations,
which is critical for nanoelectronic circuits. 


6. Experiments
6.1 Voltage-Controlled Nano-Addressing
In this section, we first verify the effectiveness of the proposed voltage-controlled nano-
addressing circuit (Fig. 8) by running SPICE simulation based on the Stanford CNFET com-
pact model (52).
In the proposed voltage-controlled nano-addressing circuit, each nanotube is gated by two N-
type MOSFET-like CNFETs. These CNFETs are of 6.4nm gate width and 32nm channel length,
as are described in the Stanford CNFET compact model. The two CNFETs in each nanotube
are given a voltage drop of Vdd = 1V. The external address voltages are Vdda1 = Vdda2 = 1V,
Vssa1 = Vssa2 = 0. As a result, the CNFETs have complementary gate voltages Vg1 + Vg2 = 1V.
Fig. 12 gives the nanotube currents in the array with different gate voltage at the first address
line. The nanotubes carry a significant current only with specific gate voltages, e.g., reaching
Iout = 5.064mA at gate voltage Vg1 = 0.495V.
With 0 and 1V external voltages, Fig. 12 gives the currents for all the nanotubes in the array.
With larger external voltages, Fig. 12 is extended to give the nanotube currents: any nanotube
with a Vga1 > 1V or Vga2 < 0V gate voltage at the first address line carries zero current.
Addressing resolution is given by the difference of addressing voltages between two adjacent
nanotubes (since MOSFETs and MOSFET-like CNFETs are limited to a < 60mV/decade in-
verse subthreshold slope). Adjusting the external address voltages minimizes any addressing
inaccuracy due to manufacturing process and system runtime parametric variations.

6.2 Comparison of CNT Crossbar based and the Existing Nano-Architecture

Let us now compare nano-circuits implemented in the proposed CNT crossbar based nano-
architecture and the existing nano-architectures. Considering DNA-guide self-assembly based
nanoelectronic architectures such as NANA (42) and SOSA (43) target the far future, and FPNI
(50) is very similar to CMOS technology by employing CMOS transistors and nanowires, I will
compare RDG-CNFET based logic implementation with molecular diode and MOS transistor
based logic implementation which is the mainstream nanoelectronic architecture in literature.5

5 Comparing CNFET and CMOS-FET circuits gives approximately 5× performance improvement (16).
400 VLSI

6
"Iout"

4
Iout (mA)

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Vg1 (V)
Fig. 12. Nanotube current Iout in mA for CNFET gate voltage Vg1 in the first address line.

Vdd a b c

en output
Fig. 13. A molecular diode/MOSFET based Boolean logic a(b + c) implementation.

As an example of a combinational logic block, a Boolean logic function a(b + c) is implemented

based on RDG-CNFETs (Fig. 6) and by molecular diodes and peripheral CMOS transistors
(Fig. 13). In the following experiments, SPICE simulation is conducted based on the latest
Stanford compact CNFET model (52), a molecular device model from a latest publication (17),
and the latest Predictive CMOS Technology Model (45).
The RDG-CNFETs are constructed based on an enhancement mode CNFET of 6.4nm gate
width and 32nm channel length, as is described in the Stanford compact model (52). The
bistable molecules at the front gate provide a resistance difference between GΩ and about
80Ω (6).6 The redox active molecules at the back gate are cobalt phthalocyanine (CoPc) (17),
which have been the basis of a NW-FET device with 1000× conductance difference (17).
The molecular diodes are based on V-shaped amphiphilic [2]rotaxane 54+ molecules (44), with
saturation current Is = 36pA, emission coefficient N = 14.66, and an on/off current ratio of
194.9. The CMOS transistors are modeled by 22nm Predictive Technology Models (45). To
balance the current difference between molecular diodes and PMOS transistors, the PMOS
transistors have a channel width/length ratio W/L = 1/10, while each molecular diode con-
sists of 10, 000 V-shaped amphiphilic [2]rotaxane 54+ molecules. As a result, the circuit has a
current on the order of nA.

6 The amorphous silicon based anti-fuse technology works with silicon based nanowires (6). Similar
technologies are expected and assumed here for carbon nanotubes.
Nanoelectronic Design Based on a CNT Nano-Architecture 401

Mo. Diode RDG-CNFET

abc Vout Pstatic Vout Pstatic
(V) (W) (V) (W)
111 0.999 1.49n 1.000 0.25n
110 0.807 0.83µ 1.000 0.33n
101 0.807 0.83µ 1.000 0.32n
011 0.497 1.45µ 0.000 15.34p
000 0.265 1.47µ 0.000 41.32p

Table 1. Output voltage and static power consumption with different inputs of RDG-CNFET
and molecular diode based Boolean logic a(b + c) implementations.

Comparing the CNFET based and the molecular diode/CMOS based logic implementations,
we have the following observations.
1. Area: The CNFET based logic implementation takes an area of 2 × 6 = 12 CNFETs
and 2 × 3 = 6 vias, while molecular diodes and MOSFET based implementation takes
an area of 2 × 4 = 8 molecular diodes and 2 MOSFETs (and two more MOSFETs if
an inverter is included at each output to restore signal voltage swing). Considering
CNFET based implementation is in a complementary logic, and the MOS transistors
do not scale well, CNFET based implementation is expected to achieve superior logic
density at a nanometer technology node.
2. Signal reliability: The CNFET based logic implementation achieves full voltage swing
at the outputs, while in the diode logic circuit, the output swing depends on the inputs,
and varies between 0.503V to 0.735V in the experiment (Table 1). Additional CMOS
circuitry (e.g., an inverter) can be included at each output to restore full voltage swing,
however, the reduced signal voltage swing in the diode logic circuit still implies com-
promised signal reliability.
3. Static power: The CNFET based logic implementation in CMOS logic achieves orders of
magnitudes of less power consumption compared with molecular diodes and MOSFET
based implementation for most input vectors (Table 1).
4. Performance: The CNFET based logic implementation achieves orders of magnitude
of timing performance improvement compared with molecular diodes and MOSFET
based implementation (Table 2).
In summary, CNFET based logic implementation achieves superior logic density, reliabil-
ity, performance, and power consumption compared with molecular diodes and CMOS-FET
based Boolean logic implementation.

6.3 Verification of RDA Circuits

In this section, we verify logic and timing correctness of robust differential asynchronous cir-
cuits by running HSPICE simulation based on 22nm Predictive Technology Models (45).
Fig. 14 gives signal waveforms for a perfect RDA circuit implementing Boolean function f =
ab in CMOS static logic. Input signal a falls from logic 1 at 100ps to logic 0 at 200ps, input
signal b rises from logic 0 at 0ps to logic 1 at 50ps. The validity signal valid(valid) rises(falls)
from logic 0(1) at 100ps to logic 1(0) at 200ps. As a result, we observe that the late arrival of
the valid(valid) signals enable the combinational logic computation for f and f¯, and filter out
402 VLSI

Mo. Diode RDG-CNFET

CL Dr Df Dr Df
( f F) (ns) (ns) (ns) (ns)
1 0.37 0.63 0.01 0.01
10 3.31 6.61 0.08 0.08
100 32.77 62.29 0.77 0.78

Table 2. Rising/falling signal propagation delays Dr /D f (ns) (from a to output) for various
load capacitance CL ( f F) of RDG-CNFET and molecular diode based Boolean logic a(b + c)
implementations. Input signal transition time varying from 1ps to 100ps leads to no consider-
able delay difference.

1500
valid
f
1000 ack.t
d.f
Voltage (mV)

valid(d)
500

−500
100 150 200 250 300
Time (ps)
Fig. 14. Signal waveforms in a robust differential asynchronous circuit with no single bit soft
error.

the internal glitches that would be observed at the f and f¯ signals if the validity signals arrive
early.
Fig. 15 gives signal waveforms for the same RDA circuit with the same input signals a and
b, while the validity signal valid is delayed by 50ps compared to the complementary validity
signal valid, representing an early false validity signal (valid) or a late arriving validity signal
(valid) due to an injected negative glitch at either the valid or the valid signal. We observe
that the differential data signals f and f¯ are clamped to the ground until both validity signals
settle to their final values valid = 1 and valid = 0. As a result, no logic malfunction is present
while all signals are delayed by 50ps.
Fig. 16 gives signal waveforms for the RDA circuit with a triangle current (of 0.1mA peak
current starting at 160ps ending at 180ps) injected to the f signal. Comparing with Fig. 14, we
observe that such a negative glitch does not lead to any logic error, instead, the arrivals of the
ack.t and ack. f signals are postponed for about 40ps, as well as all the downstream signals d.t,
d. f , and valid(d).
Fig. 17 gives signal waveforms for the RDA circuit with a triangle current (of 0.1mA peak
current starting at 120ps ending at 140ps) injected to the ack.t signal. The glitch at the ack.t
signal does not trigger the flip-flop, and the subsequent signals d.t, d. f , valid(d) and valid(d)
are not affected.
Fig. 18 gives signal waveforms for the RDA circuit with two identical triangle currents (of
0.1mA peak current starting at 120ps ending at 140ps) injected to both differential acknowl-
Nanoelectronic Design Based on a CNT Nano-Architecture 403

1500
valid
f
1000 ack.t
d.f
Voltage (mV)
valid(d)
500

−500
100 150 200 250 300
Time (ps)
Fig. 15. Signal waveforms in a robust differential asynchronous circuit with an early false
valid or a late arriving valid signal.

1500
valid
f
1000 ack.t
d.f
Voltage (mV)

500 valid(d)

−500

−1000
100 150 200 250 300
Time (ps)
Fig. 16. Signal waveforms in a robust differential asynchronous circuit with a negative glitch
injected at the f signal.

edgment signals ack.t and ack. f . We observe that the double glitches do not trigger the flip-flop
either, the subsequent signals d.t, d. f , valid(d) and valid(d) are not affected.
From these experiments, we observe that the RDA circuit achieves correct logic and correct
timing (signal arrival time sequence) at the event of a single bit soft error or common multiple
bit soft errors, by temporarily halting the circuit operation until the valid data re-appear.

7. Conclusion
In this chapter, we have studied the first purely CNT and CNFET based nano-architecture,
which is based on (1) a novel reconfigurable double gate carbon nanotube field effect transistor
(RDG-CNFET) device, (2) a multi-layer CNT crossbar structure with sandwiched via-forming
and gate-forming molecules, and (3) a novel voltage-controlled nano-addressing circuit not
requiring precise layout design, enabling manufacture of nanoelectronic systems in all existing
CMOS circuit design styles.
A complete methodology of adaptive configuration of nanoelectronic systems based on the
CNT crossbar nano-architecture is also presented, including (1) an adaptive nano-addressing
method for the voltage-controlled nano-addressing circuit, (2) an adaptive RDG-CNFET gate
matching method, and (3) a set of catastrophic defect mapping methods, which are specific (to
404 VLSI

2000
valid
f
1500 ack.t
ack.f
Voltage (mV) 1000 d.f
valid(d)

500

−500
100 150 200 250 300
Time (ps)
Fig. 17. Signal waveforms in a robust differential asynchronous circuit with a positive glitch
injected at the ack.t signal.

2000
valid
f
1500 ack.t
ack.f
Voltage (mV)

1000 d.f
valid(d)

500

−500
100 150 200 250 300
Time (ps)
Fig. 18. Signal waveforms in a robust differential asynchronous circuit with positive glitches
injected at both differential acknowledgment signals ack.t and ack. f .

the CNT crossbar nano-architecture), complete (in detecting and locating all possible catas-
trophic defects in the CNT crossbar nano-architecture), deterministic (with no probabilistic
computation), and efficient (test paths are rows or columns of CNT, or L-shaped CNT paths,
CNT open/short defect detection is separated with via/CNFET open/short defect detection,
runtime is linear to the number of defect sites). This is significant improvement compared
with the previous techniques (which either detects only a single defect (10), or is generic, ab-
stract, probabilistic, and highly complex (8; 33; 58)).
We have also examined some of the design challenges and promising techniques for CNT
and CNFET based nano-circuits. We identify significant parametric variation effects on logic
correctness and timing correctness, and propose robust (differential) asynchronous circuits by
applying Error Detection Code to asynchronous circuit design for noise immune and delay
insensitive nano-circuits.
SPICE simulation based on compact CNFET and molecular device models demonstrates su-
perior logic density, reliability, performance and power consumption of the proposed RDG-
CNFET based nanoelectronic architecture compared with previously published nanoelec-
tronic architectures, e.g., of a hybrid nano-CMOS technology including molecular diodes and
MOSFETs. Furthermore, theoretical analysis and SPICE simulation based on 22nm Predictive
Technology Models show that RDA circuits achieve much enhanced reliability in logic correct-
Nanoelectronic Design Based on a CNT Nano-Architecture 405

ness in the presence of a single bit soft error or common multiple bit soft errors, and timing
correctness in the presence of parametric variations given the physical proximity of the circuit
components.
While nanotechnology development has not enabled fabrication of such a system, this chap-
ter has demonstrated the prospected manufacturability, reliability, and performance of a
purely carbon nanotube and carbon nanotube transistor based nanoelectronic system. These
nanoelectronic system design techniques are expected to be further developed along with
nanoscale device fabrication and integration techniques which are critical to achieve and to
improve the proposed nanoelectronic architecture in several aspects, including: (1) search of
electrically bistable molecules of repeated reconfigurability and low contact resistance with
carbon nanotubes, (2) development of etching processes for electrically bistable and redox ac-
tive molecules with carbon nanotubes as masks, and (3) manufacture of nanoscale devices of
superior subthreshold slope for enhanced nano-addressing resolution, and ultra high perfor-
mance low power nanoelectronic systems.

