Thesis Verilog AMS Model

MSc Thesis
ADPLL Design for WiMAX

Wenlong Jiang
September 18, 2011

Delft University of Technology
c 2011 by Wenlong Jiang

Copyright
All rights reserved.
No part of the material protected by this copyright notice may be reproduced or utilized
in any form or by any means, electronic or mechanical, including photocopying or by any
information storage and retrieval system, without permission from this publisher.
Printed in The Netherlands
iv
September 18, 2011
DELFT UNIVERSITY OF TECHNOLOGY

FACULTY OF
ELECTRICAL ENGINEERING, MATHEMATICS AND COMPUTER
SCIENCE
The undersigned hereby certify that they have read and recommend to the Faculty of
Electrical Engineering, Mathematics and Computer Science for acceptance a thesis entitled ADPLL Design for WiMAX by Wenlong Jiang in partial fulllment of the
requirements for the degree of Master of Science.
Dated: September 18, 2011
Supervisor:
Dr. R. Bogdan Staszewski
Readers:
Dr. Ir. Wouter Serdijn
Dr. Ir. Nick van der Meijs
Ir. Frank Verwaal
Ir. Marcel van de Gevel
Dr. Xuefei Bai
Mr. Sejed Amir Reza Ahmadi Mehr
Abstract
The frequency synthesizer, which functions as a local oscillator, is a critical block in the
transceiver. It needs to meet very stringent specications and consume as less power
as possible. Design of a traditional charge-pump PLL as the frequency synthesizer in the
advanced CMOS technologies in the transceiver of advanced communication systems proves
to be not an easy work and is becoming dicult due to the supply shrink. The ADPLL
system, which denes every essential block with digital interface, proves to be an excellent
alternative.
This thesis deals with the system level design of ADPLL for the WiMAX standard. The
architecture of the ADPLL is presented, with the functional illustration for every building
block, like DCO and TDC. The ADPLL system is modeled and described in Cadence
using Verilog-AMS/Verilog. The performance of the system is analyzed in s-domain. Some
advanced algorithms have been applied to the ADPLL system. The spur mechanism in
the near-integer N cases is proposed and veried. The phase rotation algorithm and the
FREF dithering algorithm have been adopted to eectively suppress these spurs. The top
level issues of ADPLL are tackled, with emphasis on the test plan and the operation modes
of the system. The behavior level simulation results of the system are presented and the
performance summary is given.
The transistor level design of a basic DPA is presented. The layout for the important blocks
is done and the practical concerns of the DPA design are discussed. The post-extraction
simulation results are shown.
September 18, 2011
viii
September 18, 2011
Abstract
Acknowledgement
I am deeply grateful to all the people who in one way or another have helped me during
my MSc project. Without the support of others, it would be impossible for me to reach
this stage.
First and foremost I would like to express my sincere thanks to my MSc supervisor, Dr.
Robert Bogdan Staszewski. It is really an honor to work under his supervision on the eld
of the ADPLL design. Thanks for his guidance in my MSc project work and other matters.
I have learned a lot from his incomparable expertise on this eld, his strict requirement on
the design and his passion for the work. Once again, thank you for your patience.
Special thanks to my colleagues in Catena. I want to thank Ir. Frank Verwaal for the
discussions on the system design and for his life wisdom. I also want to thank Ir. Marcel
van de Gevel for the time-to-time help and proofreading of my thesis. The thanks also
go to Koen van Hartingsveldt, Gerard Lassche, Frans Sessink, Floris van der Wilt, Iqbal
Suhaib, Ernst Habekotte, Federico Bruccoleri, Bert Oude Essink, Aylin Donmez, Tom Fric
and other members in the design team for the discussion on the circuit design and the
chip creatioin; to Jerry Lit and Daniel Mitcan for the discussion on the digital design
ow; to Nicole Walford and Ivaylo Bakalski for the help on the layout; to Atze van der
Goot, Marcel van de Wiel, Frank van den Hout and other people in the CAD and IT
support group; to Cynthia Thepass and Helma Timmermans-Piersma for the help on a lot
of issues. I am especially grateful to Krass Maklev, Kave Kianush and Rien Geurtsen of
Catena Microelectronics B.V for their generous support and great assistance during my
MSc project.
I would like to express my gratitude to Popong Eendrik, Armin Tavakol and Xuefei Bai,
who work with me in this design. We have spent wonderful time together for coee,
technical discussions and cultural discussions. We have a lot of fun and I have learned a
lot from you. These memories and the friendship will be invaluable.
I am very thankful to the other members of my MSc defense committee for their insightful
September 18, 2011
questions and invaluable time, Dr. Wouter Serdjin, Dr. Nick van der Meijs and Mr. Sejed
Amir Reza Ahmadi Mehr. Moreover, I want to thank Prof. John Long for his support as
the chair of ELCA.
I would like to thank my friends here. These go to the PhD students in ELCA, especially
Morteza Alavi, Wanghua Wu and Duan Zhao for the technical discussions and the help on
other issues; go to Jianfeng Wu for the wonderful cooperation on the courses; and also go
to Fan Guo, Jing Li, Xianli Ren, Ao Ba, Kezheng Ma, Xiaoqiang Zhang, Junfeng Jiang,
Guanyu Yi, Jia Guo, Ting Zhou, Ting Yan, Zeng Zeng, etc. for the help and the fun during
the two-year study on Microelectronics in TU Delft.
Lastly, I would like to thank my parents for all that they have done for me. Only with
their support can I take the courage to study and succeed across globe. I owe my accomplishments to them.
Table of Contents
Abstract
vii
Acknowledgement
ix
1 Introduction
1-1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1-2 Introduction to Frequency Synthesizer . . . . . . . . . . . . . . .
1-2-1 Application of Frequency Synthesizer in Wireless Systems
1-2-2 Common Metrics for Frequency Synthesizer . . . . . . . .
1-2-3 Explanation for Phase Noise . . . . . . . . . . . . . . . .
1-2-4 The Impairment of Phase Noise in Frequency Synthesizer
1-3 Introduction to ADPLL . . . . . . . . . . . . . . . . . . . . . .
1-3-1 Fraction-N Charge-pump PLL and Related Issues . . . . .
1-3-2 The Simplied ADPLL Schematic . . . . . . . . . . . . .
1-4 ADPLL for this WiMAX Project . . . . . . . . . . . . . . . . . .
1-4-1 Specication Requirement for the Whole System . . . . .
1-4-2 Additional Requirement for this ADPLL Project . . . . .
1-5 Project Sketch . . . . . . . . . . . . . . . . . . . . . . . . . . .
1-6 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
3
3
4
5
8
9
9
11
13
13
14
14
15
2 ADPLL Architecture and Building Blocks

2-1 Architecture of ADPLL . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2-2 DCO and Related Dividers/Buers in ADPLL System . . . . . . . . . . . . .
17
17
19
2-3 TDC and Incrementor in ADPLL System . . . . . . . . . . . . . . . . . . . .
24
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
September 18, 2011
xii
Table of Contents
2-3-1 TDC in this ADPLL System . . . . . . . . . .
2-3-2 Retimer and High-Speed Incrementor . . . . .
2-3-3 Cooperation of TDC and Retimer+Incrementor
2-4 Digitally-Controlled RF Power Amplier . . . . . . . .
2-5 Low Speed Digital Logic . . . . . . . . . . . . . . . .
2-5-1 TDC Decoder and Normalization Block . . . .
2-5-2 Phase Detection Logic . . . . . . . . . . . . .
2-5-3 Loop Filter for PVT and Acquisition Bank . .
2-5-4 Loop Filter for Tracking Bank . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3 Modeling, Simulation and Analysis of ADPLL System

3-1 Modeling and Simulation of ADPLL . . . . . . . . . . . . .
3-2 DCO Modeling . . . . . . . . . . . . . . . . . . . . . . . .
3-2-1 Time-Domain Modeling of DCO Phase Noise . . . .
3-2-2 The Eect of Ideal Divider . . . . . . . . . . . . . .
3-2-3 Essential Verilog-AMS Code for the DCO Modeling
3-3 TDC Modeling . . . . . . . . . . . . . . . . . . . . . . . .
3-4 From System Specications to ADPLL Implementation . . .
3-4-1 Noise and Error Sources In ADPLL . . . . . . . . .
3-4-2 S-domain Analysis for the ADPLL System . . . . .
4 Advanced Algorithm for ADPLL
4-1 Zero Phase Restart . . . . . . . . . . . . . . . .
4-2 PVT Miss Mechanism . . . . . . . . . . . . . .
4-3 Spur Suppression Techniques for ADPLL . . . .
4-3-1 Spurious Tone Issue of ADPLL . . . . .
4-3-2 Phase Rotation Algorithm . . . . . . . .
4-3-3 FREF Dithering Algorithm . . . . . . . .
4-3-4 Spur Suppression with phase rotation and
5 The
5-1
5-2
5-3
ADPLL Top Level

Top-Level Schematic of ADPLL System . .
Back-Up Modules in the ADPLL System .
Test Plan for ADPLL . . . . . . . . . . .
5-3-1 System Snapshot with Digital Logic
5-3-2 DCO Open-Loop Test . . . . . . .
5-3-3 TDC Test Plan . . . . . . . . . . .
5-4 Operation Modes of ADPLL . . . . . . . .
September 18, 2011
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
FREF dithering
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
27
30
33
38
41
41
43
44
46
.
.
.
.
.
.
.
.
.
51
51
53
53
60
61
62
62
62
67
.
.
.
.
.
.
.
77
77
78
80
80
84
85
91
.
.
.
.
.
.
.
95
95
99
100
100
102
103
106
Table of Contents
5-4-1 ADPLL Start-Up . . . . . .
5-4-2 ADPLL Test Mode . . . . .
5-4-3 ADPLL Closed-Loop Mode
5-5 The ADPLL Top Level Simulation .
xiii
.
.
.
.
.
.
.
.
6 DPA Design for ADPLL

6-1 Introdution to the Class-E PA . . . . .
6-2 Introduction to the DPA Topology . .
6-3 DPA Schematic Overview . . . . . . .
6-4 Practical Concerns for the DPA Design
6-5 DPA layout overview . . . . . . . . . .
6-6 DPA Simulation Results . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
107
107
108
109
.
.
.
.
.
.
115
115
116
117
123
125
126
7 Conclusion
131
7-1 Contribution of This Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
7-2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
A Verilog-AMS (Verilog) Source Code
135
A-1 Verilog-AMS Code for DCO . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

A-2 Verilog-AMS Code for TDC . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
A-3 Verilog-AMS Code for the SPI Master . . . . . . . . . . . . . . . . . . . . . . 145
B Top-level Interface of ADPLL
153
C Register Map of the SPI Block
163
D Figures for TDC Open-Loop Test
165
E Figures for ADPLL Top Level Simulation
171
F DPA Layout
191
September 18, 2011
xiv
September 18, 2011
Table of Contents
List of Figures
1-1 Architecture of direct conversion transmitter. . . . . . . . . . . . . . . . . . .

1-2 Architecture of direct conversion receiver. . . . . . . . . . . . . . . . . . . . .
1-3 Theoretical spectrum Wvo (f ) of oscillator output vo (t). . . . . . . . . . . . .
3
4
6
1-4 Simplied block diagram of a spectrum analyzer. . . . . . . . . . . . . . . . .

1-5 Block diagram of a generic phase-noise analyzer. . . . . . . . . . . . . . . . .
7
7
1-6 Eect of LO phase noise in a receiver (reciprocal mixing). . . . . . . . . . . .
1-7 Eect of LO phase noise in a transmitter. . . . . . . . . . . . . . . . . . . . .

1-8 Simplied charge-pump PLL topology. . . . . . . . . . . . . . . . . . . . . .
1-9 Fractional-N charge-pump PLL topology. . . . . . . . . . . . . . . . . . . . .
9
10
11
1-10 Simplied schematic of ADPLL in [1](modulation part is not drawn). . . . . .
12
2-1
2-2
2-3
2-4
2-5
2-6
2-7
2-8
2-9
2-10
2-11
2-12
Bird view of ADPLL in this project. . . . . . . . . . . . . . . . . . . . . . . . 18

Zoom-in bird view for the DCO block. . . . . . . . . . . . . . . . . . . . . . 20
Frequency planning of ADPLL . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Simplied schematic for the DCO oscillator core. . . . . . . . . . . . . . . . . 22
The 1st Modulator for fractional part of tracking bank. . . . . . . . . . . 24
Flowchart of DCO operation modes. . . . . . . . . . . . . . . . . . . . . . . . 25
Zoom-in bird view for block with TDC and Retimer+Incrementor. . . . . . . . 26
Simplied timing diagram on the working mechanism of retimer, incrementor and TDC. 26
Schematic of incermentors quantization error and TDCs correction in ideal situation. 27
Pseudo-dierential TDC architecture. . . . . . . . . . . . . . . . . . . . . . . 28
Timing diagram for TDC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Schematic of proposed implementation of the retimer and incrementor. . . . . 31
September 18, 2011
xvi
2-13
2-14
2-15
2-16
2-17
2-18
2-19
2-20
2-21
2-22
2-23
2-24
2-25
2-26
2-27
2-28
List of Figures
Timing diagram for signals in a mod-8 counter. . . . . . . . . . . . . .
Timing diagram for QP and QN in the retimer. . . . . . . . . . . . . .
Timing diagram for SEL EDGE signal. . . . . . . . . . . . . . . . . .
Metastability window of sense amplier ip-op. . . . . . . . . . . . .
PHE spike due to mismatch between TDC path and incrementor path.
Generation of PHE spike. . . . . . . . . . . . . . . . . . . . . . . . .
PHE Spike with SEL EDGE signal in this design. . . . . . . . . . . . .
Generation of PHE spike in this design. . . . . . . . . . . . . . . . . .
Zoom-in birdview of DPA. . . . . . . . . . . . . . . . . . . . . . . . .
Zoom-in birdview of low speed digital logic block. . . . . . . . . . . .
Schematic of the TDC decoder and normalization block. . . . . . . . .
TDC decode logic illustration. . . . . . . . . . . . . . . . . . . . . . .
Schematic for OP and OA blocks. . . . . . . . . . . . . . . . . . . . .
High-level schematic of the GT block. . . . . . . . . . . . . . . . . . .
Schematic of a single IIR lter. . . . . . . . . . . . . . . . . . . . . .
Schematic of proportional gain path for tracking bank. . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
32
32
33
35
37
38
39
40
40
41
42
42
45
46
47
48
3-1 Composition of icker noise using multiple low-pass lters (Log-Log Scale). . .
57
3-2
3-3
3-4
3-5
3-6
3-7
58
59
61
63
66
67
Construction of icker noise with single sampling clock and oversampling. . . .

A schematic illustrating the estimation of icker noise compensation coecient.
Oscillator with an ideal divider. . . . . . . . . . . . . . . . . . . . . . . . . .
S-domain simulation result for DCO phase noise prole. . . . . . . . . . . . .
Phase noise spectrum due to the frequency resolution of the DCO tracking bank.
S-domain model for the type-II ADPLL. . . . . . . . . . . . . . . . . . . . . .
3-8 TDC Transfer function for type II PLL ( = 26 , = 215 ). . . . . . . . . . .
69
3-9 The DCO Transfer function for the type II PLL ( = 26 , = 215 ). . . . . .
70
3-10 The DCO contribution to the ADPLL phase noise in the type-II PLL. . . . . .
3-11 TDC transfer function for type II PLL with IIR lter bank . . . . . . . . . . .
3-12 DCO transfer function for type II PLL with IIR lter bank . . . . . . . . . . .
70
72
73
3-13 S-domain result for ADPLL PN (fv =3800 MHz). . . . . . . . . . . . . . . . .
74
3-14 S-domain result for ADPLL PN (fv =3300 MHz). . . . . . . . . . . . . . . . .
75
4-1
4-2
4-3
4-4
4-5
78
79
81
82
83
Transient for zero phase restart. . . . . . . . . . . . . . . . . . .

Logic for PVT tuning word generation and the timing diagram. .
Transient signal waveforms for the PVT miss algorithm. . . . . .
Timing diagram for spurs due to the TDC quantization eect. . .
ADPLL spectrum for CKV frequency of 3793.3156 MHz, RBW=1
September 18, 2011
. . .
. . .
. . .
. . .
kHz.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
List of Figures
xvii
4-6 PHE transient for CKV frequency of 3793.3156 MHz. . . . . . . . . . . . . .

4-7 Timing diagram for phase rotation. . . . . . . . . . . . . . . . . . . . . . . .
83
85
4-8 ADPLL Spectrum with phase rotation (RBW=1 kHz) . . . . . . . . . . . . .
86
4-9 ADPLL Spectrum with phase rotation (RBW=10 kHz) . . . . . . . . . . . . .
87
4-10 Schematic of modulator for FREF dithering. . . . . . . . . . . . . . . . .
88
4-11 ADPLL Spectrum with FREF dithering (RBW=1 kHz) . . . . . . . . . . . . .
89
4-12 ADPLL Spectrum with FREF dithering (RBW=10 kHz) . . . . . . . . . . . .
90
4-13 ADPLL Spectrum with phase rotation and FREF dithering (RBW=1 kHz) . .
91
4-14 ADPLL Spectrum with phase rotation and FREF dithering (RBW=10 kHz) . .
92
4-15 PHE Transient for CKV frequency of 3793.3156M (with phase rotation and FREF dithering on). 9
5-1
5-2
5-3
5-4
Top Level View of ACORE in ADPLL. . . . . . .

Top Level View of DCORE in ADPLL. . . . . . .
TDC transfer function. . . . . . . . . . . . . . . .
The TDC closed loop test (CKV frequency of 3800
. . . .
. . . .
. . . .
MHz).
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 96
. 97
. 104
. 104
5-5 TDC open-loop static test. . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

5-6 Simulation ow for the ADPLL normal operation . . . . . . . . . . . . . . . . 110
5-7 ADPLL top level test bench . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5-8 ADPLL settling transient in PVT mode and acquisiton mode(fv = 3800 MHz)
113
5-9 ADPLL settling transient in tracking mode(fv = 3800 MHz) . . . . . . . . . . 114

6-1 Topology of the Class E PA. . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6-2 Typical drain voltage and current waveform of the Class E PA[2]. . . . . . . . 117
6-3 Schematic of DPA for BT transmitter. . . . . . . . . . . . . . . . . . . . . . 118
6-4 Schematic of DPA for GSM/GPRS/EDGE transmitter. . . . . . . . . . . . . . 119
6-5
6-6
6-7
6-8
6-9
6-10
6-11
6-12
Top level schematic of DPA . . . . . . . . . . .

Schematic of DPA switch array. . . . . . . . . .
Schematic of DPA switch unit. . . . . . . . . .
Schematic of a switchable capacitor unit in DPA
Schematic of the test bench for DPA HB part .
Schematic of the test bench for DPA LB part . .
Power control capability of DPA HB. . . . . . .
Power control capability of DPA LB. . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
120
121
122
122
123
123
126
127
6-13 DPA eciency curve with respect to ACW (HB). . . . . . . . . . . . . . . . . 127

6-14 DPA eciency curve with respect to ACW (LB). . . . . . . . . . . . . . . . . 128
6-15 Phase noise performance of DPA HB. . . . . . . . . . . . . . . . . . . . . . . 128
September 18, 2011
xviii
List of Figures
6-16 Phase noise performance of DPA LB. . . . . . . . . . . . . . . . . . . . . . . 129

7-1 Primitive Pinout for this ADPLL chip . . . . . . . . . . . . . . . . . . . . . . 133
D-1 TDC static test histogram (fext is 100 MHz-203*0.5 kHz). . . . . . . . . . . . 165
D-2 TDC static test histogram (fext is 100 MHz-202*0.5 kHz). . . . . . . . . . . . 166
D-3 TDC static test histogram (fext is 100 MHz+38*0.5 kHz). . . . . . . . . . . . 167
D-4 TDC static test histogram (fext is 100 MHz+39*0.5 kHz). . . . . . . . . . . . 168
D-5 TDC static test histogram (fext is 100 MHz+288*0.5 kHz). . . . . . . . . . . 169
D-6 TDC static test histogram (fext is 100 MHz+289*0.5 kHz). . . . . . . . . . . 170
E-1
E-2
E-3
E-4
E-5
E-6
E-7
E-8
E-9
E-10
E-11
E-12
E-13
E-14
E-15
E-16
E-17
E-18
E-19
E-20
E-21
E-22
E-23
E-24
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
Phase
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
noise
result.fv =3800 MHz,Tinv =12.5 ps,fR =33.8688 MHz.RBW=10 kHz.172

result.fv =3556.214 MHz,Tinv =12.5 ps,fR =33.8688 MHz.RBW=10 kHz.176
result.fv =3803 MHz,Tinv =12.45 ps,fR =40 MHz.RBW=10 kHz. . 178
result.fv =3803 MHz,Tinv =11.5 ps,fR =40 MHz.RBW=10 kHz. . . 178
result.fv =3560.01 MHz,Tinv =12.45 ps,fR =40 MHz.RBW=10 kHz. 182
E-25 ADPLL settling transient in PVT mode and acquisiton mode(fv = 3300 MHz)
September 18, 2011
184
List of Figures
xix
E-26 ADPLL settling transient in tracking mode(fv = 3300 MHz) . . . . . . . . . . 185

E-27 ADPLL settling transient in PVT mode and acquisiton mode(fv = 4050 MHz)
186
E-28 ADPLL settling transient in tracking mode(fv = 4050 MHz) . . . . . . . . . . 187

E-29 ADPLL settling transient in PVT mode and acquisiton mode(fv = 3556.214 MHz)188
E-30 ADPLL settling transient in tracking mode(fv = 3556.214 MHz) . . . . . . . 189
F-1 Layout of the module catip adpll dpaswitcharray . . . . . . . . . . . . . . . . 192
F-2 Layout of the module catip adpll dpacaparray . . . . . . . . . . . . . . . . . . 193
September 18, 2011
xx
September 18, 2011
List of Figures
List of Tables
1-1 Specication for the WiMAX ADPLL System. . . . . . . . . . . . . . . . . .
13
2-1 Frequency resolution for PB/AB/TB in DCO core. . . . . . . . . . . . . . . .
24
3-1 Verilog-AMS abstraction levels. . . . . . . . . . . . . . . . . . . . . . . . . .

3-2 Pros and Cons of Verilog-AMS. . . . . . . . . . . . . . . . . . . . . . . . . .
3-3 DCO phase noise specication. . . . . . . . . . . . . . . . . . . . . . . . . .
53
54
71
5-1 Supply and ground signal list in the ACORE. . . . . . . . . . . . . . . . . . . 98

5-2 The ADPLL phase noise performance summary . . . . . . . . . . . . . . . . . 111
6-1 DPA HB output power with respect to frequency (ACW=15). . . . . . . . . . 129
6-2 DPA LB output power with respect to frequency (ACW=15). . . . . . . . . . 129
6-3 DPA HB output power for corner simulation. . . . . . . . . . . . . . . . . . . 129
6-4 DPA LB output power for corner simulation. . . . . . . . . . . . . . . . . . . 129
B-1
B-2
B-3
B-4
B-5
ACORE Interface . . .
DCORE Interface . .
Interface of LSD block
Interface of Sequencer
Interface of SPI block
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
153
155
156
158
159
C-1 The register map of the SPI block. . . . . . . . . . . . . . . . . . . . . . . . 163
September 18, 2011
xxii
September 18, 2011
List of Tables
Acronyms
AB
Acquisition bank
ACW
Amplitude control word
ADC
Analog-to-digital converter
ADPLL
All-digital PLL
BB
Baseband
BIST
Built-in self test
BT
Bluetooth
CCW
Capacitance control word
DAC
Digital-to-analog converter
DCO
Digitally controlled oscillator
DE
Drain eciency
DPA
Digitally-controlled power amplier
FS
Frequency synthesizer
IP
Intellectual Property
LNA
Low noise amplier
LO
Local oscillator
LPF
Low pass lter
MMD
Multi-modulus divider
NoB
Number of bits
NTW
Normalized tuning word
OFDM
Orthogonal Frequency Division Multiplexing

September 18, 2011
xxiv
List of Tables
OTW
Oscillator tuning word
PA
Power amplier
PA Driver
Power amplier driver
Power added eciency PAE

PB
PVT bank
PFD
Phase/Frequency detector
PGA
Programmable gain amplier
PLL
Phase Locked Loop
PVT
Process, voltage and temperature
RX
Receiver
SDR
Software-dened radio
SNR
Signal-to-noise ratio
SoC
System-On-Chip
TB
Tracking bank
TX
Transmitter
TDC
Time-to-digital converter
VCO
Voltage-controlled oscillator
ZPR
Zero-phase restart
September 18, 2011
Chapter 1
Introduction
1-1
Motivation
The continuous scaling down of CMOS technology has provided us far more superior computation power than ever before. With the technology node advancing from 90 nm to
65 nm to 45 nm (and now, we have microprocessor with 28 nm CMOS technology!), we
can have one mobile terminal that possesses more functionalities than old-dated desktop
computer even. This has greatly changed peoples living and working style, leading to an
era of smart phone, social networking, smart sensor, mobile business, etc. The explosive
growth of applications and services call for telecommunication systems, i.e. transceivers
with higher data throughput and safer data transmission. In high performance transceiver
design, RF and Analog front end shows as the bottleneck due to that:
1. The voltage headroom for advanced CMOS technologies is limited because of reliability issues with thin oxide core transistors. The supply voltage decreases with
technology advancement. This results in a drop of signal-to-noise ratio (SNR) and
dynamic range in the voltage domain, which makes the design of the RF/Analog
front end dicult.
2. In addition to what are provided in the standard digital process, RF and Analog
circuits usually require some extra devices. Such requirements raise the cost of the
product and prolong the development cycle as well.
3. The RF/Analog front end design takes much more eort than digital back end for
migration to new technology. Designers need to tackle various issues like reduced
supply voltage, degraded matching and lower intrinsic gain. Comparatively, the
migration of digital logic is pretty easy as long as the design ow is established.
September 18, 2011
Introduction
Therefore its indispensable to borrow the power of Digital for Analog and RF circuits.
The idea of digital assistance will bring benets like easier calibration, built-in test, or
even replacement of bulky traditional blocks with compact and exible digital blocks.
One of the most challenging design tasks in mobile RF systems is the frequency synthesizer
(FS), which is deployed as local oscillator (LO) both in the transmitter path and the
receiver path. It needs to meet a set of very stringent specications while still be low-area
and low-power. The Phase Locked Loop (PLL) is the common architecture of frequency
synthesizer for the high performance, low power wireless transceiver. The charge-pump
PLL, as the most popular traditional technique for PLL, is analog intensive. It eats up
signicant area and power and is not so scalable to new technology (need re-design). In
contrast, the All-Digital PLL (ADPLL) technology, which has been successfully applied
to Bluetooth (BT) and GSM[1][3], has shown itself as a very potential candidate for the
implementation of frequency synthesizer in more advanced communication standards, as:
1. Integrable with digital process: ADPLL technology has minimum analog and RF
circuit content and doesnt need special devices for RF/Analog application, which
reduces the cost of the chip. Moreover, it makes maximum use of the digital computation power to improve the performance of the system.
2. Easy maintenance: ADPLL has all the essential blocks with digital interface. Thus
the Intellectual Property (IP) of ADPLL is easy to maintain while being migrated
to latter technology nodes (or even not a CMOS technology, as long as the boolean
algebra still holds).
3. New modeling and simulation methodology: ADPLL can have its top level modeled
in an event-driven simulator or a time-driven simulator, like Matlab, VHDL and
Verilog-AMS. This simulation methodology accelerates the simulation and can help
identify the bottleneck of the system performance.
4. Smaller area: Since ADPLL has replaced the bulky analog blocks with digital logic,
the area can be signicantly reduced.
5. Flexibility: Since the control of ADPLL is fully digital, the parameters of ADPLL,
like loop coecients, target frequency and resolution can be easily modied according
to our need. Complex algorithms are applicable to ADPLL even after the chip is
taped out. The design cycle is greatly shortened, which stimulates the innovations,
especially at the system level.
Thus there is a huge interest in the design of ADPLL for modern communication systems. However, nothing comes for free. To implement an ADPLL which ts into tight
specications, a comprehensive eort and innovation from various aspects like algorithm,
architecture, circuit design and test has to be made. This thesis is dedicated to the design
of ADPLL for WiMAX application, especially on the exploration of system level solutions.
September 18, 2011
1-2 Introduction to Frequency Synthesizer
In addition to that, the transistor level design of a digitally-controlled power amplier

(DPA), which is used for output buer of ADPLL, will be presented. The design of this
DPA is also a preparation for the ADPLL-based transmitter or transceiver design as an
extension to our project.
1-2
1-2-1
Introduction to Frequency Synthesizer

Application of Frequency Synthesizer in Wireless Systems
The frequency synthesizer is an essential part in RF transceiver. As in both a transmitter