8. References
[1] I. Amlani, N. Pimparkar, K. Nordquist, D. Lim, S. Clavijo, Z. Qian and R. Emrick, “Au-
tomated Removal of Metallic Carbon Nanotubes in a Nanotube Ensemble by Electrical
Breakdown,” Proc. IEEE Conference on Nanotechnology, 2008, pp. 239-242.
[2] T. Austin, V. Bertacco, D. Blaauw and T. Mudge, “Opportunities and Challenges for Bet-
ter Than Worst-Case Design,” Asian South Pacific Design Automation Conference, 2005.
[3] A. Bachtold, P. Hadley, T. Nakanishi and C. Dekker, “Logic Circuits with Carbon Nan-
otube Transistors,” Science, 2001, 294(5545), pp. 1317-1320.
[4] P. Beckett, “A Fine-Grained Reconfigurable Logic Array Based on Double Gate Transis-
tors,” International Conference on Field-Programmable Technology, 2002, pp. 260-267.
[5] L. Benini and G. De Micheli, “Networks on chips: A new paradigm for component-based
MPSoC design,” Proc. MPSoC., 2004.
[6] J. Birkner, A. Chan, H. T. Chua, A. Chao, K. Gordon, B. Kleinman, P. Kolze and R. Wong,
“A Very High-Speed Field Programmable Gate Array Using Metal-To-Metal Anti-Fuse
Programmable Elements,” New Hardware Product Introduction at Custom Integrated Circuits
Conference, 1991.
[7] T. J. Brosnan and N. R. Strader II, “Modular Error Detection for Bit-Serial Multiplication,”
IEEE Trans. on Computers, 37(9), 1988, pp. 1043-1052.
[8] J. G. Brown and R. D. Blanton, “CAEN-BIST: Testing the NanoFabric,” Proc. International
Test Conference, 2004, pp. 462-471.
[9] Z. Chen, J. Appenzeller, Y.-M. Lin, J. Sippel-Oakley, A. G. Rinzler, J. Tang, S. J. Wind, P.
M. Solomon and P. Avouris, “An Integrated Logic Circuit Assembled on a Single Carbon
Nanotube,” Science, 2006, 311(5768), pp. 1735.
[10] B. Culbertson, R. Amerson, R. Carter, P. Kuekes and G. Snider, “Defect Tolerance on the
Teramac Custom Computer,” Proc. Symposium on FPGA’s for Custom Computing Machines
(FCCM), 2000, pp. 185-192.
[11] A.DeHon, “Array-Based Architecture for FET-Based, Nanoscale Electronics,” IEEE Trans.
on Nanotechnology, 2(1), pp. 23-32, 2003.
[12] A. DeHon, P. Lincoln and J. E. Savage, “Stochastic Assembly of Sublithographic
Nanoscale Interface,” IEEE Trans. Nanotechnology, 2(3), pp. 165-174, 2003.
[13] A. DeHon and M. J. Wilson, “Nanowire-Based Sublithographic Programmable Logic Ar-
rays,” Proc. FPGA, 2004, pp. 123-132.
406 VLSI

[14] C. Dwyer, L. Vicci, J. Poulton, D. Erie, R. Superfine, S. Washburn and R. M. Taylor, “The
design of DNA self-assembled computing circuitry,” IEEE Trans. on Very Large Scale Inte-
gration (VLSI) Systems, 12(11), pp. 1214-1220, Nov. 2004.
[15] D. J. Deleganes, M. Barany, G. Geannopoulos, K. Kreitzer, M. Morrise, D. Milliron, A.
P. Singh and S. Wijeratne, “Low-Voltage Swing Logic Circuits for a Pentium 4 Processor
Integer Core,” IEEE J. of Solid-State Circuits, 40(1), pp. 36-43, 2005.
[16] J. Deng, and H.-S. P. Wong, “A Compact SPICE Model for Carbon Nanotube Field Effect
Transistors Including Non-Idealities and Its Application âĂŤ Part II: Full Device Model
and Circuits Performance Benchmarking,” IEEE Trans. Electron Devices, 2007.
[17] X. Duan, Y. Huang and C. M. Lieber, “Nonvolatile Memory and Programmable Logic
from Molecule-Gated Nanowires,” Nano Letters, 2(5), pp. 487-490, 2002.
[18] D. Ernst, N. S. Kim, S. Das, S. Lee, D. Blaauw, T. Austin, T. Mudge and K. Flautner, “Ra-
zor: Circuit-Level Correction of Timing Errors for Low-Power Operation,” IEEE MICRO
special issue on Top Picks From Microarchitecture Conferences of 2004, 2005.
[19] P. Friedberg, Y. Cao, J. Cain, R. Wang, J. Rabaey, and C. Spanos, “Modeling Within-Die
Spatial Correlation Effects for Process-Design Co-Optimization,” Proc. International Sym-
posium on Quality Electronic Design, pp. 516-521, 2005.
[20] S. C. Goldstein and M. Budiu, “NanoFabrics: Spatial Computing Using Molecular Elec-
tronics,” Proc. International Symposium on Computer Architecture, 2001, pp. 178-191.
[21] B. Gojman, E. Rachlin, and J. E. Savage, “Evaluation of Design Strategies for Stochasti-
cally Assembled Nanoarray Memories,” Journal of Emerging Technologies, 1(2), 2005, pp.
73-108.
[22] J. R. Heath and M. A. Ratner, “Molecular Electronics,” Physics Today, 56(5), pp. 43-49,
2003.
[23] S. J. Kang, C. Kocabas, T. Ozel, M. Shim, N. Pimparkar, M. A. Alam, S. V. Rotkin and J. A.
Rogers, “High-Performance Electronics Using Dense, Perfectly Aligned Arrays of Single-
Walled Carbon Nanotubes,” Nature Nanotechnology, Vol. 2, pp. 230-236, April 2007.
[24] B. S. Landman and R. L. Russo, “On a Pin Versus Block Relationship for Partitions of
Logic Graphs,” IEEE Trans. on Computers, C-20, pp. 1469-1479, 1971.
[25] J. Liu, I. O’Connor, D. Navarro and F. Gaffiot, “Design of a Novel CNTFET-Based Recon-
figurable Logic Gate,” Proc. ISVLSI, 2007, pp. 285-290.
[26] B. Liu, “Reconfigurable Double Gate Carbon Nanotube Field Effect Transistor Based Na-
noelectronic Architecture,” Proc. Asia and South Pacific Design Automation Conference, 2009.
[27] B. Liu, “Adaptive Voltage Controlled Nanoelectronic Addressing for Yield, Accuracy and
Resolution,” Proc. International Symposium on Quality Electronic Design, 2009.
[28] B. Liu, “Robust Differential Asynchronous Nanoelectronic Circuits,” Proc. International
Symposium on Quality Electronic Design, 2009.
[29] B. Liu, “Defect Mapping and Adaptive Configuration of Nanoelectronic Circuits Based
on a CNT Crossbar Nano-Architecture,” Workshop on Nano, Molecular, and Quantum Com-
munications (NanoCom), 2009.
[30] International Technology Roadmap for Semiconductors, http://www.itrs.net/.
[31] A. Maheshwari and W. Burleson, “Differential Current-Sensing for On-Chip Intercon-
nects,” IEEE Trans. on VLSI Systems, 12(12), 2004, pp. 1321-1329.
[32] P. L. McEuen, M. S. Fuhrer and P. Hongkun, “Single-walled Carbon Nanotube Electron-
ics,” IEEE Trans. Nanotechnology, 1(1), pp. 78-85, 2002.
[33] M. Mishra and S. C. Goldstein, “Defect Tolerance at the End of the Roadmap,” Proc.
International Test Conference, 2003, pp. 1201-1211.
Nanoelectronic Design Based on a CNT Nano-Architecture 407

[34] S. Mitra, N. Patil and J. Zhang, “Imperfection-Immune Carbon Nanotube VLSI Logic
Circuits,” Foundations of NANO, 2008.
[35] T. K. Moon, Error Correction Coding: Mathematical Methods and Algorithms, Wiley-
Interscience, 2005.
[36] H. Naeimi and A. DeHon, “A Greedy Algorithm for Tolerating Defective Crosspoints in
NanoPLA Design,” Proc. Intl. Conf. on Field-Programmable Technology, 2004, pp. 49-56.
[37] P. Nguyen, H. T. Ng, T. Yamada, M. K. Smith, J. Li, J. Han and M. Meyyappan, “Direct
Integration of Metal Oxide Nanowire in Vertical Field-Effect Transistor,” Nano Letters,
2004, 4(4), pp. 651-657.
[38] A. Nieuwoudt and Y. Massoud, “Assessing the Implications of Process Variations on Fu-
ture Carbon Nanotube Bundle Interconnect Solutions,” Proc. Intl. Symp. on Quality Elec-
tronic Design, 2007, pp. 119-126.
[39] M. Orshansky, L. Milor, P. Chen, K. Keutzer, C. Hu, “Impact of spatial intrachip gate
length variability on the performance of high-speed digital circuits,” IEEE Trans. on
Computer-Aided Design of Integrated Circuits and Systems, 2002, pp. 544-553.
[40] S. S. P. Parkin, “Spintronics Materials and Devices: Past, Present and Future,” IEEE Inter-
national Electron Devices Meeting (IEDM) Technical Digest, pp. 903-906, 2004.
[41] N. Patil, J. Deng, A. Lin, H.-S. Philip Wong and S. Mitra, “Design Methods for Mis-
aligned and Mispositioned Carbon-Nanotube Immune Circuits,” IEEE Tran. on CAD,
2008, 27(10), pp. 1725-1736.
[42] J. P. Patwardhan, C. Dwyer, A. R. Lebeck and D. J. Sorin, “NANA: A Nano-Scale Active
Network Architecture,” ACM Journal on Emerging Technologies in Computing Systems, 2(1),
pp. 1-30, 2006.
[43] J. P. Patwardhan, V. Johri, C. Dwyer and A. R. Lebeck, “A Defect Tolerant Self-Organizing
Nanoscale SIMD Architecture,” International Conference on Architecture Support for Pro-
gramming Languages and Operating Systems, 2006, pp. 241-251.
[44] A. R. Pease, J. O. Jeppesen, J. F. Stoddart, Y. Luo, C. P. Collier and J. R. Heath, “Switching
Devices Based on Interlocked Molecules,” Acc. Chem. Res., 34, pp. 433-444, 2001.
[45] Predictive Technology Model, http://www.eas.asu.edu/∼ptm/.
[46] A. Raychowdhury and K. Roy, “Carbon Nanotube Electronics: Design of High Perfor-
mance and Low Power Digital Circuits,” IEEE Trans. on Circuits and Systems - I: Funda-
mental Theory and Applications, 54(11), pp. 2391-1401, 2007.
[47] G. S. Rose, A. C. Cabe, N. Gergel-Hackett, N. Majumdar, M. R. Stan, J. C. Bean, L. R. Har-
riott, Y. Yao and J. M. Tour, “Design Approaches for Hybrid CMOS/Molecular Memory
Based on Experimental Device Data,” Proc. Great Lakes Symposium on VLSI, 2006, pp. 2-7.
[48] J. E. Savage, E. Rachlin, A. DeHon, C. M. Lieber and Y. Wu, “Radial Addressing of
Nanowires,” ACM Journal of Emerging Technologies in Computing Systems, 2(2), pp. 129-
154. 2006.
[49] M. S. Schmidt, T. Nielsen, D. N. Madsen, A. Kristensen and P. BÃÿggild, “Nano-Scale
Silicon structures by Using Carbon Nanotubes as Reactive Ion Masks,” Nanotechnology,
16, pp. 750-753, 2005.
[50] G. S. Snider and R. S. Williams, “Nano/CMOS Architectures Using a Field-
Programmable Nanowire Interconnect,” Nanotechnology, 18(3), 2007.
[51] M. R. Stan, P. D. Franzon, S. C. Goldstein, J. C. Lach and M. M. Ziegler, “Molecular
Electronics: From Devices and Interconnect to Circuits and Architecture,” Proc. of the
IEEE, 91(11), pp. 1940-1957, 2003.
[52] Stanford CNFET Model, http://nano.stanford.edu/models.php.
408 VLSI

[53] R. Sordan, K. Balasubramanian, M. Burghard and K. Kern, “Exclusive-OR Gate with a

Single Carbon Nanotube,” Appl. Phys. Lett., 2006, 88.
[54] D. B. Strukov and K. K. Likharev, “CMOL FPGA: A Reconfigurable Architecture for Hy-
brid Digital Circuits with Two-Terminal Nanodevices,” Nanotechnology, 16(6), pp. 888-
900, 2005.
[55] J. P. Sun, G. I. Haddad, P. Mazumder and J. N. Schulman, “Resonant Tunneling Diodes:
Models and Properties,” Proc. of the IEEE, 86(4), 1998, pp. 641-660.
[56] Y.-C. Tseng, P. Xuan, A. Javey, R. Malloy, Q. Wang, J. Bokor and H. Dai, “Monolithic
Integration of Carbon Nanotube Devices with Silicon MOS Technology,” Nano Letters,
4(1), pp. 123-127, 2004.
[57] N. Tzartzanis and W. W. Walker, “Differential Current-Mode Sensing for Efficient On-
Chip Global Signaling,” IEEE J. of Solid-State Circuits, 40(11), 2005, pp. 2141-2147.
[58] Z. Wang and K. Chakrabarty, “Using Built-In Self-Test and Adaptive Recovery for Defect
Tolerance in Molecular Electronics-Based NanoFabrics,” Proc. International Test Confer-
ence, 2005.
[59] R. S. Williams and P. J. Kuekes, “Demultiplexer for a Molecular Wire Crossbar Network,”
US Patent Number 6,256,767, 2001.
[60] J. Xiong, V. Zolotov, and L. He, “Robust Extraction of Spatial Correlation,” Proc. Interna-
tional Symposium on Physical Design, 2006, pp. 2-9.
[61] J. Zhang, N. P. Patil, A. Hazeghi and S. Mitra, “Carbon Nanotube Circuits in the Presence
of Carbon Nanotube Density Variations,” Proc. Design Automation Conference, 2009.
[62] J. Zhang, N. P. Patil and S. Mitra, “Design Guidelines for Metallic-Carbon-Nanotube-
Tolerant Digital Logic Circuits, Proc. Design Automation and Test in Europe, 2008, pp. 1009-
1014.
[63] W. Zhang, N. K. Jha and L. Shang, “NATURE: A Hybrid Nanotube/CMOS Dynamically
Reconfigurable Architecture,” Proc. Design Automation Conference, 2006, pp. 711-716.
[64] G. Zhang, P. Qi, X. Wang, Y. Lu, X. Li, R. Tu, S. Bangsaruntip, D. Mann, L. Zhang and H.
Dai, “Selective Etching of Metallic Carbon Nanotubes by Gas-Phase Reaction,” Science,
314, 2006, pp. 974-977.
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 409

20
1

A New Technique of Interconnect Effects

Equalization by using Negative Group
Delay Active Circuits
Blaise Ravelo, André Pérennec and Marc Le Roy
UEB, University of Brest
Lab-STICC, UMR CNRS 3192,
France

1. Introduction
During the last two decades, technological progresses in VLSI process have brought an
outstanding development of information technology equipments and thus a great increase
in the use of communication services all over the world. As reported by both the
International Technology Roadmap for Semiconductors (ITRS) and the Overall Roadmap
Technology Characteristics (ORTC), the exponential reduction of the feature size of
H

electronic chips according to Moore’s law (Moore, 1965) still occurs together with the
exponential increase in time of the number of transistor per unit area. Combined to this
shrinking of feature sizes, the on-chip clock frequency increases continually and should
exceed 10 GHz in 2010. These ceaseless trends in VLSI circuits have led to more and more
complex interconnect systems, and thus, the implementation of metal multi-layers for intra-
chip interconnects has become a must. In the mid-1980’s, the devices were, thus, composed
of one or two layers of aluminium; in 2011, according to ITRS prediction, chips will consist
of more than ten layers of copper.
Under these conditions, owing to the higher operation speeds, the interconnect propagation
delay becomes more and more significant and dominates considerably the logic propagation
delay (Deutsch, 1990; Rabay, 1996). Because of the sensitivities of parametric variations,
clock and data flow may not be synchronized (Friedman, 1995). This explains why
interconnections are so important in the determination of VLSI system performances.
Besides, simplified models are worth being considered in order to reduce the complexity of
any study on interconnects. Since the beginning of the 1980s the modelling of propagation
delay in order to estimate the delay of an interconnect line driven by a CMOS gate has been
the subject of numerous papers (Sakurai, 1983 and 1993; Deng & Shiau, 1990). The simplest
and most used model of this delay was proposed by Elmore in 1948; it relies on the use of
only an RC-line model. Nevertheless, due to the elevation of system data rates, this model
tends to be insufficiently accurate. Therefore, more accurate models that sometimes take into
account the inductive effect (Wyatt, 1987; Ismail et al., 2000) have been proposed.
410 VLSI

To solve the problem of clock skew and propagation delays, a technique of signal integrity
enhancement based on repeater insertion was proposed by different authors (Adler and
Friedman, 1998; Ismail & Friedman, 2000). But, when the signals are significantly
attenuated, such a solution may be unable to conserve the data duration and thus,
inefficient. These considerations drove us to recently propose a new technique for
interconnect-effect equalization (Ravelo, Perennec & Le Roy, 2007a, 2008a, 2009 and 2009)
through use of negative group delay (NGD) active circuits. As shown in Fig. 1, it consists
merely in cascading these NGD circuits at the end of the interconnect line. The possibility of
signal recovery with a reduction of signal rise/fall-, settling- and propagation-delays was
theoretically demonstrated and evidenced through simulations in (Ravelo et al., 2007a,
2008a) and confirmed by experiments in (Ravelo et al., 2009).

Fig. 1. Interconnect line driven by a logic gate ended by an NGD circuit.

In fact, evidences of the NGD phenomenon have been provided through theoretical
demonstrations and experiments with passive-electronic devices (Lucyszyn et al., 1993;
Eleftheriades et al., 2003; Siddiqui et al., 2004 and 2005) and active ones (Solli & Chiao, 2002;
Kitano et al. 2003; Nakanishi et al., 2002; Munday & Henderson, 2004). As described in
several physics domains (Wang et al., 2000; Dogariu et al., 2001; Solli & Chiao, 2002), in the
case of a smoothed signal propagating in a device/material that generates NGD, the, the
peak of the output signal and its front edge are both in time advance compared to the input
ones. Then, confirmations that this counterintuitive phenomenon is not physically at odds
with the causality principle have been provided (Wang et al., 2000; Nakanishi et al., 2002). A
literature review shows that the first circuits that exhibited NGD at microwave wavelengths
displayed also significant losses, whereas the baseband-operating ones were intrinsically
limited to low frequencies. To cope with these issues, we, recently, reported on the design,
test and validation through simulations and experiments of a new and totally integrable
topology of NGD active circuit (Ravelo et al., 2007b, 2007c and 2008b); this topology relies
on the use of a FET and showed its ability to compensate for losses at microwave
frequencies over broad bandwidth. Transposition of this NGD topology to baseband
frequencies allowed us to develop new structures that demonstrated their ability to
simultaneously generate an NGD and gain for broad and baseband signals (Ravelo et al,
2008). Then, the idea put forward by Solly and Chiao (Solly, 2002) to compensate
degradations introduced by passive systems such as interconnect lines by using NGD
devices became possible.
As a continuation of these investigations, this chapter deals with further developments of
this technique. Section 2 gives insight into the way this technique works, and briefly
explains the role of the NGD circuit. The theory of interconnect modelling and the definition
of the propagation delay are both recalled in Section 3. Analytical approach and
experimental validations of RC-model equalization are presented in Section 4. The feasibility
of the proposed technique, when the inductive effects are taken into account, is dealt in
Section 5. In Section 6, a completely original and fully-integrable topology is proposed by
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 411

getting rid off inductance to cope with their implementation issue. Thus, the results of
simulations, which provided a very good validation of the performances expected from
theory, are analysed and discussed. A summary of this chapter is given in the last Section
together with proposals about possible future developments.

2. The basic principle of the proposed compensation technique

Figure 2 illustrates the general case of the distortion undergone by a numerical signal
degraded by the interconnect circuitry as introduced in Fig. 1. Compared to the input signal,
vi(t), the degradation of the output one, vl(t), can be assessed from the signal attenuation as
well as the rise-, fall- and settling-times and the 50% propagation delay. It is worth recalling
that this last parameter, denoted here by Tp50% or merely Tp, is defined as the time needed by
the output signal, vl, to reach 50% of the unit-step input of amplitude, VM, here. The delay
required to reach 50% of the logic swing is traditionally referred as the delay time.

Fig. 2. Time-domain responses of the ideal system shown in Fig. 1 for a periodical input
voltage, vi(t).

In frequency domain, the degradation between the input, vi and the output, vl corresponds
to a transfer function denoted, Gl(s) whose gain magnitude and group delay usually verify
the following inequalities:

Gl ( j )  1 and  l ( )  Gl ( j ) /   0 . (1)

The output Laplace transform of this interconnect line can be written as:

vl(s) = Gl(s) . vi(s). (2)

As shown by Figs. 1 and 2, this study was aimed at finding a relevant configuration or
circuit able to provide a compensated output, vN, (black thick curve) as close as possible to
the input signal, vi, (red dashed curve). It means that the following mathematical
approximation can be made:

vN(t) ≈ vi(t). (3)

412 VLSI

In theory, for well-matched circuits, the transfer system to be found, Gx(s), must be
associated to Gl(s) so that equation (4) is verified:

vN (s)  Gl (s).Gx (s ).vi (s) . (4)

According to the circuit and system theory, through use of equations (3) and (4), one gets
the adequate transfer function:

Gl (s ).G x (s ).vi (s )  v i (s )  G x (s )  1 / G p (s ) . (5)

Consequently, in the frequency domain, the system gain and group delay must be such that:

Gx ( j ) dB  Gl ( j ) dB , (6)
 x ( )   l ( ) . (7)

So, on condition to take into account the condition expressed in equation (1), the gain and
the group delay must be respectively such that |Gx(jω)|dB > 0 and τx(ω) < 0. Technically,
these conditions require the cascade of a system able to simultaneously exhibit Gain and an
NGD in baseband. As described by the block diagram of Fig. 3, the whole cascaded system
is characterized by its transfer function G(s) where Gl is the interconnect transfer function
and Gx = GNGD;  is the whole group delay.

Fig. 3. Block diagram: interconnect passive system, Gl(s) cascaded by the active NGD
circuit, GNGD(s): G(s) = Gl(s).GNGD(s).

This process constitutes, then, a technological solution of equalization technique. In this

case, the gain and group delay generated by the compensation system should be
respectively the reverse and the opposite of those of the interconnect circuitry as depicted in
Fig. 4. The principle of interconnect loss compensation and group delay reduction is
illustrated in Fig. 4. At this stage, it is worth noting that the compensator must contain a
system able to exhibit a group delay negative at base band frequencies not only with an
opposite value of the interconnect one but also over the corresponding frequency band. A
similar remark can be made about the gain compensation. The need for active circuits drove
us to propose, here, a topology based on the use of a FET, chosen because of its biasing
simplicity, the different models available (simple for analytical theory and more accurate
ones for final simulations), its faculty to operate at tens of GHz as well as its easy
integration.
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 413

Fig. 4. Illustration of the compensation principle: frequency responses of the magnitude in

dB (a) and group delay (b).

In conclusion, the NGD circuit consists in the achievement of a group delay and attenuation
compensatory function. Fig. 4 depicts the compensation principle by considering the general
frequency behaviour of interconnects or transmission lines. So, prior to conducting a
feasibility study of this technique with concrete systems, let us briefly recall the theory on
commonly used models of interconnects, i.e. RC- and RLC-circuits, and on propagation
delay assessments.

3. Theory on the interconnect modelling and the propagation delay

approximation
In most of microelectronic device interconnect models, the conductance effect can be
neglected compared to the per-unit length resistance, Rl, the inductance, Ll, and the
capacitance, Cl. Therefore, let us consider the interconnect structure presented in Fig. 5: it
consists of an RLC-line model of length, d, driven by a gate with an output resistance, Rs. As
previously mentioned, vl is the circuit output voltage and vi, the input one.
414 VLSI

Fig. 5. A gate with output resistance, Rs, driving a RLC-model interconnect.

For the time-domain analysis of the structure under study, the input voltage, vi, is assigned
as a Heaviside unit step function, Г(t), of amplitude, VM:

vi(t)=VMГ(t). (8)
As defined previously, the 50% propagation delay, Tp , is defined as the root of equation (9):

vl(Tp) = VM/2. (9)

Before the calculation of this propagation delay, it is worth focusing on the system transfer
function, which is defined as Gl(s) = Vl(s)/Vi(s). It was established (Ajoy et al., 2004) that,
according to the configuration of Fig. 5, this quantity is expressed as:

1
Gl ( s )  , (10)
(1  sRs ) cosh(d )  ( Rs / Zc ) sinh(d )
where
Z c  ( R l  L l s ) /C l s , (11)
and
  ( Rl  Ll s)C l s , (12)
are, respectively, the characteristic impedance and the propagation constant of the line. This
transfer function is analysed through use of polynomial expansion as done for classical
linear systems. For simplification, let us deal with the normalized transfer function, gl(s)
which is determined by Gl(s)/Gl(0), and that can be expressed by the m-order linear
expression:

1  a1 s  a 2 s 2  ...  an s n
g l (s )  , (13)
1  b1 s  b 2 s 2  ...  bm s m
where the coefficients, ai (i = {1,...,n}) and bj (j = {1,...,m}) are real numbers, and m and n are
integers. According to the literature (Elmore, 1948; Wyatt, 1987; Ismail et al., 2000), use of
this expression allows one to estimate the 50% propagation delay expressed in equation (9).
Among the existing approximation, it is worth recalling that the simplest and the most used
in the industrial context is the one proposed by Elmore in 1948. It is based on the first-order
consideration of equation (13). Indeed, this estimation of the propagation delay is merely
defined by:
TpElmore = b1 - a1. (14)
One should note that this propagation delay is exactly equal to the group delay of the
system under consideration at very low frequencies (Vlach et al. 1991). Nevertheless, this
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 415

formula has proven to become less and less accurate towards the elevation of the operating
frequency. A new formula was proposed by Wyatt (Wyatt 1987): it differs from the previous
approach by only the values of the coefficients, ai and bi, which are defined by the reciprocal
of the dominant pole of the system transfer function:

n
a1  z
i 1
1
i , (15)

m
b1  p
i 1
1
i , (16)

where the real numbers, zi and pi, are, respectively, the zeros and the poles of gl(s). It is
worth noting that this approach provides an exact expression of Tp in the case of the RC-
model (Ll = 0).

3.1 Recall on RC-line model

It was established (Sakurai , 1983; Deng & Shiau, 1990) that the first-order approximation of
the transfer function in equation (10) leads to the following expression:

Gl (s )  1 /1  [( Rs  Rl d )  C l d ]s  Grc (s ) . (17)

This allows the modelling of the interconnect circuitry presented in Fig. 5 as a simple
equivalent circuit (Fig. 6).

Fig. 6. Simplified representation of the structure shown in Fig. 5 by considering the first-
order approximation of the transfer function.

Therefore, the driver gate loaded by the distributed transmission line can be equivalent to a
lumped RC-circuit. To make easier the analytical calculation, let us consider the equivalent
parameters of the system under study, Rt = Rs + Rld and C = Cld. So, the Elmore propagation
delay of this well-known circuit is given by:
TpRC = Rt C . (18)
But, it can be shown by calculation through the unit step response that the exact value of
this quantity is expressed as:
TpRC  RtC ln(2 ) . (19)
Then, the rise time, denoted trRC, which is defined as the time needed by the output signal to
pass from 10% to 90% of the final output value, is written as:
trRC = RtC ln(9) . (20)
416 VLSI

3.2 Summary on RLC-line theory

As previously mentioned, nowadays, interconnect line modelling requires, in most cases, to
thoroughly consider the inductance effect. This parameter is usually taken into account
through use of a second-order system, such as an RLC-network with the following canonical
transfer function:

g l (s )  n2 /(s 2  2 n s   n2 ) . (21)

where
 n  1 / Lt C , (22)
and
  ( Rt / 2 ) C / L t , (23)
are, respectively, the undamped natural angular frequency and the damping ratio. From this
second-order transfer function, an accurate propagation delay, Tpd, was established by
Ismail & Friedman (2000):

1.35
Tp  ( e -2.9  1.48 ) /n . (24)
To reduce this propagation delay, and as first envisaged by Solli and Chiao (Solli et al, 2002),
it could be worth cascading the interconnect line with an NGD active circuit. But a
preliminary to this proposal is a detailed study of the resulting device in order to get a
confirmation of the compensation principle efficacy. This will be the focus of the next
Section.

4. Compensation of a first-order interconnect model (RC-circuit)

To compensate for interconnect spurious effects, especially losses and time delay, we
recently developed and tested a new equalization technique based on the use of an NGD
active circuit (Ravelo et al., 2007a, 2008a and 2009). In this section, we will briefly recall the
theoretical fundamentals and validate the technique through experiments carried out in
frequency- and time-domains in the case where a first-order circuit, i.e. an RC-circuit, is
used to model the interconnect effects.

4.1 Theory
As explained above, the equalization principle consists in ending the circuit to be
compensated, here an RC one, with an NGD active cell as depicted in Fig. 7. To simplify the
analytical approaches, the FET is modelled by a controlled voltage current source with a
transconductance, gm, and the drain source resistance, Rds.
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 417

Fig. 7. RC-NGD circuit: RC-circuit cascaded with a basic cell of the NGD active circuit (FET
in feedback with an RL series network) and the corresponding equivalent circuit.

In a first step, let us recall (Ravelo et al., 2007a, 2008a, 2008b, 2009) the transfer function and
the group delay expressions of the NGD cell alone (in the dash box in Fig. 7) at low
frequencies:

(1  g m R ) Rds
G NGD (0)  , (25)
R  Rds
L(1  g m Rds ) .
 NGD (0)  (26)
( R  Rds )(1  g m R)
On condition that:
1  gm R  0 , (27)
the NGD cell exhibits a negative group delay (τNGD(ω  0) < 0) at very low frequencies.

Then, the transfer function and the gain at low frequencies of the whole RC-NGD circuit
presented in Fig. 7 are, respectively, expressed as:

Rds (1  g m R )  g m Rds Ls ,
G (s)  (28)
R  Rds  Rt (1  g m Rds )   Rt C ( R  Rds )  L  s  Rt CLs 2
Rds (1  g m R) .
G (0)  (29)
R  Rds  Rt (1  g m Rds )
One should note that, because of the unmatched effect between the RC- and NGD-circuits,
the transfer function, G(s), is not equal to Grc(s).GNGD(s) (Ravelo et al., 2009). The same
remark applies to the group delays:

G( j )  Grc ( j ) . GNGD ( j ) , (30)

 ( )   rc ( )   NGD ( ) . (31)

(i) Demonstration of the 50% propagation delay reduction:

From equations (18) and (28), the Elmore propagation delay (or the group delay at low
frequencies) of the compensated RC-NGD circuit can be expressed as follows (Ravelo et al.,
2008a):
418 VLSI

2
[1  gm ( Rt  Rds ) gm Rt Rds ]L  TpRC ( R  Rds )(1  gm R )
Tp  . (32)
( R  Rt  Rds  gm Rt Rds )(1  gm Rds )
So, the delay is reduced (Tp < TpRC,) on condition that:

L  (1  g m R )Rt2C /( 1  g m R ) . (33)
But this inequality is automatically verified if the condition expressed in equation (27) is
true, i.e. if the active circuit generates NGD. According to this study, the optimum values of
the NGD circuit can be synthesised as a function of the RC-values.

(ii) Synthesis of an NGD cell according to the RC-model values

As the aim of the equalization principle is the generation of an output signal equal or close
to the input one, the transfer function magnitude and group delay should be, respectively
almost equal to unity and null. For simplification purpose, let us apply these conditions at
very low frequencies:

|G(0)| ≈ 1 and τ(0) ≈ 0. (34)

From equations (31) and inversion of (29) and (32), one gets the synthesis relations
expressed in equations (35) and (36):

R  [ 2 Rds  ( gm Rds  1)Rt ] /( gm Rds  1) , (35)

( g m R  1)( R  Rds )RtC
L 2
. (36)
g m Rds Rt  g m ( Rds  Rt )  1
These expressions are meaningful when the losses displayed by the interconnect are less
than the maximal gain magnitude of an NGD cell, Gmax :

Gmax  g m Rds . (37)

Otherwise, several stages of NGD cells are needed to generate a whole gain of about unity;
then, as reported in (Ravelo et al., 2008a), it is advised to use equation (38) for the calculation
of the optimal number of cells, n:

n  1  int[  ln( 1  e T /( Rt C ) ) / ln(Gmax )] . (38)

T is the time duration of the considered input square pulse. For a real x, the value of the
greatest integer given by the function, int(x), is equal to or lower than x. It is worth
underlining that the implementation of, at least, two or an even number of transistors is
preferable when the application under study requires avoiding signal inversion.