(TX) and a receiver (RX), it is deployed as LO to perform frequency translation between
baseband (BB) and RF.
Figure 1-1 shows a simplied direct-conversion transmitter.
/3)
,
'$&
/25),
'LJLWDO
%DVHEDQG
3$
/3)
/25)4
4
'$&
Figure 1-1: Architecture of direct conversion transmitter.
In a direct-conversion transmitter, the in-phase (I) and quadrature-phase (Q) pulse-shaped

digital baseband signals are converted into analog signals via the digital-to-analog converter
(DAC). Then the low pass lter (LPF) can lter out the alias in frequency domain due
to sampling in DACs 1 . After that, the baseband analog signals are up-converted to RF
frequency by a single-sideband modulator, which is usually implemented as a quadrature
mixer as shown in Figure 1-1. A power amplier (PA) acts as the last stage in the transmitter path to provide enough output power to the antenna.
Figure 1-2 shows a simplied direct-conversion receiver2 .
1
Also out-of-channel noise is ltered. Usually thats not a big issue in transmitter.
Direct-conversion receiver is often called zero-IF receiver, which means its intermediate frequency (IF)
is ideally DC.
2
September 18, 2011
Introduction
/3)
,
3*$
$'&
/25),
/1$
/25)4
'LJLWDO
%DVHEDQG
/3)
4
3*$
$'&
Figure 1-2: Architecture of direct conversion receiver.
The signal received from the antenna will rst be amplied by a low noise amplier (LNA)3 .
Then its down-converted to baseband via a quadrature mixer. The following LPFs will
lter unwanted frequency components (both interference and noise) and the programmable
gain amplier (PGA) can bring signal to the required level for the analog-to-digital converter (ADC). After being converted to the digital domain, the I and Q signals are fed into
the digital baseband for further processing.
1-2-2
Common Metrics for Frequency Synthesizer
As a critical part of the RF transceiver, the frequency synthesizer has to meet specications
which vary for dierent applications. Yet there are some common metrics that are shared
in common and are considered in our project. A tour of these metrics is the key to
understanding the design, simulation and test presented in this thesis.
Since the frequency synthesizer is used to translate the signal frequency from BB to RF
or from RF to BB, the frequency accuracy of its output is very important. When the
frequency of the LO deviates from the desired value, for RX, after quadrature mixing,
some unwanted frequency components will not be attenuated by LPF and some useful
signal will be ltered away, which leads to degradation of SNR of the whole system. For
advanced communication systems like WiMAX and WiFi, Orthogonal Frequency Division
Multiplexing (OFDM) technology has been widely deployed. Then this frequency error
will undermine the orthogonality of subcarriers, if the digital baseband processor cannot
correct for this. Still the detail of this is more of an issue for telecommunication standards.
In this project we get the specication of tolerated frequency error from customers.
3
Sometimes we have a band-select lter in front of LNA. That depends on the tradeo of noise and
linearity, and always cost.
September 18, 2011
In older telecommunication standard like GSM and BT, the frequency range for the whole
band is small. While for more recent communication standards, the range can be pretty
wide. Some standards even have multi bands. Thus its essential to have a good frequency
planning at the frequency synthesizer level or the transceiver level to make sure LO can
cover the whole frequency range.
As can be seen in Figure 1-1 and Figure 1-2, LO needs to provide the I/Q signal for
quadrature mixer. The amplitude and phase mismatch of the I/Q signal would degrade
the image rejection of quadrature mixer. Say and G are respectively phase and
amplitude mismatch of the I/Q signal, as follows:
I = A(1 + G) cos(LO t + )
(1-1)
Q = A sin(LO t)
(1-2)
If we assume an RF signal above the LO which is cos((LO + IF )t) and an image signal
below the LO which is cos((LO IF )t), due to the mismatch of the I/Q signal, the image
signal will generate some undesired output at the frequency of interest IF . The magnitude
of the image rejection, as the power ratio of the desired output to the undesired output,
can be approximated as[4]
4
(1-3)
IRR
2
+ G2
Yet note that not all transceiver will need both I and Q signals. As in the traditional
polar transmitter which is used for constant-envelope modulation, maybe just one-phase
signal (single-ended or dierential) is needed. However, for some transmitter architectures
of advanced modulation, maybe phases more than just I and Q are needed.
The above metrics are mostly static, which means they are measured when the frequency
synthesizer has settled to a certain frequency. In some modern communication systems,
the frequency synthesizer is required to settle to another frequency in very short time. In
that case the settling time and the settling dynamic of the frequency synthesizer are also
very important.
The more complex specication for frequency synthesizer is phase noise. A detailed explanation for it is given below.
1-2-3
Explanation for Phase Noise
The output of a generic oscillator vo (t) with a sinusoidal wave shape and a nominal oscillation frequency fo hertz is[5]:
vo (t) = [A + a(t)] cos[2fo (t) + (t)]
(1-4)
Here A is the mean amplitude of the oscillator output, a(t) is the zero-mean amplitude
noise, and (t) contains all phase and frequency departures from the nominal oscillation
September 18, 2011
Introduction
frequency fo and phase 2fo t. Phase disturbance (t) (in radians) includes the zero-mean
phase noise, the initial phase, and the integrated eects of frequency oset and frequency
drift.
For generic oscillators that contain an amplitude-control mechanism, amplitude uctuations are greatly suppressed. Besides that, the signal is often converted into a square wave
somewhere in the system, which also clips o amplitude noise. Thus the eects of phase
noise far overshadow the eects of amplitude noise. This applies to our ADPLL design in
which a DCO is followed by dividers, as can be seen in Chapter 2. We will ignore the eect
of amplitude noise for the discussion below.
Since vo (t) as in Equation 1-4 is a random signal, we shall consider the Fourier transform
of its autocorrelation function, which is the double-sided spectrum of vo (t). We convert
this to the single-sided spectrum Wvo (f ), by multiplication of two. Ideally, in the absence
of phase noise, the single-sided spectrum Wvo (f ) would be a single line located at f = fo .
Due to the phase noise, the spectrum would spread into the vicinity of fo . The more noisy
the signal is, the greater the spreading is, as can be seen in Figure 1-3. However, the total
power of the signal, which equals the integral of Wvo (f ) over all frequencies f = 0 to ,
is A2 /2 volts2 .
No phase noise
Moderate noise
Larger noise
Figure 1-3: Theoretical spectrum Wvo (f ) of oscillator output vo (t).
The normalized version of Wvo (f ), L(f ), is dened as

L(f ) =
Wvo (fo + f )
A2 /2
(1-5)
which means the noise power within a bandwidth of 1 Hz in a single sideband at a frequency
oset of f from the center frequency fo , relative to the total power. Usually we express
L(f ) as 10 log[L(f )] dBc/Hz.
September 18, 2011
Wvo (f ) and thus L(f ) can be measured via a RF spectrum analyzer. Figure 1-4 gives a
simplied block diagram of one kind of spectrum analyzer. The signal with the frequency
fo is mixed with the signal of a swept local oscillator that has a frequency of fLO . The
frequency dierence fo fLO is applied to a bandpass lter that has the center frequency
fIF and the resolution bandwidth RBW. The output of this bandpass lter is fed into a
square-law detector and then a lowpass smoothing lter with video bandwidth VBW. This
smoothing-lter output either goes directly to a display that shows power or goes through a
logarithmic converter to a display that shows power on a dB scale, which is PRF (f ). After
scaled to RBW and normalized to the integral of the power spectrum, we can estimate
L(f ) from PRF (f ).
3RZHU
7R
'LVSOD\
6LJQDO,QYRW
0L[HU
%DQGSDVV
)LOWHU
6TXDUH/DZ
'HWHFWRU
6PRRWKLQJ
)LOWHU
/RJ
G%
IR
I/2
I,)5%:
9%:
)UHTXHQF\
6ZHSW
2VFLOODWRU
Figure 1-4: Simplied block diagram of a spectrum analyzer.
The problem with the RF spectrum analyzer is that it needs to handle the total power of
signal while still can detect the weak sidebands due to spurs or phase noise. Such diculty
has led to the widespread use of W (f ), i.e. the low-pass, single-sided spectrum of the
phase-noise modulation (t). A conceptual block diagram for the measurement of W (f )
is shown in Figure 1-5. The measurement instrument consists of a phase demodulator
which reproduces a magnitude-scaled version of (t), a low-frequency spectrum analyzer
to produce W (f ), and a logarithmic converter for the display purpose.
6LJQDO,Q
YRW
$FRV>IRWW@
3KDVH
'HPRGXODWRU
6SHFWUXP
$QDO\]HU
W
/RJ
:I
7R'LVSOD\
ORJ:I
Figure 1-5: Block diagram of a generic phase-noise analyzer.
If the phase-noise amplitude is small enough, it can be shown that L(f ) W (f )/2, so
we can obtain W (f ) using a phase-noise analyzer and minus 3 dB to get the corresponding
L(f ). Since our specications for phase noise are lower than -90 dBc/Hz, as can be seen in
Table 1-1, this approximation works very well. In the system simulation shown in Chapter
5, L(f ) is derived just in this way. We will touch that later.
September 18, 2011
Introduction
1-2-4
The Impairment of Phase Noise in Frequency Synthesizer
The impairment due to phase noise of LO in a receiver is usually illustrated by the phenomenon called reciprocal mixing. This happens when a receiver receives two signals at its
antenna, one is the small desired signal and one is an undesired large interference featuring
a frequency close to that of desired signal. If the LO has a signicant amount of phase noise
in it, when two signals are mixed at the mixer, the noise of LO from frequency synthesizer
is superimposed on both of the down-converted signals, as seen in Figure 1-6 4 . Then at
the output of the mixer, the small desired signal is corrupted by the LOs noise which has
been down-converted by the large interference. This is named reciprocal mixing because
the RF port of the mixer now acts like an LO port (the large interference becomes the LO
signal) and the LO port has become the RF port (the so called RF signal is actually the
noisy LO signal).
6LJQDOGXHWRFRQYHUVLRQRIWKHLQWHUIHUHU
/2
,QWHUIHUHU
6LJQDOGXHWRFRQYHUVLRQRIWKH
:DQWHG
6LJQDO
ZDQWHGVLQJDO
VNLUW
,)
,)
Figure 1-6: Eect of LO phase noise in a receiver (reciprocal mixing).
The real situation is more complicated as the specications for most standards would
include both single-tone interference and the modulated interfering signals. To have a good
estimation of the requirement for LOs phase noise, comprehensive measures of theoretical
calculation, simulation and eld test are often taken.
In a transmitter, the phase noise of LO will spread the ideal spectrum of the output
signal, causing spurious emissions. For a receiver that wants to detect a small desired
signal from some transmitter, the tail of a large signal in nearby channel from another
transmitter would signicantly degrade SNR of the received signal5 , as shown in Fig 1-7.
Thus communication standards usually specify stringent mask for spectrum of transmitter
output.
In addition to the simple analysis above, for modern wireless systems that use OFDM,
the bandwidth of one channel can be in the order of several MHz or even larger than 10
4
5
It may seem strange to put the spectrums of deterministic signal and stochastic signal in one picture.
Be aware that this is dierent from reciprocal mixing eect in the receiver.
September 18, 2011
1-3 Introduction to ADPLL
1HDUE\
7UDQVPLWWHU
:DQWHG
6LJQDO
F
F
Figure 1-7: Eect of LO phase noise in a transmitter.
MHz (and sometimes, the bandwidth of the channel is variable, which strategy WiMAX has
adopted). The channel is divided into several orthogonal subcarriers with bandwidth on the
order of 100 kHz. In that case, the close-in phase noise would spread out every subcarrier,
cause inter-subcarrier interference and jeopardize the orthogonality of subcarriers. The
close-in spot noise of the frequency synthesizer at a certain frequency oset, which means
L(f ), will indicate the degradation
resulted from the adjacent subcarriers. The integrated

single-sided phase noise, L(f )df , instructs the interference to any subcarrier introduced
by the interaction of all other subcarriers and the LO phase noise. Here the boundaries of
the integration are dened according to the bandwidth of the communication standard.
1-3
1-3-1
Introduction to ADPLL
Fraction-N Charge-pump PLL and Related Issues
Traditionally a great majority of frequency synthesizers for wireless applications are based
on the charge-pump PLL topology. As shown in Fig. 1-8, the output clock of the voltagecontrolled oscillator (VCO) is divided by N. The divided clock FDIV is compared with
reference clock FREF. The phase error (actually, the time dierence) of the edge of the
two clocks will be detected by the Phase/Frequency Detector (PFD) and it will generate
either an UP or a DOWN pulse proportional to the detected time dierence. That UP or
DOWN pulse will control the on/o of current source IP and IN . In the loop lter, this
September 18, 2011
10
Introduction
current ow will be converted to a VCO tuning voltage which will control the frequency of
FVCO and thus closing the loop. The frequency of FVCO shall be N times the frequency
of FREF when the loop is stable and settled.
&KDUJH
3XPS
8S
)5()
/RRS)LOWHU
,3
7XQLQJ
9ROWDJH
9&2
)9&2
3)'
&
'RZQ
,1
5
&
)',9
1
)UHTXHQF\
'LYLGHU
Figure 1-8: Simplied charge-pump PLL topology.
As we have mentioned above, a ne frequency resolution is desired in modern wireless

systems. If N is just integer, which means the frequency resolution is FREF, a.k.a. IntegerN PLL, then the reference frequency is on the order of hundreds of Hz or even tens of Hz.
However, the bandwidth of PLL is usually no more than one tenth of FREF for stability
concern. In that case the settling of PLL is pretty slow and cannot correct for the drift of
VCO in time. Besides, a very narrow bandwidth will result in a loop lter that takes too
much area.
The Fractional-N PLL can alleviate the tradeo between frequency resolution and bandwidth. The division ratio of frequency divider N can be fractional thus the resolution will
be determined by the eective bits of fractional part of N. The most common method
to generate a fractional division ratio is to use a multi-modulus divider (MMD), i.e. the
division ratio of the divider would toggle between two integer values and the average eect
results in a fractional value. This toggling behavior can incur spur or noise within the
bandwidth. To remove them out of bandwidth, usually a Sigma-Delta Modulator (SDM)
is to generate the command word for MMD.
The Fraction-N charge-pump PLL architecture, as shown in Fig. 1-9, has issues being
integrated into current deep-submicron CMOS technology, such as the leakage of loop lter,
distortion due to non-idealities of charge-pump, special mask needed from performance
concern. There are some challenges brought up when used in modern wireless systems[6]:
September 18, 2011
1-3 Introduction to ADPLL
11
1. Wideband wireless systems demand larger tuning range, which may results in a higher
gain or a larger swing in VCO control. The high gain of VCO tuning voltage, a.k.a.
KV CO , will make PLL noise-sensitive. Thus PLL generates higher spur and noise
due to the hostile environment in the System-On-Chip (SoC). Or a larger swing will
incur non-linearity of VCO and needs more eort for calibration.
2. There is a trend to choose the polar transmitter architecture, which encompasses
PLL and the power amplier driver (PA driver) as the essential blocks, for complex
communication system. To add a modulation to PLL, measures like either predistortion or two-point modulation are needed. However, these measures are all
subject to the non-idealities in the analog domain, like variation with respect to
process, voltage and temperature (PVT). The accuracy is limited. The performance
degrades due to these non-idealities.
&KDUJH
3XPS
8S
)5()
/RRS)LOWHU
,3
7XQLQJ
9ROWDJH
9&2
)9&2
3)'
5
&
'RZQ
&
,1
)',9
11
00'
EW
)&:
6LJPD'HOWD
0RGXODWRU
Figure 1-9: Fractional-N charge-pump PLL topology.
1-3-2
The Simplied ADPLL Schematic
ADPLL is a very promising candidate to replace charge-pump PLL in advanced CMOS

process and modern wireless applications. Fig 1-10 is the simplied schematic of an ADPLL
as in [1]. The DCO, an oscillator whose frequency is controlled by the digital tuning
word rather than the tuning voltage, lies in the heart of the ADPLL. FCW, which is
the abbreviation of Frequency Command Word, is the ratio of the desired frequency fV
September 18, 2011
12
Introduction
divided by the reference frequency fR . Since this FCW is digital signal, the cost of ner
frequency resolution is just more number of bits (NoB) for the fractional part of FCW.
Ideally one FREF cycle shall contain FCW cycles of the desired frequency. Thus the
reference phase accumulator in the gure gives the phase of the desired signal divided by
2. The oscillator phase accumulator, which is actually a counter clocked at DCO positive
edges, indicates the phase of DCO output in a resolution of 1 cycle (2). The time-todigital converter (TDC) can compensate for the fractional error of the oscillator phase
accumulator. Then the phase detector, which is merely an arithmetic subtractor, produces
faithfully the phase error PHE[k] between the desired signal and the feed-back signal from
DCO. PHE[k] is fed to the digital loop lter and the loop lter generates the normalized
tuning word (NTW). The oscillator tuning word (OTW) is obtained by multiplying NTW
DCO (an estimation of DCO gain, which
with the normalization ratio of fR divided by K
is dened as the frequency change due to change of one LSB in the DCO tuning word).
Therefore, we have a closed-loop ADPLL with full digital signals at the top level.
It shall be noted that for digital logic operations, a synchronous clock is needed. However,
the output clock of DCO (CKV) is asynchronous with the reference clock (FREF). A
retimer, which is drawn as a ip op, is used to generate the clock in the digital domain
(CKR). The idea is basically to oversample FREF with the high-frequency clock CKV.
Therefore the edge of CKR is aligned with the edge of CKV. Also the digital logic is active
only when TDC has nished the activity and been quiet.
5HIHUHQFH3KDVH
$FFXPXODWRU
'&2
55>N@
)&:
3+(>N@
'LJLWDO
/RRS)LOWHU
17:
'&2*DLQ
1RUPDOL]DWLRQ
27:

6DPSOHU
59>N@
5HWLPHU
)5()
2VFLOODWRU3KDVH
$FFXPXODWRU
59>L@
&.5
>.@
7'&
'&2SHULRG
QRUPDOL]DWLRQ
)UDFWLRQDO(UURU
&RUUHFWLRQ
Figure 1-10: Simplied schematic of ADPLL in [1](modulation part is not drawn).

September 18, 2011
7R3$
1-4 ADPLL for this WiMAX Project
13
Through these eorts, all blocks in ADPLL can be seen as ASIC cells. Say for DCO the
input is the digital tuning word(s) and the only information we care about in the DCO
output is the timing of edge. All issues with analog-intensive blocks in the charge-pump
PLL are avoided. The digital logic is highly immune to noise and if well calibrated, will not
be hurt by the DCO non-linearity. The bandwidth of ADPLL is well dened compared to
analog peers. Whats more, the digital signals are easier to store and maintain than analog
signals. For all these reasons, the ADPLL technology has advanced to so-called duty-cycled
PLL [7] and RF built-in self test (BIST) [8], and even further to transmitter synthesized
from standard cell library [9] and is expected to be the corner stone for software-dened
radio (SDR) [10].
There are now two basic architectures for ADPLL. The rst one resembles the traditional
fractional-N charge pump PLL. A programmable frequency divider is used in the feedback
path and the TDC simply replaces the combination of the phase detector and the charge
pump. The phase dierence is fed into the digital loop lter instead of the analog loop lter.
Thus its called digital fractional-N PLL. The second one, which is a simplied version
of Fig 1-10, doesnt need programmable divider and truly works in phase domain. Therefore
its named divider-less ADPLL6 . Both types have shown rather good performance in recent
papers[11][12][13]. In this project we choose the second architecture because comparatively
it consumes less power for the same performance in the literature and more techniques have
been proposed for this architecture[14].
1-4
1-4-1
ADPLL for this WiMAX Project

Specication Requirement for the Whole System
The specications of the frequency synthesizer from the customer for WiMAX standard
are shown in Table 1-1.
Table 1-1: Specication for the WiMAX ADPLL System.
Parameter
Frequency Bands
Frequency Step Size
Integrated SSB Noise (1 kHz - 10 MHz)
Spot Noise @ 10 kHz
Spot Noise @ 100 kHz
Far-out Noise
Target
2.3-2.7,3.3-3.8 GHz
25 Hz
-39 dBc
-90 dBc/Hz
-95 dBc/Hz
-150 dBc/Hz
According to Table 1-1, this frequency synthesizer is to deliver two frequency bands: High
Band (HB) which is from 3.3 to 3.8 GHz and Low Band (LB) which is from 2.3 to 2.7 GHz.
6
Note actually some xed-ratio divider like divide-by-2 or divide-by-4 divider can be introduced in the
feedback path of this architecture to ease the requirement of phase accumulator as well as TDC.
September 18, 2011
14
Introduction
Therefore a frequency plan shall be proposed for ADPLL to cover dual band. The frequency
step size, i.e. the frequency resolution of ADPLL, shall be within 25 Hz. The phase noise
specications, including the integrated phase noise, spot close-in noise and far-out noise,
shall be met.
In measurement, sometimes the RMS phase error rather than the integrated SSB phase
noise is given. There is a simple formula for the conversion of these two values:
180
RMS phase error (in degree) = (
)

2 L(f )df
(1-6)

and L(f )df is just integrated SSB noise. From Equation 1-6, we can know an integrated
SSB phase noise of -39 dBc/Hz corresponds to a RMS phase error of 0.91 .
1-4-2
Additional Requirement for this ADPLL Project
In ADPLL, the bulky analog loop lter in the charge-pump PLL is replaced by a digital
loop lter, which results in a signicant chip area save. Active area in recent ADPLL
papers is within 0.5 mm2 [12] [13] [11]. In this project, the whole area of chip would likely
be pad-limited due to extra pads for test and measurement. We also plan to add SRAM in
the chip for the test. This adds to extra cost of chip area also. Still, we expect the active
area for functional part to be within 0.5 mm2 .
Low power consumption is also an important consideration, especially for mobile applications. In this project, the goal of the total power consumption for the core part, excluding
FREF slicer, DPA output buer and Digital test blocks, is expected to be less than 10 mW.
Unlike the traditional charge pump PLL which shows a trade-o between bandwidth and
settling time, ADPLL can control its bandwidth easily in the digital domain so that it can
achieve both fast-settling and narrow bandwidth. We aim at 10-20 us settling time in this
project to show the potential use in the frequency-hopping application.
1-5
Project Sketch
This ADPLL project is a cooperative project in which three MSc students are involved.
My part lies in the system level work as well as the DPA design, as shown above. My
colleague Popong Eendrik is mainly responsible for the design of TDC and Armin Tavakol
is responsible for the implementation of the digitally controlled oscillator (DCO) with the
related dividers. This project is conducted in Catena Microelectronics BV and supervised
by Professor Robert Bogdan Staszewski. The whole eort is to explore how to use ADPLL
for the WiMAX standard and converge to a feasible implementation.
September 18, 2011
1-6 Outline of the Thesis
1-6
15
Outline of the Thesis
This thesis focuses on the system level design of ADPLL for the WiMAX application,
as well as a simple DPA transistor-level implementation as the output buer of ADPLL.
Chapter 2 shows the architecture of the ADPLL system and the building blocks. Chapter
3 presents the modeling and description of the whole system in Cadence using the VerilogAMS/Verilog languages. Also the performance analysis of the system is given. In Chapter
4, some advanced algorithms for ADPLL are discussed to overcome the non-idealities of the
system. Chapter 5 goes back to the top level of the system, with emphasis on the test plan,
the operation mode and the simulation results for the system. Chapter 6 demonstrates a
simple implementation of the DPA block for ADPLL. In the last chapter, some conclusions
are drawn to summarize the contribution of this thesis and to present the future work.
The appendix includes the important data not shown in the bulk part of the thesis, like
the code samples, the register map, the gures showing the system performance, etc.
September 18, 2011
16
September 18, 2011
Introduction
Chapter 2
ADPLL Architecture and Building
Blocks
As most complicated systems go, the design of ADPLL needs comprehensive cooperation
both at system level and circuit level. Especially when the design is to explore the possibility and do innovations, the methodology of either top down or bottom up will not fulll
the task. The design team has to go through multiple cycles of iteration and negotiation
to make the system feasible. To present the work, we choose to rst deliver the whole
architecture of ADPLL and then dive into the essential blocks.
Our ADPLL is a derivative structure as presented in [3]. The dierence with previous one
shown in Fig 1-10 is that in this structure the feedback phase information is dierentiated
into frequency information. After that frequency dierence with desired frequency is detected and then accumulated into phase error as input for digital loop lter. Theoretically
the function is the same with the previous one except for the extra delay. However, with
dierentiation and accumulation separated, one can easily freeze the value of PHE signal
or reset that to zero. This facilitates the implementation of some algorithm and can help
avoid some undesired perturbations. We will return to that in Section 4-1 and 4-2.
2-1
Architecture of ADPLL
A bird view of ADPLL in this project is shown in Fig 2-1. In this gure, our ADPLL
system is clearly divided into two divisions: red blocks that demand custom design either
due to the analog nature (like DCO) or due to that it is a very high speed logic (like
incrementor in ADPLL feedback path); blue blocks that work at low frequency and are
expected to be synthesized in digital ow. Therefore it is a mixed-signal system and only
September 18, 2011
18
ADPLL Architecture and Building Blocks
with eort and cooperation from analog/RF designer, digital designer and system engineer
can this chip be made possible.
The dash lines in this gure separate the whole chip into ve parts. The most right
part is DPA which acts as the output buer of ADPLL. The black blocks above and
below DPA block are virtual load intended for future integration into receiver, i.e. they
are not contained in the design. The part next to DPA block has DCO and the related
dividers/buers. This part will receive the control from the loop and generate the desired
signal for virtual receiver block / DPA block / feedback path in ADPLL. The left part
which contains TDC and Retimer+Incrementor receives the feedback signal from DCO
and feeds the raw phase information to digital loop logic. The blue part in the middle is
the loop logic which processes the phase information from TDC/Incrementor and delivers
the control word to the DCO block. The top part with SPI Slave, Sequencer and memory
in it will specify the working mode of ADPLL and store snapshot of the system for test.
With all these ve parts in place, we demonstrate a controllable ADPLL that can close
loop with buers for input (output) signal to go in (out of) chip, in short, a system feasible
in silicon.
63,B5(6(7
0,62
026,
66BQ
6&/.
([WB6HTB&/.
This chapter is mainly dedicated to the rst four parts in this gure. The SPI slave,
sequencer and memory will be discussed in detail in Chapter 5. Some advanced algorithms
have been applied to ADPLL system and will be discussed in Chapter 4.
63,6ODYH
'LYLGHU
'LYLGH
9''BGLJLWDO
9''BVOLFHU
966B'&2
966B+6'
966BGLJLWDO
966BVOLFHU
9''BJOREDO
966BJOREDO
6WDWLFFRQWUROZRUGV
*BZRUG
(QDEOHB,,5
6HTB&/.
0(0B7,0(B
6HTB&ONB6(/
&RUHB5(6(7
9''B+6'
([WB6HTB&/.
&25(B5(6(7
,QWB6HTB&/.
9''B'&2
6HTB&/.
,QWB6HTB&/.
6HTXHQFHU
3LSH
3LSH%XIIHU
&.5
'\QDPLFFRQWURO
ZRUGV
2VFLOODWRUB397
397*DLQ
5HIHUHQFH3KDVH
35
35
)&:
'&2B,1B3B*
0(0B*$,1B3

*3
23
G3+5
08;
7'&
7'&GHFRGH 3+,B,
DQG
QRUPDOL]DWLRQ 3+)

$QDORJ)5()B1
)5()GLWKHU
G3+)

)5()

)5()B6(/
G3+(

*$
781(B$

G3+9B603
3+9B603
&.9'

&.5
7UDFNLQJ*DLQ
781(B7,

&.9'
&.9B7'&
&.9B'&2
)5()
'&2B,1B7)
/RZIUHTXHQF\B&.9
&.9'
+%3$%XIIH
+%B%8)B(1
+%B3$B',9B(1
3LSH%XIIHU
&.9BS

3KDVH
5RWDWRU
'
'&2BWDLOUHV
Figure 2-1: Bird view of ADPLL in this project.
September 18, 2011
'LYLGHU
'LYLGH
,
4
,
4
$QDORJ7HVWB6(/
$QDORJ7HVWB2XWSXW
EXIIHU
35BHQ
&.9B'&2
',9BBB(1

&.5
&.9B'&2
&.9B7'&
'LYLGHU
'LYLGH
&.5
781(B7) 2VFLOODWRUB7UDFNLQJIUDF
27)
6(/B('*(
&.9B6(/
GLIIHUHQWLDO
/%B3$B',9B(1

*7
&.5B7*6
'LYLGHU
'LYLGH
'&2
'&2B,1B7
27,
0(0B*$,1B7
&.5
7HVWB'LYLGHUBUVWBD
/%B%8)B(1
2VFLOODWRUB7UDFNLQJLQW
=35__5HVHW
2$

GLIIHUHQWLDWRU

5HWLPHU
,QFUHPHQWRU
3+(
'&2B,1B$
'LYLGHU
'LYLGH
'LYLGHU
'LYLGHU
&.9B
2VFLOODWRUB$FTXLVLWLRQ
$FTXLVLWLRQ*DLQ
0(0B*$,1B$
/%3$%XIIHU
&.5
397BPLVV
&ONB*DWLQJB(QDEOH
6OLFHU'LWKHULQJ
'
EXIIHU
&.5
'LJLWDOB)5()
$QDORJ)5()B3

'&2B,1B3B'6
781(B3
&.9BS
&.9B
6PDOO
'LYLGHU
'LYLGH
3LSH
2-2 DCO and Related Dividers/Buers in ADPLL System
2-2
19
DCO and Related Dividers/Buers in ADPLL System
As mentioned in section 1-3, DCO has played an important role in ADPLL as the foundation to perform digital-to-frequency conversion (DFC). Despite the analog nature of an
oscillator, it is encapsulated as an ASIC cell with digital I/O interface. The input tuning
words OTW control the output frequency of DCO. The edge transition instances of DCO
output signals will contain all the information we want, like frequency, phase noise and
I/Q phase mismatch. Therefore, the modeling of DCO from the system point of view can
be greatly simplied. The specications and consideration for DCO from system level are
also pretty straight forward.
It shall be noted that in ADPLL DCO need to connect to other blocks. DCOs outputs
will be divided and fed back to TDC and high-speed counter in ADPLL. Then the loop
can correct for DCOs imperfections such as frequency drift and frequency pushing. The
outputs of DCO are also supplied to DPA and then go o-chip. Also the quadrature signals
shall be generated for the potential quadrature mixers in receiver.
All these connections are shown in zoom-in bird view of ADPLL system for DCO block,
as in Fig 2-2. To avoid the I/Q mismatch during signal propagation, dierential signal
with 2 times the desired frequency are needed for the local divide-by-2 quadrature signal
generation within receiver. So DCO need to deliver signal with frequency of 4.6-5.4 GHz
and 6.6-7.6 GHz. The ratio of highest frequency for HB and lowest frequency in LB is
almost octave. Thus its hard for oscillator to directly cover this range. The frequency
plan in this project is to generate signal with frequency at 6.6-7.6 GHz using a divide-by2 divider and generate signal with frequency at 4.6-5.4 GHz using a divide-by-3 divider.
Therefore the oscillator only needs to cover frequency range of 13.2-16.2 GHz.
It shall be noted that to take frequency deviation of PVT corner into consideration, DCO
core need to have about 1.5 GHz margin besides the frequency range specied above,
i.e. 11.7-17.7 GHz. Still the only dierence shown in system simulation is the deviation
of central frequency and can be easily compensated by calibration. Thus the frequency
deviation of DCO core is not taken into account into the system level simulation.
Fig 2-3 shows the schematic for frequency planning. The divide-by-3 function is implemented as a cascade of divide-by-2 and divide-by-1.5 so that divide-by-2 divider can be
shared for HB and LB. Another divide-by-2 divider in the feedback path is adopted. Thus
feedback CKV is just one fourth of output frequency of DCO core (3.3-4.05 GHz). The
design diculty of TDC and the high-speed incrementor is reduced.
The dividers and buers in Fig 2-2 are pretty simple. In Fig 2-2, there are ve control
signals to turn on/o each divider or buer. DIV 1 5 EN is to enable/disable the divide-by1.5 divider for LB TX and RX. LB BUF EN(HB BUF EN) turns on and o the buer for
LB(HB) output. LB PA DIV2 EN and HB PA DIV2 EN will enable/disable the divideby-2 divider for LB TX and HB TX. The redundancy of control signals provides both the
exibility and the way to test the power consumption of every divider/buer.
September 18, 2011
20

3LSH
%XIIHU
'&2B,1B3B*
'&2B,1B3B'6
EXIIHU

&.9B
'&2B,1B$
GLIIHUHQWLDO

'&2B,1B7
'LYLGHU
'LYLGH
'LYLGHU
'LYLGHU
'LYLGHU
'LYLGH
/%B%8)B(1
',9BBB(1
'&2
/%B3$B',9B(1
EXIIHU
'LYLGHU
'LYLGH

'&2B,1B7)
+%B%8)B(1
+%B3$B',9B(1

'&2BWDLOUHV
3LSH
%XIIHU
&.9BS
6PDOO &.9B
'LYLGHU
'LYLGH
Figure 2-2: Zoom-in bird view for the DCO block.
The focus is the DCO core as in Fig 2-4, which is a push-pull oscillator. The cross-coupled
transistors provide negative resistance needed in the oscillator. Compared with normal
cross-couple based oscillator, an obvious dierence is that some resistors are added to lower
down the icker noise within DCO. As we will see later, the icker noise in DCO is upconverted to 1/f 3 in phase noise spectrum and may hurt close-in phase noise performance
of ADPLL. Reduction of this eases the eort to meet system specications. The inductance
value in LC tank of oscillator core is xed. OTW will control the capacitance banks in LC
tank. In that way, the frequency of the oscillator core is digitally controlled. The 5 bit
binary DCOtailres signal at the tail of oscillator can specify the bias current of the core.
The oscillator core is turned o when DCOtailres value is 0. Larger value of DCOtailres
means more current bias and thus larger negative resistance from cross-coupled transistors
for the oscillator. Then the amplitude of oscillator output signal is larger and the phase
noise performance also improves. By modifying this signal, the situation that oscillator
cannot start up will be prevented just by pumping in more current, and we can have control
on the trade-o between performance and power consumption.
DCO in this project has three banks with dierent dynamic range and resolution in the
frequency domain: PVT bank (PB), acquisition bank (AB) and tracking bank (TB). They
together guarantee both the frequency range the DCO core need to cover and the resolution
DCO is to deliver. The control words for PB and AB are separately 7 bits and 6 bits binary
code. For TB, 64 bit thermometer-code control word is chosen to satisfy the matching
September 18, 2011
21
7R5;/%
*+]
7R'3$/%
*+]
'LYLGH
'&2FRUH
'LYLGH
'LYLGH
*+]
*+]
'LYLGH
'LYLGH
7RIHHGEDFNSDWK
*+]
7R'3$+%
*+]
7R5;+%
*+]
Figure 2-3: Frequency planning of ADPLL
September 18, 2011
22
9''
RXWS
'&2B,1B3B*!
'&2B,1B3B'6!