To validate these theoretical predictions, proof-of-concept (POC) circuits were designed,

fabricated and measured in frequency- and time-domains as explained hereafter.
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 419

4.2 Experimental validations

The starting values of the NGD circuit come from the synthesis relations knowing the
characteristics of the chosen FET and the RC-model values. Then, the most accurate
available models were used to run simulations of the POC circuits on ADSTM simulator
software from AgilentTM ; a sensitivity analysis was also performed. The different POC
circuits presented in this section were implemented in hybrid planar technology (association
of microstrip and surface mount chip/package components).

(i) Design Process

This demonstrator was aimed at illustrating visually the recovery of the degraded signal
and the delay reduction. The circuits were designed and implemented separately; they
consisted of an RC-circuit, an NGD-one and the combination of both into an RC-NGD circuit
(Fig. 8(b)). To perform a broadband biasing and to avoid the eventual disruptions caused by
the bias network at low frequencies, the FET was biased through an active-load technique.

(a)

(b)
Fig. 8. Schematics of the (a) RC- and (b) RC-NGD-circuits (including the biasing network in
thin lines) using a PHEMT FET (ATF-34143, Vgs = 0 V, Vd = 3 V, Id = 110 mA), for Rt = 33 Ω, C
= 680 pF, R = 56 Ω, Ro = 10 Ω, L = 220 nH, Cb = 100 nF and Z0 is the output reference load.

The FET parameters required for the synthesis equations were extracted from the non-linear
model and found to be gm = 226 mS and Rds = 27 . For the RC circuit, the RL values of the
NGD circuit were synthesised from equations (33) and (34). Then, accurate frequency
responses were obtained through combination of circuit simulations (lumped components
and non linear FET model) with electromagnetic co-simulations of the distributed parts by
Momentum Software (from AgilentTM). The values of the available lumped components were
used in a final slight optimisation procedure. The layout of the hybrid planar circuit shown
in Fig. 9 was printed on an 800-µm-thick FR4 substrate of relative permittivity, εr = 4.3; then
the surface-mount chip passive components were set onto the substrate.
420 VLSI

Fig. 9. Layout of the implemented RC-NGD circuit.

(ii) Simulation Results and Sensitivity Analysis

The simulations of the circuit pictured in Fig. 9 and of the RC and NGD ones were
independently run in frequency- and time-domains under the specifications and conditions
mentioned above. A sensitivity Monte-Carlo analysis over 10 trials was carried out on using
tolerance values of ± 5% around their nominal values for the NGD elements (R = 56 Ω and
L = 220 nH). All the results are presented in Figs. 10 and 11.

- Frequency-domain analysis: Fig. 10(a) shows that the magnitude |G(f)| of the whole RC-
NGD circuit is kept at about 0 dB up to 40 MHz, and the corresponding group delay |(f)|
is, as expected, strongly decreased from DC to 15 MHz compared to that of the RC circuit.
Moreover, the frequency responses are only slightly sensitive to lumped-component
tolerance.

(a) (b)
Fig. 10. (a) Magnitude and (b) group delay frequency responses issued from simulations
with a Monte-Carlo analysis of the NGD circuit elements R and L (± 5%).

- Time-domain analysis: It is worth underlining that, at first, the signal was measured at the
output of a Rhode & Schwarz Signal Generator SMJ 100A at the highest available rate, i.e. 25
Msym/s, and then further used in simulations as the input signal. Then, excitation of the
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 421

simulated RC- and RC-NGD-circuits with this input square wave pulse (magnitude, VM = 1
V) led to the time-domain simulation results presented in Fig. 11, where the dotted curve
indicates the degradation induced by the RC-circuit alone; moreover, the signal recovery is
evidenced by the thick black curve. It clearly appears that the compensation is significant,
but incomplete. Because of the drain-source current inversion, the output signal is reversed
compared to the input voltage. This is why the plot presented in Fig. 11 is that of -VN.

Fig. 11. Results of time-domain simulations (with a Monte Carlo analysis of the NGD circuit
elements, R and L (± 5%)) for an input square wave pulse (VM = 1 V with a 40 ns data
duration).

It is worth noting that the 50% propagation delay is shortened from 16.5 ns (for the RC
circuit) to about 3 ns (for the RC-NGD circuit), i.e. a relative reduction of, at least, 81%. A
Monte-Carlo sensitivity analysis with a ± 5%-variation around the R and L nominal values
showed nearly no change of the RC-NGD circuit output signal (including the front and
leading edges), but the final value is somewhat slightly affected.

(iii) Measured results and discussions:

To check for the validity of the results of theoretical analysis and simulations, three circuits
were fabricated and tested in both frequency- and time-domains.

- Frequency-domain results: These measurements were made with a vector network

analyzer (Rhode & Schwarz ZVRE 9kHz - 4GHz), which provides the scattering matrix (S-
parameters) of the devices under test. Then, the transfer function is calculated through use
of the classical passage relationships in order to get the magnitude and the group delay
(calculated from the transfer function phase response). The fabricated devices consisted of
the single RC-circuit, the NGD one and a third one, denoted RC-NGD circuit, which
combined both functions in order to avoid any connector mismatch liable to occur if the first
two ones were simply cascaded. Figure 12(a) shows that the magnitude and group delay
responses of the three circuits are in good agreement with the simulation results presented
in Figs. 10. As expected from equations (30) and (31), the total magnitude and group delay
values are different from those issued from the addition of the individual ones because of
the existence of mismatch between the RC and NGD parts (GRCNGD ≠ GRC.GNGD). The
measured magnitude, |G(f)|dB, is close to 0 up to 40 MHz and gets down to -10 dB at 80
MHz.
422 VLSI

(a) (b)
Fig. 12. Measurements of (a) magnitude and (b) group delay produced by the RC-, NGD-
and RC-NGD-circuits.

One should note that, though the RC-NGD circuit group delay, (f), is not fully cancelled
(Fig. 12(b)), it is strongly reduced below 20 MHz thanks to the NGD circuit; moreover, it is
kept below 4 ns up to 80 MHz. In theory, a higher absolute value of NGD could be obtained,
but a compromise between NGD value, NGD bandwidth and gain flatness has to be found
to minimize the overshoot or the ripple in time-domain
- Time-domain experimental results: Throughout the time-domain simulations and
measurements, the circuit load, Z0, is set at a high impedance value (Z0 = 1 M).

Fig. 13. Time-domain responses with an input square pulse (25-Msym/s rate, 2 ns rise- and
fall-times) and zoom on twice the symbol duration.
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 423

A square wave input pulse of amplitude VM = 1 V was delivered at a rate of 25-Msym/s

(corresponding to a 40-ns data duration) by the baseband data output of an R&S SMJ 100A
vector signal generator and recorded on a 2 Gs/s LeCroy digital oscilloscope. The input
pulse, Vi, was thus monitored as well as the output ones of the RC- and RC-NGD-circuits,
VRC and -VN, prior to their resynchronization through use of the same reference signal.
Figure 13 shows that the output leading edge provided by the RC-circuit is degraded and
characterised by trRC  35 ns as rise time and a 50% propagation delay of TpRC  18.50 ns.
Compared to VRC, the output waveform, -VN, is reshaped and less distorted. Hence, the
NGD circuit compensation allowed a reduction of both parameters down to tr  10 ns and
Tp  2.50 ns, that is relative reductions of 71.4% (1-tr/trRC) and 86.5% (1-Tp/TpRC), respectively.
As shown in Fig. 13, the trailing edge is also strongly improved. This point is worth being
noted in the case of an enhancement of the data rate.
The proposed interconnect equalization technique was validated through modelling of the
interconnect line degradation by an RC-circuit, which is the most current model in use.
However, in order to supplement this study, the inductive effects of interconnect lines will
be taken into account in the next Section.

5. Equalization under conditions of interconnect line inductive effects

As underlined at the beginning of this chapter, the ceaseless increase of operating
frequencies, the inductive effects can no longer be neglected in the modelling of interconnect
systems. In this study, the line is modelled by the RLC network shown in Fig. 14(a) and
driven by a logic gate with Rs as output resistance. In order to check whether the proposed
compensation technique still works, let us consider the whole circuit presented in Fig. 14(b):
it consists of the interconnect model cascaded by the compensation circuit, which is
composed of a FET fedback by an RL-series network. To simplify the theoretical study, the
FET is replaced by its low frequency equivalent-circuit model as introduced in Fig. 5. The
terms, Rt = Rs + Rld, Lt = Lld and C = Cld, introduced in the previous study of the RC-network
are still used.

(a)
(b)
Fig. 14. (a) line model: RLC network driven by a logic gate with output resistance, Rs; (b) the
whole circuit composed of the line model compensated by an NGD cell.

5.1 Theory
According to the procedure used in the previous section, the transfer function of the whole
system is:
424 VLSI

G(s )  Rds [1  g m ( R  Ls )] /( 0   1 s   2 s 2   3 s 3 ) , (39)

where
 0  R  Rt  Rds (1  gm Rt ) , (40)
 1  RtC ( R  Rds )  gm Rds Lt  L  Lt , (41)
 2  [LRt  Lt ( R  Rds )]C , (42)
 3  LLtC . (43)
Then, at very low frequencies (ω  0), the corresponding gain, G(0), and the Elmore
propagation delay, Tp, are respectively expressed as:

G(0 )  [ Rds (1  gm R )] /[ R  Rt  Rds (1  gm Rt )] , (42)

( gm R - 1)[ RtC ( R + Rds ) + Lt (1 + gm R )] - (1 + gm Rds )(1 + gm Rt )
Tp = . (43)
[ R + Rt (1 + gm Rds ) + Rds ]( gm R - 1)
Application of the equalization objectives expressed in (32) to equations (42) and (43) allows
one to extract the following synthesis relations:

R  [ 2 Rds  ( g m Rds  1)Rt ] /( g m Rds  1) , (44)

L  ( g m R  1)( R  R ds )CRt  (1   2
g m R ds )Lt /[ g m R ds Rt  g m ( R ds  Rt )  1] . (45)
To evidence the relevance of this RLC effect equalization, simulations under realistic
conditions were run.

5.2 Validations by simulations

In order to solve the problem of the output voltage sign and to take into account inductive
effects, a two-stage NGD circuit was simulated through use of ADSTM for two types of
interconnect lines models: i) a model with lumped RLC elements and an input signal of 1 ns
in data duration, and ii) an RLC distributed line and an input signal of 5-ns data duration.
Moreover, using two NGD cells allows widening the NGD frequency bandwidth (Ravelo et
al, 2007c) and thus, input signals with higher data rates and smaller rise-/fall-times can be
considered.

(i) Compensation of a lumped RLC-circuit for an input signal with 1 ns-data duration
Fig. 15 depicts the whole compensated RLC-NGD circuit under study. It consists in a
lumped RLC-circuit ended with a two-stage NGD active circuit. The FETs used by the latter
(EC-2612 with gm = 98.14 mS and Rds = 116.8 Ω) are from Mimix BroadbandTM.

Fig. 15. Two-stage NGD active circuit compensating the RLC network (Rt = 100 Ω, Lt = 6 nH,
C = 4.3 pF, R1 = 86 Ω, R2 = 89 Ω, L1 = 5.2 nH, L2 = 2.6 nH and FET/EC-2612).
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 425

One should note that the time- and frequency-domain results presented here were obtained
by using the FET linear model provided by this manufacturer and recorded at the biasing
point Vds = 3 V and Ids = 25 mA.

- Frequency-domain results: Figs. 16 (a) and (b) present the magnitudes and the group
delays of the RLC- and NGD-circuits separately, and then, both cascaded (noted RLC-NGD),
as depicted in Fig. 15. Fig. 16(a) shows that the transfer-function magnitude of the overall
circuit (black thick curve) is kept within -4 and 0 dB up to 3 GHz.

(a) (b)
Fig. 16. Simulated frequency responses of the RLC, NGD and RLC-NGD circuits: (a)
magnitude and (b) group delay.

Once again, one should note that the value of this whole transfer-function magnitude is
different from the sum (in dB) of the individual magnitudes (GRLCNGD ≠ GRLC.GNGD) because
of mismatch between both parts. Figure 16(b) evidences that, thanks to the NGD effect (thin
blue curve), the total group delay of the whole circuit (black thick curve) is less than 68 ps. A
comparison of the group delays respectively produced by the RLC and the RLC-NGD
circuits shows a significant reduction up to about 0.8 GHz with the latter.

- Time-domain results: Figure 17 presents the results of transient simulations run in the case
of a 2-ns periodic input signal, Vi, whose rise-/fall-times are about 92 ps (thin red curve).
The response at the output of the RLC circuit alone, VRLC, is depicted by the degraded
dashed curve, and the corresponding 50% propagation delay, TpRLC, is equal to 304 ps.
Further to the insertion of the NGD circuit, the black thick curve representative of VN, i.e. at
the output of the RLC-NGD circuit, shows improvement with a relative reduction of more
than 85% of propagation delay since Tp is about 44 ps. Moreover, by comparison to Vi, VN
presents neither attenuation, nor overshoot, and the leading and trailing edges are both
improved.
426 VLSI

Fig. 17. Time-domain responses produced by simulations of the circuit in Fig. 15 for a 2-ns
periodic trapezoidal input, rise-/fall-times, tr = 92 ps and 50%-duty ratio with zoom on a
half-period (bottom).

As the accuracy of the lumped RLC model of interconnect line used in these simulations
might be insufficient for certain VLSI configurations, the simulation described in the next
paragraph dealt with a distributed RLC-model (Ravelo et al, 2007a).

(ii) Compensation of a distributed RLC-line in the case of an input signal of 5-ns data-
duration

Fig. 18. RLC-line driven by a gate of output resistance, Rs = 90 Ω, compensated by a two-

stage NGD circuit (FET/EC-2612, R1 = 73 Ω, L1 = 99 nH, R2 = 102 Ω and L2 = 17 nH) loaded
by another gate of input capacitor CL = 30 pF.

Fig. 18 shows the circuit under study; the interconnect line is modelled by an RLC
distributed circuit. The classical RLC-line is driven by a gate of output resistance, Rs,
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 427

compensated by a two-stage NGD circuit loaded with another gate of input capacitor, CL.
The transmission line parameters were taken from ITRS roadmap so that Rl = 76 Ω/cm,
Ll = 5.3 nH/cm and Cl = 2.6 pF/cm for a 0.8-cm-long line. For these values, the synthesis
relations were used together with an optimization process to get an output, VN, close to the
input, Vi. Finally, the component values for the two NGD cells were: R1 = 73 Ω, L1 = 99 nH,
R2 = 102 Ω and L2 = 17 nH.

- Frequency-domain results: Figs. 19(a) and (b) respectively show the transfer function
magnitude and the group delays produced by simulations of the three circuits (RLC-line,
NGD- and RLC-NGD-circuits). In Fig. 19(a), the important attenuation displayed by the
RLC-line is compensated by the NGD cells so that the total gain is kept at about 0 dB up to
about 500 MHz. With respect to the group delay produced by the RLC-line, the one by the
RLC-NGD circuit (thick black curve) is strongly reduced as expected thanks to the NGD
contribution (thin blue curve). Indeed, at baseband frequencies, this latter provides a
minimal NGD value around -0.5 ns.