27:
RXWQ
$%
'&2B,1B3B'6
'&2B,1B$!

'&2B,1B3B*
3%
'&2B,1B$
7%
'&2B,1B7!
'&2B,1B7)
7)%
'&2B,1B$
'&2B,1B7
&VHU
&VHU
&SDO
'&2WDLOUHV!

Figure 2-4: Simplied schematic for the DCO oscillator core.
September 18, 2011
&SDO
23
requirement. As shown in Fig 2-4, in PB and AB, the input tuning words control NMOS
switches to turn on/o capacitors. The resolution we can achieve via this method is limited
by the smallest capacitance value the process can provide and the parasitic of the switches,
which is not good enough. To achieve a ner resolution, in TB the switch is in parallel with
relatively large capacitors (compared to the capacitor in series with the switch). When
switch is turned on, the large capacitors are shorted and only series capacitors show up.
When switch is turned o, what we see from the terminals are small capacitors in series
with big capacitors. Say the capacitance value for the series capacitor is Cser and the
capacitance for the parallel capacitor is Cpal , the turn-on capacitance value for the cap
Cser Cpal
Cpal
while the turn-o value for the cap unit is 12 Cser
= Cser
. Thus
unit is Cser
2
+Cpal
2 Cser +Cpal
capacitance change can be really ne by using huge parallel capacitors. However, here
the analysis is simplied and for practical design, the eect from not-ideal switches (nite
resistance both on and o) and parasitic capacitance kick in. The accurate values of
capacitance change should be extracted via the transistor-level analysis and simulation.
For the nominal case, the capacitance change corresponding to on/o of LSB of PVT bank
(CP ) is 2.35 fF, the capacitance change corresponding to on/o of LSB of acquisition bank
(CA ) is 160 aF and the capacitance change corresponding to on/o of LSB of tracking
bank (CT ) is 10 aF. The central capacitance is set to be 361 fF and the inductance is
325 pH.
The change of capacitance will correspond to a change of frequency. For the total capacitance of Ctot and inductance value of L, the oscillation frequency is
f=
2 LCtot
(2-1)
the derivative would give

f =
1
1
C
C
= f
= 2 2 Lf 3 C
2 2 LCtot Ctot
2Ctot
(2-2)
As is shown the ratio of f and C is proportional to f 3 . For oscillator with narrow tuning
range, f 3 will not change that much. However, in our case (fM AX /fM IN )3 = (16.2/13.2)3 =
f
1.85. The variation of C
respect to frequency shall be taken into consideration. Table 2-1
shows the frequency change of DCO core for LSB in PB/AB/TB at the highest frequency,
the lowest frequency and the central frequency. When we consider the frequency resolution
of the feedback CKV, i.e. Kdco , it is one fourth of the values in this table.
The fractional part of tracking bank (TFB) utilizes the high-speed dithering to further
increase the frequency resolution. This dithering is achieved by using a digital SigmaDelta modulator with a high speed clock. The average of high-rate toggling 0/1 bit stream
approaches low-speed fractional input. In our design, a 1st order Sigma-Delta modulator
is adopted. As shown in Fig 2-5, it is just a clocked adder. The high speed clock is
the CKVD8 signal in Fig 2-12, which is CKV divided by 8. The 5 bit fractional input
September 18, 2011
24

Table 2-1: Frequency resolution for PB/AB/TB in DCO core.
Frequency resolution
PVT bank
acquisition bank
tracking bank
Lowest frequency
13.2 GHz
34.67 MHz
2361 kHz
148 kHz
Central frequency
14.7 GHz
47.89 MHz
3261 kHz
204 kHz
Highest frequency
16.2 GHz
64.10 MHz
4364 kHz
273 kHz
corresponds to a frequency resolution of 215 204kHz

= 1.6kHz in the central frequency. This
4
simple structure reduces the power compared to higher-order modulator and the matching
requirement is lower as we only have 1 bit output. One may worry that the 1st order
modulator cannot fully randomize the output stream and will result in too high spurs. Yet
the input of this modulator is not constant due to the noise of DCO and TDC. And the
clock of this modulator works at a rather high speed. So the far-out spur due to dithering
is spread out and is suciently low, though still observable, as shown in Sec. As this
modulator is expected to be synthesized, we will come back to it later.
Figure 2-5: The 1st Modulator for fractional part of tracking bank.
The cooperation of PB, AB, TB and TFB provides both a sucient frequency range and
a sucient frequency resolution. As ADPLL needs to settle to a certain frequency, it will
traverse the modes of using PB (PVT mode), AB (acquisition mode), TB & TFB (tracking
mode), as in Fig 2-6. The mode switchover is controlled by the sequencer discussed in
Chapter 5.
2-3
TDC and Incrementor in ADPLL System
A zoom in version of the block at the left part of Fig 2-1 is shown in Fig 2-7. As is
shown, TDC and Incrementor+Retimer lie at the heart of this block. They together
provide the sampled feedback phase information for comparison in digital phase detector.
September 18, 2011
2-3 TDC and Incrementor in ADPLL System
397PRGH
&DSDFLWDQFHUDQJHI)
&DSDFLWDQFHUHVROXWLRQI)
$FTXLVLWLRQPRGH
&DSDFLWDQFHUDQJHI)
&DSDFLWDQFHUHVROXWLRQD)
7UDFNLQJPRGH
&DSDFLWDQFHUDQJHD)
&DSDFLWDQFHUHVROXWLRQD)
I)
25
397WXQLQJUDQJH
I)
$FTXLVLWLRQWXQLQJUDQJH
7UDFNLQJWXQLQJUDQJH
Figure 2-6: Flowchart of DCO operation modes.
A simplied timing diagram in Fig 2-8 illustrates how phase information can be measured
via cooperation of the retimer, the incrementor and TDC. Say the reference clock is FREF
and the feedback clock is CKV. Their positive edges are shown in the timing diagram.
The function of retimer is just to clock FREF with CKV. Then the retimed output CKR
features aligned edges with CKV and average frequency same with FREF. If the frequency
of CKV is 2.25 times the frequency of FREF and we assume the phase starts from 0,
the phase of CKV at FREF rising edges (measured in cycles) shall be sequentially 2.25,
4.5, 6.75, 9, 11.25, 13.5, 15.75, 18 and etc. The high-speed incrementor counts the cycle
numbers of CKV and its output (PHV SMP) is sampled by CKR. Then we have PHV SMP
with value of 3, 5, 7, 9, 12, 14, 16, 18 and etc. From this we can see that PHV SMP is
a quantized output of CKV phase with resolution of one cycle (2). On the other hand,
the timing dierence between rising edges of FREF and CKR is measured in TDC and
then normalized to one cycle of CKV to correct for the quantization error of PHV SMP.
These two combined together deliver exactly the feedback phase information. Fig 2-9 is
a schematic of the quantization error from the incrementor and the correction from TDC
with respect to phase (cycles of CKV) in the ideal situation.
There are other modules in Fig 2-7. The phase rotator receives quadrature feedback signals
and executes the phase rotation algorithm to suppress spurs in near integer-N cases, which
will be discussed in detail in Chapter 4. The signal CKV SEL controls a multiplexer to
choose whether the CKV for TDC is from DCO as in normal ADPLL operation or from o
chip generated signal for TDC test. The feedback CKV from DCO is divided by 16. The
signal AnalogTest SEL chooses o-chip analog signal between reference clock FREF and a
signal which is CKV divided by 16. If the divided-by-16 CKV is fed o-chip, we can check
September 18, 2011
26
'LJLWDOB)5()
&ONB*DWLQJB(QDEOH
$QDORJ)5()B3
08;
)5()
6OLFHU'LWKHULQJ
7'&
7'&GHFRGH 3+,B,
DQG
QRUPDOL]DWLRQ 3+)

$QDORJ)5()B1

)5()B6(/
)5()GLWKHU
6(/B('*(

3+9B603
5HWLPHU
,QFUHPHQWRU
&.9'
&.5
7HVWB'LYLGHUBUVWBD
&.9B7'&
&.9B'&2
'LYLGHU
'LYLGH
&.9'
6(/B('*(
&.9B6(/
)5()
35BHQ
,
4
,
4
&.9B'&2
&.9B'&2
&.9B7'&
$QDORJ7HVWB6(/
/RZIUHTXHQF\B&.9
$QDORJ7HVWB2XWSXW
&.9BS

3KDVH
5RWDWRU
Figure 2-7: Zoom-in bird view for block with TDC and Retimer+Incrementor.
)5()

&.9

3+)

&.5
3+9B603

Figure 2-8: Simplied timing diagram on the working mechanism of retimer, incrementor
and TDC.
September 18, 2011

27
4XDQWL]DWLRQHUURURI,QFUHPHQWRU

3KDVH
LQ&.9F\OHV
&RUUHFWLRQIURP7'&
Figure 2-9: Schematic of incermentors quantization error and TDCs correction in ideal
situation.
the output frequency of ADPLL and the close-in phase noise without DPA turned on. If
FREF is fed o-chip, we can check whether reference clock is working as we expected.
They are discussed in detail in Chapter 5.
FREF slicer in Fig 2-7 receives the o-chip analog dierential sinusoidal signal from crystal
oscillator and then generates the square wave clock signal. The generated FREF signal
must have a low enough phase noise so as not to degrade the ADPLL performance. In our
design, the FREF signal will be dithered to suppress spurs in near integer-N situations and
we will come back to this topic in Chapter 4. In case that the slicer cannot work properly,
a lower quality signal digital FREF is fed to the chip to make sure that at least ADPLL
can still work. This digital FREF signal is also utilized in the TDC test plan, as in Section
5-3-3. A multiplexer chooses which of these two FREF signals to be the FREF signal in
TDC and the incrementor.
TDC and the high-speed incrementor are rather complex modules and are critical to the
performance of the whole system. Therefore they are expected to be fully-custom designed.
This thesis will treat them more from system-level view. The detailed transistor-level design
is beyond the scope.
2-3-1
TDC in this ADPLL System
As is stated above, TDC in the general sense is to deliver the time dierence between the
edges of two signals. In the ADPLL system, more specically, the time dierence between
the rising edge of FREF and CKR is needed. Because the edge of CKR and CKV are
aligned, in the implementation, TDC measures the time dierence between FREF and
CKV edges.
Here we adopt the original pseudo-dierential TDC topology composed by cell with inverters and ip ops as in Fig 2-10[15]. The inverter acts as a delay element with a resolution
September 18, 2011
28
of Tinv between 10.5 ps to 12.5 ps. So the CKV signal will be delayed by kTinv after it
propagates through k inverters. The ip op is clocked by FREF and will sample the value
of delayed CKV at the FREF rising edge1 . Notice that after every inverter, the polarity of
the signal (high or low) will be reversed. To make the output right, in Cell #1, the positive
input of the ip op is connected to the delayed version of CKV and the negative input
of ip op is connected to the delayed version of CKV. While for Cell #2 the connection
is reversed. Such a toggling connection goes on in the whole TDC chain.
&HOO

&.9
'
'
'
'
)5()
&.9
4
4
4
4
Figure 2-10: Pseudo-dierential TDC architecture.
For the analysis of the TDC module, just look at the positive input of ip ops, it is
equivalent to that CKV is delayed by buers with delay of Tinv and then sampled by FREF.
We name the positive input of kth ip op as D[k] and then we have timing diagrams for
TDC as in Fig 2-11. In Fig 2-11 (a), the last CKV edge before the FREF rising edge is
falling edge while in g 2-11 (b), the last CKV edge before the FREF rising edge is rising
edge. In the traditional charge pump PLL, the case in (a) is called FREF leads CKV
while the case in (b) is called FREF lags CKV. Also we assume that one CKV cycle TCKV
is exactly 8 Tinv to simplify the illustration (In reality, TCKV is more than 20 Tinv and we
will talk about that later.). For (a) the output of ip ops Q(1:9) is 001111000. The
transition of 0 1 happens between Q(2) and Q(3). Then the timing dierence between
the FREF rising edge and the last CKV falling edge earlier than FREF rising edge, which
is shown as TF , is quantized as two inverter delays (2Tinv ). Similarly, since the transition
of 1 0 happens between Q(6) and Q(7), the timing dierence between the FREF rising
edge and the last CKV rising edge earlier than the FREF rising edge, which is shown as
TR , is quantized as 6Tinv . For the case in (b), we can also conclude that TR is 2Tinv and
TF is 6Tinv .
Therefore the time dierence between CKV edges and FREF edges are quantized in this
TDC by the resolution of Tinv . Yet the desired information is actually the timing dierence
1
Here ip op is simplied as an ideal sampler. Practical ip op has setup time and hold time
requirement. We will talk about that later.
September 18, 2011
29
)
)
&.9
&.9
Figure 2-11: Timing diagram for TDC.
September 18, 2011
30
between the FREF rising edge and the rst CKV rising edge no earlier than FREF rising
edge, which is shown in Fig 2-11 as TCKV TR . It will be normalized to TCKV , transferred
to phase domain as PHF in Fig 2-8 and combined with incrementor output for phase
comparison in digital logic. Assume that TR is quantized as kTinv , we have the phase
domain output as:
TCKV TR
kTinv
Tinv
=1
=1k
(2-3)
TCKV
TCKV
TCKV
inv
Thus its important to estimate TTCKV
, which we call Ktdc with enough accuracy. In Fig
2-11, the quantized output for TR and TF are called TDC RISE and TDC FALL. Notice
that the dierence of TR and TF is half cycle of CKV, so the dierence between TDC RISE
and TDC FALL can be an estimation of 12 TTCKV
. We have
inv
TCKV
2(|TDC FALL-TDC RISE|)
Tinv
(2-4)
The inverse is just Ktdc .

This estimation is rather coarse as its limited by TDC resolution. A way to approach
enough accuracy is to accumulate multiple cycles of the estimated value and then do
averaging. We will return to this issue later. For now, one can just assume that this
normalization is accurate enough.
To catch both the RISE edge and the FALL edge of CKV, TDC shall guarantee to cover
one full cycle of CKV. The number of TDC cells L can be calculated as below:
L
max(TCKV )
min(Tinv )
(2-5)
From Fig 2-3 we know that the lowest frequency of CKV is 3.3 GHz, corresponding to 300
ps for max(TCKV ). min(Tinv ) is 10.5 ps from corner simulation result[15]. Then L shall
be no less than 29. To save some margin, L in this design is chosen as 32, i.e. TDC RISE
and TDC FALL will be 5 bits. To minimize the mismatch, 4 cells have been added before
and after the eective cell chain. Therefore we have 40 cells in this TDC as in Fig 2-10.
2-3-2
Retimer and High-Speed Incrementor
As the retimer and incrementor is not implemented at the transistor level yet, for modeling
purpose, we adopt the topology as in Figure 4.27 of [14]. The schematic is redrawn in Fig
2-12.
The retimed clock CKR is generated by clocking FREF with CKV, which can be done via
an ideal ip op as is shown in Fig 1-10. However, in reality the ip op has metastability,
i.e. when the rising edge of FREF and CKV are too close, it may take very long time for
the ip op to resolve the value of CKR. Then it is not proper to rely on CKR as the clock
September 18, 2011
31
3+9>@

3+9B603>@
0RGFRXQWHU
4;
4;
41
&.5
&.9'
+LJKHU3+9ELWV
&.9'
4;
41)
&.9'
3+9>@

3+9>@

6(/B('*(
3+9>@

&.5B1
)5()

43
&.5
&.5B3
&.9
&.9'
&.9'
Figure 2-12: Schematic of proposed implementation of the retimer and incrementor.
for the system. The way to solve this problem is to sample FREF both at the CKV rising
edge and the CKV falling edge, as shown in Fig 2-12, the output are named respectively
QP and QNF. Since at least one of the two CKV edges are far away from the FREF rising
edge, either QP or QNF will deliver the correct sampled value within a short time. QNF
is sampled by the CKV rising edge and the output is QN. Then QN and QP are aligned
with CKV.
The top part of Fig 2-12 is a mod-8 counter for the 3 least signicant bits in PHV. As
shown in the timing diagram of Fig 2-13, PHV[2:0] does increase by 1 at every CKV cycle.
The signal CKVD8 can be used as clock in the synchronous counter for the higher bits
of PHV. Therefore the higher bits can just work at lower frequency to reduce the power
consumption.
PHV[2:0] are both sampled by QP and QN in Fig 2-12. For QP and QN, corresponding
to the two situations as in Fig 2-11 (a) and Fig 2-11 (b), we have timing diagram shown
in Fig 2-14(a) and Fig 2-14 (b). In this timing diagram QN lags QP for one CKV cycle
in the case of Fig 2-11 (a) while QN and QP are aligned in the case of Fig 2-11 (b). Thus
we need to compensate for the case (a) when SEL EDGE is 1. One can observe that the
dierence between Fig 2-11 (a) and Fig 2-11 (b) is the value of Q(1). Q(1) is 0 in case (a)
while and Q(1) is 1 in case (b). So we have the logic for PHV compensation as:
When SEL EDGE is 0 and Q(1) is 0, PHV SMP shall minus 1 for correct operation.
September 18, 2011
32
&.9
4;
4;
4;
4;
4;
4;
3+9>@
Figure 2-13: Timing diagram for signals in a mod-8 counter.
&.9
&.9
)5()
)5()
43
43
41)
41)
41
41
Figure 2-14: Timing diagram for QP and QN in the retimer.
September 18, 2011
2-3-3
33
Cooperation of TDC and Retimer+Incrementor
Now we have a functional TDC and a functional Retimer+Incrementor. Its indispensable

to inspect their cooperation so that they can deliver the desired phase information. There
are two important issues that are of concern:
1. The generation of SEL EDGE signal.
One may notice that for the operation of Retimer+Incrementor, there is an important
signal SEL EDGE. This is to choose CKR P or CKR N as the retimed clock CKR. As
seen in Fig 2-7, this signal is generated from the TDC decoder logic by utilizing the
redundancy of the TDC output. Assume that TCKV = 20Tinv , which more resembles
our design than the situation shown in Fig 2-11 with TCKV = 8Tinv , if FREF rising
edge and CKV rising edge are too close, we have timing diagram as in Fig 2-15. One
can conclude that for both cases, the value of Q(2) to Q(8) are all 0. Actually as
shown in the gure, Q(k) is the value of CKV that is kTinv ahead of FREF sampling
moment. Thus if Q(k) is 0 for k that is around 14 TTCKV
(in Fig 2-15 the value is 5), it
inv
means that CKV falling edge is far away from FREF rising edge. Then SEL EDGE
can be made low to choose the falling edge. Also if Q(5) is 1, then the CKV rising
edge is far away from the FREF rising edge and SEL EDGE can be made high to
choose the rising edge.
)5()
)5()
&.9
&.9
'
'
'
'
'
'
'
'
'
'
'
'
'
'
'
'
'
'
Figure 2-15: Timing diagram for SEL EDGE signal.
In our design
TCKV
Tinv
can vary with desired frequency as well as TDC delay variation

September 18, 2011
34

caused by process, voltage and temperature (PVT). It may change between 20 and
30. Still, we can just choose Q[5] as the SEL EDGE signal.
There is a doubt with the metastability of the SEL EDGE signal itself. Notice that
when SEL EDGE is metastable, it means that the rising edge of D[5] is close to the
rising edge of FREF, which has the hidden meaning that neither the rising edge nor
the falling edge of CKV will be close to the rising edge of FREF. In that sense we can
choose either CKR P or CKR N in Fig 2-12 as CKR for the digital domain operation.
The only concern is that we shall make sure that SEL EDGE is a stable 0 or 1 before
the rising edge of CKR P or CKR N to avoid the possible glitch. That is achieved
by inserting two columns of ip ops clocked at CKVD8 as in Fig 2-12. As CKVD8
has a maximum frequency of 500 MHz (4.05 GHz/8), this adds a minimum delay
of 2 ns for the SEL EDGE signal to resolve the metastability. From Figure 5-11 of
[15] (redrawn here as Fig 2-16), the extrapolation of the curve gives a metastability
window tw of less than 3 1026 s for the delay of 2 ns. The error probability for
SEL EDGE is
Pe
tw
= fR = tw fv fR < 31026 4.05109 40106 = 4.86109 (2-6)
1 second
Tv
The mean time between failures is 6.48 years, which is just the inverse of the value
above. Then it is almost impossible for the metastability of SEL EDGE to propagate
into the whole system. If we want higher reliability for operation of more than 5 years,
another column of ip ops clocked at CKVD8 can be added.
2. The PHE spike problem. The PHE spike problem can be illustrated in an extreme
situation. Say FREF rising edge and CKV rising edge are just at the same time,
when we observe at that time, would incrementor increase by 1 or dont increase?
Would TDC deliver a compensation value of 0 or 1?
If the incrementor increases by 1 and the compensation of TDC is 1, then the accumulated output phase is correct. This also applies when the incrementor doesnt
increase while TDC compensates 0. However, if the incrementor increases by 1 and
the compensation of TDC is 0, the output phase is one cycle larger than the correct
value. Then PHE will show a spike of -1. If the incrementor doesnt increase and the
compensation of TDC is 1, the output phase is one cycle smaller than the correct
value. PHE will show a spike of +1.
The spike of 1 also happens when TDC path and Incrementor path has any time
mismatch. This may be because of the CKV delay dierence for the two modules due
to layout imperfections or the setup time dierence of ip ops in the two modules.
Fig 2-17 shows the consequences if there is some time mismatch between two modules.
Here we assume FREF signals of the two modules are exactly the same while CKV
signals are not. 2 We name the CKV signal for TDC as CKV TDC while the
2
Actually the essential reason is the mismatch of time dierence between FREF edge and CKV edge
in the two modules.
September 18, 2011
35
Figure 2-16: Metastability window of sense amplier ip-op.
September 18, 2011
36

CKV signal for Retimer+Incrementor as CKV inc. In Fig 2-17 (a) CKV inc lags
CKV TDC by a very small fraction of CKV cycle . As shown in the timing diagram,
for most of the time, when the FREF rising edge is not between CKV inc rising
edge and CKV TDC rising edge, the compensation due to TDC is smaller than
quantization error of incrementor by . That kind of oset is of no signicance to the
operation of ADPLL. However, if FREF edge is between CKV TDC and CKV inc,
the compensation due to TDC is larger than the quantization error of the incrementor
by 1 . So comparatively, we have a PHE spike of +1. Similarly, for Fig 2-17 (b)
CKV TDC lags CKV inc by just TCKV . For most of the time, the quantization error
of incrementor is smaller than the TDC compensation by while if FREF rising edge
is between CKV inc rising edge and CKV TDC rising edge, the quantization error
of incrementor is larger than the TDC compensation by 1 . Then PHE spike of
-1 can be expected.
It may be easier to explain this phenomenon in Fig 2-18. Unlike the ideal situation
in Fig 2-9, Fig 2-18 (a) corresponds to Fig 2-17 (a) and Fig 2-18 (b) corresponds to
Fig 2-17 (b). Here is chosen as 0.1 and one can easily see the periodical PHE spike
as in deep green curve.
In previous works on ADPLL, this phenomenon is also mentioned [16, 17]. The
method proposed in these papers is basically to detect this spike in the digital logic
and then compensate for that. However, notice that an intentional frequency hopping
can also cause a sudden change of PHE. Besides that, some unexpected change of
temperature or supply voltage may also lead to such eect. To distinguish that from
this spike, the detection algorithm has to be robust, which may add to the complexity
of the logic and introduce extra delay in the loop.
For our design we have SEL EDGE which will choose the CKV falling edge to retime
FREF when the CKV rising edge and the FREF rising edge is too close. That
means for the incrementor we are actually using the CKV rising edge that follows
the falling edge. Then the quantization error can be larger than one CKV cycle. The
consequences of time mismatch between TDC module and Retimer+Incrementor
module is shown in Fig 2-19. Compared to Fig 2-17, one may notice that for case
(a), when FREF is between rising edge of CKV TDC and CKV inc, the oset is the
same as the normal situation. However, for case (b), we still have an oset of 1.
It is much more clear when we consider Fig 2-20. Case (a) and case (b) both have
time that a spike of +1 happens (that is a spike of -1 for PHE). Now we can know
this can happen and only happens when SEL EDGE chooses CKV falling edge for
retiming and FREF rising edge happens before CKV TDC rising edge. One can nd
that it is equivalent to SEL EDGE==0 and TDC Q(1)==0.
Thus we have another signal PHI I as the TDC decoder output in Fig 2-1. It is high
when and only when SEL EDGE is low and TDC Q(1) is low. (Notice that here the
input for the TDC decoder is from the eective TDC chain with 32 TDC delay cells.
The dummy cells before and after the eective chain are not fed to decoder input.)
September 18, 2011
37
)5()
7'&
3+)
&.9B7'&
&.9
&.9BLQF
5HWLPHU

,QFUHPHQWRU
3+9B603
&.9B7'&
&.9B7'&
&.9BLQF
&.9BLQF
)5()
)5()
&RPSHQVDWLRQRI7'&
4XDQWL]DWLRQHUURURI
FRXQWHU
4XDQWL]DWLRQHUURURI
FRXQWHU
&RPSHQVDWLRQRI7'&
)5()
7'&
&.9B7'&
3+)
&.9
&.9BLQF
5HWLPHU

,QFUHPHQWRU
3+9B603
&.9B7'&
&.9B7'&
&.9BLQF
&.9BLQF
)5()
)5()
4XDQWL]DWLRQHUURURI
FRXQWHU
&RPSHQVDWLRQRI7'&
&RPSHQVDWLRQRI7'&
4XDQWL]DWLRQHUURURI
FRXQWHU
Figure 2-17: PHE spike due to mismatch between TDC path and incrementor path.
September 18, 2011
38