(a) (b)
Fig. 19. Frequency responses produced through simulations of the RLC-, NGD-, and RLC-
NGD circuits shown in Fig. 18: magnitude (a) and group delay (b).

- Time-domain results: Time-domain simulations were run on using at the input a 10-
ns-period signal with 0.60 ns as rise-/fall-times.

Fig. 20. Time domain output responses of the RLC-line and of the RLC-NGD compensated
structure for a trapezoidal input voltage at a 200 Mbit/s rate.
428 VLSI

Fig. 20 evidences a marked degradation of the output waveform, Vl, of the RLC-line
compared to the input waveform. Once again, the insertion of the NGD active circuit allows
great enhancement illustrated by a notable signal regeneration, a reduction of the
propagation delay of about 1.60 ns (from 1.86 ns to 0.26 ns) accompanied with an
enhancement of the signal leading- and trailing-edges. These results confirm the efficacy
and the reliability of this technique to equalize the degradation induced either by lumped or
distributed RLC-model of interconnect lines.

6. Improvement of the NGD topology toward an integration process

Application of these innovative cells to the compensation of VLSI interconnect-induced
degradation requires compatibility with VLSI integration process. So, here, the main issue is
the integration and manufacturing of inductive elements such as the inductance of the FET
feedback and the biasing Inductance (choke self with a high value). The latter can be
replaced with, for example, an active bias which, moreover, provides a possibility of
operating bandwidth enhancement. To get rid of the feedback inductance, let us consider a
new topology of NGD active circuit with no inductance. It simply consists of a FET cascaded
with a shunt RC-network. To match the cell output to the following stage input, an R2
resistor is connected in shunt at its end. As described in Fig. 21, this new topology is
cascaded at the end of an RC circuit, which models the interconnect line to be compensated.

Fig. 21. RC-circuit compensated by a NGD cell without inductance.

Let us focus, at first, on the analytical study of the proposed cell according to the RC circuit
values prior providing evidence, through simulations, of the compensation of the RC effects,
i.e. signal recovery and delay reduction, by using the proposed NGD topology

6.1 Theory
According to (Ravelo et al, 2008), the transfer function of the RC-NGD device described in
Fig. 21 is:

 Gmax R2 ( 1  R1C1s )
G( s )  , (48)
( 1  Rt Cs )[ R1  R2  Rds  R1C1s( R2  Rds )]
where Gmax is the maximal gain value expressed in equation (37). The minus sign is
explained by the intrinsic FET model, which entails naturally the inversion of the output
voltage direction compared to the input one. Otherwise, with the same approach as in
Section 4.1, from this transfer function it can be established that the gain at very low
frequencies and the Elmore propagation delay are given by:
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 429

 Gmax R2 ,
G( 0 )  (49)
( R1  R2  Rds )
[( R2  Rds )T pRC  R1C1 ( Rt  R1 )]
Tp  . (50)
( R1  R2  Rds )

From expression (49), loss compensation (|G(0)| > 1) is effective on condition that the
following condition between Gmax and the resistance values be met:

R1  Rds .
Gmax  1  (51)
R2

In addition, from expression (50), it can be found that, whatever the values of the RC- and
the NGD-circuit parameters, TP is always lower than TpRC. In other words, the RC-
propagation delay is always reduced in the configuration of Fig. 21. At this stage, it is worth
pointing out that equation (50), despite its usefulness, is an approximation of the 50%
propagation delay as proposed by Elmore. Indeed, according to Ismail and Friedman, a
relative inaccuracy of about 30% is possible. Another limitation appears when the following
condition is satisfied:

R1C1 ( Rt  R1 )
Rt C  . (52)
R2  Rds

In this case, expression (50) provides a negative value for Tp, which leads to an unrealistic
behaviour that would contradict the causality principle. Indeed, calculation of the exact
expression for Tp from the root of the equation, vN(Tp) = VM/2, gives always a positive value
for the total propagation delay.
Despite these restrictions, equations (49) and (50) are particularly useful for a first analytical
approximation and permit the extraction of the synthesis relations needed to determine the
values of the NGD cell components:

R1  R2 (Gmax  1)  Rds , (53)

R2  ( R1  Rds ) /(Gmax  1) , (54)

and by using the equation, τ(0) ≈ 0:

C 1 = CR( R1 + R2 + Rds ) / R12 . (55)

The synthesis relations (53) and (54) are physically realistic under the following conditions:

Gmax  1 , (56)
R2  Rds /(Gmax  1) . (57)
430 VLSI

6.2 Discussion on the simulation results

The previous synthesis relations were used to design and simulate, with ADS software, the
circuit presented in Fig. 22(a). The thin lines indicate the DC bias network. The RC-circuit is
compensated by a two-stage NGD circuit, and the whole circuit is excited at the rate of
1 Gbit/s by an input signal, Vi, with rise and fall times of about 0.2 ns.

(a)
(b)
Fig. 22. (a): The whole circuit composed of an RC-model compensated by a two-stage NGD
circuit in active biasing; FET (EC-2612, Vd = 3 V, Ids = 30 mA), R = 68 , C = 10 pF, R1 = 142
, R2 = 32 , C1 = C2 = 10 pF, R3 = 100 , R4 = 51  and (b) the corresponding simulation
results.

An active load bias with the same FETs (EC-2612) was applied to the circuit. No inductance
element is needed in this circuit. Transient simulations run with ideal components gave the
results displayed on Fig. 22(b). Once again, the comparison of the RC- and RC-NGD-outputs
highlights a signal recovery, characterised by a gain of about 1.85 dB at t = T/2 and a
reduction of propagation from TpRC ≈ 47 ps to Tp ≈ 12 ps or ΔT/TpRC ≈ 74.4% in relative value.
Time-domain measurements are scheduled and will indicate if further improvements are
required prior to the integration in a final VLSI device.

7. Conclusion and future works

A novel and innovative technique of interconnect effect equalization in electronic systems
was developed through theoretical studies and experimentally validated. It relies on the use
of active circuits able to simultaneously generate gain and negative group delay in baseband
over broad bandwidth.
The properties of these circuits were used to propose a new compensation approach
consisting in an equalization of both the positive group delay and attenuation induced by
interconnects by an equivalent negative group delay and gain.
The theory on commonly used circuits to model interconnect effects were briefly recalled.
Then, a circuit composed of a first-order interconnect model (i.e. an RC-circuit) cascaded
with an NGD cell was theoretically studied in order to determine the conditions required to
compensate for both the degraded propagation delay and the attenuation and to express the
synthesis relations to be used in the determination of the values of the NGD cell
components. This NGD cell simply consists in a FET fedback by an RL series network. Then,
for this first modelling of interconnect line, a proof-of-concept circuit implemented in hybrid
planar technology was fabricated to demonstrate the efficiency of this technique. The
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 431

experimental results in frequency- and time-domains were in very good agreement with
simulations and validated the compensation technique in the case of an input signal with a
25 Mbit/s data rate. Indeed, the reductions of the rise time and the 50% propagation delay
were 71 and 86%, respectively.
In many VLSI systems and particularly in long wires and/or for high data rates or clocks,
the inductive spurious effects can no longer be neglected. So, a more elaborated system
composed of an RLC interconnect model compensated with NGD cells was also studied
analytically in order to check for the validity and efficacy of the equalization technique and
to determine the synthesis relations to be used in further applications. To validate the
approach, a first series of simulations was run with an RLC lumped model for an input
signal at 1 Gbits/s-rate; then , the model used in the second set of simulations was an RLC
distributed line for an input signal at a rate of 200 Mbit/s. These simulations under realistic
conditions confirmed the compensation approach with reduction of the propagation delay
of the same order as previously. Moreover, as observed with the RC-model, the front and
trailing edges both showed great enhancements indicative of a good recovery of the signal
integrity.
Finally, to be able to compensate for interconnect effects in VLSI systems, the proposed
circuits must be compatible with a VLSI integration process. This requirement drove us to
propose improvements of the proposed topology in order to cope with inductance
integration and manufacturing prerequisites. So, a topology with no inductance, but with
the same behaviour and performances as previously was proposed. A theoretical study
provided evidence of its ability to exhibit a negative group delay in baseband together with
gain. Then, time-domain simulations of a two-stage NGD device excited by a 1 Gbits/s-rate
input signal were run to validate the expected compensation approach and check for the
signal recovery.

The implementation of this equalization technique in the case of a VLSI integration process
is expected to allow compensation for spurious effects such as delay and attenuation
introduced by long inter-chip interconnects in SiP and SoC equipments or by long wires and
buses. A preliminary step would be the design and implementation of such a circuit in
MMIC technology and especially by using distributed elements. At this stage, even if
experimentally the NGD cells were not particularly sensitive to noise contribution, it would
be worth comparing this approach and repeater insertion under rough conditions, i.e. long
wires with a significant attenuation, in order to gain key information on their respective
behaviour under conditions of significant noise. As identified in ITRS roadmap, the power
consumption is now one of the major constraints in chip design and has been identified as
one of the top three overall challenges over the last 5 years. Faced to these constraints,
further investigations are needed to accurately evaluate the consumption of the presented
NGD active circuits.
432 VLSI

8. References
Adler, V. & Friedman, E. G. (1998). Repeater Design to Reduce Delay and Power in Resistive
Interconnect, IEEE Trans. Circuits Syst. II, Analog and Digital Signal Processing, Vol.
54, No. 5, pp. 607-616.
Bakoglu H. B. & Meindl J. D. (1985). Optimal Interconnection Circuits for VLSI, IEEE Trans.
On Electron. Devices, Vol. 32, No. 5, pp. 903-909.
Barke E. (1988), Line-to-ground capacitance calculation for VLSI: a comparison, IEEE Trans.
of Computer Added Design, Vol. 7, No. 2, pp. 295-298.
Deng, A. C. & Shiau, Y. C. (1990). Generic linear RC delay modeling for digital CMOS
circuits, IEEE Tran. on Computer-Aided Design, Vol. 9, No. 4, pp. 367-376.
Deutsch., A. (1990). High-speed signal propagation on lossy transmission lines, IBM J. Res.
Develop., Vol. 34, No. 4, pp. 601-615.
Deutsch, A.; Kopcsay, G. V.; Restle, P. J.; Smith, H. H.; Katopis, G.; Becker, W. D.; Coteus, P.
W.; Surovic, C. W.; Rubin, B. J.; Dune, R. P.; Gallo, T. A.; Jankis, K. A.; Terman, L.
M.; Dennard, R. H.; Asai-Halsz, G.; Krauter, B. L. & Knebel, D. R. (1997). When are
the transmission line effects important for on-chip interconnections, IEEE Trans. on
MTT, Vol. 45, No. 10, pp. 1836-1846.
Deutsch, A.; Kopcsay, G. V.; Surovic, C. W.; Rubin, B. J.; Terman, L.M.; Dunne, Jr. R. P.;
Gallo, T. A. & Dennard, R. H. (1995). Modeling and characterization of long on-chip
interconnections for high speed-performance microprocessors, IBM J. Res. Develop.,
Vol. 39, No. 5, pp. 547-667.
Dogariu, A.; Kuzmich, A. & Cao, H. & Wang, L. J. (2001). Superluminal light pulse
propagation via rephasing in a transparent anomalously dispersive medium, Optics
Express, Vol. 8, No. 6, pp. 344-350.
Eleftheriades, G. V.; Siddiqui, O. & Iyer, A. K. (2003). Transmission line for negative
refractive index media and associated implementations without excess resonators,
IEEE MWC Lett., Vol. 13, No. 2, 51, pp. 51-53.
Elmore, W. C. (1948). The transient response of damped linear networks, J. Appl. Phys., Vol.
19, pp. 55-63.
Friedman, E. (1995). Clock distribution networks in VLSI circuits and systems, New York
IEEE Press.
Grover, F. (1945). Inductance Calculations Working Formulas and Tables, Instrum. Soc. of
America.
Ismail, Y. I. & Friedman, E. G. (2000). Effects of inductance on the propagation, delay and
repeater insertion in VLSI circuits, IEEE Tran. VLSI Sys., Vol. 8, No. 2, pp. 195-206.
Ismail, Y. I.; Friedman, E. G. & Neves, J. L. (2000). Equivalent Elmore delay for RLC trees,
IEEE Tran. Computed-Aided Design, Vol. 19, No. 1, pp. 83-97.
Kitano, M.; Nakanishi, T. & Sugiyama, K. (2003). Negative group delay and superluminal
propagation: an electronic circuit approach, IEEE Journal of Selected Topics in
Quantum Electronics, Vol. 9, No. 1, pp. 43-51.
Krauter, B. & Mehrotra S. (1998). Layout based frequency dependent inductance and
resistance extraction for on-chip interconnect timing analysis, IEEE Design
Automation Conference, Proceedings of the ACM, pp. 303–308.
Lucyszyn, S.; Robertson, I. D. & Aghvami, A. H. (1993). Negative group delay synthesiser,
Electron. Lett., Vol. 29, pp. 798-800.
A New Technique of Interconnect Effects Equalization
by using Negative Group Delay Active Circuits 433

Moore, G. E. (1965). Cramming more components into integrated circuits, in Electronics, Vol.
38, No. 8, pp. 114-117.
Munday, J. N. & Henderson, R. H. (2004). Superluminal time advance of a complex audio
signal, Appl. Phys. Lett., Vol. 85, pp. 503-504.
Nakanishi, T.; Sugiyama, K. & Kitano, M. (2002). Demonstration of negative group delays in
a simple electronic circuit, American Journal of Physics, Issue 11, Vol. 70, pp. 1117-
1121.
Palit, A. K.; Meyer, V.; Duganapalli, K. K.; Anheier, W. & Schloeffel, J. (2004). Test pattern
generation based on predicted signal integrity loss through reduced order
interconnect model, 16th Workshop Test Methods and Reliability of Circuits and
Systems.
Rabay, J. M. (1996). Digital integrated circuits, a design perspective, Englewood Cliffs, NJ:
Prentice-Hall.
Ravelo, B. (Dec. 2008). Negative group delay active devices: theory, experimental
validations and applications, Ph.D. thesis (in French), Lab-STICC, UMR CNRS 3192,
University of Brest, France.
Ravelo, B.; Pérennec, A. & Le Roy, M. (2007a). Equalization of interconnect propagation
delay with negative group delay active circuits, 11th IEEE Workshop on SPI, Genova,
Italy, pp. 15-18.
Ravelo, B.; Perennec, A.; Le Roy, M. & Boucher Y. (2007b). Active microwave circuit with
negative group delay, IEEE MWC Lett., Vol. 17, Issue 12, pp. 861-863.
Ravelo, B.; Perennec, A. & Le Roy, M. (2007c). Synthesis of broadband negative group delay
active circuits, IEEE MTT-S Symp. Dig., Honolulu (Hawaii), pp. 2177-2180.
Ravelo, B.; Pérennec, A. & Le Roy, M. (2008a). Application of negative group delay active
circuits to reduce the 50% propagation delay of RC-line model, 12th IEEE Workshop
on SPI, Avignon, France.
Ravelo, B.; Pérennec, A. & Le Roy, M. (2008b). Negative group delay active topologies
respectively dedicated to microwave frequencies and baseband signals, Journal of
the EuMA, Vol. 4, pp. 124-130.
Ravelo, B.; Pérennec, A. & Le Roy, M. (2009). Experimental validation of the RC-interconnect
effect equalization with negative group delay active circuit in planar hybrid
technology, 13th IEEE Workshop on SPI, Strasbourg, pp. 1-4.
Ruehli, A. & Brennan, P. (1975). Capacitance models for integrated circuit metallization
wires, J. of Solid-State Integrated Circuits, Vol. 10, No. 6, pp. 530-536.
Sakurai, T. (1983). Approximation of wiring delay in MOSFET LSI, IEEE J. of Solid State
Circuits, Vol. 18, No. 4, pp. 418-425.
Sakurai, T. (1993). Closed-form expressions of interconnection delay, coupling and crosstalk
in VLSI’s, IEEE Tran. on Electron. Devices, Vol. 40, No. 1, pp. 118-124.
Siddiqui, O. F.; Erickson, S. J.; Eleftheriades, G. V. & Mojahedi, M. (2004). Time-domain
measurement of negative-index transmission-line metamaterials, IEEE Trans. MTT,
Vol. 52, No. 5, pp. 1449-1453.
Siddiqui, O.; Mojahedi, M.; Erickson, S. & Eleftheriades, G. V. (2003). Periodically loaded
transmission line with effective negative refractive index and negative group
velocity, IEEE Tran. on Antennas and Propagation, Vol. 51, No. 10.
Solli, D. & Chiao, R. Y. (2002). Superluminal effects and negative delays in electronics and
their applications, Physical Review E, Issue 5.
434 VLSI

Standley D. & Wyatt, J. L. Jr., (1986). Improved signal delay bounds for RC tree networks,
VLSI Memo, No. 86-317, Massachusetts Institute of Technology, Cambridge,
Massachusetts.
Vlach, J.; Barby, J. A.; Vannelli, A.; Talkhan, T. & Shi. C. J. (1991). Group delay as an estimate
of delay in logic, IEEE Tran. Computed-Aided Design, Vol. 10, No. 7, pp. 949-953.
Wang, L. J.; Kuzmich, A. & Dogariu, A. (2000). Gain-assisted superluminal light
propagation, Nature 406, pp. 277-279.
Wyatt, J. L. Jr., & Yu, Q. (1984). Signal delay in RC meshes, trees and lines, Proceedings of the
IEEE International Conference on Computer-Aided Design, pp. 15-17.
Wyatt, J. L. Jr. (1985). Signal delay in RC mesh networks, IEEE Tran. Circuits and Systems,
CAS-32(5), pp. 507-510.
Wyatt, J. L. Jr. (1987). Signal propagation delay in RC models for interconnect, Circuit
Analysis, Simulation and Design, Part II: VLSI Circuit Analysis and Simulation, A.
Ruehli, ed., Vol. 3 in the series Advances in CAD for VLSI, North-Holland.
Yu, Q.; Wyatt, J. L. Jr.; Zukowski, C.; Tan, H-N. & O'Brien, P. (1985). Improved bounds on
signal delay in linear RC models for MOS interconnect, Proceedings of the IEEE
International Symposium on Circuits and Systems, pp. 903-906.
Book Embeddings 435

21

Book Embeddings
Saïd Bettayeb
University of Houston Clear Lake
Houston, TX 77058, USA

1. Introduction
Graph embeddings play an important role in interconnection network and VLSI (Very Large
Scale Integration) design. Simulation of one interconnection network by another can be
represented as a graph embedding problem. Determining the number of layers required to
build a VLSI chip, also called book-embedding, is another application of graph embeddings.
In this chapter, we explore the latter problem. After an overview of the results on book
embedding of the hypercube, we present results on book embedding of the k-ary hyperube,
a variant of the hypercube. We also present recently obtained results on the book
embedding of the torus graph.
Graph embeddings have been studied in the literature extensively for the important role it
plays in interconnection network and VLSI (Very Large Scale Integration) design (Bernhart
& Kainen, 1979; Bettayeb et al., 1989; Bettayeb & Sudborough., 1992; Bettayeb &
Sudborough, 1989; Chung et., 1987; Yannakakis, 1989). Simulating an interconnection
network, say A, by another, say B, is represented as a graph embedding problem where the
nodes and edges of A are mapped to nodes and paths of B. A book embedding of a graph G
is the mapping of the nodes of G onto the spine of a book and the edges of G onto pages so
that the edges assigned to the same page do not cross. Determining the minimum number of
pages required for such an embedding is the focus of this chapter. The minimum number of
pages in which a graph G can be embedded is called the pagenumber of G, pg(G).
Determining the pagenumber of an arbitrary graph has been shown to be NP-complete
(Chung et al, 1987; Yannakakis, 1989). It remains NP-complete to determine if an arbitrary
graph can be embedded in two pages. In a 1980 article, Garey et al (Garey et al., 1980)
proved that determining the pagenumber of an arbitrary graph remains NP-complete even
if we assume that the node embedding part is fixed, i.e. the layout of the nodes is given. An
equally challenging task is the problem of determining the pagerwidth or geometric
thickness of an arbitrary graph G which defined to be the minimum number of layers in a
planar drawing of an arbitrary G.
Researchers have been drawn to study this problem and variations of this problem (Chung
et al., 1987; Galil et al., 1989; Yannakakis, 1989) because of its many and diverse applications
such as fault tolerant computing (Chung et al., 1987), graph drawing, and graph separators
(Galil et al., 1989). Other problems remain to be solved such as the relationship of the
436 VLSI

pagenumber of a graph and other invariants. Enomoto and Miyauchi(Enomoto & Miyauchi,
1999) considered the case where edges may use more than one page.
The pagenumber of a graph has strong implications in VLSI design. The pagenumber is the
minimum number of layers required to produce a VLSI chip. Another area of VLSI design
that can be described in terms of book embedding is the configuring of processors in the
presence of faults. Given an array of processors, some of which may be faulty, we lay them
in a line. This could be either physical or logical. Running parallel to the line of processors
are bundles of wires. As we scan the line of processors, we activate switches connecting the
good processors to a bundle of wires and bypassing the bad processors. The bundle of wires
act like a stack in that, when a processor � requests a connection to another processor, � is
connected to a particular bundle and pushes the other processor connections down to a
wire. When our scan reaches the processor to which processor � connects, it is popped off
the bundle since the wire is no longer needed, and the other connections are returned to
their original positions. The desired property in this case, is the minimization of the number
of bundles required to interconnect all of the good processors in the desired layout. This is
used in the Diogenes method of fault tolerant design as described by Chung, Leighton, and
Rosenberg (Chung et al., 1987). If we take each bundle of wires and represent it as a page,
we have a book embedding. Chung, Leighton, and Rosenberg (Chung et al., 1987) have
studied the book embedding problem for a variety of graphs including trees, grids, X-trees,
Benes networks, complete graphs, and the binary hypercube.

2. Definitions and Terminology

Let � � ��, �� be an undirected graph. A linear layout � of the vertices of G is a mapping of
V to the set ��, 2, �, � , ��, where � �   |�|. Let ��, �� and ��′, �′� be edges in � such that
���� � ���� and ���′� �  ���′�. Then ��, �� and ��′, �′� intersect with respect to L if ���� �
���′� � ���� � ���′�. If, in the other hand, ���� � ��� ′ � � ��� ′ � � ���� or  ��� ′ � � ���� �
���� � ���′� the edges ��, �� and ��′, �′� are said to nest with respect to �.
A book embedding, also referred to as stack embedding, is a linear layout of the vertices and
an assignment of the edges to pages such that no page contains a pair of edges that intersect.
The pagenumber of a graph, also referred to as book thickness, is the minimum number of
pages achieved among all possible book embeddings.
When each edge has to be drawn as a straight line segment, the problem is referred to as the
geometric thickness or real linear thickness (Kainen, 1990; Dillencourt et al, 2000)
Another variation studied in the literature is the so-called queue embedding. A queue
embedding is a linear layout of the vertices of a graph and an assignment of the edges to
queues such that no queue contains a pair of edges that nests (Bettayeb et al. 2010).

A d-dimensional torus is the d-dimensional mesh with wraparound edges. The wraparound
edges connect the first and last vertex in each dimension. The notation ���   �   ��   � � ��  �
denotes the d-dimensional torus where �� , �  � �  � �,  is the size of the � �� dimension.
The binary hypercube of dimension �, denoted by �� ��� has 2� vertices labeled by the
binary representation of integers between 0 and 2� � �. Two vertices are connected by an
edge if and only if their labels differ in exactly one bit position. The k-ary hypercube of
dimension �, denoted by �� ��� has � � vertices labeled by the k-ary representation of
Book Embeddings 437

integers between 0 and � � � �� Two vertices are connected by an edge if and only if their
labels differ in exactly one position by one (modulo k).

3. Book Thickness of Graphs

In a 1979 article, Bernhart and Kainen (Bernhart & Kainen, 1979) introduced the problem of
book embedding. They proved that the pagenumber of a graph G, �����   �  � if and only G
is a subgraph of some planar graph. They also conjectured that planar graphs have
unbounded pagenumber. Heath and Istrail (Heath & Istrail, 1992) disproved this conjecture
and proved that the pagenumber of graphs with genus � is Ο(��). The genus of a graph is
defined as follows. A graph G can be embedded with no edge crossing on a compact
orientable two-manifold surface on which a number of handles have been placed. A handle
is used by an edge that would otherwise cross other edges. The genus of a surface is the
number of handles on that surface. The genus of a graph is the minimum genus of all
possible surfaces on which G can be embedded. So planar graphs have genus 0 since no
handle is needed.
In (Kainen & Overbay, 2003, the authors studied the pagenuumber in terms of block-
cutpoint tree. A block-cutpoint graph of a graph �  �   ��� �� is a bipartite graph �’ � ��’� �’�
where �’  �  � � �� There is a vertex � �� � for each block of G and a vertex � �� � for each
cutpoint in �� The vertices x and y are connected by an edge if and only if the block
represented by x contains the cutpoint y. A graph � is connected if and only if its block-
cutpoint is tree. They showed that the pagenumber of a graph is the maximum of the
pagenumber of its blocks. They also showed that a graph is planar if and only if it
homeomorphic to a graph with pagenumber at most two.

4. Book Embedding of Planar Graphs

A graph � � ��� �� has pagenumber 0 if and only if ���   �  0, �� �. � consists of isolated
vertices. Trees are shown to have pagenumber 1. A graph is said to be outerplanar if it can
be drawn in the plane so that all of its vertices lie on the same face. Outerplanar graphs have
pagenumber one. The following theorem is due to Kainen and Overbay (Kainen & Overbay,
2003):
Theorem: A graph G has pagenumber � with vertex ordering �� � �� � � � �� if and only
� �   ��  � ��  �  � ��  , where each �� is an outerplanar graph embedded with vertex ordering
��  � ��  � � � �� .
In (Bernhart & Kainen, 1979), the authors conjectured that planar graphs have unbounded
pagenumber but this was disproved by Heath and Istrail (Heath & Istrail, 1992). Buss and
Shor (Buss & Shor, 1984) gave an algorithm that embeds all planar graphs in nine pages.
Heath and Istrail (Heath & Istrail, 1992)gave an improvement that achieves seven pages.
Yannakakis (Yannakakis, 1989) brought the number to four and showed that indeed four
pages are sufficient and necessary.

4.1 Outline of Yannakakis Algorithm

Let C be a cycle C of the graph G that contains no external chords. Denote the vertices of C
by �� �� � � �. The edge ��� �� is called the long edge, and the rest of edges of C are called
438 VLSI

short edges. Let Gc be the graph obtained from C by adding its chords. If F and F’ are two
inner faces of Gc where the long edge of F is a short edge of F’ then all vertices of F must
appear between two consecutive vertices of F’. First, Let K be the cycle bounding the inner
face of Gc that contains the long edge of C. Its short edges must be the long edges of the
other inner cycles. Layout the interior of K. Then, expand recursively the inner cycles.
As pointed out in (Yannakakis, 1989), the vertices of a planar graph can be partitioned into
levels. All edges connect either vertices of the same level or vertices of adjacent levels. The
former are called level-edges and the latter binding edges. Level 0 consists of the vertices
forming the cycle K. Laying out the interior of K is accomplished by first laying level 1
vertices and coloring the short edges of K and the binding edges between levels 0 and 1.
Expand recursively the cycles formed by level 1 vertices.

5. Book Embedding of the Torus and the k-ary hypercube

In (Bettayeb & Hoelzeman, 2009), we established the upper bound and lower bound on the
pagenumber of the k-ary hypercube. The k-ary hypercube is a generalization of the binary
hypercube. It is defined as follows. The k-ary hypercube of dimension d is an undirected
graph of kd vertices labeled by the integers 0 through kd-1. Two nodes x and y are connected
if and only the k-ary representations of their labels differ in exactly one position by one
modulo k. We showed that the pagenumber of the �-dimensional �-ary hypercube is at least
�� �� �
2���   and the upper bound is  ( ��� ) where a =���. In (Chung et al., 1987), Chung et al.
showed that the binary hypercube of dimension �, �� ���, admits an � � � page embedding
which is within a factor of 2 of optimal because the lower bound can easily be shown to be
�
�
.  However, Heath, Leighton and Rosenberg (Heath et al., 1992) have shown that the
pagenumber of the ternary hypercube of dimension � has a lower bound �3��� �. It seems
that when k is even the pagenumber grows linearly with the number of dimensions while it
grows exponentially when k is odd. In (Bettayeb et al., 1020) we show that the pagenumber
the k-ary hypercube depends on the parity of k. When k is even the pagenumber of �� ���
grows linearly with the dimension � but when k is odd, the pagenumber grows
exponentially with �. With the d-dimensional torus ���   �   ��   � � ��  �, if all �� , � � � � �, are
even the pagenumber is grows linearly with the number of dimensions �. In that paper, we
also describe a layout technique with sequential corrections of the order of vertices that
mitigates the problem for the case when dimensions are odd. Basically, the technique
modifies the standard layout of alternating left-to-right and right-to-left segments with an
amortizations of corrections that allow the reverse order to be realized without paying the
penalty of all edges in the first-to-last wraparound connections to be simultaneouslu
crossing. This would reduce the number of pages.
Recently, we showed that for any d-dimensional torus the pagenumber is bounded by
2���  –  3 (Bettayeb et al., 2010). It had been shown ( Bettayeb & Hoelzeman, 2009; Chung et
al., 1987; Yannakakis, 1989 ) that the pagenumber of a graph is at least as large as the
minimum number of outerplanar graphs into which it can decomposed.
For the queue embedding problem, the queue number for the torus grows linearly in the
number of dimensions, regardless of the parity of the sizes of its dimensions. The queue
number of the k-ary hypercube �� ��� also grows linearly with the dimension � and does not
depend on the parity of �. It is indeed 2� � �.
Book Embeddings 439

6. Conclusion
In this chapter, we presented results on book embedding and queue embedding of graphs.
The upper bound for the book-embedding of the torus we achieved is 2���  –  3. It is
interesting to know if this could be improved. Heath, Leighton and Rosenberg (Heath et al.
1992) showed that the ternary hypercube has a lower bound �3��� �. It follows that the
lower bound for torus is exponential in the number of dimensions when the sizes of its
dimensions are odd. The authors in (Heath et al., 1992) conjectured that family of graphs
with large queue number and small page (or stack) number do not exist. In (Bettayeb et al.,
2010), we describe a class of graphs, namely the �-dimensional �-ary modified hypercubes
which have pagenumber ����. We conjectured that the queue number for such graphs grow
more rapidly than anu linear function of the dimension �.