5HPQDQWGXHWR

PLVDOLJQPHQW

3KDVH
LQ&.9F\OHV

&RUUHFWLRQIURP7'&

5HPQDQWGXHWR

PLVDOLJQPHQW

3KDVH
LQ&.9F\OHV
&RUUHFWLRQIURP7'&
Figure 2-18: Generation of PHE spike.
That just corresponds to our previous assertion for PHV compensation. In that way
the PHE spike is eliminated by only inspecting TDC output. The complex logic for
spike detection is avoided and we can distinguish it from the sudden change of PHE
due to frequency hopping or voltage/temperature change.
2-4
Digitally-Controlled RF Power Amplier
In our ADPLL design, DPA works as the output buer of ADPLL. It receives the ADPLL
output and feeds them to o-chip measurement devices. Besides this purpose, we add
some amplitude control signals to control the output power. This is to prepare for future
implementation of the transmitter for WiMAX. There is no strict requirement for the
resolution of this power control and also no hard limit for the eciency of this DPA.
However, the noise contributed by DPA shall be sucient low to avoid overshadowing the
phase noise of ADPLL output signal.
We have also a zoomed in bird view for DPA block as in Fig 2-21. The detailed design of
DPA is given in Chapter 6.
September 18, 2011
2-4 Digitally-Controlled RF Power Amplier
39
)5()
7'&
3+)
&.9B7'&
&.9
&.9BLQF
5HWLPHU

,QFUHPHQWRU
3+9B603
&.9B7'&
&.9B7'&
&.9BLQF
&.9BLQF
)5()
)5()
&RPSHQVDWLRQRI7'&
4XDQWL]DWLRQHUURURI
FRXQWHU
4XDQWL]DWLRQHUURURI
FRXQWHU
&RPSHQVDWLRQRI7'&
)5()
7'&
&.9B7'&
3+)
&.9
&.9BLQF
5HWLPHU

,QFUHPHQWRU
3+9B603
&.9B7'&
&.9B7'&
&.9BLQF
&.9BLQF
)5()
)5()
4XDQWL]DWLRQHUURURI
FRXQWHU
&RPSHQVDWLRQRI7'&
&RPSHQVDWLRQRI7'&
4XDQWL]DWLRQHUURURI
FRXQWHU
Figure 2-19: PHE Spike with SEL EDGE signal in this design.
September 18, 2011
40

5HPQDQWGXHWR

PLVDOLJQPHQW

D

3KDVH
LQ&.9F\OHV

&RUUHFWLRQIURP7'&

5HPQDQWGXHWR

PLVDOLJQPHQW

&RUUHFWLRQIURP7'&
Figure 2-20: Generation of PHE spike in this design.
'3$B$&:B/% '3$B&&:

/%3$%XIIHU
$'3//B/%
/%
3$
3$B&DS%LDV
966B3$
$'3//B+%
+%
3$
+%3$%XIIHU
'3$B$&:B+%

'3$B&&:
Figure 2-21: Zoom-in birdview of DPA.
September 18, 2011

E
3KDVH
LQ&.9F\OHV
2-5 Low Speed Digital Logic
2-5
41
Low Speed Digital Logic
As shown in Fig 2-22, outputs of TDC and the incrementor are fed to low speed digital
logic. The logic rst extracts the feedback phase information and does dierentiation to
obtain the feedback frequency. This is compared with desired frequency and the result is
accumulated as phase error information PHE. PHE signal will be fed to individual loop
lter for control of three capacitor banks (PB, AB and TB as in section 2-2) of DCO.
Therefore DCO can be stabilized at desired frequency.
2VFLOODWRUB397
397*DLQ
5HIHUHQFH3KDVH
35
'&2B,1B3B*
0(0B*$,1B3

35
*3
23
781(B3
G3+5
)&:
&.5
&.5
397BPLVV
2VFLOODWRUB$FTXLVLWLRQ
$FTXLVLWLRQ*DLQ
0(0B*$,1B$

7'&B4

G3+)

7'&GHFRGH 3+,B,
DQG
QRUPDOL]DWLRQ 3+)

G3+(

3+(
'&2B,1B$
*$
781(B$
G3+9B603
2$
&.5

2VFLOODWRUB7UDFNLQJLQW
GLIIHUHQWLDWRU

3+9B603
'&2B,1B3B'6
7UDFNLQJ*DLQ
=35__5HVHW
781(B7,

*7
'&2B,1B7
27,
0(0B*$,1B7

&.5
781(B7) 2VFLOODWRUB7UDFNLQJIUDF

27)
'&2B,1B7)
&.5B7*6
&.5
&.9'
Figure 2-22: Zoom-in birdview of low speed digital logic block.
2-5-1
TDC Decoder and Normalization Block
Fig 2-23 shows the schematic of the TDC decoder and normalization block. The outputs
of TDC core TDC Q are Q(k) which are shown in Fig 2-10, where k is from 1 to 40.
This information has to be decoded to TDC RISE and TDC FALL as in Fig 2-11. Notice
that Q(1)Q(4) and Q(37)Q(40) are outputs of dummy cells and shall not be taken into
account, thus the eective input has 32 bits. The decode logic is pretty straightforward:
TDC RISE corresponds to the rst 1 0 transition and TDC FALL corresponds to the
rst 0 1 transition. Fig 2-24 exemplies how the TDC decode logic works.
For the low speed digital logic, we use CKR as its universal clock. As in Fig 2-12, CKR
edge and FREF edge can be very close. On the other hand, the TDC output also changes
when the FREF rising edge happens. It may lead to some issues if we want to sample the
new TDC output and send that to the decoder at CKR rising edges.
There are two easy ways to work around that. One is to choose sampling TDC outputs
at CKR falling edge. The other one is to implement TDC decoding logic as a latch. The
September 18, 2011
42
3+,B,
7'&B4
/DWFKHG
'HFRGHU

3+)B)
7'&B5,6(
7'&B)$//
0XOWLSOLHU

6(/B('*(
&.5
1RUPBVWDUW
1RUPBVWRS
$FFXPXODWRU
.WGF
7&.97LQY
2II&KLS
&DOFXODWLRQ
Figure 2-23: Schematic of the TDC decoder and normalization block.
Figure 2-24: TDC decode logic illustration.
September 18, 2011
43
decoder always works as long as CKR is high. The benet of this method is that, as we
will see later, we have more time for the TDC normalization. With a latched decoder, the
normalization of TDC RISE can borrow some part of the time when CKR is high, as it
doesnt take much time for the decoder to deliver the correct output. Therefore the timing
constraint of normalization logic is relaxed and the power consumption can be reduced.
As mentioned in Equation 2-3, to normalize the time dierence of edges indicated by TDC
to the CKV cycle, we need to estimate Ktdc and multiply with k, i.e. TDC RISE. Equation
2-4 shows that with the information of TDC RISE and TDC FALL, TTCKV
can be estimated.
inv
To achieve an enough accuracy for this estimation, |TDC FALL-TDC RISE| is averaged
over 256 cycles. Actually |TDC FALL-TDC RISE| is accumulated for 256 cycles and then
shift by 7 bits concerning the proportional coecient of 2 in Equation 2-4. The start and
stop signal is provided by the sequencer, which we will discuss in Chapter 5.
. A generic divider can be quite
However, division logic is needed to get Ktdc from TTCKV
inv
complex and power hungry. Thus in this design we implement the division operation oinv
is read out from register and then the calculated TTCKV
is
chip, i.e. the estimation for TTCKV
inv
written into some register as the normalization coecient. For the future ADPLL design
with baseband, the digital signal processor (DSP) can be utilized to fulll this division
task.
Finally, a multiplier is needed to execute TDC RISE Ktdc in real time. With a latched
TDC decoder, we have actually more than half FREF cycle for this operation. Ideally, the
output only has fractional bits. Yet chances are that the output may be more than 1 due
to the inaccuracy of Ktdc and the uctuation of TDC RISE. Then one extra bit is needed
for the output.
To exemplify how normalization and multiplication works in xed-point logic, we assume
that the summation of |TDC FALL-TDC RISE| over 256 cycles is 2740. Then this number
inv
is shifted by 7 bits and inversed to obtain TTCKV
. To provide enough resolution and cover
Tinv
dynamic range of Ktdc , TCKV has LSB corresponding to 216 and MSB corresponding to
24 . The xed-point representation for Ktdc is actually
223
1
16
= 3062
2
=
2740/27
2740
(2-7)
If TDC RISE=5, as LSB of PHF is 221 , we have

3062
f ixedpoint
PHF = TDC RISE Ktdc
=
5 16 221 = 5 3062 22116 = 489920
2
which is about 0.234 for oating equivalent value.
2-5-2
(2-8)
Phase Detection Logic
The feedback phase information provided by PHV SMP, PHF and PHI I will be compared
with the desired value in the phase detection logic. As mentioned above, actually the
frequency error is detected and its accumulation gives the phase error.
September 18, 2011
44
As in Fig 2-22, we have two signals to control the accumulator. One signal, Reset||ZPR is
to reset accumulator output to zero for reset and zero phase restart (ZPR) algorithm. We
will discuss that in Section 4-1. Another signal, PVT miss, is to freeze the accumulator
output to avoid a wrong frequency detection result. This will be explained in Section 4-2.
This dierentiation-then-accumulation way provides the exibility to control the loop dynamic. However, an extra delay of 1 CKR cycles is introduced to the loop. Then the
loop stability is hurt when loop gain is rather high. For TB operation, this is not a big
issue because the loop gain is pretty small. However, for PB and AB, the proportional
coecient, , shall not be made too large to make sure the loop is stable.
The 25 Hz frequency resolution and reference frequency of around 30 to 40 MHz set the
fractional bit of FCW to be 21 bits. As for the integer part of FCW, sucient number of
bits is needed to avoid aliasing. Feedback frequency of 3.3-4.05 GHz corresponds to about
7 bits. Actually the integer part of FCW is 11 bits wide for an enough dynamic range
margin3 .
This sets the word length of integer part and fractional part for signals in the digital logic.
PHF has 22 bits as the extra bit is to avoid overow. Its dierentiated output dPHF has
23 bits to include the sign bit. The frequency detection result, dPHE, has 11 integer bits
and 21 fractional bits. This also applies to the accumulated phase error result PHE.
2-5-3
Loop Filter for PVT and Acquisition Bank
In Fig 2-22, the phase error result from the accumulator is fed to three gain circuits
separately for PB, AB and TB (GP, GA and GT). The output of gain circuit is sent to
the oscillator control circuit (OP and OA correspond to GP and GA, OTI and OTF is for
integer and fractional part of GT output respectively). Most of the system specications
are related to how the tracking bank loop lter is constructed, i.e. the algorithm in GT,
OT and OTI. We will leave the discussion of these complex blocks to next subsection and
start from the relatively simple loop lter for PB and AB.
In the gain circuits, GP and GA, the arithmetic shift and the multiplication are executed.
The arithmetic shift is to implement proportional coecients P and A in PVT and acquisition loop, as we will discuss in Section 3-4. The proportional coecients are chosen
as integer powers of 2, say 22 , 23 , to simplify calculation. The output of the arithmetic
shift, NTW, is just the ratio of the desired frequency change of DCO divided by reference
frequency. From Table 2-1, every bank of DCO has its own frequency resolution (we call
that Kdco,P , Kdco,A and Kdco,T ) and the resolution even changes with oscillating frequency.
Then NTW has to be multiplied with the values K fR and K fR for DCO Gain Normaldco,p
dco,p
dco,p and K
dco,a are respectively the
ization in PVT bank and Aquisition bank, where K
estimation of CKV frequency resolution for PVT bank and acquisition bank.
3
Say, if we want to use a reference clock with lower frequency. The loop shall still work properly.
September 18, 2011
45
The oscillator control circuits, OP and OA, are to control which mode in Fig 2-6 DCO is
traversing. The descriptive schematic of OP and OA is shown in Fig 2-25. The topology
of OP and OA block are almost the same, except for the PVT miss logic in OP block. We
will ignore that for now and come back to this in Section 4-2.

5(*B'&2B3
'&2B,1B3B*

'&2B,1B3
781(B3
&.5
&7/B6567B3

'&2B,1B3B'6

&7/B$567=
397PLVV
ORJLF
0(0B'&2B3
23B=35
&7/B3//B3
&.5

5(*B'&2B$

'&2B,1B$
781(B$
&7/B6567B$
&.5
&7/B$567=

0(0B'&2B$
&7/B3//B$
2$B=35
&.5
Figure 2-25: Schematic for OP and OA blocks.
The schematic of OP block shows quite a few control signals. CTL PLL P is to control
whether PVT mode is active or already frozen (say, when DCO has advanced to acquisition or tracking mode). When CTL PLL P is high, the output of GP block TUNE P
is selected. When CTL PLL P is low, the tuning word at the last CKR cycle is again
selected as the input, which means PVT tuning word is xed and PVT mode is frozen.
Thus CTL PLL P changes from high to low when we switch from PVT mode to acquisition
mode. Then OP ZPR will have a high pulse to activate the ZPR mechanism detailed in
Section 4-1. CTL SRST P is a synchronous reset and CTL ARSTZ is an asynchronous reset. MEM DCO P reads a start value for PVT bank tuning word to speed up the frequency
settling process. The REG DCO P signal can be stored in some register for ADPLL test.
Notice that TUNE P is a signed digital word and ranges from 26 to 26 1. However, the
September 18, 2011
46
number of capacitor unit turned on for PB ranges from 0 to 27 1, which is an unsigned

word. So we shall adds 26 to TUNE P and that results in DCO IN P signal. This can be
achieved by just reversing the signed bit (MSB) of TUNE P. For example, TUNE P is 0
and then DCO IN P is 64. That is the central capacitance value for DCO, as we expect.
2-5-4
Loop Filter for Tracking Bank
The eective processing of phase error signal in the loop lters for PVT and acquisition
bank are only arithmetic shift and multiplication. No integration is included. This is called
type-I PLL. The purpose of PVT mode and acquisition mode is to make sure that the CKV
frequency can settle to the range of tracking bank within a rather short time, say a few
microseconds. The type-I PLL is appropriate for the fast frequency settling. Nevertheless,
for tracking bank, the phase noise and the settling time are both important considerations.
As more constraints are added, the architecture of the loop lter for tracking bank is more
complicated than that of PVT bank and acquisition bank.
*BZRUG
3URSRUWLRQDOBRXWSXW
0(0B*$,1B7
781(B7,
3+(

*6BHYHQW
/DPEGD
/DPEGD
/DPEGD
/DPEGD

/6'B(QDEOHB,,5
Fig 2-26 is a high-level schematic for the GT block. The PHE signal has to pass through
an IIR lter bank before the arithmetic shift operation for proportional gain. The IIR
lter bank contains four IIR lters in series. Fig 2-27 shows the schematic of single IIR
lter stage. The enable signal selects the output signal to be just the input signal or the
IIR lter output so as to enable/disable this IIR lter. SRST will reset the lter output
to be zero. The signal Lambda controls the coecient of this IIR lter 2 .
6567
&.5
6567
&.5
,,5ILOWHUEDQNV
6567
&.5
3URSRUWLRQDO
/RRSJDLQ
17:7
)LOWHUBRXWSXW
7UDFNLQJEDQN
.GFR
1RUPDOL]DWLRQ

781(B7)

,QWHJUDO
/RRSJDLQ

6(4B7
5KR
Figure 2-26: High-level schematic of the GT block.
The expression for y[k] when the IIR lter is enabled is

y[k] = (1 2 )y[k 1] + 2 x[k]
September 18, 2011
(2-9)
47
Figure 2-27: Schematic of a single IIR lter.
Similar to the analysis of Equation 3-10, we know that this lter has a corner frequency of
1
f .
2 2 R
In Fig 2-26, the 4-bit LSD Enable IIR signal provides the enable signal to every IIR lter
separately. The 2-bit Lambda0 signal is to choose of the rst IIR lter from 0, 1, 2 and
3. The same function applies to Lambda1, Lambda2 and Lambda3. They together provide
a great exibility for the whole IIR lter bank.
The proportional coecient for tracking bank (T ) is also realized as an arithmetic shift.
T is of high importance to the phase noise of the ADPLL system. An intuitive observation
is that when T is smaller, the eect of noise due to TDC and reference clock is reduced
while the loop has weaker capability to correct for DCO frequency drift. Vice versa. T
has to be chosen carefully for the tradeo and some exibility is desired. In this design T
can range from 23 to 26 as controlled by 2-bit G word signal.
We asserted above that larger T means stronger capability to correct for DCO frequency
drift. It also means larger T can shorten the frequency settling time when ADPLL has
entered the tracking mode. So the idea is to rst adopt a larger T (say, 23 or 24 ) and
when the frequency is pretty close to desired frequency, T will be changed to smaller value
(say, 25 or 26 ). Therefore we can minimize the settling time while still maintain a good
phase noise performance, which is not so easy to be done in traditional charge-pump PLL.
However, during the transition of T value, the control word for tracking bank shall not
show an abrupt change, i.e. a smooth curve of NTW T is desired. To achieve that, a
gear-shifting circuit which is similar to the IIR lter is used. Its schematic is shown in
Fig 2-28. Here we have a xed arithmetic shift of 3 bits and an extra arithmetic shift
which is controlled by 2-bit G word signal. In total we can shift the input by 3 to 6 bits,
corresponding to 23 to 26 as we mentioned before. Notice that the main dierence of the
gear-shifting circuit with the IIR lter is that the feedback ip op is triggered by GS event
September 18, 2011
48
signal rather than CKR signal. Thus the feedback value only changes when GS event goes
from low to high. In the sequencer we can synchronize the GS event rising edge with the
change of G word. Then a smooth transition of T is achieved.
In the design practice, since its easier for the tool to just use one clock domain for synthesis,
one can use CKR as clock of ip op. A detection logic for GS event rising edge is to choose
either to use the new ip op output or the previous GS FB signal (in Fig 2-28) as the
updated value of GS FB. As this change only needs some simple AND, NOT and MUX
logic, the schematic is not drawn here.
Figure 2-28: Schematic of proportional gain path for tracking bank.
Besides the proportional path, a path featuring integral loop gain is added in parallel. This
path has an accumulator in it. When the path is enabled by the rising edge of SEQ T2
signal, the ADPLL becomes a type-II PLL. We will see in Section 3-4 that type-II PLL
can suppress in band phase noise contributed by up-converted icker noise of DCO.
For a type-II PLL, the average value of PHE will be forced to zero for a xed output
frequency. However, before this integral path is activated, for the type-I PLL, in most
situations, PHE has a non-zero average value. Thus it takes some time for PHE to settle
to around zero at transition transient from type-I to type-II. To avoid this settling transient,
before type transition starts, PHE is sampled as a residue for the integration. We call this
residue [k0 ]. Then after integral path is turned on, the accumulated value is PHE[k][k0 ]
rather than PHE[k].
Thus we need a residue latching circuit and an accumulator circuit for integration. The
schematic is exactly the same as Fig 4.44 in [14] and is not redrawn here. For the integral
loop gain , similar to proportional gain T , its implemented as an arithmetic shift to
simplify the logic. The shift bit can range from 8 to 15, meaning that can change from
28 to 215 . A xed shift of 8 bits is rstly done. Then 3-bit Rho signal can select the
extra shift bits(from 0 to 7). Therefore the value of can also be variable. However,
compared with T , will not be changed during the procedure of frequency settling and
no gear-shifting circuit is needed.
September 18, 2011
49
The proportional path and integral path output are summed together as NTW T and that
dco,T is the
is multiplied with K fR for DCO gain normalization of tracking bank, where K
dco,T
estimation of the frequency resolution for CKV in tracking bank. The multiplier output,
TUNE T, is split into the integer part and the fractional part for tracking bank. TUNE TI,
the integer part, goes to the OTI block in Fig 2-22. OTI resembles OP and OA block with
a few dierences. Since the tracking mode is the main working mode of ADPLL, there is no
ZPR generation circuit in OTI block. The 6-bit binary control signal is decoded to 64-bit
thermometer code to control DCO TB, as shown in Fig 2-4. TUNE TF, the fractional
part, is rst converted to an unsigned number DCO TF and then fed to the 1st order
modulator in Fig 2-5. Notice that since both TUNE TI and TUNE TF are converted to
the unsigned value and stored for ADPLL test, we can merge OTI and OTF together as
OT block and leave out high-speed modulator for DCO TFB as an individual block in the
implementation stage, as shown in Fig 5-2.
September 18, 2011
50
September 18, 2011
Chapter 3
Modeling, Simulation and Analysis of
ADPLL System
In Chapter 2 ADPLL system and the building blocks are presented to the detail. To assess
the performance of the whole system, every block needs to be modeled or described at
system level. Then a system-level simulation shall be done and data will be processed
and analyzed. In this chapter, the modeling and simulation methodology of ADPLL is
illustrated. Then we come up with model of critical blocks like DCO and TDC. After that,
we go back to system analysis with insight from s-domain analysis and processed data from
time-domain simulation.
3-1
Modeling and Simulation of ADPLL
Simulation of the whole PLL system is always a very challenging work. Its mostly due
to the dynamic range of frequency in PLL. The oscillator is working at the frequency of
several GHz or tens of GHz while the frequency accuracy requirement for the system is tens
or hundreds of Hz. To get the response of the oscillator with enough accuracy, simulator
like SPICE or SPECTRE will execute the simulation in step of picosecond (ps) or even
femtosecond (fs). On the other hand, to characterize an accurate and smooth in-band phase
noise the simulation needs to take several milliseond (ms). If we are to include the start-up
transient for the LC tank as well as the decoupling network, and include parasitic from
post-layout extraction, chances are the circuit will not converge and the server simulates
for one month and then crashes.
So its indispensable to have a sucient abstraction for the detailed circuit and verify the
whole PLL system with these abstractions. In the traditional charge-pump PLL design, this
is called the behavioral level simulation. Various tools like Matlab, Verilog-A[18] or other
September 18, 2011
52
Modeling, Simulation and Analysis of ADPLL System
tools such as CppSim[19] are deployed for this job. In the ADPLL design the situation is
dierent. As all essential blocks have digital interfaces, the hardware description language
(HDL) can be used for the modeling of the whole loop for its great power of digital logic
description. HDL with real-value modeling capability (allowing oating-point signals) is
preferred. At the high abstraction level, using real-value signals is much more convenient
and exible than using xed-point signals in simulation. Previous ADPLL designs have
adopted VHDL as the simulation language[14]. However, ADPLL of this project will be the
IP of Catena so it is nice to choose a language compatible with current design ow in the
company. Since the company has chosen Verilog for digital logic description, here VerilogAMS is chosen. Verilog-AMS is well compatible with Verilog and provides the real-value
signal as type wreal. Thus we can have very abstracted behavioral level real-valued model
as well as almost near-synthesizable, xed-point model in Verilog-AMS description. Whats
more, the Verilog-AMS/Verilog co-simulation in Cadence is seamless. Thus when NoB for
signals has been decided, Verilog-AMS code of digital logic can be transferred to Verilog
without special eort. Actually, the simulator of Verilog-AMS in Cadence can invoke
SPECTRE for analog circuit simulation at the transistor level. However, to achieve that,
analog-to-digital and digital-to-analog interface have to be set carefully or even specically
modeled. Also the simulation speed is slowed down. So, in the ADPLL system simulation
within the scope of this thesis, analog blocks are simulated in the model of Verilog-AMS
rather than SPECTRE.
Note that besides Verilog-AMS, Matlab is used at the very primary stage of system modeling in this project, as scripting in Matlab is easier. It can act as a fast way to verify the
eectiveness of proposed algorithms. We will determine and verify the basic structure in
Matlab and then shift to Verilog-AMS in Cadence, which is superior in digital logic description capability and closer to nal implementation (can handle xed-point calculation
and can be easily synthesized).
The Verilog-AMS model in this project has 4 abstraction levels, like VHDL abstraction
level in [14]. Level 1 code could use both wreal and integer type signals. The description
is quite similar to the code in Matlab yet the system has much more clear division of
each block due to the module based nature of Verilog-AMS. Level 2 code has xed-point
interface. The eective number of bits for signals in the simulation of level 1 is taken into
consideration so that the truncation and overow (rounding in the digital domain) are
to be avoided in the normal case. Level 3 will be the fully synthesizable RTL code. As
nowadays synthesis tools are very powerful, sometimes level 2 code can be just used for
synthesis, i.e. level 2 and level 3 are just the same code. When level 3 code is done, timing
constraint is specied and standard cell library is ready, synthesis tool will automatically1
deliver a gate-level netlist with back annotated timing information. Thus a simulation with
much more precise timing can be done. From this we can know that modeling of analog
parts will just stop at level 2.
Despite the benets oered by Verilog-AMS, it has some disadvantages, most of them
1
With instructions, of course.
September 18, 2011
3-2 DCO Modeling
53
Table 3-1: Verilog-AMS abstraction levels.
Level1
Level2
Level3
Level4
Model with wreal and integer type signals. Behavioral model.

Modules with xed-point interfaces. Truncation and rounding are considered.
Fully synthesizable RTL code.
Gate-level netlist with back annotated timing information.
from the old Verilog standard. There are two disadvantages that are worth notice in our
modeling:
1. The word length of integer signal in Verilog-AMS is 32 bits. It may not be sucient
to cover the dynamic range of the frequency in PLL, say the ratio of the highest
frequency in PLL model (usually pertained to signals in oscillator) and the lowest
frequency in PLL model(can be frequency accuracy for PLL), or some internal signals
in the digital logic. In that case, even for level 1 code, a xed-point representation
rather than integer number is needed.
2. The minimum time accuracy of Verilog-AMS simulation is 1 fs. One may argue that
1 fs is a very small time interval even for oscillator. However, notice that
t
f
=
f
t
(3-1)
One can easily gure out that for 1 GHz signal, a deviation of 1 fs could lead to
frequency deviation of 1 kHz. So this accuracy needs to be taken care of in the
model.
Besides that, Verilog-AMS is not compatible with the up-to-date SystemVerilog language.
Although there is an eort to drive the merge of SystemVerilog and Verilog-AMS, a.k.a.
SystemVerilogAMS[20], it does take some time to get the support from main EDA vendors.
Table 3-2 is a list of pros and cons for Verilog-AMS language as used in modeling.
This thesis will not explain Verilog-AMS in detail. The modeling methodology for every
block shown below can, in principle, be applicable to other languages. Still, there are some
practical issues specically with Verilog-AMS in this project and they will be discussed
within certain context.
3-2
3-2-1
DCO Modeling
Time-Domain Modeling of DCO Phase Noise
We have made a default assertion that the phase noise of DCO is self-contained in the
rising edge transition timing information of its output signals. Here we will show the DCO
September 18, 2011
54

Table 3-2: Pros and Cons of Verilog-AMS.
Pros
Fully compatible with Verilog
Support mixed-signal simulation
with SPICE/SPECTRE as analog
simulator
Support co-simulation with VHDL,
VHDL-AMS and etc.
Event driven simulation mechanism
to avoid unnecessary oversampling
Digital design ow friendly
Cons
Incompatible with SystemVerilog
yet
Need to dene analog/digital interface carefully
Limited timing accuracy to 1 fs
phase noise can be modeled in time domain and with that, we will go to a complete DCO
model in Verilog-AMS given later.
Measurements have shown the continuous phase-noise spectra of oscillators can be approximated by[5]:
h4 h3 h2 h1
W (f ) 4 + 3 + 2 +
rad2 /Hz
(3-2)
+ h0
f
f
f
f
where f is the frequency oset from the carrier and hn (n=0,1,2,3,4) are certain coecients.
From this equation, a complete asymptote of the phase noise shall contain ve segment
straight-lines with slope of -40 dB/decade, -30 dB/decade, -20 dB/decade, -10 dB/decade
and 0 dB/decade in the log-log scale plot2 . Yet for the oscillator in PLL, usually the terms
of fh33 , fh22 and h0 are of concern. From the classical Leeson model[21], the rst term is
due to up-conversion of the icker noise inside the oscillator and the second term is due to
up-conversion of the white noise inside the oscillator. A simple illustration is that the noise
will add to the time-domain uncertainty of the zero-crossing point (for dierential output
as in Fig 2-4, that would be the time V (outp) = V (outn)). However, as the oscillator
have no memory about what happened in the previous cycle, these uncertainties would
add up, which is just an integrator eect and adds to an additional 1/s transfer function
in S domain3 . This eect as reected in the power spectrum is an extra -20 dB/dec slope.
The 0 dB/dec phase noise is mostly due to the contribution of dividers/buers. In contrast with oscillator, dividers/buers receive the input signal with clear timestamp of edge
transition. Although the noise also causes time uncertainty for zero-crossing points, this
doesnt accumulate. The close-in icker noise is usually buried in the up-conversion phase
2
We will use dB/dec in short of dB/decade. Notice another terminology of dB/octave, or in short
dB/oct, is also widely used. The translation between the two slopes is very simple: -10 dB/dec = -3
dB/oct.
3
One may argue that the integrator eect is kind of a memory. However, that means oscillator only has
a memory about the edge of the last cycle yet no memory about what exactly happened in the last cycle.
September 18, 2011
3-2 DCO Modeling
55
noise caused by the oscillator and only the white noise shows up at oset frequency thats
far enough.
Thus the at 0 dB/decade phase noise is modeled as non-accumulative jitter, while the -20
dB/decade phase noise and -30 dB/decade are modeled as accumulative wander[14].
For the non-accumulative jitter of 0 dB/dec, assuming the input signal are ideal oscillating
signal with kth positive edge exactly at kT0 , where T0 is the ideal oscillation cycle time.
Then the timestamp of divider/buer output is:
tj [k] = kT0 + tj [k]
(3-3)
tj [k] is a zero-mean random value with Gaussian distribution and the standard deviation
is:
T0
Lf0
(3-4)
tj =
2
where f0 is the ideal oscillation frequency and L is the phase noise as dened in subsection
1-2-3.
It shall be noted here for the 0 dB/dec phase noise, a.k.a. the noise oor, the magnitude
of L depends on the sampling rate. When only the rising edge information is considered,
as in our case, the noise is sampled at a rate of f0 and folded within the frequency range
of f20 f20 . Nevertheless, in the spectrum analyzer, both the rising edges and falling
edges are captured. Then the noise is folded to the frequency range of f0 f0 . Thus
the magnitude of the noise oor shall be 3 dB lower than the case with only the rising
edge information. The spectre simulator may also sample at both the rising edges and the
falling edges, which needs to be veried. Therefore one needs to be aware of whether to
add (or minus) 3 dB for the simulation result of spectre and the measurement result of the
spectrum analyzer when executing the simulation.
Say the phase noise for the 0 dB/dec segment is L = 157 dBc/Hz, the desired frequency
is f0 = 3.65 GHz, and the period is T0 = 274 ps, tj would be about 37 fs.
As tj [k] can be either positive or negative, the edge transition of the divider/buer
output may happen before the edge event of the input signal, which causes a diculty
for the modeling in event-driven simulator like VHDL/Verilog-AMS. Thus in this design
tj [k] is precalculated and we merge that jitter into the period information as:
p[k] = T0 + tj [k] tj [k 1]
(3-5)
where p[k] is the time of period for the kth cycle. We will illustrate how to build up the
DCO model using period information later.
For the accumulative wander corresponding to the -20 dB/dec segment in the phase noise
prole, from previous discussion on its origin, one can infer that it shall be modeled as:
p[k] = T0 + tww [k]
(3-6)
September 18, 2011
56
tww [k] is also a zero-mean random value with Gaussian distribution and its standard
deviation is:
f
tww =
T0 L(f )
(3-7)
f0

Since the standard deviation is proportional to f L(f ), as long as the -20 dB/dec
slope holds, it doesnt matter which frequency oset (f ) you exactly choose in the phase
noise prole for the calculation.
Say at the oset frequency f = 10 MHz, we have L(f ) = 140 dBc/Hz, and the
desired frequency f0 = 3.65 GHz, the period T0 = 274 ps, the standard deviation tww is
about 4.54 fs.
For the accumulative wander corresponding to the -30 dB/dec segment in the phase noise
prole, we have a similar equation
p[k] = T0 + twf [k]
(3-8)
the dierence from the modeling of up-converted white noise is that twf [k] is caused by
the 1/f noise rather than the white noise.
The construction of the 1/f noise is achieved by passing a white noise through several rst-order low-pass lters. As shown in Fig 3-1, when the envelope formed by the
DC gain and the corner frequency of these lters, say the line composed by the points
(A1 , fc,1 ), (A2 , fc,2 ), (A3 , fc,3 ), . . . , (An , fc,n ) has a slope of -10 dB/dec, the composite lter
will approaches a response with -10 dB/dec.
f
When the ratio of corner frequency for the adjacent lter r = c,k+1
is 10, which is usually
fc,k
good enough to approximate this -10 dB/dec prole, the ratio of the DC gain would be
A=
Ak
= 1010dB/20
Ak+1
(3-9)
Then the kth lter is modeled as a rst-order IIR lter with the equation
yk [i] = (1 ak )yk [i 1] + ak A(k1) x[i]
(3-10)
This translated to z-domain, we have the transfer function

H(z) =
ak A(k1) z
z (1 ak )
(3-11)
The DC gain can be known as Ak = A(k1) . For the corner frequency, as z = ej2f /fs ,
where fs is the sampling frequency, when f
fs , we have z 1 + j2f /fs . Then the
denominator of equation 3-11 will be ak + j2f /fs . For corner frequency of fc,k , we have
ak = 2
September 18, 2011
fc,k
fs
(3-12)
3-2 DCO Modeling
57
$PSOLWXGH
$
$
+M

+M
$
$
G%GHF
+M
G%GHF
$Q
+QM
U
IF IF
IF
IFQ
)UHTXHQF\
Figure 3-1: Composition of icker noise using multiple low-pass lters (Log-Log Scale).
September 18, 2011
58
In this design, ve lters with respective corner frequency of 100 Hz, 1 kHz, 10 kHz, 100
kHz and 1 MHz are chosen. From DCO simulation results, the up-converted icker noise
is overshadowed by the up-converted white noise when the oset frequency is beyond 1
MHz. Thus the highest corner frequency is just 1 MHz. All the lters are sampled at the
same rate, which is 1/400 of DCO core oscillating frequency. Say if the DCO core works
at 14 GHz, the sampling rate is 35 MHz. Then the combined output will be oversampled
by the DCO core clock to avoid aliasing. This modeling procedure is illustrated in Fig 3-2.
The anti-aliasing lter in the gure can be done with oversampling, or more specic, the
zero-order hold interpolation.