7. References
Bernhart, F.; Kanien P.C., (1979). The Book Thickness of a Graph, Journal of Combinatorial
Theory, Series B (27), 1979, pp. 320-331.
Bettayeb, S.; (1995). On the K-ary Hypercube. Journal of Theoretical Computer Science 140,
1995, pp. 333-339.
Bettayeb, S.; Heydari, H.; Morales, L.; Sudborough, I.H., (2010). Stack and Queue Layouts
for Toruses and Extended Hypercubes. To appear
Bettayeb, S.; Hoelzeman, D., (2009). Upper and Lower Bounds on the Pagenumber of the
Book Embedding of the k-ary Hypercube. Journal of Digital Information
Management, 7 (1), 2009, pp. 31-35.
Bettayeb, S.; Miller, Z.; Sudborough, I.H., (1992). Embedding Grids into Hypercubes. Journal
of Computer and System Sciences, 45 (3), 1992, pp. 340-366.
Bettayeb, S.; Sudborough, I.H., (1989). Grid Embedding into Ternary Hypercubes. Proc. Of
the ACM South Central Regional Conference, 1989, pp. 62-64.
Buss, J.F.; Shor, P.W., (1984). On the Pagenumber of Planar Graphs. Proc. Of the 16th Annual
Symposium on Theory of Computing, 1984, pp. 98-100.
Chung, F.R.K; Leightonm F.T; Rosenberg, A.L. (1983). DIOGENES: A Methodology for
Designing Fault Tolerant Processor Arrays. 13th Conf. on Fault Tolerant Computing,
1983, pp. 26-32.
Chung, F.R.K; Leightonm F.T; Rosenberg, A.L. (1987). Embedding Graphs in Books: A
Layout Problem with Application to VLSI Design. SIAM Journal of Algebraic
Discrete Methods, 8 (1), 1987, pp. 33-58.
Dean, A.M.; Hutchinson, J.P. (1991). Relations among Embedding Parameters for Graphs.
Graph Theory, Combinatorics, and Applications, vol. 1, Wiley Interscience Publ., New
York, 1991, pp. 287-296.
Dean, A.M.; Hutchinson, J.P. ; Sheinerman, E.R. (1991). On the Thickness and Arboricity of a
Graph. Journal of Combinatorial Theory Series B, vol. 52, 1991, pp. 147-151.
Dillencourt, M.B.; Epstein, D.; Hirschberg, D.S. (2000). Geometric Thicknes of Complete
Graphs. Journal of Graph Algorithms and Applications. vol. 4, 2000, pp. 5-17.
Enomoto, H.; Miyauchi, M.S., (1999). Embedding Graphs into a Three Page Book with O(M
log N) crossings of edges over the spine. SIAM Journal of Discrete Math. 12, 1999, pp.
337-341.
440 VLSI

Heath, L.S; Istrail, S. (1992). Comparing Queues and Stacks As Mechanisms for Laying out
Graphs. Journal of the Association for computing Machinery, 39, 1992, pp. 479-501.
Heath, L.S; Leighton, F.T.; Rosenberg, A.L. (1992). Comparing Queues and Stacks As
Mechanisms for Laying out Graphs. SIAM Journal of Discrete Mathematics, 5 (3),
1992, pp. 398-412.
Kainen, P.C.; Overbay, S. (2003). Book Embeddings of Graphs and a Theorem of Whitney.
Technical Report GUGU-2/25/03, http:// www.georgetown.edu/faculty/kainen/pbip3.pdf .
Kainen, P.C. 1990. The Book Thickness of a Graph. Congr. Numer., 71, 1990, pp. 127-132.
Yannakakis, M. (1989) Embedding Planar Graphs in four Pages., Journal of Computer and
System Sciences, 38 (1), 1989, pp. 36-67.
VLSI Thermal Analysis and Monitoring 441

22
X

VLSI Thermal Analysis and Monitoring

Ahmed Lakhssassi and Mohammed Bougataya
Université du Québec en Outaouais
Canada

1. Introduction
The rapid evolution in the industry of the integrated circuits during the last decade was so
quick that currently it is possible to integrate complex systems on a single SoC (System on a
Chip). Due to aggressive technology scaling, VLSI integration density as well as power
density increases drastically. As thermal phenomena research activities on micro-scale level
are gaining popularity due to an abundance of SoC and MEMS-based applications, various
measurement techniques are needed to understand the thermal behaviour of VLSI chip. In
particular, measurement techniques for surface temperature distributions of large VLSI
systems are a highly challenging research topic. Hence, surface peaks thermal detection is
necessary in modern VLSI circuits; their internal stress due to packaging combined with
local self heating becomes serious and may result in large performance variation, circuit
malfunction and even chip cracking.

One of the important questions in the field of thermal issues of VLSI systems and micro-
systems is how to perform the thermal monitoring, in order to indicate the overheating
situations, without complicated control circuits. Traditional approach consists of placement
of many sensors everywhere on the chip, and then their output can be read simultaneously
and compared with the reference voltage recognized as the overheating level.

The idea of the proposed method is to predict the local temperature and gradient along the
given distance in a few places only on the monitored surface, and evaluate obtained
information in order to predict the temperature of the heat source. Therefore, in the case of
an SoC devices, there is no place on the layout for the complicated unit performing
computations, but there is also no need for it, as we only want to detect the overheating
situations. These peaks are essentials during thermal die monitoring to avoid a critical
induced thermo-mechanical stress. Moreover, in most cases, the overheating occurs only in
one place.

Due to aggressive technology scaling, VLSI integration density as well as power density
increases drastically. For example, the power density of high performance microprocessors
has already reached 50W/cm2 at 100nm technology and it will reach 100W/cm2 at 50nm
technology (ITRS, 2003). This evolution towards higher integration levels is motivated by
the needs of advanced high performance, lighter and more compact systems with less
442 VLSI

power consumption. Meanwhile, to mitigate the overall power consumption, many low
power techniques such as dynamic power management (Wu et al., 2000), clock gating (Oh &
Pedram, 2001), voltage islands (Puri et al., 2003), dual Vdd/Vth (Srivastava et al., 2004) and
power gating (Kao et al., 1997), (Long & He, 2003) are proposed recently. These techniques,
though helpful to reduce the overall power consumption, may cause significant on-chip
thermal gradients and local hot spots due to different clock/power gating activities and
varying voltage scaling. It has been reported in (Gronowski et al., 1998) that temperature
variations of 30◦C can occur in a high performance microprocessor design. The magnitude of
thermal gradients and associated thermo-mechanical stress is expected to increase further as
VLSI designs move into nanometer processes and multi-GHz frequencies.

Nevertheless, the growth of power density dissipated brought a number of critical thermo-
mechanical problems. Thus the heat produced in the structure of the VLSI devices is
directed towards its edges where it is dissipated by radiation, conduction, or convection.
The principal effect of the absence of a good dynamic thermal management is the gradual
and continuous degradation of the quality of performance as well as some other direct
effects on the life cycle of the electronic systems (Buedo et al., 2002). Thus, an algorithm for
the detection and the localization of the thermal peaks is extremely important in order to
manage the thermal stress on high density semiconductor devices. These peaks are
essentials during thermal die monitoring to avoid a critical induced thermo-mechanical
stress.

This study presents a VLSI thermal peak monitoring approach, using GDS, for development
of SPTDA algorithm. The proposed algorithm using only two sensor cells will be formulated
in a manner to facilitate the development of modular architectures using minimum silicon
space in regards of their implementation in VLSI. The architecture selected will be modelled
in high level languages, simulated in order to evaluate its performances, and then
implemented on a FPGA (Field-Programmable Gate Array). A closed loop of simulation is
used in order to evaluate the performances of the architectures proposed at each stage. The
proposed architecture of the algorithm will be designed in a modular perspective after the
separation of the different elementary functions of the algorithm. Hence the design of
SPTDA with flexible modular-based architecture will be presented. The architecture is
designed in high level languages such as Matlab™ – Simulink®, simulated, tested using
VHDL and synthesized using Xilinx™ ISE (Xilinx, 2009) and Altera™ DSP (Altera, 2009)
tools. The simulation and hardware implementation results obtained will be compared to a
finite element method (FEM) temperature prediction of the entire GDS method
configuration cells.

This chapter will present a packaged VLSI thermal analysis by FEM, thermal monitoring
approach using GDS (Gradient Direction Sensors) method. Also the design of surface peaks
thermal detector algorithm (SPTDA) with flexible modular-based architecture will be
presented.
VLSI Thermal Analysis and Monitoring 443

2. Thermal monitoring by Gradient Direction Sensors method

This part presents the first approach to the method of predicting the temperature of a single
heat source on the VLSI device surface. The second approach used concern prediction of the
temperature gradient along the given distance in a few places only on the monitored chip-
surface. The information obtained will be used to evaluate thermal peak position in order to
achieve the temperature of the heat source. This means that we know temperature values in
some places on the chip and we try to find the temperature values of the heat sources and
eventually predict thermo-mechanical stress distribution in the whole structure.

The proposed algorithm is based on the GDS method for evaluating a single heat source on
the chip surface. This method principle of work is explained in details in (Wójciak &
Napieralski, 1997). For two sensors, A and C placed in the distance a (Fig. 1), the difference
between their output voltages is proportional to the changes of the temperature value along
the distance a (Wójciak & Napieralski, 1997). This is true only when the heat source is
directly on the line AC for any other cases the values of the angle α has to be taken into
account for the proper calculation of T (1).
T TC  T A V  VA (1)
  C
r a. cos  a. cos 
where r is the distance from the heat source. In figure 1, we have:
b  AD, AD  TB  TA  VB  VA (2)
b  c  AE, AE  TC  TA  VC  VA (3)
In order to obtain the information about angle  we should apply the third sensor. In the
simplest case the GDS contains only three temperature sensors placed in distance a (fig. 1)

Fig. 1. The 3 sensors cell  (0, 30) (Wójciak & Napieralski, 1997)

On the basis of eqs. (2) and (3) and the geometrical dependencies from (fig. 1), eq. (4) is
obtained:
2  VB  V A 1 
tan   .   (4)
3  VC  V A 2 
Using the cell we can obtain information on the temperature distribution and partly on the
position of the heat source. In order to obtain the temperature value of a single punctual
heat source we have to calculate the distance between the sensor and the heat source.
444 VLSI

Isoth A Isoth C
Isoth B
C
A Heatflow from
Equivalent
single heat source
B
Cell 1 (3 sensors)
R2
Cell 2 (3 sensors)
H
R1
Isoth D
E

F
D Isoth E
Isoth F

Fig. 2. Problem description and the distribution of the sensors cells

Two sensor cells are required for this purpose (fig2). The cells are placed in a given distance
(H) and each of them gives information about the angle α (α1 and α2) in the direction of the
heat source. Hence, the heat source and cells from a triangle in which the length of one side
and values of the angles adjacent to this side are known. This means that we can calculate
the distances between the heat source and sensors. Now we can calculate the temperature
gradient along the known distance. By adding it to the temperature of the sensor we obtain
the temperature of the heat source. Two sensor cells A,B,C and D,E,F are placed in two
corners of a monitored layout in the distance H. Hence, the temperature of the heat source
can be obtained by equation (5).

TS  VS 
H
 VC  VA  
tan2 1  1 3  tan 2    VA (5)
a 3  1  tan 1  tan 2   tan 1  tan 2 
Figure 2 shows the proposed distribution of the 6 sensors divided into 2 cells located
whether on or outside the chip.

3. Algorithmic Design Methodology

3.1 Presentation of algorithmic division
The general flowchart of the proposed SPTDA algorithm is shown in figure 3. It shows some
parallelism in computation between the internal variables represented by tanα1 and tanα2.
This parallelism will be used later in the architectural design in order to optimize the speed
of the running implementation. Since successive evaluation of heat source temperature is
needed for fast heat-source changes. The algorithm will create 2 triplets of values coming
form the detectors. These 2 triplets will be used to calculate the tangents values and estimate
R1, R2 and the temperature of the heat source. By dividing the algorithm into its basic
arithmetic elementary operations (AEO) we will be able to formulate its data flow graph in
order to model the algorithm.
VLSI Thermal Analysis and Monitoring 445

Input : values reported

by detectors.
Initialisation Module

Computation of tanα1
and tan α2.

Computation Computation
of r1 of r2
Estimation of Ts

Thresholds (Tmax, Length, Width)

Delay

Delay

Verification Module

Loop

Fig. 3. Flowchart of the SPTDA algorithm

3.2 Architecture modeling using Simulink®

Before modelling the system using Xilinx ISE® or Altera DSP® tools we have to perform a
quantification analysis in order to determine the combination of fixed bits that will lead to
the minimum quantification error. Due to the fixed number of bits in the fixed point
representation, overflow and underflow problems are serious and common. Thus, a set of
simulations is required to determine the best fixed point representation. Figure 4 show that
a quantification combination of a signed 11.4 and 12.4 bits had great differences with the
main function; 13.4 was the turning point, and increasing the number of bits presented no
great enhancement to the result. However, it is important to note that the quantification
process induce a certain loss of precision. The fixed point specification has to be applied to
every operation and even every constant block in a model.
446 VLSI

Different fixed point representations [15].[4]

250 [14].[4]
[12].[4]
200 [11].[4]
[13].[4]

150

100
Temperature

50

-50

-100

-150
0 20 40 60 80 100 120 140 160 180 200
Matching intervals

Fig. 4. Results generated using different fixed-point representations

4. Thermal analysis and computational results

4.2 WSI physical structure
In this part we will present steady state and transient thermal analysis of WSI (Wafer Scale
Integration) chip junction by FEM approach. It will be used to build models to validate
thermal peaks prediction by GDS method. Hence, the geometrical coordinates of the
investigated source can be obtained by applying the gradient direction sensors. This way,
the possibilities to minimize the thermal peaks in the critical surface areas for BGA (Ball
Grid Array) packaged WSI devices can be explored.

Convection coefficient

E :AL level

D : Molding compound level

Z C : Chip level

B : Sold

A : Substrat level

X Heat source

Fig. 5. Cross section view of the simulated WSI structure

VLSI Thermal Analysis and Monitoring 447

F
R1
D
E
B HS1 HS2

A
R2
C

Fig. 6. Physical position of cell sensors and heat sources on the FEM model

As illustrated in figure 5 (dimensions not respected), the WSI device studied is multilevel
structure with simple Si (silicon) substrate covered with different layers, Al (Aluminum),
solder ball, and molding compound. The packaging assembly was a ceramic BGA. Once the
geometry of the device had been determined and the heat transfer mechanisms quantified, it
was possible to model the system using finite element analysis. Using the computer code
NISA (Numerical Integrated Elements for System Analysis), a 3-D model was created with
more than 100 000 isoparametric thermal shell elements. For the heating computations, this
element models the 3-D state of heat flow. For the thermal part, the element has the
temperature (T) as the only degree of freedom at each node. Figure 6 shows heat sources
and sensors emplacement in the surface of processor. That will enable us to establish in-situ
sensor network to achieve the most homogeneous thermo-mechanical cartography. In this
study we present a case of WSI structure with internal heat generation. This configuration
will enable us to simulate device intense activity and to construct thermal control unit that
will ensure a suitable cooling. Moreover, we have to make sure that the temperature device
structure variation remains suitable for appropriate induced thermo-mechanical stress.
Figure 7 shows position of cell sensors on WSI device for temperature and stress results.