V

V

V

Figure 3-2: Construction of icker noise with single sampling clock and oversampling.
Still we need to know the standard deviation of the input white noise x[j] for the parallel
lter banks. In [14], an estimated value has been proposed as

fc,1
T0 2L(c,1 )
(3-13)
T,1/f =
f0
and some coecient (5.5 dB) to compensate for the correlation is given. However this
value is quite mysterious and cannot be used in our modeling. So a more straight forward
method is applied here to obtain the desired compensation coecient.
From the phase noise prole, one can easily observe the region where the up-converted
icker noise dominates and the region where the up-converted white noise dominates. The
September 18, 2011
3-2 DCO Modeling
59
extrapolation of the two segments intersect at a certain frequency fcorner . If we plot the
power spectrum of tww [k] and twf [k], they shall also just intersect exactly at this frequency. As the standard deviation of tww [k] (tww ) can be calculated from Equation
3-7, it is easy to construct a white noise with this standard deviation in Matlab. With
another uncorrelated white noise as x[j] in Fig 3-2, a icker noise can also be generated.
Then we can adjust the standard deviation of x[j] (x[j] ) so that its power spectrum intersect the power spectrum of tww [k] at fcorner . From that, we just x the compensation
coecient needed. This method can achieve the accuracy of 1 dB, mainly constrained by
the simulation time and the memory needed.
Fig 3-3 illustrates how the coecient is estimated. Assume that the phase noise at 100
Hz frequency oset (the -30 dB/dec segment) is -12.3 dBc/Hz while at 10 MHz frequency
oset (the -20 dB/dec segment) is -140.9 dBc/Hz, the schematic says the corner is about
79 kHz. The extropolation of -30 dB/dec segment and -20 dB/dec at this corner frequency
is respectively -99.2 dBc/Hz and -98.9 dBc/Hz. These two values are rather close, which
means the compensation coecient is estimated very well. The icker noise prole also
reveals the sinc ltering eect due to the oversampling operation.
150
white noise corresponds to upconverted white noise in phase noise
flicker noise corresponds to upconverted flicker noise in phase noise
X: 1.111e+04
Y: 165.4
X: 1.012e+05
Y: 174.5
X: 7.901e+04
Y: 173.1
X: 1.01e+06
Y: 185.7
200
250
300 3
10
10
10
10
10
10
10
10
10
Figure 3-3: A schematic illustrating the estimation of icker noise compensation coecient.
September 18, 2011
60
3-2-2
The Eect of Ideal Divider
Quite often the phase noise of signals are not measured or simulated directly but done
after the signal has been fed to some dividers. On one hand, as seen in Fig 2-2, dividers
are usually deployed with the oscillator. On the other hand, in the measurement its easier
to analyze the phase noise of the divided signal as the frequency is lower and in simulation
we need to consider the contribution of dividers to the phase noise of the whole block as
well.
While in the modeling of DCO, it is convenient to have ideal dividers/buers with an
oscillator which contributes all the noise. We have seen that the period information of
oscillator can model the 0 dB/dec, -20 dB/dec and -30 dB/dec phase noise. The modeling
of ideal dividers and buers are pretty trivial in the event-driven simulator. The required
values (tj ,tww and x[j] ) for the divided clock can be easily obtained via inspecting the
measurement or simulation output. Thus the job is just to infer the corresponding values
when we refer these noises to the oscillator core.
For the signals before and after ideal divider as in Fig 3-4, the non-accumulative jitter due
to tj and the accumulative wander due to tww are treated in [14] and we have
tj,0 = tj,1
(3-14)
1
tww,0 = tww,1
N
(3-15)
However, the accumulative wander due to twf is not mentioned in [14]. A good estimation
is that
1
x[j],0 = x[j],1
(3-16)
N
That is because:
1. Its obvious from Equation 3-11 that tww [k] and tww [k + 1] has a very strong
correlation. Even without oversampling, some trials in Matlab show that Equation
3-16 is a good estimation.
2. Actually the icker noise is constructed via the zero-order hold oversampling. Say the
oversampling rate is 400, then after the ideal divide-by-4 divider, the only dierence
is that the oversampling rate has dropped to 100. Notice the bins in the zero-order
hold window share the same value of tww so the accumulative wander of four cycles
is just four times the wander value in one cycle. Thus Equation 3-16 holds.
September 18, 2011
3-2 DCO Modeling
61
I
,GHDO'LYLGHU
'LYLGH1
I
Figure 3-4: Oscillator with an ideal divider.
3-2-3
Essential Verilog-AMS Code for the DCO Modeling
The code of the DCO block is given in Section

A-1. The instantaneous period is calculated
at line 266. Ideally the period shall be 2 LC. The inductance value DCO L is xed
and the instantaneous capacitance value DCO Cinst are controlled by tuning words as
modeled at line 262. tpp and pretpp model the non-accumulative jitter as tj . tdev1
models accumulative wander corresponding to -20 dB/dec segment as tww and tdev2
models accumulative wander corresponding to -30 dB/dec segment as twf . The upconverted icker noise twf is modeled from line 231 to line 248 exactly as we discuss
before.
At line 268 and line 282, for every half cycle the dierential output DCOcorep and DCOcoren of the DCO core toggle between high and low, i.e. a representation of DCO core
oscillation in the digital domain The divide-by-2 division is modeled from line 295 to line
302. The divide-by-3 division is modeled in line 276 to 280 and line 290 to 294. The
divide-by-4 quadrature output Div4 Ip, Div4 Qp, Div4 In and Div4 Qn are assigned as
the feedback CKV to phase rotator input from line 153 to 156. Line 222-228 model the
enable signals for dividers and buers.
We shall notice that the ideal period 2 LC at line 262 has an innite resolution. However,
as we mentioned in Section 3-1, Verilog-AMS has a limited time resolution of 1 fs. So the
delivered frequency in the DCO model may deviate from the frequency in the real situation
by several kHz. This results in a fake violation of the frequency resolution specication
of 25 Hz in Table 1-1. The solution here is to model the jitter and wanders of DCO with
the resolution of 1 attosecond (as). The jitter and wanders are to perturb the ideal period
with the unit of 1 as. Then the resolution of the average delivered frequency in the model
is not limited by the 1 fs unit from Verilog-AMS.
Note such solution has some premises:
1. The rms value of the jitter and the wanders shall be sucient larger than 1 fs. If the
oscillator is too clean, the time domain noise from the jitter and the wanders are not
enough to perturb the ideal period. Then the frequency resolution of the delivered
frequency is still not well characterized. Here in our design the jitter from the noise
oor is 37 fs, which is far enough for this purpose.
2. The resolution of 1 as can still leads to a fake frequency deviation of a few Hz. So
September 18, 2011
62

if the requirement of the frequency resolution is even ner than that, we shall assign
higher resolution to the jitter and the wanders.
3-3
TDC Modeling
To describe TDC as in the whole system, we need to rst determine what factors in TDC
performance are important and cannot be ignored.
Lets rst consider the time-domain noise of TDC[15]. Just similar to ADC, which has the
quantization noise in the voltage domain, TDC has quantization noise in the time domain.
The noise power is
T2
q2 = inv
(3-17)
12
Besides that, we have the jitter due to noise in the inverter and comparator. As stated from
the transistor-level simulation of TDC[15], the rms jitter is no more than 500 fs. Compared
to min(Tinv ) = 10.5ps, the jitter adds less than 0.25 dB to the whole time-domain noise. So
such jitter can be ignored in the modeling and we only need to care about the quantization
eect of TDC.
Then the mismatch between the delay of every cell is considered, which is mainly caused by
the layout imperfections. That leads to the non-linearity of TDC transfer curve, just like
INL and DNL in ADC. This non-linearity causes the spurs in ADPLL output spectrum,
which will be shown in Chapter 4. Therefore its important to model the TDC non-linearity
in the model.
These considerations result in the Verilog-AMS code for the model of TDC presented in
section A-2. The TDC is decomposed into 40 cells. As in listing A.2, the cell is just an
inverter with a specied delay and an ideal ip op. In listing A.3, 40 cells are connected.
The delays cells are normalized to the average TDC delay and the normalized ratios are
shown in the lines with localparam. Then by modifying the value of ratio in those lines,
the mismatch can mimic the transistor-level simulation result.
3-4
3-4-1
From System Specications to ADPLL Implementation

Noise and Error Sources In ADPLL
As we look at the whole ADPLL system, the digital logic will not give rise to any noise or
error as long as the estimation of Ktdc and Kdco is suciently accurate. There are mainly
four sources for the phase noise of ADPLL:
September 18, 2011
3-4 From System Specications to ADPLL Implementation
63
1. The uncorrected DCO phase noise. As we will see in s-domain analysis, ADPLL
can correct for the phase noise and the frequency drift of DCO that is within the
bandwidth. For the phase noise at rather far-away frequency oset region, ADPLL
cannot track and correct for that. Therefore it is still inherited in ADPLL output.
For the description of the DCO phase noise, three parameters are needed, as discussed
in section 2-2: the phase noise of one frequency oset located in the -30 dB/dec segment, the phase noise of another frequency oset located in the -20 dB/dec segment
and the value of the white noise. In s-domain, the three segments are extrapolated
and added together as an estimation of DCO phase noise prole. Say we have the
DCO phase noise simulation result with -72.3 dBc/Hz at 10 kHz oset, -98.9 dBc/Hz
at 100 kHz oset, -140.9 dBc/Hz at 10 MHz oset and -157 dBc/Hz for the noise
oor. In the modeling we assume that the DCO phase noise prole is fully in the -30
dB/dec segment for 10 kHz oset and fully in the -20 dB/dec segment for 10 MHz
oset. Then the extrapolated s-domain result for the DCO phase noise is shown in
Fig 3-5. One can see that it diers from the simulation result at 10 kHz oset and 100
kHz oset for less than 1dB. Thus the assumption is not 100% true yet the deviation
is quite small and can be ignored.
40
60
X: 1.001e+004
Y: 71.75
PN[dBc/Hz]
80
X: 1.002e+005
Y: 98.55
100
120
140
X: 1e+007
Y: 140.8
160 3
10
10
10
freq[Hz]
10
10
Figure 3-5: S-domain simulation result for DCO phase noise prole.
2. The inaccuracy of reference clock. In the previous discussion we assume that the
reference clock FREF is a very accurate frequency source and doesnt show the phase
noise itself. The truth is that for commercial products, cost factor has to be taken into
consideration and the reference clock source, e.g. crystal oscillator, will always have a
September 18, 2011
64

not-so-low phase noise. This also adds to the in-band phase noise of ADPLL, with a
multiplicative ratio of FCW since the desired phase is FCW times the reference phase.
The situation will become severe for the high performance application, especially in
the OFDM system, as it requires a very low in-band phase noise.
The dominating component in the phase noise prole of the reference clock is the
white noise. Therefore the phase noise of the reference clock is pretty easy to model
and only needs one parameter RMS jitter. Say we have 1.5 ps RMS jitter and the
reference clock is 35 MHz, then we have -145.1 dBc/Hz noise oor for reference clock.
With FCW equal to 100, a resonable value in our design, the noise oor would be
-105.1 dBc/Hz.
3. The nite resolution of TDC. In Fig 2-9, the compensation of TDC completely
matches the quantization error of incrementor. However, that only happens when
we have innitesimal resolution for TDC. In our design, TDC can have a resolution
of between 10.512.5 ps. Thus the phase error for digital logic is hurt by the TDC
resolution.
If the input for TDC is fully random, then the error from the quantization can
be simplied as a quantization noise, like the case in ADC. The phase resolution
inv
. Then the quantization noise in phase
corresponding to TDC resolution is 2 TTCKV
domain is
(2)2 Tinv 2
2
=
)
(3-18)
(
12 TCKV
The quantization noise is sampled by reference frequency, thus the phase noise contribution due to TDC resolution is4
L=
(2)2 Tinv 2 1
(
)
12 TCKV fR
(3-19)
For our design, the worst case occurs when Tinv is the maximal value (12.5 ps), TCKV
is the minimal value (1/4.05 GHz) and fR is the low bound of reference frequency
range (30 MHz), the TDC phase noise is -95.5 dBc/Hz.
4. The nite frequency resolution of the DCO capacitor bank. The digital loop lter
can, in theory, give an innite-length NTW word as the DCO tuning word. However,
DCO capacitor banks have only limited resolution. Therefore some information of
NTW will be lost as we transfer that to OTW, which may add to the phase noise of
the whole system.
The eect of the nite frequency resolution in the tracking bank is of interest since
ADPLL normally works in the tracking mode. The resolution in the tracking mode
is dened by the modulator as in Fig 2-5. This has been discussed in [14]. The
nite fractional bits of tracking bank and the shaped quantization noise from the
4
For the single-sided spectrum of the phase W we have W =
September 18, 2011
fR /2 .
Since L =
W
2 ,
we have L =
fR .
65
modulator both contribute to the phase noise. The equivalent phase noise at the
DCO output caused by the nite resolution and the modulator is
f 2
1 Kdco 2 1
f 2n
f 2
1 Kdco 2 1
(sinc
) + (
(2 sin
) (sinc
)
( W
)
)
12 2 F f fR
fR
12 f f
f
f
(3-20)
where WF is the fractional bit of the modulator input, f is the working frequency
of the sigma-delta modulator, n is the order of modulator and Kdco is the frequency
resolution of the tracking bank.
L(f ) =
In Equation 3-20, the rst term represents the quantization error due to the nite
fractional bits of tracking bank. The second term represents the noise, i.e. the
noise shaped quantization error due to the modulator. The sinc function in both
terms represents the eect of the zero order hold. In the rst term the value is clocked
by CKR and in the second term the value is clocked by CKVD. So the divisors of
sinc fuction in two terms dier. The sin function in the seond term represents the
noise shaping of the nth order modulation.
IN our design WF =5 and n=1. Kdco and f both vary with the DCO working
frequency. For the frequency oset region we care about, f
f , then the part
of Equation 3-20 at the right hand side of + symbol can by estimated as
2
1 Kdco 2 1
f 2
f 2
1 Kdco 2 1
f 2 2 Kdco
(
)
)
(2 sin
) (sinc
) (
(2
) =
(3-21)
3
12 f f
f
f
12 f f f
3 f
As we know, Kdco is proportional to f 3 , thus the value in the above equation reaches
its largest value for the highest CKV frequency. Table 2-1 says at the highest frequency Kdco = 273kHz
68kHz, f = 4050MHz
500MHz. Then the result of the
4
8
above equation is -159 dBc/Hz, which is fairly enough for our specication.
For the evaluation of the second term of Equation 3-20, as this part shows a -20
dB/dec prole, it would be easier to compare that with the DCO phase noise. We
choose Kdco to be 68 kHz and fR as 30 MHz.The phase noise incurred at the 100 kHz
frequency oset is -119 dBc/Hz and at the 10 kHz frequency oset is -99 dBc/Hz.
Thats much smaller than the contribution from the DCO phase noise.
Fig 3-6 plots the phase noise spectrum due to the frequency resolution of the DCO
tracking bank. The red line corresponds to the rst term in Equation 3-20 and the
blue line represents the second term in Equation 3-20. The green line is the sum
of the two terms. One can see that for the in-band and transition-band frequency
region (f from 10 kHz to 10 MHz), the green line is much lower compared to the
DCO phase noise prole as in Fig 3-5. For the far-out frequency region (f larger
than 10 MHz), the green line is rather lower than the specication of -150 dBc/Hz.
To give a summarization, the frequency resolution of DCO and the 1st modulation will not degrade the performance of ADPLL. In the discussion below, we will
ignore the inuence of this factor.
September 18, 2011
66
50
Phase Noise due to finite bit of the tracking bank tuning word (WF)
Phase Noise due to noise shaping
The Total Phase Noise due to finite frequency resolution of the DCO
X: 1.007e+004
Y: 99.07
Phase Noise Contribution (dBc/Hz)
100
X: 9.987e+004
Y: 119
X: 1.005e+006
Y: 139
X: 1.021e+004
Y: 159.1
150
200
250
300
3
10
10
10
10
10
Frequency Offset f (Hz)
10
10
10
10
Figure 3-6: Phase noise spectrum due to the frequency resolution of the DCO tracking bank.
September 18, 2011
3-4-2
67
S-domain Analysis for the ADPLL System
When we look back at Table 1-1, the frequency range is covered in the analog domain by a
careful frequency planning and in the digital domain by enough integer bits for the FCW
signal. The requirement for the frequency resolution is satised via enough fractional bits
for FCW signal. To gain an insight on how ADPLL can satisfy the phase noise requirement,
the s-domain analysis is performed for a quick estimation.
For the phase noise, we are mostly concerned with the main working status of ADPLL,
which means when ADPLL is in tracking mode and has switched to a type-II PLL. A
s-domain diagram for ADPLL as in Fig 4.45 of [14] is redrawn in Fig 3-7.
3KDVH'HWHFWRU
1RUPDOL]HG'&2
Q'&2
/RRS)LOWHU
1
5>UDG@
(>F\FOHV@
17:
27:
I 5 /6%
.
'&2
V
9>UDG@
. '&2>+]/6%@ u S

I5V
Figure 3-7: S-domain model for the type-II ADPLL.
The DCO features a gain of KDCO 2 as

V = 2f dt = 2(f0 + OT W KDCO )dt
(3-22)
The normalization circuits in the low speed digital logic block are included in Normalized
DCO (nDCO) part in s-domain schematic. The input of normalization circuit is NTW,
which is the sum of the proportional path and the integral path. We estimate KDCO as
DCO and in the normalization circuit, NTW is multiplied with fR .
K
K
DCO
For the integral path, we have the expression of accumulator:

y[k] = y[k 1] + x[k]
(3-23)
1
Y (z)
z 1
=
=
X(z)
1 z 1
z1
(3-24)
In z-domain, we have
September 18, 2011
68
As z 1 + j2f /fR = 1 + s/fR , a simple calculation gives that the s-domain transfer
function for the accumulator is fR /s. 5
For the phase detector, the desired phase information and the feedback information is compared and their dierence is normalized to the CKV cycle. The desired phase information
is FCW times the phase of reference clock R . Here we use N instead of FCW for more
general analysis. The normalization of the phase dierence is just a division by 2.
DCO = KDCO , the open
Assuming we have an accurate enough estimation of KDCO , i.e. K
loop transfer function for ADPLL is
Hol (s) = ( +
fR s + fR2
fR fR
)
=
s s
s2
(3-25)
The closed-loop transfer function of the oscillator phase nosie is

Hcl,v (s) =
s2
s2 + fR s + fR2
(3-26)
The closed-loop transfer function of the reference phase noise is

Hcl,F REF (s) = N
fR s + fR2
s2 + fR s + fR2
(3-27)
The closed-loop transfer function of the TDC phase noise is

Hcl,T DC (s) =
fR s + fR2
s2 + fR s + fR2
(3-28)
From [5, 14], we know that the loop gain K = fR 6 , the natrual frquency n = fR and
the damping factor = 12 . Say the reference frequency is 35 MHz, is 26 and = 215 ,
we have K=546.9 k rad/s=87.0 kHz, n =193.3 k rad/s=30.8 kHz, =1.414.
Here we have the TDC transfer function and the DCO transfer function with fR of 33.8688
MHz. This particular reference frequency is chosen as some commercial available crystal
oscillator just operates at that frequency. The TDC transfer function is plotted in Fig 3-8.
Its almost at when the frequency oset is smaller than loop gain K and then rolls down
in an asymptotes with a slope of -20 dB/dec. The DCO transfer function is plotted in Fig
3-9. Its almost at for the frequency oset higher than 200 kHz. For the close-in oset
frequency, the phase noise rolls up at an asymptotes of 40 dB/dec.
The 40 dB/dec slope of DCO transfer curve with the -30 dB/dec up-converter icker noise
prole in DCO together show a 10 dB/dec slope, which means the up-converted icker
5
An intuitive explanation is that the accumulation rate is fR . Then we have a coecient of fR with
1/s for the integration.
6
In [5] this loop gain concept is mentioned. That can be a more reasonable estimation for the bandwidth
of type II PLL.
September 18, 2011
69
10
amplitude[dB]
0
X: 1.002e+005
Y: 3.327
10
20
30
40
50 3
10
10
freq[Hz]
10
freq[Hz]
10
10
10
10
10
phase[degree]
20
40
60
80
100 3
10
10
Figure 3-8: TDC Transfer function for type II PLL ( = 26 , = 215 ).
noise in DCO is suppressed in the type-II PLL. The contribution of DCO to ADPLL phase
noise in type-II ADPLL is shown in Fig 3-10. One can clearly see the 10 dB/dec rolling
up for the close-in frequency oset region and the -20 dB/dec rolling down for the far-out
frequency oset region, which is just a up-converted white noise prole in the DCO phase
noise.
Then the intrinsic tradeo of all PLL system shows up if phase noise specication is to
be satised. When we have larger and , the bandwidth of PLL is larger and the DCO
phase noise is more suppressed. Thus the design eort for DCO is relaxed. However, at the
same time, the at frequency oset region for TDC and the reference phase noise transfer
curve extends, which indicates that we need less noise contribution from TDC and the
reference to meet the requirement for the in-band phase noise and the integrated phase
noise. This adds to the design diculty of TDC, a FREF slicer and also adds to the cost
of a high-quality reference.
In the design practice, both TDC and DCO take a lot of eort and are power hungry. Its
essential to select optimal coecients for ADPLL so as to ease design pressure for the two
blocks. The traditional rule of thumb is to select the oset frequency where (reference
phase noise+TDC phase noise) and DCO phase noise meet as the PLL bandwidth. Here
we have the worst case TDC phase noise of about -95.5 dBc/Hz, which is very close to
the phase noise requirement of -95 dBc/Hz at 100 kHz frequency oset. Whats more, to
avoid the violation of the integrated phase noise specication, the bandwidth cannot be
September 18, 2011
70
0
X: 1.002e+005
Y: 1.87
amplitude[dB]
10
20
30
40
50
60 3
10
10
freq[Hz]
10
freq[Hz]
10
10
10
10
10
phase[degree]
200
150
100
50
0 3
10
10
Figure 3-9: The DCO Transfer function for the type II PLL ( = 26 , = 215 ).
90
100
PN[dBc/Hz]
110
120
130
140
150 3
10
10
10
freq[Hz]
10
10
Figure 3-10: The DCO contribution to the ADPLL phase noise in the type-II PLL.
September 18, 2011
71
too large. For the worst case of the TDC resolution (slow corner, high temperature and
low voltage), the bandwidth is expected to be less than 100 kHz.
Besides the bandwidth, for type-II PLL, we have another value which is the damping factor
. For our case, a over-damped type-II PLL ( 1) is preferred since it avoids a signicant
gain peaking 7 . For =1.414, as we look at Fig 3-8 and Fig 3-9, there is no peaking for
DCO transfer function and only less than 1dB gain peaking for TDC transfer function.
This is especially useful for our case as we are actually stressed by the phase noise due to
not-so-ne TDC resolution. One can see that at the 100 kHz frequency oset where most
tough spec lies, TDC and DCO transfer function both show a little bit attenuation.
Up till now, we have left the IIR lter bank in ADPLL logic untouched. For an IIR lter
as shown in Fig 2-27, we have its s-domain equation [14]
Hiir (s) =
1 + s/fR
1 + s/(2 fR )
(3-29)
The intuition tells us that the contribution to the far-out phase noise from TDC and
reference clock would be reduced. However, as the DCO far-out phase noise is already
uncorrected, its hard to see the inuence of the lter bank to the DCO phase noise
contribution in the ADPLL system. Here s-domain simulation is done and we get the
TDC transfer function in Fig 3-11 and the DCO transfer function in Fig 3-12. Here rst
and second lter in the bank are turned on. Lambda0 is 1 and Lambda1 is 2.
Comparison between Fig 3-8 and Fig 3-11 shows that IIR lters do add attenuation to
the far-out phase noise due to TDC and the reference clock. Comparison between Fig 3-9
and Fig 3-12 reveals that the DCO transfer function has almost no change, expect a very
small peaking at out-of-band frequency region. Thus the total far-out noise can be reduced
without degrading the close-in phase noise performance. With IIR lters in the IIR lter
bank all turned on and all chosen as 3, the attenuation in TDC transfer function is much
more obvious yet with signicant peaking for DCO transfer function. In our design, with
only Lambda0=1 and Lambda1=2, the extra ltering is already enough.
Through s-domain analysis and simulation, one can have a rough estimation of DCO phase
noise requirement. The specication for DCO design is shown below
Phase noise specication
@10 kHz oset (dBc/Hz)
@100 kHz oset (dBc/Hz)
@10 MHz oset (dBc/Hz)
Low frequency (3300MHz)

-71.7
-98.0
-140.2
High frequency (3800MHz)

-72.3
-98.9
-140.9
Table 3-3: DCO phase noise specication.

7
The over-damped type-II PLL is sometimes called the 1st order type-II PLL as it features minimum
gain peaking and suppression of DCO close-in phase noise.
September 18, 2011
72
amplitude[dB]
20
0
20
40
60 3
10
10
freq[Hz]
10
freq[Hz]
10
10
10
10
10
phase[degree]
50
100
150 3
10
10
Figure 3-11: TDC transfer function for type II PLL with IIR lter bank
September 18, 2011
73
amplitude[dB]
20
0
20
40
60 3
10
10
freq[Hz]
10
freq[Hz]
10
10
10
10
10
phase[degree]
200
150
100
50
0 3
10
10
Figure 3-12: DCO transfer function for type II PLL with IIR lter bank
September 18, 2011
74
Notice in the s-domain simulation the highest frequency is 3.8 GHz rather than 4.05 GHz,
as the frequency range of 3.8-4.05 GHz is only intended for the LB output. Since the HB
outputs and the feedback CKV signal are DCO core outputs divided by 4 while LB outputs
are DCO core outputs divided by 6, the eective division ratio of the LB output compared
to the CKV signal is 1.5. Thus the LB output phase noise is 3.5 dB lower than HB output
phase noise. There is no need to worry about the LB output phase noise as long as the
requirement for the HB output phase noise is met.
The ADPLL phase noise is simulated in s-domain with the value in Table 3-3. The reference
clock is 33.8688 MHz and TDC resolution is the worst case value 12.5 ps. RMS jitter for
reference clock is 1.5 ps. One can see that in Fig 3-13 and Fig 3-14 for the high frequency
and the low frequency the ADPLL phase noise complies with the specication.
Summed Phase Noise of DCO & TDC & FREF; Locked_frequency is 3800.000000MHz; Integrated Phase Noise is 41.2147 dBc
90
X: 1.001e+004
Y: 91.02
X: 1.002e+005
Y: 96.4
100
110
PN[dBc/Hz]
X: 1.002e+006
Y: 116.7
120
130
140
X: 1e+007
Y: 140.5
150
3
10
10
10
freq[Hz]
10
Figure 3-13: S-domain result for ADPLL PN (fv =3800 MHz).
September 18, 2011
10
75
Summed Phase Noise of DCO & TDC & FREF; Locked_frequency is 3300.000000MHz; Integrated Phase Noise is 41.421 dBc
90
X: 1.001e+004
Y: 91
95
X: 1.002e+005
Y: 96.72
100
105
PN[dBc/Hz]
110
X: 1.002e+006
Y: 117
115
120
125
130
135
X: 1e+007
Y: 139.9
140
3
10
10
10
freq[Hz]
10
10
Figure 3-14: S-domain result for ADPLL PN (fv =3300 MHz).
September 18, 2011
76
September 18, 2011
Chapter 4
Advanced Algorithm for ADPLL
4-1
Zero Phase Restart
When ADPLL goes from the PVT mode to the acquisition mode, CTL PLL P becomes
low and the control word for PB is frozen. Thus in the view of the acquisition bank, DCO
has a new central frequency determined together by the original DCO frequency and the
frozen PB control word. Notice that in the PVT mode, the phase error information is
obtained on the condition that NTW A is 0 and NTW T is 0. If MEM DCO A (as in Fig
2-25) and MEM DCO T are zero, then the acquisition bank and the tracking bank are
xed to the middle capacitance value (half capacitance units turned on and half turned
o) in PVT mode. To make sure that the DCO frequency settles smoothly, PHE shall
be reset to zero once ADPLL switches to the acquisition mode. In this way NTW A and
NTW T are set to 0 at the beginning of the acquisition mode. The DCO output frequency
is consistent before and after the mode switchover. Similarly, the PHE value shall also be
reset during the switchover from the acquisition mode to the tracking mode.
We call this reset at the mode switchovers as zero phase restart (ZPR). The implementation
is rather straightforward with the dierentiation-then-accumulation phase detection logic
as in Section 2-5. One only needs to set the Reset||ZPR signal (as in Fig 2-22) high to
reset the accumulator at the mode switchover event. At Fig 2-25, the detection logic of the
mode switchover event is presented. Once CTL PLL P goes from high to low, the previous
registered value is high and inversion of the current value is also high. Then OP ZPR
becomes 1. At the next CKR cycle, OP ZPR will becomes low again as the registered
CTL PLL P becomes low. So a positive pulse of OP ZPR would indicate that ADPLL is
switching from PVT mode to acquisition mode.
Similarly, OA ZPR signal sends a positive pulse when ADPLL is going from acquisition
mode to tracking mode. We feed the system reset signal, OP ZPR and OA ZPR signal
September 18, 2011
78
to an OR gate and its output, Reset||ZPR will reset the PHE signal in the occasions we
expect.
Fig 4-1 shows the transient with the ZPR mechanism. One can easily see the OP ZPR and
OA ZPR pulse at the mode switchover. Then PHE signal is reset to zero. In this gure
the pulse of signal OP ZPR is wider than OA ZPR. That is because in the implementation
the falling edge of CTL PLL P is detected and then OP ZPR pulse is stretched to last for
several CKR cycles. Then the DCO can settles to the frequency indicated by the frozen
word for PVT bank before it starts the acquisition mode. The stability is improved. As in
the switchover from the acquisition mode to the tracking mode we dont see the stability
problem, one CKR cycle is enough for the OA ZPR pulse.
&XUVRU IV
%DVHOLQH
&XUVRU%DVHOLQH IV
7LPH$ IV
IV
IV

H
H
3+(>@
&7/B3//B3
&7/B3//B$
&7/B3//B7
23B=35
2$B=35
'd
H

1
0
0
0
0
Figure 4-1: Transient for zero phase restart.
4-2
PVT Miss Mechanism
In Fig 2-4 there are two tuning words for PVT bank: DCO IN P G and DCO IN P DS,
one connected to the gate of the switch in the capacitor units and one connected to the
drain and source of the switch in the capacitor units via inverters. They are used to prevent
the choking of the DCO core.
The capacitance of the PVT bank plays a signicant part in the total capacitance in DCO.
When drastic capacitor switching events happen in PVT mode, e.g. the toggling event of
September 18, 2011
4-2 PVT Miss Mechanism
79
the tuning word between 0111111 to 1000000, chances are that the oscillator may pause
working for a short time before it resumes the oscillation and settles to the new frequency.
During this transient, the divider will not capture the expected edge. The output of TDC
and incrementor also fails to deliver the correct phase information.
The transistor level simulation reveals that DCO works ne when the switches are turned
o. Yet the choking does happen for the case when switches are turned on. A valid solution
obtained from the transistor level design exploration is to pull up DCO IN P DS signal
before pulling up DCO IN P G. Fig 4-2 gives a digital implementation for this purpose.
One can see that for the rising edge of DCO IN P, DCO IN P G rises one CKR cycle later
while DCO IN P DS has no delay; for the falling edge of DCO IN P, both DCO IN P G
and DCO IN P DS are delayed. Since DCO IN P G is to turn on/o the switch, this
circuit introduces one extra delay to the PVT mode of ADPLL.
'&2B,1B3B*
'&2B,1B3
&.5
'&2B,1B3BQBGHOD\HG
'&2B,1B3B'6
'&2B,1B3BQ
&.5
'&2B,1B3B*
'&2B,1B3B'6
&.5
'&2B,1B3
'&2B,1B3B*
'&2B,1B3BQ
'&2B,1B3BQBGHOD\HG
'&2B,1B3B'6
Figure 4-2: Logic for PVT tuning word generation and the timing diagram.
September 18, 2011
80
Still one can easily see that when DCO IN P keeps toggling between 1 and 0,
DCO IN P DS will be always high, i.e. the drain and source of switches are always biased
at voltage of logic 0 (VSS), which is undesired. We avoid this situation by the mechanism
asserted below:
If DCO IN P changes at one CKR rising edge, it shall keep this value for the next CKR
rising edge. In the other word, once changed, DCO IN P shall be constant for at least two
cycles.
To achieve this, PHE needs to be frozen at some certain time, which means dPHE is
missed at these moments. We call this the PVT miss mechanism. It can be done by
observing TUNE P, the input of OP block. Ideally, TUNE P changes one cycle ahead of
DCO IN P. Thus when the logic detects the change of TUNE P while ADPLL is in the
PVT mode1 , PVT miss signal will be made high to freeze the accumulator output PHE.
Fig 4-3 shows the signal waveforms when PVT miss algorithm is activated. TUNE P will
hold a certain value for at least two cycles before it changes and when it changes the value,
PVT miss signal sends a high pulse to freeze the PHE signal. The timing diagram for
DCO P, DCO P G and DCO P DS is also shown in this gure.
4-3
4-3-1
Spur Suppression Techniques for ADPLL

Spurious Tone Issue of ADPLL
For the previous s-domain analysis in Section 3-4, we have made two assumptions: 1. Every
TDC cells will show the same delay. There is no mismatch between TDC cells, i.e. TDC
has zero DNL and INL. 2. TDC input is fully random. Therefore the quantization error
can be approximated as the quantization noise as in Equation 3-19.
Nevertheless, that is not a true image. For the static ADPLL without any modulation,
when the frequency is locked, the feedback phase increases linearly. TDC input (timing
dierence) will follow a saw tooth trajectory with modulo of one CKV cycle. This gives rise
to the periodicity of PHF and PHE, which causes spurs at the ADPLL output spectrum.
This spur issue has been treated in a universal way in [13], where the quantization eect
is viewed as a subset of TDCs nonlinearity. Its asserted that when FCW=(N+), where
| < 1|, the spur may show up at frequency osets which are integer times ||FREF,
depending on the Fourier coecients of the nonlinearity for the input-output transfer
curve. As the loop lter attenuates the frequency component of PHE that are not within
the PLL bandwidth, the spur performance is worst when ||
1, which is also called
near-integer N case.
1
The detection logic just compares TUNE P of the previous cycle and TUNE P of the current cycle,
i.e. TUNE P[k]==TUNE P[k-1].
September 18, 2011
4-3 Spur Suppression Techniques for ADPLL
81
:DYHIRUP6LP9LVLRQ
3DJHRI
&XUVRU IV
%DVHOLQH
&XUVRU%DVHOLQH IV
/6'B&.5
IV
7LPH$ IV
IV
IV
IV
IV
IV

1
H
H
3+(>@
H

781(B3>@

'&2B3>@

'&2B3B*>@

397BPLVV
'&2B3>@
'&2B3B*>@
'&2B3B'6>@
3ULQWHGRQ)UL$XJ&(67
0
0
0
0
3ULQWHGE\6LP9LVLRQIURP&DGHQFH'HVLJQ6\VWHPV,QF
Figure 4-3: Transient signal waveforms for the PVT miss algorithm.
September 18, 2011
82
However, the time-domain simulation result suggests that its better to consider the TDC
mismatch and the quantization eect separately, as they raise signicant spurs at dierent
frequency osets.
For the TDC mismatch eect, the spurs would mostly reside at the frequency oset of
k*||FREF, where k is from 0 to 5. For the quantization eect, when ||
1, the timing
dierence between FREF and CKV rising edges would drift only a little bit every FREF
cycle. Then the time error indicated by the phase error signal PHE increases in a saw
tooth way with the modulo of one TDC cell delay, as in Fig 4-4. The timing dierence of
the CKV edge and the FREF edge sampled by CKR diers from the previous sampling
||
CKV
N
1
F REF CKV
value by t cycle = | CKV
F REF
| = | N CKV
| = CKV
. Then it takes Tinv||
F REF
cycles of CKR sampling to sweep one TDC cell delay, which corresponds to the period of
inv CKV
Tcycle = T||F
. We could expect spur at harmonics of
REF
fq,spur =
||F REF
Tinv CKV
(4-1)
LQY
UHI
TF\FOH
Figure 4-4: Timing diagram for spurs due to the TDC quantization eect.
For example, when fR = 33.8688 MHz and fv = 3793.3156 MHz, we have N=112.0002953.
The desired frequency deviates only 10 kHz from the integer-N frequency. The TDC
resolution is 12.5ps. The mismatch of TDC is modeled in the system as in Section 3-3.
We shall expect the spurs at frequency osets which are integer times 10 kHz due to the
TDC mismatch and which are integer times
10kHz
= 210.9kHz
12.5ps 3793.3156MHz
(4-2)
due to the TDC quantization eect.

Here in Fig 4-5 the phase noise from the time-domain simulation of CKV with 3793.3156
MHz and FREF with 33.8688 MHz is given. There are two strong spurs located at the 10
kHz oset and the 210 kHz oset, as well as some weaker spurs at harmonics of these two
frequencies, as expected from the analysis above. The transient of PHE signal is given in
September 18, 2011
83
Phase noise of output signal. FREF is 33.8688MHz, Locked frequency is 3793.315600MHz

FCW is 112.00030, rbw is 1.000000e+03
Phase Noise @10kHz offset is 64.26 dBc/Hz, @30kHz offset is 73.89 dBc/Hz, @100kHz offset is 85.04 dBc/Hz, @1MHz offset is 118.70 dBc/Hz
Integrated DSB Phase Noise from 10KHz to 10MHz is 31.82 dBc
60
X: 2.099e+05
Y: 67.7
X: 9877
Y: 64.26
80
100
120
140
160
180 2
10
10
10
10
10
10
10
10
10
10
Figure 4-5: ADPLL spectrum for CKV frequency of 3793.3156 MHz, RBW=1 kHz.
:DYHIRUP6LP9LVLRQ
&XUVRU IV
%DVHOLQH
&XUVRU%DVHOLQH IV
3DJHRI
IV
IV
IV
IV
IV
IV
7LPH$ IV
IV
IV

3+(>@ d5

Figure 4-6: PHE transient for CKV frequency of 3793.3156 MHz.
September 18, 2011
84
Fig 4-6. The signal shows a period of about 5 us, which corresponds to the spurs due to
the TDC quantization eect.
An obvious solution to solve this spur issue is to improve the TDC resolution, which
consumes more power and increases the design complexity[16]. Thus other ways to work
around this issue have been proposed, e.g. calibration or dithering[22, 13, 23]. In our
design, the algorithms of phase rotation and FREF dithering have been adopted for the
spur suppression. These two algorithms are chosen because they are simple yet eective
for our case.
4-3-2
Phase Rotation Algorithm
As shown in Fig 2-2, the feedback clock is generated by the divided-by-2 divider, which
means that the four clock phases are naturally available. As stated in [24], the feedback
frequency can be oseted fR /4 via the phase rotation algorithm.
Fig 4-7 is a timing diagram similar to Fig.5 in [24], where the output of phase rotator will
shift to the next phase every FREF cycle. To exemplify how this phase rotation algorithm
works, we assume FCW=3.0. In Fig 4-7 (a), the phase will rotate in the sequences of
Ip Qp In Qn Ip. When the transitions happen at the time indicated by solid
line, the phase increment seen by the incrementor and TDC, as normalized to CKV cycle,
will be 2.75. This is equivalent to a frequency translation of fR /4. In Fig 4-7 (b), the
phase will rotate in the sequence of Ip Qn In Qp Ip and the phase increment
seen by the incrementor and TDC, as normalized to CKV cycle, will be 3.25. This is
equivalent to the frequency translation of +fR /4.
This algorithm aims at breaking the near-integer N condition for FCW, as for Fig 4-7 (a)
and (b) we shall add -0.25 and 0.25 to the original FCW value. The new FCW value will
not be close to an integer any more.
For Fig 4-7 (a), notice that if the phase transition of Ip Qp happens at the time indicated
by dashed lines, either leads or lags the solid line, the incrementor would see one extra
rising edge. Then the phase error would deviates from the expected value by 1. One can
either ignore this error value in the integer part of PHE or correct for this error, yet with
the same disadvantage we have stated for the PHE spike due to the mismatch between
the TDC path and the incrementor path as in Section 2-3. In another way, [24] proposes
to detect the falling edge of CKR as to start up phase rotation. As for the transition of
Ip Qp, when we gate the rotation event start-up with the rising edge of Qp, the timing
for transition is set to be in the regime of solid line and the possibility of an extra rising
edge is eliminated. However, for this method, the next phase is involved and some extra
logic is needed.
In contrast to the potential problem in Fig 4-7 (a), Fig 4-7 (b) will not give rise to this extra
edge issue and is adopted in our design. The phase rotation is performed by a multiplexer
in Fig 2-7, which is controlled via a modulo-4 counter clocked at CKR falling edges. To
September 18, 2011
85
,S
4S
,Q
4Q
D
,S
4S
,Q
4Q
E
Figure 4-7: Timing diagram for phase rotation.
avoid the possible glitch in the modulo-4 counter logic, the Gray code counter shall be
used. At the same time we shall compensate for the equivalent frequency translation in
the phase detection logic, which is achieved in PR module shown in Fig 2-22. When PR
signal is high, dPHR is the xed-point value of ( ffRv + 0.25) rather than just FCW.
Fig 4-8 shows the ADPLL spectrum for CKV=3793.3156 MHz but now with the phase
rotation activated. RBW is 1 kHz. Compared to Fig 4-5, the spurs have been signicantly
suppressed, especially for these located at the frequency oset of 10 kHz, 20 kHz and
30 kHz. However, we still see signicant spurs at around 40 kHz and 200 kHz. This
phenomenon is not hard to explain. For TDC, it sees an equivalent FCW as (N + 0.25 + ).
Due to subsampling, the 4th order harmonic component of (0.25+) falls in the bandwidth.
That gives rise to the spur at 40 kHz. For the spur at 200 kHz, it is still due to the TDC
quantization. However, the frequency location is subtle. In the spectrum with RBW=10
kHz as Fig 4-9, spurs are still noticeable, which means the improvement brought by only
phase rotation is not that satisfactory.
4-3-3
FREF Dithering Algorithm
In [23, 25, 26] dithering has been used to reduce the spur in ill-behaved near-integer N
situation. Its purpose is to introduce randomized delay of up to several Tinv to FREF so
as to avoid the slow slope in Fig 4-4. In that way the spurs due to TDC quantization are
suppressed.
September 18, 2011
86

FCW is 112.00030, rbw is 1.000000e+03
60
X: 4.012e+04
Y: 75.8
X: 2e+05
Y: 78.61
80
100
120
140
160
180
200 2
10
10
10
10
10
10
10
10
10
10
Figure 4-8: ADPLL spectrum for CKV frequency of 3793.3156 MHz with phase rotation
algorithm activated, RBW=1 kHz.
September 18, 2011
87

FCW is 112.00030, rbw is 1.000000e+04
80
90
100
110
120
130
140
150
160
170 3
10
10
10
10
10
10
10
10
10
Figure 4-9: ADPLL spectrum for CKV frequency of 3793.3156 MHz with phase rotation
September 18, 2011
88
In our design, the randomized dithering technique proposed in [25] has been adopted.
The pseudo-random signal is used as the input of a 3rd order Sigma-Delta modulator shown
in Fig 4-10, which just consists of three adders and several ip ops clocked at the negative
edge of FREF. The sum value of d1d7, ranging from 0 to 7, indicates how long the delay
shall be added to the FREF. The delay resolution in the system simulation is assumed to
be 9ps to guarantee the delay range spans several TDC cell delay. For example, if d1d7
is 0011001, their sum is 3 and the delay for the FREF is 27 ps.
&
=
&
=
&
=
&
G
180B,1
G

&
5
=
&
=
&
=
&
G
G
G
G
5
&

&

&

&
G
5

)5()
)5()
)5()
Figure 4-10: Schematic of modulator for FREF dithering.
The input of the modulator, NUM IN, is randomized to avoid possible tones in the modulator output2 . Similar to the analysis in Section 3-4-1, this random input adds to the phase
noise of TDC and thus will degrade the close-in phase noise of ADPLL. To reduce this
eect, the distribution range of NUM IN is limited here. Rather than having a uniform
distribution in [0,63] the 6 bit NUM IN is uniformly distributed in [13,44]. Through this
the noise contribution from FREF dithering is lowered down by 6 dB and the modulator
output is still free from tones.
In the schematic of Fig 2-7 the FREF dithering is absorbed into the slicer. In real silicon,
that can be done either in a FREF slicer as in [26] or in an individual DTC in [23], which is
not implemented yet. More accurate model extracted from the transistor-level simulation
shall be substituted into the system simulation to determine the performance.
The spectrum of ADPLL for CKV=3793.3156 MHz with FREF dithering activated only
is shown in Fig 4-11. RBW is 1 kHz. The spurs due to the TDC quantization are well
suppressed yet we still can see spurs at 10 kHz / 20 kHz / 30 kHz / 40 kHz, which is what
we expect as the FREF dithering does not help much on the mismatch of the whole TDC.
In the ADPLL spectrum with RBW of 10 kHz in Fig 4-12, the inband noise is very high
due to the spurs.
2
One can use LFSR to create pseudo-random number
September 18, 2011
89

FCW is 112.00030, rbw is 1.000000e+03
60
80
100
120
140
160
180 2
10
10
10
10
10
10
10
10
10
10
Figure 4-11: ADPLL spectrum for CKV frequency of 3793.3156 MHz with FREF dithering
Algorithm Activated, RBW=1 KHz.
September 18, 2011
90

FCW is 112.00030, rbw is 1.000000e+04
70
80
90
100
110
120
130
140
150
160
170 3
10
10
10
10
10
10
10
10
10
Figure 4-12: ADPLL spectrum for CKV frequency of 3793.3156 MHz with FREF dithering
September 18, 2011
4-3-4
91
Spur Suppression with phase rotation and FREF dithering
From Fig 4-8 and Fig 4-11, one can conclude that the phase rotation is more eective
against spurs due to the TDC cell mismatch and FREF dithering is more eective against
spurs due to the TDC quantization eect. With both algorithm activated, the spurs shall
be suciently suppressed. Fig 4-13 and Fig 4-14 just present the spectrum for RBW=1
kHz and RBW=10 kHz when both algorithm are turned on. All the spurs are eectively
suppressed and the only noticeable spur in Fig 4-13 is at 40 kHz oset. From Fig 4-14 we
can see it doesnt degrade the systems performance much and with even the worst TDC
resolution, the spectrum ts the specication requirement.
FCW is 112.00030, rbw is 1.000000e+03
80
90
100
110
120
130
140
150
160
170
180 2
10
10
10
10
10
10
10
10
10
10
Figure 4-13: ADPLL spectrum for CKV frequency of 3793.3156 MHz,

Phase rotation and FREF dithering algorithm activated, RBW=1 kHz.
The PHE transient with the two algorithms activated is shown in Fig 4-15. The PHE signal
is well randomized compared to Fig 4-6, which illustrates why spurs are well suppressed.
Whats more, the channel center frequency step size for the WiMAX application is 250 kHz
or 100 kHz[27]. Our ADPLL loop bandwidth is within 100 kHz. Thus for every integer N,
only one or two channels featuring a FCW value close enough to the integer may give rise
September 18, 2011
92

FCW is 112.00030, rbw is 1.000000e+04
80
90
100
110
120
130
140
150
160
170 3
10
10
10
10
10
10
10
10
10
Figure 4-14: ADPLL spectrum for CKV frequency of 3793.3156 MHz

phase rotation and FREF dithering algorithm activated, RBW=10 kHz.
:DYHIRUP6LP9LVLRQ
3DJHRI
&XUVRU IV
%DVHOLQH
&XUVRU%DVHOLQH IV
IV
IV
IV
IV
IV
IV
7LPH$ IV
IV
IV

3+(>@
d2

Figure 4-15: PHE Transient for CKV frequency of 3793.3156M (with phase rotation and
FREF dithering on).
September 18, 2011
93
to the spurs. Phase rotation and FREF dithering will be applied to these channels. For
most cases, the two algorithms can be deactivated and we have less in-band phase noise as
well as simpler loop topology.
September 18, 2011
94
September 18, 2011
Chapter 5
The ADPLL Top Level
In previous chapters, the ADPLL architecture and algorithms have been introduced, with
the insight into building blocks, simulation methodology, performance analysis and advanced algorithms. To forge these fancy ideas into a feasible solution (nally a working
chip) takes signicant eort at the top level of the system. In this chapter, we present the
top-level schematic and then deal with various issues of the ADPLL top level. Extensive
top level simulations are also performed to make sure the system works as we desire.
5-1
Top-Level Schematic of ADPLL System
The top level of ADPLL is divided into three parts: analog core (ACORE), digital core
(DCORE) and AnalogIO. The ACORE part contains all the blocks to be handcrafted, i.e.
their schematic and layout are designed manually. The DCORE part has all the blocks
to be synthesized automatically. The AnalogIO part contains the IO rings for ACORE,
which is not covered in the scope of this thesis and is ignored here.
Fig 5-1 is a schematic of ACORE and Fig 5-2 is a schematic of DCORE. Due to the limited
space, the ACORE/DCORE interfaces are simplied here. Their interface signals are listed
exhaustively in Table B-1 and Table B-2 as a reference.
In Fig 5-1, the Phase Rotator block is the implementation of the phase rotation algorithm
as discussed in Section 4-3. It contains a simple Gray counter and a multiplexer with the
I/Q signals from DCO (Ip,Qp,In and Qn). For its interface, PR en is to enable/disable
the phase rotation algorithm and PR rst a will reset the output of the phase rotator to
the default input signal (Qn). The CKV test module block has a divide-by-16 divider and
a multiplexer for the feedback CKV which are shown at the bottom left corner of Fig 2-7.
CKV SEL is to choose the feedback CKV signal between the Phase Rotator output and
September 18, 2011
'&2
9''B'&2
966B'&2
',9B%8)B(1
Figure 5-1: Top Level View of ACORE in ADPLL.
September 18, 2011
9''B+6'
966B+6'
'LJLWDO)5()
$QDORJ)(5)B1
'&2BWDLOBUHVLVWRU
966B'LYLGHU
5HWLPHU

,QFUHPHQWRU
3+9B603!
'&2B,1B
9''B'LYLGHU
7'&
&.5
&.9BS!
)5()B6(/
)5()
&.9B7'&
&.9'
&.9B6(/
35BUVWBD
35B&/.
7HVWB'LYLGHUBUVWBD
3$LQSXWB/%
3$LQSXWB+%
3KDVH
5RWDWRU
966B6OLFHU
)5()0X[
&.9WHVW
PRGXOH
35BHQ
)5()GLWKHU
7'&4!
'3$B&&:!
&ONB*DWLQJB(QDEOH
'3$B$&:B/%!
'3$
)5()6OLFHU
'LWKHULQJ
&.9'
'3$B$&:B+%!
/RZIUHTXHQF\B&.9
3$B&DS%LDV
3$B+%
3$B/%
9''BEXIIHU
966B3$
$QDORJ)(5)B3
The ADPLL Top Level
9''B6OLFHU
96
6(/B('*(
5-1 Top-Level Schematic of ADPLL System
6\V0HP2XW
7'&B4!
3+9B603!
)5()
6LJPD
'HOWD
0RGXODWRU
6(/B('*(
&.9'
'&2B,1B7)
97