2 sensor cells Heat source

X1
Y X2
X3
X4

X
Fig. 7. Schematic position of sensors cell and heat source on WSI device

4.3 WSI Thermal monitoring

In (Lakhsasi et al., 2006) a method has been presented in details, which can be used for
mixed fluid-heat transfer approach for VLSI steady state thermal analysis used to analyze IC
package problem. However, during thermal analysis the temperature of the chip is
determined for typical packages and power levels using fluid boundary conditions to
448 VLSI

evaluate equivalent convection coefficient to be applied as a thermal junction BC’s.

Therefore, the knowledge of the complete temperature field implies the knowledge of the
temperature gradients at all times, which is of significance for reliability issues. For the large
VLSI device dissipating multiple localized heat sources the single heat source can be
considered. As shown in figure 8, fast variations of temperatures in space and in time
influence the local peak temperature. Hence, in the case of excessive heating or large
amount of power dissipated during short time, the complete transient finite element
computations are strongly recommended.

Fig. 8. VLSI device transient thermal analysis dissipating multiple localized heat sources

Fig. 9. VLSI device steady state thermal analysis dissipating single localized heat source

In this study, investigations are done for the simplest case, only six temperature sensors
(A,B,C,D,E and F, Figure 2) in the form of two sensor cells and one single power heating
source, in order to validate prediction with the 3-D FEM model. The simulations have been
carried out for a one source placed at the junction surface level. As expected, the peak
temperature profile is located at the centre of the heat source (figure 9). There is subsequent
relaxation on temperature gradients through the structure leading to an essentially uniform
temperature variation ΔT. The sensor cells can be placed in any way out of the monitored
area (different distance H between cell 1 and cell 2), but in some cases adequate placement
can simplify the thermal control unit design. In this part the results of thermal peaks can be
very useful for indicating overheating situations and critical thermo-mechanical stress
occurring in the device structure. Hence, table 1 display a comparison between the
temperatures peaks on surface with different implementation under the same conditions.
VLSI Thermal Analysis and Monitoring 449

Thus, the FEM results obtained (Figure 10 and 11) are in full agreement with the GDS
predictions.

In this study we use NISA program to construct 3D thermo-mechanical model of WSI

device. Furthermore, the application of the finite element method to determine the gradient
of temperature that arises in WSI structure is not always simple especially for multi
interconnection components such as ball grid array and flip-chip packages. This thermal
investigation is based on power loss density distribution (thermal cartography) combined
with steady state finite element analysis to predict spatial thermo-mechanical stress at
different location in the WSI structure. Therefore, analytical GDS thermal prediction will be
compared with FEM method computations.

Detectors Values (oC)

Detector Set 1 Set 2 Set 3
VA 68.655 98.655 75.324
VB 70.248 100.248 76.324
VC 71.325 101.325 77.055
VE 69.365 99.365 75.603
VF 70.325 100.325 76.540
VG 72.318 102.318 78.649
Temperature peaks estimated on surface (oC)
Estimated by Set 1 Set 2 Set 3
FEM 82.115 115.213 84.632
SPTDA Float 85.320 115.300 85.730
SPTDA Fix 84.060 114.100 84.190
SPTDA Altera 85.630 115.600 85.310

Table 1. Temperature peaks comparisons by FEM and SPTDA

During implementation the fixed-point representation of the SPTDA algorithm was used.
After implementation, the same input is directed to the algorithm simultaneously in
Simulink® and on the FPGA board. The results are routed back to Simulink®, multiplexed,
and projected on the same 2-D graph in order to compare the output. As many simulations
and co-simulations have preceded the implementation, the result was expected to be the
same as the comparison between the floating and the fixed point comparison. An optimal
frequency close to 100 MHZ has been achieved. Furthermore, the VHDL TB (Test Bench)
was constructed and the force-file was used to stimulate inputs and to compare with the
algorithm predictions. Hence, the set of simulations revealed that the estimations generated
by the SPTDA algorithm presented a great concordance with the predictions generated by
the Finite Element Method (FEM) presented in (Lakhsasi et al., 2006; Bougataya et al., 2006).
450 VLSI

Fig. 10. Temperature distribution according to A-C-axis [Co]

Fig. 11. Temperature distribution according to B-axis [Co]

4.4 Thermal boundary conditions issue

One of the most challenging issues in creating compact thermal models is to use an
appropriate set of boundary conditions for generating ‘data’ with a detailed finite element
model representing the thermal envelope of the application. The thermal analysis depends
on:
- The cooling option (radiators top and/or bottom) applied,
- The location/vicinity and power of its heat-dissipating neighbours,
- The thermal conductivity of the VLSI materials: PCB, heat sink, package, substrate
and heat spreader.
In this study we use the NISA finite element program to predict thermal behaviour of the
VLSI device structure. A wide variety of boundary conditions can be applied using the FEM
software. However, the boundary condition on the vertical sides of the simulation region is
somewhat problematic. Placing a fixed boundary condition on these surfaces produces an
VLSI Thermal Analysis and Monitoring 451

incorrect result, unless a very large simulation region is used at the expense of very long
simulation run times. A more natural boundary condition is a zero flow condition across
these tiny surfaces (adiabatic boundary conditions) as shown in figure 12. The remaining
boundary conditions to be defined are on the bottom and top surfaces of the VLSI device,
representing the heat sink interfaces. Because the VLSI devices is relatively thin and silicon
and solder are good thermal conductors, heat flows happens mainly in the vertical direction,
so the boundary conditions in both horizontal directions can be considered adiabatic. The
uniform heat removal at the bottom and top is modelled by heat flux exchange coefficient h
[W/m2*oK]. The power dissipated by the device is modelled by heat flux produced into the
components. The problem formulation is presented graphically in figure 12.

Heat flux
(Surface Power Heat source
density)

Adiabatic

Adiabatic

Adiabatic

Adiabatic
Cooling

Fig. 12. VLSI device: inside thermal boundary conditions (TBC’s) (BC).

5. Thermal stress prediction

If we know the peak temperature along with materials property and boundary conditions
we can evaluate the thermal stress using the proposed approach. The algorithm presented
herein might prove to be practical comparatively with thermal stress sensors methods
especially for applications where the temperature has to be known only in a limited number
of points, e.g. the determination of hot spot temperature or on-line temperature monitoring.
The thermal stress (excluding the intrinsic one induced by packaging process) in thin films is
given by the following expression:
σth = ( E / (1 – ν )) Δ α.ΔT

Where E / (1 – ν) is the composite elastic constant for the different layers and Δα the
difference in the coefficients of thermal expansion (CTE) between different level of
packaging (Kobeda et al., 1989). In (Lakhsasi & Skorek, 2002) a method has been presented
in details for calculation of the compressive stress in the silicon level given by:

σxx, max ≈ -9.63 ( Eα)Si .ΔTin (Kobeda et al., 1989)

σxx, max ≈ -9.63 (0.15 x 2.8)Si .(102.3-81.1)
σxx, max ≈ -8.574 MPa

However, the induced compressive thermal stress will be combined to the intrinsic one due
to the fabrication processes, and the stress due to the mechanical clamping mechanism.
452 VLSI

Table 2 gives the thermo mechanical parameters for different materials used in the
computation.

Materials Chip M BT Solder Ball Al Die

(die) Substrate attach
Young’s Modulus 131 16 26(xy), 11(z) 17 70 100
(GPa)
Poison’s Ratio 0.3 0.25 0.39(xz.yz), 0.11(xy) 0.4 0.33 0.25
CTE ppm/0C 2.8 15 15(xy), 52(z) 21 22.4 0.8
Thermal 150 65 0.3 50 160 33
Conductivity
W/m.0C
Density Kg/m3 2330 1660 1660 8460 2700 3300
Specific heat J/kg. 712 1672 1672 957 960 1100
0C

Table 2. Device Materials thermo-mechanical Properties

M e t h o d Distance Cell 1– Source Temp Distance Cell 1- Distance Cell 2-

cell 2, H (µm) Tmax (oC) source 1 (µm) source 1 (µm)
Analytical 8000 100.2 10376 10809
results by GDS
3-D FEM model 8000 102.3 10148 10984
results.
Table 3. Results comparison between GDS method and 3-D FEM model

The nodal temperatures stored in the thermal part were used to perform a thermal stress
analysis of the device structure. It is assumed that the structure is stress-free at 25 oC. The
presence of a temperature variation throughout the device structure causes deformation due
to thermal contraction and expansion. A thermal stress analysis is performed to calculate
these deflections and associated stresses due to the thermal loading. The computation is
extended to the whole volume of the device structure.

The present approach may be suitable for large VLSI devices packaging, because it
potentially allows the combination and integration of both thermal and mechanical control,
in a single unit. As an example, the maximum level of stress generated by intense heating of
WSI devices have been evaluated and examined by GDS method. During design and finale
packaging of large VLSI devices, in-situ thermo-mechanical control unite must be
implemented to ensure the safe fulfillments of their operating conditions. The nature of such
interface materials must therefore be considered very carefully during WSI design and
packaging. Further research should be focused on elaborating software tools for
optimization of temperature sensor positions within the allowed area for an IC designer.

Figure 13 and 14 shows thermal stress profile according to X1-axis. The peak compressive
stress of σxx= -8.5 MPa is reported around the region of the heat source.
VLSI Thermal Analysis and Monitoring 453

Fig. 13. VLSI thermal stress profile according to X1-axis, σxx= -8.5 MPa Maximum

Fig. 14. VL VLSI thermal stress profile according to X2-axis, σxx= -3.3 MPa Maximum
454 VLSI

VLSI layers are very thin and any imperfection in structure may lead to cracking and
subsequent shear-initiated delamination at the silicon level. The nature of such interface
materials must therefore be considered very carefully in VLSI devices design for intense
applications. Depending on the technology requirements and definition of failure, the
mechanism of failure may take several forms.

6. Discussion
One of the important questions in the field of thermal issues of VLSI systems and
microsystems is how to perform the thermal monitoring, in order to indicate the overheating
situations, without complicated control circuits. Traditional approach consists of placement
of many sensors everywhere on the chip, and then their output can be read simultaneously
(Szekely, 1994) and compared with the reference voltage recognized as the overheating
level.
The idea of the proposed algorithm is to predict the local temperature and gradient along
the given distance in a few places only on the monitored surface, and evaluate obtained
information in order to predict the temperature of the heat source. Therefore, in the case of
an SoC devices, there is no place on the layout for the complicated unit performing
computations, but there is also no need for it, as we only want to detect the overheating
situations. These peaks are essentials during thermal die monitoring to avoid a critical
induced thermo-mechanical stress. Moreover, in most cases, the overheating occurs only in
one place.
In this chapter, a methodology to evaluate and predict a thermal peak of large VLSI circuit
was presented. The important factors contributing to the device's thermal heating were
characterized. The monitoring approach reported in this paper can be applied to predict the
thermal stress peak of multilevel structures. Surface peaks thermal detection is necessary in
modern VLSI circuits; their internal stress due to packaging combined with local self heating
becomes serious and may result in large performance variation, circuit malfunction and
even chip cracking. Also, in this paper GDS technique was used to develop SPTDA
algorithm.

7. Conclusion
This study presented an approach to thermal and thermo-mechanical stress monitoring of
VLSI chip. The possibility of evaluating thermal peaks and associated thermal stress
distribution all over the monitored area by using GDS and FEM has been shown.
Furthermore, this study presented an approach to the application of inverse problem for
thermo-mechanical analysis. The thermal peaks of investigated source can be obtained by
applying the gradient direction sensors method. The adequate placement of sensors can
accurately evaluate the thermal peaks and simplify the thermo-mechanical control unit
design. That will enable the chip designer to establish the most homogeneous thermo-
mechanical cartography during operation. The cost of the thermal management of WSI
device depends heavily upon the efficiency of the chip design. However, heat sources
spatial distribution has a significant effect on the WSI device operation. Hence, in-situ
thermo-mechanical control unite must be implemented to prevent unexpected pitfalls.
VLSI Thermal Analysis and Monitoring 455

Another aspect presented in this study is related to surface peaks thermal detection in
modern VLSI circuits; because their internal stress due to packaging combined with local
self heating becomes serious and may result in large performance variation, circuit
malfunction and even chip cracking. As an example, in this study GDS technique was used
to develop SPTDA algorithm. Several approaches were implemented to achieve a better
performance for the SPTDA operation. In (Wójciak & Napieralski, 1997), physical circuit
architecture was proposed. However, complicated control components had to be deployed
in the circuit which is not suitable for today’s constraints regarding area, especially if the
control algorithm needs to be placed on-chip, for example, on a sensor network node.
Thereby, deploying the SPTDA in sensor network can be made feasible in the future.

8. References
A. Lakhsasi, A. Skorek," Dynamic Finite Element Approach for Analyzing Stress and
Distortion in Multilevel Devices ", SOLID-STATE ELECTRONICS, PERGAMON,
Elsevier Science Ltd., Volume 46/6 pp. 925-932, May 2002.
A. Lakhsasi, M. Bougataya, D. Massicotte: Practical approach to gradient direction sensor
method in very large scale integration thermomechanical stress analysis, J. Vac. Sci.
Technol. A 24(3), pp. 758-763, May 2006.
A. Srivastava, D. Sylvester, and D. Blaauw, “Concurrent Sizing, Vdd and Vth Assignment
for Low-Power Design,” in Proc. Design, Automation and Test in Eurpoe, vol. 1,
Feb 2004, pp. 718–719.
Altera corp (http://www.altera.com), 2009.
C. Long and L. He, “Distributed sleep transistor network for power reduction,” in Proc.
Design Automation Conf., 2003.
E Kobeda and al ‘’ in situ measurements during thermal oxidation of silicon ‘’ J.vac.sci
Technol B7 (2), Mar/aprl 1989
ITRS, International Technology Roadmap for Semiconductors (ITRS), 2003.
J. Kao, A. Chandrakasan, and D. Antoniadis, “Transistor sizing issues and tool for multi-
threshold CMOS technology,” in Proc. Design Automation Conf., 1997.
J. Oh and M. Pedram, “Gated clock routing for low-power microprocessor design,” IEEE
Trans. on Computer-Aided Design of Integrated Circuits and Systems, vol. 20, pp.
715–722, Jun 2001.
M. Bougataya, A. Lakhsasi and D. Massicotte: Steady State Thermo-mechanical Stress
Prediction for Large VLSI circuits using GDS Method. IEEE CCECE06 Proceeding
SBN:1-4244-0038-4, pp.917-921
P. Gronowski et al., “High performance microprocessor design,” IEEE J. Solid-State Circuits,
vol. 33, pp. 676–686, May 1998.
Q. Wu, Q. Qiu, and M. Pedram, “Dynamic power management of complex systems using
generalized stochastic Petri nets,” in Proc. Design Automation Conf., Jun 2000.
R. Puri, L. Stok, J. Cohn, D. Kung, D. Pan, D. Sylvester, A. Srivastava, and S. H. Kulkarni,
“Pushing ASIC Performance in a Power Envelope,” in Proc. Design Automation
Conf., 2003.
Sergio Lopez-Buedo, Javier Garrido, and Eduardo I. Boemo: ‘Dynamically Inserting,
Operating, and Eliminating Thermal Sensors of FPGA-Based Systems’, IEEE TRAN
456 VLSI

ON COM AND PACK TECHNOLOGIES, VOL. 25, NO. 4, DECEMBER 2002, pp.
561-566.
V. Szekely, Thermal monitoring of microelectronic structures, Microelectronic J., 25 (1994)
157-170.
W. Wójciak and A. Napieralski ‘’Thermal monitoring of a single heat source in
semiconductor devices – the first approach’’ Microelectronics Journal 28 (1997) p
313-316.
Xilinx corp (http://www.xilinx.com), 2009.