7'&0HP2XW

6\V0HPB6(/

7'&0HPB6(/
/6'B&.5

$567
63,B5(6(7

'&2B,1B3B*
3+(B29Q
'&2B,1B$

63,
3+(B29S
/6'
7\SH,,B29Q
7\SH,,B29S

&RUHB5(6(7
66BQ
$GGUB7'&0HP
/6'B567
'&2B,1B3B'6
35BUVWBD
026,
6&/.
$GGUB6\V0HP

'&2B781(B7)
'&2B,1B7
0,62
5(*B'&2B3

5(*B'&2B$
&.5
&RUHB5(6(7
VHTFONJHQ

5(*B'&2B7

7YBDYJ
,QWB6HTB&/.
6HTB&ONB6(/
6\V0HPB6(/

&RUHB5(6(7
,QWB6HTB&/.
6HTB&/.
([WB6HTB&/.
3+(
G3+(
)LOWHUBRXWSXW
3URSRUWLRQDOBRXWSXW
&.5
([WB6HTB&/.
6\V0HP,Q

$GGUB6\V0HP
6\V0HP2XW
6\V0HP

9''B'LJLWDO
966B'LJLWDO

/6'B567
7'&0HP&ONB6(/
&.5
)5()
6HTXHQFHU
$567

7'&0HPB6(/
6HTB&/.
7'&B5,6(
7'&B5,6(BWHVW
/6'B567
7'&0HPB&/.
&RUHB5(6(7
7'&0HP,Q

7'&0HP
7'&0HP2XW

$GGUB7'&0HP

/6'B567
Figure 5-2: Top Level View of DCORE in ADPLL.
September 18, 2011
98
The ADPLL Top Level
the o-chip input signal Lowfrequency CKV. All other blocks in the ACORE have been
already introduced. Table 5-1 lists the supplies and grounds in the ACORE block. All
VSS signals are 0 V and all VDD signals are 1.2V. The PA CapBias voltage is 0.6 V. The
idea of using so many supplies/grounds is to make sure sensitive blocks can have a clean
supply for the performance requirement (like the DCO core) and to be able to measure the
power consumption of every block separately. However, since the transistor-level design of
the blocks in ACORE is not nished and the decoupling capacitors are not added, here in
Table 5-1 the specications of the tolerable voltage drop and the tolerable supply noise are
not given.
Supply/Ground
VDD DCO, VSS DCO
VDD Divider, VSS Divider
VDD buer
VSS PA
PA CapBias
VDD Slicer, VSS Slicer
VDD HSD, VSS HSD
Blocks/Modules connected
The DCO core in the DCO block
The buers and dividers in the DCO block; Phase Rotator
The buer stages in the DPA block
The PA stages and the buer stages in the DPA block
The capacitor banks in the DPA block
The FREF slicer + FREF dithering implementation
The TDC block and the Retimer+Incrementor block
Table 5-1: Supply and ground signal list in the ACORE.
In the DCORE schematic of Fig 5-2, the low speed digital logic as in Section 2-5 is encapsulated as the LSD block. Its interface signals are listed in Table B-3. The fractional part
of the tracking bank tuning word goes to the Sigma-Delta modulator which is clocked by
the high-speed clock CKVD. Besides these two blocks, DCORE has the sequencer block,
the SPI block, the SysMem block and the TDCMem block. The last two blocks are mainly
memories for test and we will refer to them later. All digital blocks have a supply of 1.1
V.
The sequencer block provides signals to trigger the events in the LSD block after the
ADPLL loop starts up, like mode switchover, Ktdc normalization, gear-shifting, transition
from type-I PLL to type-II PLL and etc. In the ADPLL normal working status, the CKR
signal is valid. As shown in Fig 5-2, it is divided by 16 in the seqclkgen block and the
output signal Int Seq CLK, which has a frequency of around 2 MHz, will be chosen as the
sequencers clock (Seq CLK) by the signal Seq Clk SEL in the sequencer clock multiplexer.
The sequencer has a 7 bit counter which counts the positive edge of the sequencer clock
after start-up. This counter reserves value 0 for the ADPLL open loop operation and starts
from value 1. We will refer to that later. When the counter value reaches a certain value
specied by a certain input of sequencer from the SPI block, some output signal(s) will
be activated/deactivated to trigger a planned event. For example, the sequencer input
MEM TIME A2T signal is 5. Then when the counter value is 5, which is about 2 us after
ADPLL starts up, the sequencer decides to tell the LSD block to switch from the acquisition
mode to the tracking mode. At that time, the CTL PLL A signal as in Fig 2-25 goes from
high to low and the CTL PLL T signal goes from low to high. The CTL SRST T signal
September 18, 2011
5-2 Back-Up Modules in the ADPLL System
99
becomes low to disable the synchronous reset of the OT block. ARST will also be low
to enable the operation of the Sigma-Delta modulator. When the counter value reaches
the full scale, i.e. 127, it maintains this value rather than wraps around. Via this, the
sequencer has a time resolution of around 0.5 us and a time span of around 60 us. The
interface signals of the sequencer are listed in table B-4.
Its worth notice that only when the DCO block starts oscillation and the retimer works
can we have a valid CKR. To provide a clock for the sequencer when these conditions are
not satised, say before ADPLL starts up, here in Fig 5-2 an o-chip signal Ext Seq CLK is
provided as a clock. The signal Seq Clk SEL from the SPI block is made high to select this
external clock when CKR is not ready. If a valid CKR is delivered, the signal Seq Clk SEL
is made low to select Int Seq CLK as its clock. In this way the sequencer is always with a
valid clock and can control the operation in the LSD block.
The SPI block is an implementation of the slave device in the Serial Peripheral Interface
(SPI) Bus[28, 29] with some registers. The SPI Bus is chosen as it is simple, exible and
can have a high data transmission rate (can work at MHz or tens of MHz). The SPI block is
clocked by the external clock SCLK. Before the start-up of ADPLL, the SPI block will read
the desired information from the o-chip controller via the signal line MOSI and then save
the information to registers. The registers specify the select/enable/reset/timing signals
to the sequencer block or the LSD block or the blocks in the ACORE part. Therefore the
working mode and the status of ADPLL can be exible. During the transient of ADPLL
operation, some signals within the ADPLL system will be saved to registers or memories.
The SPI block can read the saved signals and send them to the external controller via the
signal line MISO, either for monitoring purpose or for test purpose. Its importance will be
revealed as we go through this chapter. The interface signal list of the SPI block is given
in Table B-5. One extra input signal SPI RESET from pad is used to reset the registers
associated with the SPI block. Then the ADPLL is working in the default mode.
5-2
Back-Up Modules in the ADPLL System
In our ADPLL design, some back-up modules are introduced to make the system more
robust, which helps reduce the risk of this prototype design.
For the ADPLL input, as seen in the ACORE schematic, except for the normal analog
FREF signal from the FREF slicer we also have Digital FREF signal generated o chip.
This Digital FREF can be noisier than Analog FREF signal and when its chosen as FREF
for ADPLL loop, the in-band phase noise of the system may degrade. Nevertheless, if
the o-chip crystal oscillator or FREF slicer doesnt work normally, Digital FREF is still
available and can be chosen as FREF for the system by pulling down FREF SEL, which
is the control signal of FREF multiplexer. In that way the system can still work.
For the ADPLL output, a divided-by-16 divider in the CKV test module block in the
ACORE schematic is used for the feedback CKV. The frequency of divided output is no
September 18, 2011
100
The ADPLL Top Level
more than 250 MHz and it shall go o-chip without special eort1 . If the output frequency
is still too high to go o-chip, the division ratio of the divider can be made larger than 16,
say 32 or 64. A spectrum analyzer can be connected to this o-chip signal and read out the
frequency of CKV. The jitter of this divider only has a signicant inuence on the far-out
noise. Thus we can estimate of the close-in phase noise performance for the CKV signal
via this divider output. In this way some useful information about the ADPLL output will
be retrieved even when DPA is not functional or when DPA adds some strong interference
to the system.
5-3
Test Plan for ADPLL
The test plan is an essential part to a successful design on silicon. Under the constraints
of limited pads, one shall nd a way to diagnose the whole system, to gather enough
information for the system evaluation, for system debug, and even as clues on how to
improve the system.
As the ADPLL features digital I/O for every essential block, some digital features are
included in its test plan. Registers and memories are deployed here to take advantage of
the high logic/memory density in 40 nm CMOS process. Then the SPI block is utilized to
read/write the contents of these registers and memories.
5-3-1
System Snapshot with Digital Logic
For ADPLL, the compared phase/frequency signal and the DCO capacitor bank control
signals are digital. These signals render the insight on the system status and facilitate the
system debug, if that is needed.
For example, if there are signicant spurs at the output spectrum, one can look at the
transient of the PHE signal to see whether it has a similar shape as in 4-6. If so, we can
apply the phase rotation and FREF dithering algorithms to suppress these spurs. Also,
one can look at the PHE or dPHE signal to make sure there is no PHE spike event due
to the mismatch of the TDC path and the incrementor path. Moreover, by observing the
tuning word for tracking bank, NTW T as in Fig 2-26, one can know whether gear-shifting
circuits / IIR lters / the type-I to type-II transition circuitry work as we expect or not.
Compared to analog peers, probing into digital signals will not perturb the normal working
of system. They can be accessed either out-of-chip in a real time way or processed by the
BIST circuitry which is often used in SoC or stored in the memory as a snapshot. Here
in our project, most digital signals feature long word length (32 for PHE and dPHE). A
real-time monitoring will take a lot of pins, increase the chip area and also increase the
1
For example, current mode logic (CML) is often used for really high speed signal at on-chip to o-chip
interface.
September 18, 2011
5-3 Test Plan for ADPLL
101
test cost. Also this chip has no DSP for digital processing and BIST introduces too much
overhead on complexity. Therefore we prefer to save a snapshot of the ADPLL system in
memory and then read it out slowly via the SPI block.
The signal chosen for the system snapshot is chosen from PHE, dPHE, Filter output and
Proportional out. The latter two signals are show in Fig 2-26. The 2-bit control signal,
SysMem SEL, select which signal to be forwarded to the memory for system snapshot
(SysMem). All these signals are updated per CKR cycle. Thus the clock for the memory
of SysMem is also CKR. The SysMem block can store 4096 cycles of one signal, i.e. the
address signal for SysMem, Addr SysMem, is 12 bit. As the frequency of CKR is about
30-40 MHz, the snapshot has the transient of the chosen signal for a duration of more
than 100 us. That is sucient to cover the frequency settling transient of ADPLL and is
larger than the time range controlled by the sequencer.2 Even spurs located at pretty low
frequency oset (on the order of several tens of kHz) can be observed if PHE is the stored
signal.
The content in SysMem gives the detail information about the system working status. After
the memory is lled up by the chosen signal, the external SPI controller tells the SPI block
the address of the word in SysMem to be read out, which is the signal Addr SysMem. This
word is selected, read in by the SPI block to some register and then sent to the external
controller via the signal line MISO. This reading operation will take several cycles of SCLK.
So to get the whole content of SysMem it may take several minutes. Thus it would be
nice to know about some brief system information before we spend time on reading out
SysMem content. In our design some registers are added to serve this purpose.
When ADPLL has settled, the tuning word of capacitor banks (PB, AB and TB) can
be saved in registers. The sequencer will write REG DCO P and REG DCO A as in Fig
2-25 (also there is a REG DCO T signal with 6 bit integer part and 5 bit fractional part)
to registers at the time specied by the signal MEM TIME REGDCO. These words can
be read out and used as MEM DCO P and MEM DCO A next time when ADPLL is to
output a frequency close to current working frequency. In this way the frequency settling
transient is shortened and loop stability is also improved.
There are some overow registers as the outputs of the LSD block (PHE OVp, PHE OVn,
typeII OVp, typeII OVn). The accumulation operations from dPHE to PHE as in Fig 2-22
and in the integral path of tracking bank loop lter as in Fig 2-26 are monitored. When
these accumulation operations cause a overow, either in positive direction or negative
direction, the corresponding register will record this event (the output goes from low to
high). In this way we have a brief estimation on the status of the ADPLL system. For example, when PHE OVp is high, most probably the ADPLL system is locked to a frequency
lower than the desired frequency.
The values in these registers are directly connected the SPI block. So the o-chip controller
can specify the SPI block to read these values out after ADPLL has settled. That operation
2
If the SysMem clock is too area consuming for the chip, one can reduce the address signal to 11 bits.
That almost satises our requirement to monitor the system start up transient.
September 18, 2011
102
The ADPLL Top Level
is much faster than the SysMem reading out procedure as it only needs several cycles of
SCLK.
5-3-2
DCO Open-Loop Test
The DCO block lies at the heart of the ADPLL system and its performance is critical to
the whole system. Thus it is desirable that DCO can be test alone. To achieve that, we
need to be able to open the ADPLL loop and congure the DCO tuning words. These
are rather straight forward in an all-digital environment. The timing of the LSD block
is specied via the SPI block and controlled via the sequencer outputs. As mentioned in
Section 5-1, the timing of the sequencer is relied on an internal counter, which starts from
1 rather than 0. When we are to open the ADPLL loop, all signals with timing information
from SPI are set to 0. Thus the counter output can never be equal to any of the timing
values specied by the SPI block. The OP, OA and OT blocks are always reset and the
outputs of loop lter are nullied. Yet the tuning words of DCO can be changed as desired
via the SPI output MEM DCO P, MEM DCO A and MEM DCO T. As in Fig 2-25, if the
loop lter outputs are nullied, these signals will determine the DCO tuning words alone.
The tuning word can be modied on the y to sweep the DCO frequency curve. One can
refer to the measurement set-up at Page 58 of [30]. There the SPI is controlled by the
FPGA, which is connected to the PC. The PA output signal goes to the spectrum analyzer
and its output is also fed to the PC via the GPIO bus. First the PC delivers the tuning
word information to the SPI via FPGA. The DCO receives these new tuning words and
settles to the new frequency. The spectrum analyzer has the phase noise and the frequency
information and then sends it to the PC. In that way, the DCO test is done automatically.
Also the DCO tail resistor can be modied to measure the DCO phase noise performance
with respect to the bias current.
Actually in our ADPLL the DCO open-loop test would be performed before the measurement of ADPLL closed-loop performance. As we mentioned in Section 2-5, in the blocks of
GP/GA/GT the signals will be multiplied with some coecient for the Kdco normalization. Then we need to have Kdco in advance, as can be easily obtained here. Say, here we
x the values of MEM DCO P and MEM DCO A and then measure the output frequency
of the ADPLL as the integer part of MEM DCO T goes from 0 to 63 (the fractional part
MEM DCO T is set to 0). Thereafter we can extrapolate the slope for the curve of fre dco = -51.3 kHz/LSB and the
quency versus tuning word. Say the slope gives a slope of K
fR
reference clock frequency is fR = 33.8688 MHz. Then K is about 660.21. In the ADPLL
dco
closed-loop operation, this information is provided via the signal MEM GAIN T, which
only has 11 bits for integer part. Thus MEM GAIN T will be set to 660 when the ADPLL
is to settle to the frequency within the range of this curve.
September 18, 2011
5-3-3
103
TDC Test Plan
TDC is another essential block in the ADPLL system. From Section 3-4 we know that its
resolution has a signicant contribution to the in-band phase noise of the ADPLL system.
From Section 4-3, we know the TDC mismatch and the quantization eect are the reasons
for spurs in the ADPLL output spectrum. Therefore, it would be nice if the TDC resolution
and its mismatch can be measured.
The TDC resolution can be obtained during the normalization operation of the ADPLL
system. In Section 2-5, we accumulate |TDC FALL-TDC RISE| and then do averaging
to estimate Ktdc . Notice that Ktdc is just the xed-point representation of the estimation
inv
for TTCKV
, as the desired frequency is known, the TDC resolution Tinv is also known. For
example, say the desired frequency is 3800 MHz and the accumulation output for TDC
27
inv
normalization is 2740, then we know TTCKV
2740
. We have the TDC resolution Tinv =
1
27
12.29 ps. This estimation achieves an accuracy of 2%.
3800 MHz 2740
To characterize the TDC mismatch, one needs to deliver the TDC transfer function as in
Fig 5-3 [15, 31]. An easy way is the density test method, which is done in the ADPLL
closed-loop operation. When we specify a desired frequency which is not a near-integer
N case, the word TDC RISE as in Fig 2-23 will sweep the TDC codes. If there is no
mismatch between every TDC cell, the probability of TDC RISE for every TDC code is
the same. Then all TDC codes shall have almost the same density in the histogram of the
TDC RISE output. Nevertheless, when the mismatch kicks in, the density of the TDC
codes shall dier. The density ratio between these codes reects the mismatch of the TDC
cells. For example, the average density of the TDC code is 25000 and the density of code 5
is 27500. Say the TDC normalization operation indicates that the TDC resolution is 12.5
ps. Then we estimate the delay of the fth cell in the eective TDC chain (the 9th cell in
the whole TDC chain) as 27500
12.5 ps = 13.75 ps.
25000
For this density test method, enough cycles of TDC RISE word shall be recorded for the
histogram statistics. Here the TDC RISE signal is stored into the TDCMem block for
65536 CKR cycles, i.e. the address signal for TDCMem (Addr TDCMem) is 16 bits. Then
the content of TDCMem is read out from the SPI block. The TDC RISE value for the
rst 5536 cycles are thrown away to make sure we use the TDC RISE value only when the
ADPLL has settled to the desired frequency. This measurement is done for 9 times and
then we do the density test for all the TDC RISE words stored in these measurements.
The histogram, as in Fig 5-4, has an accuracy of around 2% when checked against the
TDC modeling in list A-2.
This density test method is very convenient as it can be done on the y. However, it has
several disadvantages. Such test is based on the premise that ADPLL functions correctly
in the closed-loop mode, which means other modules like DCO, Retimer+Incrementor and
the digital logic shall be all OK. Another reason is a lot of measurements need to be done
for a very accurate characterization of the TDC mismatch, say with 1%. That adds to the
test cost. The most important reason is, as we look at Fig 5-4, the histogram only tells us
September 18, 2011
104
The ADPLL Top Level

TDC Transfer Function
40
35
30
Code
25
20
15
10
5
0
0
50
100
150
200
250
300
350
HCLK to FREF Delay [ps]
400
450
500
Figure 5-3: TDC transfer function.
x 10
2.5
Histogram Statistics
1.5
0.5
10
15
TDC_RISE word
20
25
30
Figure 5-4: The TDC closed loop test (CKV frequency of 3800 MHz).
September 18, 2011
105
the mismatch from code 1 to code 21, which is cell #5 to cell #25 in the whole TDC chain
as shown in Fig 2-10. Since the delay of the eective TDC chain, from cell #5 to cell #36,
is longer than one CKV cycle, the closed-loop operation cannot cover all the TDC codes.
Therefore some open-loop test method is needed to solve the above problems.
The open loop test set-up is shown in Fig 5-5. In the gure the pins of DigitalFREF
and Lowfrequency CKV are utilized. So we shall let the multiplexer for FREF to choose
DigitalFREF rather than AnalogFREF and let the multiplexer in the CKV test module to
choose Lowfrequency CKV rather than the feedback CKV. Thus the ADPLL loop is open.
The FREF pulse signal is a low frequency clock signal. In the simulation it is set to be
33.8688 MHz yet can be another frequency. The Ext CLK signal is a clock signal with the
frequency higher than the FREF pulse signal yet still much lower than the feedback CKV
signal. Here we choose its frequency to be 100 MHz.
'HOD\OLQH
/RZIUHTXHQF\B&.9
$'3//&KLS
)5()BSXOVH
'LJLWDO)5()
([WB&/.
Figure 5-5: TDC open-loop static test.
In this test structure, FREF pulse is sampled by the Ext CLK signal. Thus the rising
edge of DigitalFREF signal lags the rising edge of the output of the second ip op by
one Ext CLK cycle Text . On the other hand, the rising edge of Lowfrequency CKV lags
the rising edge of the output of the second ip op by the delay of the delay line . When
Text = , the rising edge of DigitalFREF and Lowfrequency CLK just occur at the same
time. So when the delay line is ne tuned so that is a little bit smaller than Text , say
200 ps, the rising edge of Lowfrequency CKV leads that of DigitalFREF by 200 ps. This
timing dierence will be detected by the TDC and from Q(5) to Q(36) a transition from 1
to 0 will be detected.
Nevertheless, such transition cannot be processed properly by the normal TDC decoding
logic as in Fig 2-23. The reason is that now FREF and CKV have the same frequency and
the oversampling circuit in Fig 2-12 will not function. Thus there is no valid CKR3 and
the latch in Fig 2-23 doesnt work. The way to route such problem is to use negative edge
of FREF to clock the TDC decoding logic rather than latching it. The decode output will
3
Thus at this time the sequencer shall be clocked by the Ext Seq CLK signal.
September 18, 2011
106
The ADPLL Top Level
be sent to TDCMem which is clocked by the rising edge of FREF, which is the reason of
using a multiplexer between CKR and FREF for the clock of TDCMem.
As is noticed, assuming that the DigitalFREF and Lowfrequency CKV signals have no
jitter, once the frequency of Ext CKV fext and the delay of the delay line are xed, the
timing relationship of the rising edge of these two signals are also xed. The TDC decoded
output shall also be unchanged. So it is called a static test. However, since all signals
have jitter and the delay line has a variation, the TDC decoded output changes. We save
the TDC decoded output for enough cycles (65536 cycles would be far enough) and then
choose the code with the highest density as the averaged TDC output, which is called
majority voting algorithm.
To sweep the timing dierence between the two rising edges for characterizing the TDC
transfer function, the frequency of Ext CKV will increase/decrease in a uniform step. Then
Text changes and nally the averaged TDC output also changes. Here fext goes in a step
= t
, when fext is around 100 MHz, the time step of Te xt corresponding
of 500 Hz. As f
f
t
to the frequency step is 50 fs, which is less than 0.5% of one TDC cell delay. Then as we
sweep fext , transition events of the averaged TDC output would indicate the delay of a
certain TDC cell and thus the resolution as well as the mismatch of TDC.
For example, in our simulation, from Fig D-1 and Fig D-2, one can know that the TDC
average output goes from 17 to 16 as fext goes from (100 MHz-203*0.5 kHz) to (100 MHz202*0.5 kHz). From Fig D-3 and Fig D-4, one can know that the TDC average output goes
from 16 to 15 as fext goes from (100 MHz+38*0.5 kHz) to (100 MHz+39*0.5 kHz). Then
1
1
the delay of cell #20 can be estimated as 1002020.50.001
100+390.50.001
us = 12.060 ps.
The model of TDC in list A-2 says the delay of cell #20 is 0.968*12.5 ps=12.1 ps. The
accuracy is around 0.3%. Similarly, we have the TDC average output transition from 15
to 14 when fext goes from (100 MHz+288*0.5 kHz) to (100 MHz+289*0.5 kHz). Then the
delay of cell #21 is estimated as 12.480 ps. The estimation accuracy is 0.16%.
To summarize, this open loop static test method can achieve an accuracy of 1% for the
TDC transfer function characterization. Then we can get a very accurate estimation of
the resolution and mismatch of TDC. It shall be noted that we can modify the frequency
step of the measurement to either increase the accuracy or reduce the measurement time.
5-4
Operation Modes of ADPLL
In Chapter 2 and Chapter 4, we have already touched into the operation of the ADPLL
system, mainly with an architecture view. For example, for the frequency settling transient,
the ADPLL system needs to go through the PVT mode, the acquisition mode and the
tracking mode. It needs to perform ZPR, gear-shifting and transition from the type-I PLL
to the type-II PLL at the certain time. Here we illustrate the operation modes of ADPLL
from a top level view. The system is treated as a chip and the insight is given on making
September 18, 2011
5-4 Operation Modes of ADPLL
107
the system an entity with the building blocks, the algorithms and the test plans mentioned
in this thesis.
5-4-1
ADPLL Start-Up
Before the ADPLL system starts up, the SPI RESET signal is high to reset the registers
in the SPI block. The registers contain all the information to control the ADPLL system
operation, as listed in Table C-1. Thus the whole system are reset: the DCOtailres signal
in Fig 2-4 is 0 so that DCO doesnt oscillate; the sequencer chooses Ext Seq CLK as its
clock with its counter reset to 1; the operation of the LSD block is reset by the control
signals from the sequencer; the SysMem block and the TDCMem block are also reset and
no valid signals are dumped into the memory.
After that silent period, the SPI RESET signal is set low. Then via the MOSI signal
line, the external controller writes the values of the registers in the SPI block to start up
the system. Firstly, the needed analog modules in ACORE will be wake up. DCOtailres
is set non-zero to turn on the DCO core and deliver a certain bias current for it. The
phase rotator is either enabled or disabled via PR en in Fig 5-1. This also applies to the
FREF dithering, which is controlled by the FREFdither signal. The certain signals for
controlling the dividers/buers in the DCO block and the tuning words of DPA are set as
we desire. For example, if the LB output of ADPLL is needed, we shall enable DIV 1 5 EN,
LB BUF EN and LB PA DIV2 EN in Fig 2-2 as well as DPA ACW LB signal in Fig 2-1.
Yet we can also disable DPA ACW LB signal and choose the CKVD16 signal as in Fig 5-1
to measure the ADPLL output frequency without turning on DPA. Whats more, if one is
to perform the open-loop static test of TDC, the SPI block just needs to write the certain
bits of the register ckvfrefsel as in Table C-1 to FREF SEL and CKV SEL signals low. In
that way, the Lowfrequency CKV signal and the DigitalFREF signal are chosen, ready for
the TDC open-loop test.
After the SPI block invokes the analog modules needed, the registers of TIME* (TIMEP,
TIMEP2A, TIMEKtdc, etc.) will be loaded with values that specify the timing of the
loop logic to the sequencer. The values of these registers are sent to the corresponding
inputs of the sequencer, like MEM TIME P, MEM TIME P2A, MEM TIME Ktdc Norm,
etc. Then the internal reset signal Core RESET as in Fig 5-2 is made low to bring the
counter of the sequencer to work. Then the loop logic is enabled and the whole system
starts up.
5-4-2
ADPLL Test Mode
From previous discussions one can know that the ADPLL loop needs to be open for the
test of DCO or TDC. And that is achieved by setting all the values of TIME* registers
to 0 before Core RESET signal is made low. Then the sequencer will not enable the
PVT/acquisition/tracking path in the digital logic. The phase accumulator as in Fig 2-22
September 18, 2011
108
The ADPLL Top Level
is reset by the Reset||ZPR signal. Therefore the digital logic is reset and the ADPLL loop
is open. If the MEM DCO* signals (MEM DCO P, MEM DCO A and MEM DCO T) are
congured via the corresponding registers MEMDCO* in the SPI register map and the
ADPLL output is monitored, the DCO open loop test is just performed. If the FREF SEL
signal chooses DigitalFREF and the CKV SEL signal chooses Lowfrequency CKV, the
TDC open loop test can be done with the test set up as in Fig 5-5.
As in the ADPLL normal operation, during the settling transient, the system will traverse
the PVT mode, the acquisition mode and the tracking mode. In the tracking mode, the
gear-shifting event is triggered and ADPLL changes from type-I PLL to type-II PLL. It
helps if we can stop the system at a certain phase rather than let it go through all the
events. For example, when the ADPLL system stays as a type-I PLL, we will see how
the up-converted icker noise of DCO hurts the in-band phase noise performance of the
whole system. Or we can enable/disable the gear-shifting event and compare the phase
noise performance for the two cases. This purpose can be served by setting some TIME*
registers to 0. Say we set TIME / TIMEP2A / TIMEKtdc / TIMEA2T as non-zero values
while TIMEGS / TIMELambda / TIMEType are set to 0, the gear-shifting mechanism is
disabled and ADPLL stays as a type-I PLL.
5-4-3
ADPLL Closed-Loop Mode
When all TIME* register are congured as non-zero values, ADPLL works in a closed-loop
way and will traverse all the events to the nal normal working status (tracking mode,
Gear-shifting triggered, IIR lter banks enabled, type-II PLL). Yet some preparations
shall be done for a correct simulation or measurement. Thus we have a simulation ow for
the ADPLL normal operation.
The simulation ow for the ADPLL normal operation is shown in Fig 5-6. We have
mentioned above that Kdco can be measured in the ADPLL open-loop mode. In the
simulation, it is done in another way. When the desired frequency is xed, the frequency
resolution Kdco can be calculated as the result of equation (2-2) divided by 4. Then we can
know the coecients for DCO Gain Normalization K fR , K fR and KfR with a certain
dco,p
dco,a
dco,t
reference clock. For example, if the desired frequency is 3550 MHz, the frequency resolution
is
dco,t = 1 2 2 Lf 3 C = 1 2 2 325pH (3550MHz 4)3 10 aF = 45.9 kHz (5-1)
K
4
4
MHz
then for reference clock at 33.8688 MHz, we have KfR = 33.8688
737.9. Thus during
45.9 kHz
dco,t
ADPLL start-up transient, the value of MEM GAIN T will be written as 738. These values
from the calculation will be written into the corresponding registers MEMGAIN*.
Still, Ktdc is also needed. As stated in Section 5-3-3, that is done by the TDC normalization
circuitry, which is triggered when the ADPLL system has entered the tracking mode. This
September 18, 2011
5-5 The ADPLL Top Level Simulation
109
normalization takes 256 FREF cycles, which is around 8 us. Thus before the long simulation
for characterzing the ADPLL phase noise, a short simulation shall be performance for
the TDC normalization. Here a simulation for 150 us is done. When the simulation
ends, the SPI block reads the value from the Tv avg register which is the sum value of
|TDC FALL-TDC RISE| for 256 cyels. Then we do a calculation like Equation 2-7. The
result is save to the registers Ktdc 0 and Ktdc 1. Then we have an accurate Ktdc and do
a rather long simulation of 5 ms for the ADPLL phase noise performance.
The timestamp information of the CKV rising edges in the simulation are stored in a le.
After the simulation end, Matlab just reads the timestamp information and do a rst order
polynomial curve tting for these timestamps. Note the very rst part of the timestamp
information is thrown away as at that time the system has settled to a frequency. The
slope of the tting is just the average period of CKV. The CKV frequency is know and
compared with the desired frequency to get the frequency deviation. The deviation of
every timestamp point from the tted polynomial indicates an instantaneous time error.
The error divided by the average period times 2 delivers the phase error. Then we can do
FFT to get the spectrum of phase error as W (f ) in Section 1-2-3. Then the phase noise
spectrum L(f ) is just that minus 3 dB.
5-5
The ADPLL Top Level Simulation
With the top level issues wrapped up in a nutshell, the ADPLL system is simulated and
the performance is extracted from the timestamp information as stated in last section. Fig
5-7 shows the ADPLL top level test bench. The reference clock signal AnalogFREF P and
AnalogFREF N are modeled in a similar way as the CKV signal in the DCO model. One
dierence is that the oscillating frequency is at fR rather than fv . The other dierence
is that we only model the white noise oor for the reference signals. Digital FREF and
Lowfrequency CKV can be modeled in the way of Fig 5-5 for the TDC open-loop test. For
the ADPLL normal operation we dont care how they are modeled. The Ext Seq CLK
models a clock signal with the frequency of several kHz. The jitter performance of it is of
no importance as it only provides the clock for the counter before the frequency synthesizer
starts up. The SPI RESET signal is high for the rst 10 us of the simulation to reset the
SPI block and then becomes low. After that the SPI master writes the control words to
the SPI block via the MOSI signal line to congure the operation modes of the ADPLL
system. The example code of the SPI master is in Section A-3. The edges of ADPLL
output signals, including CKVD16, DPA HB and DPA LB, can be detected and stored in
some le for the post-processing.
Here we simulate for dierent cases: the desired frequency is at high frequency (around
3800 MHz), middle frequency(around 3550 MHz) or low frequency (around 3300 MHz);
the TDC resolution under slow corner (Tinv is around 12.5 ps), under typical corner (Tinv is
around 11.5 ps) or under fast corner (Tinv is around 10.5 ps); the reference clock frequency
fR is 33.8688 MHz or 40 MHz. The near-integer N cases for the two reference frequencies
September 18, 2011
110
The ADPLL Top Level
Figure 5-6: Simulation ow for the ADPLL normal operation
September 18, 2011
111
63,B5HVHW
66BQ
6&/.
026,
0,62
([WB6HTB&/.
63,0DVWHU
&.9'
$QDORJ)5()B3
$QDORJ)5()B1
$'3//
&KLS
'LJLWDO)5()
'3$B+%
'3$B/%
/RZIUHTXHQF\B&.9
Figure 5-7: ADPLL top level test bench
are also simulated. fv is 3556.214 MHz for the reference frequency of 33.8666 MHz and
3560.01 MHz for the reference frequency of 40 MHz. The phase rotation algorithm and
the FREF dithering algorithm are enabled for the simulation of near integer-N cases. The
phase noise proles from the simulations are shown from Fig E-1 to Fig E-24. In these
plots the resolution bandwidth RBW is 10 kHz. That is ne enough for the phase noise at
frequency oset no less than 100 kHz. The spectrums with RBW of 5 kHz give the phase
noise performance at frequency osets closer than 100 kHz, which are not shown due to
the limited space of this thesis.
The phase noise performance is summarized in Table 5-2.
Specication
Frequency deviation
Spot noise @ 10 kHz
Spot noise @ 30 kHz
Spot noise @ 100 kHz
Spot noise @ 1 MHz
Integrated SSB Noise (1 kHz - 10 MHz)
Best performance
5.57 Hz
-96.06 dBc/Hz
-95.08 dBc/Hz
-98.84 dBc/Hz
-120.97 dBc/Hz
-43.61 dBc
Worst performance
14.36 Hz
-90.96 dBc/Hz
-91.22 dBc/Hz
-94.45 dBc/Hz
-112.15 dBc/Hz
-40.13 dBc
Table 5-2: The ADPLL phase noise performance summary
From this table it is clear that the frequency step size specication and the integrated
SSB noise specication as listed in Table 1-1 are met. The phase noise specication for
the 10 kHz frequency oset is also met. The phase noise specication for the 100 kHz
frequency oset violates the specication just by 0.55 dB in the worst case.
September 18, 2011
112
The ADPLL Top Level
The frequency settling transient of ADPLL is also investigated. The timestamp information
of CKV at the very rst part of the simulation when ADPLL is settling to the desired
frequency is processed by Matlab. The instantaneous period of CKV is compared with the
desired period. The dierence dT is plotted versus the time, which indicates how ADPLL
settles to the specied frequency. However, as the CKV signal is noisy in time domain,
the noise on the timestamp information may mask the frequency settling behavior when
the CKV frequency get close to the desired frequency. Thus we also do averaging for the
period of 100 CKV cycles. The CKV period is label as Tv [k], k is just the index. Then we
have the averaged sequence Tv,avg [k] = (Tv [k] + Tv [k + 1] + + Tv [k + 98] + Tv [k + 99])/100.
The dierence of the average period value with the desired period value would lter out
the noise and reveal the detail of the frequency settling.
Fig 5-8 and Fig 5-9 show the ADPLL settling transient in the PVT mode/acquisition mode
and in the tracking mode separately. The upper subgure in each plot is the dierence
of the instantaneous period and the desired period. The bottom subgure in each plot is
the dierence of the average period and the desired period. The frequency settling begins
when there is signicant change of dT, say 36.48 us in Fig 5-8. After about 1 us, ADPLL
switches from the PVT mode to the acquisition mode. The ZPR algorithm is triggered.
We know that the reset pulse of OP ZPR lasts several cycles. Thats the reason why the
dT value is almost xed at 37.41 us. When OP ZPR goes low, the dT value continues to
approach zero and after around another 1 us, ADPLL enters the tracking mode.
In the bottom subgure of Fig 5-9, one can see that when ADPLL switches to the tracking
mode, the period deviation dT approaches zero quickly and then jumps up and down
around zero before gear-shifting, which means CKV is quite noisy at that time. After we
trigger the gear-shifting event and change ADPLL to a type-II PLL, the variation of dT
is smaller and dT slowly settles to zero. From this gure we know at about 49 us ADPLL
just settles to the desired frequency. The time for frequency settling is around 13 us.
We have also simulated the settling transient for low frequency (fv is 3300 MHz), high
frequency (fv is 4050 MHz) and the near integer-N case (fv is 3556.214 MHz) as from Fig
E-25 to Fig E-30. The settling behavior is similar and the settling time in all cases is
within 15 us.
September 18, 2011
397EHJLQV

113
;
<

G7IV

397PRGHHQGV
$FTXLVLWLRQPRGH
EHJLQV

$FTXLVLWLRQPRGHHQGV
7UDFNLQJPRGHEHJLQV
;
<
;
<

7LPHXV

;
<
G7IV

;
<
;
<

7LPHXV

Figure 5-8: ADPLL settling transient in PVT mode and acquisiton mode(fv = 3800 MHz)
September 18, 2011
114
The ADPLL Top Level

G7IV

;
<

7LPHXV

$FTXLVLWLRQPRGHHQGV
7UDFNLQJPRGHEHJLQV

G7IV

)DVWWUDFNLQJHQGV
*HDUVKLIWLQJHYHQW
;
<
)UHTXHQF\
VHWWOHV
;
<
;
<

7LPHXV

Figure 5-9: ADPLL settling transient in tracking mode(fv = 3800 MHz)
September 18, 2011

Chapter 6
DPA Design for ADPLL
Design of DPA in our ADPLL project serves two purposes: 1. Function as a low noise output buer for ADPLL; 2. Explore the possibility of the WiMAX transmitter architecture
with ADPLL and DPA.
Thus unlike the core blocks (TDC, DCO) of ADPLL, the specication of DPA is not
very tight. One important consideration is that it should not degrade the far-out noise
specication of DPA (-150dBc/Hz). This, as we will see in the simulation result, is not a
tough requirement. To simplify the design eort, we have chosen the single-ended DPA
with a topology similar to Fig 6-4.
The idea of this DPA topology is rooted in the traditional Class-E PA. We will discuss
from that and converge to our design.
6-1
Introdution to the Class-E PA
This DPA design in the ADPLL project is based on the basic Class E PA concept, which
is proposed by Sokal in 1975[32]. The architecture of Class E PA is shown in Fig 6-1.
The transistor in Class E PA functions mainly as a switch. Therefore we prefer the input
signal to be a rectangular wave so as to turn the switch fully on and o. The RF choke
works more like a constant DC current source. So basically the current would charge the
capacitor Cx when the switch is o and the voltage at the drain of transistor would go up.
When the switch is on, the capacitor is discharged and the voltage would decrease to zero.
Therefore, the DC current from RF choke is converted to the RF current.
Lo and Co form a tank resonating at the fundamental frequency of the input signal to make
sure the harmonic component in the current owing to the load is almost ltered out. Lx ,
September 18, 2011
116

9GG
5)&KRNH
/R
&R
/[
,QSXW
&[
Figure 6-1: Topology of the Class E PA.
the residue inductance, together with Cx will dene the waveform of drain voltage and
current. There are three objectives for the waveform of drain voltage and current[33]:
1. The rise of the voltage across the transistor at turn-o should be delayed until after
the transistor is o.
2. The drain voltage should be brought back to zero at the time of transistor turn-on.
3. The slope of drain voltage should be zero at the time of turn on.
When these conditions are met, at anytime, either the drain current or the drain voltage
of the switch transistor will be zero to make sure there is no power dissipated on the
transistor. (A typical waveform of drain voltage and current of the Class E PA is shown
in Fig 6-2.) If all passive components are lossless, the Class E PA can achieve the ideal
drain eciency of 100%. So we call that optimum class E. An amplier which cannot
meet these three criteria is called sub optimum class E.
6-2
Introduction to the DPA Topology
The DPA concept proposed in [34, 35] is a near Class-E Amplier. The DPA schematics
of these two papers are shown below as Fig 6-3 and Fig 6-4. Via the digital control bits
(or called amplitude control word, ACW), we can control AND gates to pass or stop the
input oscillating signals, thus to enable/disable the switch transistor array. Therefore the
output power will be digital controllable.
September 18, 2011
6-3 DPA Schematic Overview
117
Figure 6-2: Typical drain voltage and current waveform of the Class E PA[2].
Actually, this DPA architecture deviates from the Class E PA in that the switch transistor
array does show nite resistance rather than zero when turned on. If ACW is large, i.e.
most of the AND gates will pass the input oscillating signal, the drain voltage and current
waveform ressembles those of a typical Class E PA. When ACW is small, the resistance of
switch is so large that the charge build up in the parallel capacitor cannot be discharged to
near zero when the switch is on. Therefore, the amplitude of voltage swing at the drain of
transistor is smaller and the output power is smaller. The switch will also dissipate some
portion of power since the voltage is not zero when its on. Then the drain eciency of
DPA will be lower.
6-3
DPA Schematic Overview
The top-level schematic for DPA is shown in Fig 6-5. In the schematic there is a module
named catip adpll dpaswitcharray which functions as a resistance-controllable switch array,
a module named catip adpll dpacaparray which is to tune the parallel capacitor value in
the classical DPA toplogy, and a buer in front of catip adpll dpaswitcharray.
The schematic of DPA switch array is shown in Fig 6-6. The Signal input signal is just
the output of the previous buer as in Fig 6-5. It is fed to an array of DPA switch unit
modules, which are binary controlled by the ACW signal. The schematic of switch unit
module is given in Fig 6-7. By using the NAND gate and the inverter stage, an AND
September 18, 2011
118
Figure 6-3: Schematic of DPA for BT transmitter.
September 18, 2011
119
Figure 6-4: Schematic of DPA for GSM/GPRS/EDGE transmitter.
September 18, 2011
120
Figure 6-5: Top level schematic of DPA
function of ACW and Signal input is realized. These two stages also function as a buer
for the DPA switch stage. The scaling guarantees that the rising time and falling time of
signal is within 20-30ps, which is about 10% of the cycle of the input periodical signal.
Thus the immunity of DPA to the supply noise is improved.
The NAND gate is just a basic CMOS logic NAND gate. Compared to a transmission-gate
AND (or NAND) gate, the CMOS logic is regenerative and has driving strength. Thus
the buer chain length for the switch stage is reduced. Also the requirement of the far-out
noise for WiMAX is not as tough as GSM. The noise of a CMOS NAND gate is acceptable.
The module of DPA cap array is a switchable-capacitor array that is binary controlled
by the capacitance control word (CCW). The schematic of one switchable-capacitor unit
is shown in Fig 6-8. The Capcontrol signal is a certain bit of the CCW word (CCW0
CCW3) and it controls a NMOS switch between the 50 fF MOM capacitor and the ground.
The capacitor unit is connected to the output signal at the drain of the switch transistors.
Thus by changing the CCW value, we can tune the capacitance value parallel to the switch
transistors, which is Cx in Fig 6-1. By this we can adapt the DPA block to dierent working
frequency and dierent package parasitic values. The resistor Rbias provides a bias voltage
for the drain of the switch. Thus this node will not be oating and the switch is fully turned
o when the Capcontrol signal is low.
September 18, 2011
121
9''BEXIIHU
'3$BGUDLQ
6LJQDOBLQSXW
$&:
'3$6ZLWFKXQLW
966B3$
9''BEXIIHU
$&:
'3$6ZLWFKXQLW
966B3$
9''BEXIIHU
$&:
'3$6ZLWFKXQLW
966B3$
9''BEXIIHU
$&:
'3$6ZLWFKXQLW
966B3$
Figure 6-6: Schematic of DPA switch array.
September 18, 2011
122
9''BEXIIHU
'3$BGUDLQ
Q Q
6HOBLQSXW
Q Q
QQ
Q Q
6LJQDOBLQSXW
QQ
Q Q
6HOBLQSXW
Q Q
966B3$
966B3$
Figure 6-7: Schematic of DPA switch unit.
&DSXQLWBVLJQDO
I)
&DS%LDV
Q Q
5ELDV N
&DSFRQWURO
&DS966
Figure 6-8: Schematic of a switchable capacitor unit in DPA
September 18, 2011
966B3$
6-4 Practical Concerns for the DPA Design
6-4
123
Practical Concerns for the DPA Design
DPA functions as the output buer for ADPLL. So it will interface with package parasitic
like bondwire / bondpad / lead frame. It also uses o-chip components as the RF choke
and the matching network, which is of course, never ideal. Besides that, DPA does deliver
a signicant amount of power. Thus the reliability needs to be taken into consideration.
Schematics of the test benches for DPA HB (3.3-3.8GHz) and LB (2.3-2.7GHz) mimicking
the practical situation are shown in Fig 6-9 and Fig 6-10.
9''
9''BEXIIHU
Q+
9
/ERQG
Q+
/OHDG
S+
/WUDFH
S+
/OHDG
S+
/FDS
S+
&FDSS)
3$BLQSXWB+%
FDWLSBDGSOOBGSDKE
9''
9LQ
,2FHOO
%RQGSDG
'3$B&&:!
&FRQWDFW
I)
&OHDG
I)
'3$B$&:B+%!
*1'
2KP
I)
*1'
*1'
966
966B3$B+%
S+
*1'
Figure 6-9: Schematic of the test bench for DPA HB part
9''
9''BEXIIHU
Q+
9
/ERQG
Q+
/OHDG
S+
/OHDG
S+
/WUDFH
S+
/FDS
S+
&FDS
S)
/LQG
Q+
3$BLQSXWB/%
FDWLSBDGSOOBGSDOE
9''
9LQ
'3$B$&:B/%!
,2FHOO
%RQGSDG
'3$B&&:!
&OHDG
I)
*1'
966
&FRQWDFW
I)
*1'
&LQG
I)
I)
2KP
*1'
966B3$B/%
S+
*1'
Figure 6-10: Schematic of the test bench for DPA LB part

September 18, 2011
124
In the test benches, the parasitic introduced by the IO cell and the bondpad are extracted
out. We assume the bondwire inductance value to by a typical value of 1.5 nH. The
parasitic of the lead frame is modeled by Llead1, Llead2 and Clead. The inductance value
for the RF choke is 10 nH. The inductance due to PCB trace is 500 pH. The parasitic
capacitance for the lead frame contact is 100 fF. For DPA HB, we just use a o-chip
capacitor component for matching network. The capacitance value Ccap and the parasitic
inductance value Lcap are 1 pF and 700 pH respectively. For DPA LB, an o-chip capacitor
and an o-chip inductor are used. The capacitance value Ccap and the parasitic inductance
value Lcap are 1.5 pF and 700 pH respectively while the inductance value Lind and the
parasitic capacitance value Lcap are 2.5 nH and 350 fF respectively.
There some are issues to consider:
1) Not ideal ground. The ground on a chip can never be ideal. VSS of every block will
be connected to the o-chip ground via some bondwire. When there is a current spike
going through the bondwire, which is modeled as an inductor, the potential of the on-chip
ground will shift. This is especially worse for singled-ended, high frequency, high current
component, i.e. our DPA.
So here the ground inductance is minimized by choosing two parallel down-bonding bondwire to the die paddle. The estimation is about 400 pH. The ground bounce peak voltage
for HB and LB are about 200mV.
2) Only low-Q tank. High-Q tank assumption of the Class E PA is important to make sure
that the output waveform is purely sinusoidal at fundamental frequency. However, due to
the bondwire-bondpad-leadframe network, the Q value of our DPA cannot be very high.
Besides that, the band of WiMAX is wider compared to BT and GSM. With a very high
Q tank, the output power would vary too much versus frequency. So in this design, the
on-o chip interface and o-chip components form a low-Q matching network. The output
waveform also deviates a lot from the ideal sine wave.
As we see in Fig 6-9 and Fig 6-10, the parasitic of o-chip components are also modeled by
checking the SRF of inductor and capacitor. For the HB case, the optimized inductance
value for an o-chip inductor is rather small. Thus the matching network doesnt have an
o-chip inductor and the performance would not degrade too much.
3) Transistor reliability. For traditional Class E PA, the drain voltage can go to over 3 times
VDD. The nominal working voltage for core transistor is 1.1V. From [35], voltage swing
at the drain must be controlled by the matching network to satisfy that the equivalent
dc voltage on the drain resulting from one RF cycle is smaller than 2 times VDD. In this
design, two measures are adopted for reliving the stress on switch transistors:
a) For DPA switch stage, use a separate supply voltage of 0.6V. The drain voltage is,
to rst order estimation, proportional to supply voltage connected to RF choke. So low
supply voltage will result in low peak value of drain voltage. b) Intentionally increase the
capacitance value of the capacitor parallel with switch transistors. Drain voltage goes up
in one RF cycle when switch transistors are o and current charges parallel capacitor. So
September 18, 2011
6-5 DPA layout overview
125
using larger capacitor value will damp the drain voltage waveform, which will reduce the
DPA eciency.
The simulated peak voltage between drain and source for switch transistor is smaller than
1.4V. Thats far enough to guarantee the transistor works properly.
4) ESD issue. As the drain of switch transistors will be connected to the IO cell and
the bondpad, ESD is also an important concern. In our design, the buer stage (including
NAND gates and inverters) will have a 1.2V supply. That on-chip supply will be connected
to VDD of the IO cell for PA output. Since we know that peak voltage between local ground
and PA output is smaller than 1.4V. The diode in IO cell will not be turned on and will
not inuence the performance of DPA.
5) Current handling capability. For on-chip power amplier, since signicant amount of
current will ow through the switches, we do need to investigate current capability of
transistors. The design kit of the process species DC current maximum value (to prevent
electro-migration) and RMS current maximum value (to prevent the excessive heat kill
the chip). Basically its the metal of drain and source that may have problem. In the
layout the metal lines are made wide enough to satisfy the requirement of current handling
capability.
6-5
DPA layout overview
The layout of the critical modules of DPA, catip adpll dpaswitcharray and
catip adpll dapcaparray, are shown in Fig F-1 and Fig F-2. Fig F-1 shows a 7*5 array
of switch units. The insider 5*3 array are the eective switch units and the most outside
circle of switch units are dummy cells for better matching performance. The switch units
for the adjacent columns are ipped so that they can share the same VDD (or VSS). The
metal lines for the ground and supply signal are made wide to handle the current. In Fig
F-2 the similar topology is applied for the switchable capacitor array. Nevertheless, there
is no dummy cell since the matching is not a critical concern. The connection to the MOM
capacitor in the capacitor unit is manually made wider to handle the RMS current in the
capacitor.
As the nal oorplan and the package are not xed, also due to the limited time, the
layout of the whole DPA (for HB and LB) is not done. The output signals, the ground and
the supply shall go to the pads via the thickest metal to minimize the parasitic resistance.
The ground signal and the supply signal shall go close to each other to minimize the
parasitic inductance and gain some decoupling capacitance for free. Small capacitors can
be inserted into catip adpll dpaswitcharray to localize the high frequency current. Some
large capacitor can be put close to the whole array to localize the low frequency current.
We may need to put additional buer(s) before the inverter in 6-5 if the path from the
DCO block output to the DPA input is too long.
September 18, 2011
126
6-6
DPA Simulation Results
Some simulations with extracted parasitics from the layout of modules have been done to
evaluate the DPA performance. The power control capability of DPA for HB and LB are
presented in Fig 6-11 and Fig 6-12. The output power doesnt change much as long as
ACW is large, i.e. the switch resistance is negligible. It changes quite a lot when ACW is
small, as we expected.
6
Output power (dBm)
8
0
6
7
8
9
ACW for DPA HB
10
11
12
13
14
15
Figure 6-11: Power control capability of DPA HB.
The drain eciencies (DE) and the power added eciencies (PAE) of DPA for HB and
LB are plotted in Fig 6-13 and Fig 6-14. For large ACW values, DE is around 60% and
PAE is around 40%. The PAE value rst increases and then becomes a little lower as we
sweep the ACW value. That is because the output power doesnt increase too much yet
the power consumption for the buer stages increase faster when more switch units are
turned on.
The phase noise performance for HB and LB are plotted in Fig 6-15 and Fig 6-16.
The DPA output power with respect to the frequency is shown in Table 6-1 and Table 6-2.
For the corner simulation, the typical, slow and fast corners at 27 C are simulated
as well as the typical corner with dierent temperature (27 C, 85 C and -40 C).
The corner simulation result for DPA HB and LB are listed in Table 6-3 and Table
tab:DPALBOutputPowerCorner.
September 18, 2011
6-6 DPA Simulation Results
127
Output power (dBm)
5
0
6
7
8
9
ACW for DPA LB
10
11
12
13
14
15
14
15
Figure 6-12: Power control capability of DPA LB.
65
Drain efficiency for DPA HB
Power added efficiency for DPA HB
60
55
Efficiency %
50
45
40
35
30
25
20
15
0
6
7
8
9
ACW for DPA HB
10
11
12
13
Figure 6-13: DPA eciency curve with respect to ACW (HB).

September 18, 2011
128
60
Drain efficiency for DPA LB

Power added efficiency for DPA LB
55
Efficiency %
50
45
40
35
30
25
0
6
7
8
9
ACW for DPA LB
10
11
12
13
14
15
Figure 6-14: DPA eciency curve with respect to ACW (LB).
130
Phase Noise (dBc/Hz)
135
140
145
150
155
160
165 4
10
10
10
Frequency offset (Hz)
10
Figure 6-15: Phase noise performance of DPA HB.
September 18, 2011
10
6-6 DPA Simulation Results
129
135
Phase Noise (dBc/Hz)
140
145
150
155
160
165
170 4
10
10
10
Frequency offset (Hz)
10
10
Figure 6-16: Phase noise performance of DPA LB.

Table 6-1: DPA HB output power with respect to frequency (ACW=15).
Frequency (GHz)
Output power (dBm)
3.3
5.527
3.4
5.427
3.5
5.293
3.6
5.14
3.7
4.94
3.8
4.786
Table 6-2: DPA LB output power with respect to frequency (ACW=15).
Frequency (GHz)
Output power (dBm)
2.3
3.725
2.4
3.343
2.5
2.882
2.6
2.377
2.7
1.853
Table 6-3: DPA HB output power for corner simulation.
Corner
Output power (dBm)
TT 27 C
5.219
FF 27 C
5.631
SS 27 C
4.666
TT 85 C
5.8
TT -40 C
5.237
Table 6-4: DPA LB output power for corner simulation.
Corner
Output power (dBm)
TT 27 C
2.882
FF 27 C
3.13
SS 27 C
2.59
TT 85 C
2.872
TT -40 C
2.902
September 18, 2011
130
September 18, 2011
Chapter 7
Conclusion
7-1
Contribution of This Thesis
This work consists of the system level design of ADPLL for the WiMAX standard and a
simple transistor level design of DPA. The main contributions of this work are:
1. The architecture and building blocks of the ADPLL system are presented, with details
on the function of every module. Modeled and described in Verilog-AMS/Verilog,
the whole system is simulated in time domain and the performance is presented. The
phase noise performance ts into the specication except for only 0.55 dB violation
at the frequency oset of 100 kHz. The settling time of the frequency synthesizer is
around 15 us.
2. The spur mechanism for the near-integer N cases of ADPLL is analyzed. The spur
suppression techniques, the phase rotation algorithm and the FREF dithering algorithm, are applied to the system and the spurs are eectively suppressed.
3. The top level issues of the ADPLL system are taken care of. The test plan for the
whole system and the critical modules, DCO and TDC, are proposed. The operation
modes of ADPLL are specied, with the requirement for the SPI block, the sequencer
block and the memory modules.
4. A simple DPA as the output buer of ADPLL is designed in the 40 nm process.
The layout of catip adpll dpaswitcharray and catip adpll dpacaparray, the essential
parts of the DPA block, are presented. The primitive simulation results of the DPA
circuitry are given.
September 18, 2011
132
7-2
Conclusion
Future Work
The expected future work can be categorized into three aspects:

1. The system design aspect. It is worth to be noticed that in the scope of this thesis,
the TDC normalization and the DCO normalization are both done in an o-chip
way. As we dont have a powerful digital signal processor, the division in the TDC
normalization as in Fig 2-23 is done o-chip. The DCO normalization is done via
the DCO open loop test as in Section 5-3. A commercial chip shall realize the
normalizations on-chip. For the TDC normalization algorithm, a correlation method
is proposed in [11] and shall be a good research topic. For the DCO normalization
algorithm, the LMS calibration algorithm has been proposed [36], which may be
possible to be applied in our ADPLL system also.
Whats more, the mismatch of DCO is not modeled in the work of this thesis. A loopup table of the DCO capacitor value can be built from the data of the DCO transistorlevel simulation result and used in the system simulation. A polar transmitter system
is expected to be built based on the system. When the modulation for the transmitter
operation is added to ADPLL, the DCO mismatch may lead to the performance
degradation.
2. The DPA design aspect. Due to the limited time, the DPA implemented here is very
basic. It needs further verication and optimization. The monte-carlo simulation
should be done to investigate the inuence of the mismatch on the DPA performance.
When the package is determined, the value of the o-chip components (inductors and
capacitors) shall be chosen to maximize the output power within the corresponding
band (HB or LB). The inuence of the variation of the parasitics and the component
values on the DPA performance should also be explored.
To be eligible for integration into a polar transmitter of non-constant amplitude
modulation, we may need to add extra bits to ACW to increase the tuning granularity.
Impedance transformation may be needed in the matching network so as to change
the terminal load to be less than 50 Ohm to increase the output power. Some other
topologies like inverse Class D power amplier can also be a option for the future
improved design[37].
3. The chip creation aspect. A primitive pinout intended for this ADPLL chip is presented in Fig 7-1. Besides the essential modules that are under design also nished
(TDC, DCO, DPA), there are some red blocks here, which means they are not taken
care of. Among them, the design of FREF slicer (with the FREF dithering) and the
retimer+incrementor will take a signicant eort. SRAM is red due to the temporary
lack of memory compiler for 40 nm process.
After the sizes and the interfaces of these blocks can be delivered, we will go to
oorplan and determine the package we need. Then the chip will be assembled and
ready for the top level chip creation steps.
September 18, 2011
7-2 Future Work
133
'LJLWDO)5()
$QDORJ)5()B1
'3$B/%
&.9WHVW
PRGXOH
966B+6'
(GJH
$OLJQHU
966B3$
966B3$
$QDORJ)5()B3
966B6OLFHU
9''B6OLFHU
/RZ)UHTXHQF\B&
.9
&.9B',9
)5()6OLFHU
3$B/%
)5()08;
7'&
3KDVH
5RWDWRU
5HWLPHU
&RXQWHU
9''B+6'
'3$B+%
6HTXHQFHU
'&25(
([WB6HTB&ON
3$B+%
026,
9''B3$BEXIIHU
+LJK
6SHHG
6'0
966B'LYLGHU
63,
0,62
'&2
65$0
6&/.
9''B'LYLGHU
66
9''B'LJLWDO
966B'LJLWDO
5HVHW
9''B'&2
966B'&2
Figure 7-1: Primitive Pinout for this ADPLL chip
September 18, 2011
134
Conclusion
To summarize, there is enough space for innovation, improvement and optimization based
on current progress of the ADPLL project.
September 18, 2011
Appendix B
Top-level Interface of ADPLL
Table B-1: ACORE Interface

Signal
VDD Slicer
VSS Slicer
VDD HSD
VSS HSD
VDD Divider
Word
length
1
1
1
1
1
Input/Output/Supply
Comments
Supply
Supply
Supply
Supply
Supply
VDD for FREF slicer

VSS for FREF Slicer
VDD for TDC and Incrementor
VSS for TDC and Incrementor
VDD for buers and dividers in
DCO block
VSS for buers and dividers in
DCO block
VDD for DCO core
VSS for DCO core
VDD for buer stage in DPA
block
VSS for PA stage and buer
stage in DPA block
Voltage bias for DPA capacitor
bank
DPA HB output signal
DPA LB output signal
CKV divided by 16, for test
Retimed FREF
CKV divided by 8. Generated in
incrementor. Used in modulator for fractional bits in tracking bank.
TDC raw output
Incrementor output
Reference clock
VSS Divider
Supply
VDD DCO
VSS DCO
VDD buer
1
1
1
Supply
Supply
Supply
VSS PA
Supply
PA CapBias
Supply
PA HB
PA LB
CKVD16
CKR
CKVD8
1
1
1
1
1
Output to pad
Output to pad
Output to pad
Output to DCORE
Output to DCORE
TDCQ
PHV SMP
FREF
40
10
1
Output to DCORE
Output to DCORE
Output to DCORE
September 18, 2011
154

Signal
Lowfrequency CKV
Word
length
1
Input/Output/Supply
Comments
Input from pad
O-chip signal may used as CKV

for TDC test
Reference clock signal generated
from device
Dierential reference clock signal
generated from crystal oscillator
Dierential reference clock signal
generated from crystal oscillator
SEL EDGE signal for retimer
and incrementor
Signal to enable/disable FREF
dithering algorithm
Select FREF signal between DigitalFREF and AnalogFREF
Choose CKV from phase rotator
output or external signal
Amplitude control word for DPA
HB
LB
Control word of capacitor bank
for DPA
Enable signal for Phase Rotator
Asynchronous reset for Phase
Rotator
Tuning words for drain and
source terminal of switches in
PVT bank
Tuning words for gate terminal
of switches in PVT bank
Tuning words for Acquisition
bank
Tuning words for integer part of
tracking bank
Tuning word for fractional part
of tracking bank
Control word for tail resistor in
DCO core
Enable/Disable divide-by-3 divider after DCO core
Enable/Disable divide-by-2 divider which generates input signal of DPA HB
Enable/Disable divide-by-2 divider which generates input signal of DPA LB
Enable/Disable buer of HB signal
DigitalFREF
Input from pad
AnalogFREF P
Input from pad
AnalogFREF N
Input from pad
SEL EDGE
Input from DCORE
FREFdither
Input from DCORE
FREF SEL
Input from DCORE
CKV SEL
Input from DCORE
DPA ACW HB
Input from DCORE
DPA ACW LB
Input from DCORE
DPA CCW
Input from DCORE
PR en
PR rst a
1
1
Input from DCORE

Input from DCORE
DCO IN P DS
Input from DCORE
DCO IN P G
Input from DCORE
DCO IN A
Input from DCORE
DCO IN T
64
Input from DCORE
DCO IN TF
Input from DCORE
DCO tail resistor
Input from DCORE
DIV 1 5 EN
Input from DCORE
HB PA DIV2 EN
Input from DCORE
LB PA DIV2 EN
Input from DCORE
HB BUF EN
Input from DCORE
September 18, 2011
155
Signal
LB BUF EN
Word
length
1
Input/Output/Supply
Comments
Input from DCORE
Enable/Disable buer of LB signal

Asynchronous reset signal for
divide-by-16 divider used for test
Enable/Disable clock gating signal in TDC block (expected feature)
Test Divider rst a
Input from DCORE
Clk Gating Enable
Input from DCORE
Table B-2: DCORE Interface

Signal
VDD Digital
VSS Digital
CKVD
Word
length
1
1
1
Input/Output/Supply
Comments
Supply
Supply
Input from ACORE
VDD for the whole DCORE

VSS for the whole DCORE
CKV divided by 8. Used in
modulator for fractional bits in
tracking bank.
Retimed reference clock
TDC raw output
Incrementor output
SEL EDGE signal for retimer and
incrementor
dithering algorithm
Reference clock used for TDC test
HB
Amplitude control word for DPA LB
Control word of capacitor bank for
DPA
Asynchronous reset for Phase Rotator
Tuning words for drain and source
terminal of switches in PVT bank
Tuning words for gate terminal of
switches in PVT bank
Tuning words for Acquisition bank
Tuning words for integer part of
tracking bank
Tuning word for fractional part of
tracking bank
CKR
TDC Q
PHV SMP
SEL EDGE
1
40
10
1
Input from ACORE

Input from ACORE
Input from ACORE
Output to ACORE
FREFdither
Output to ACORE
FREF SEL
Output to ACORE
FREF
CKV SEL
1
1
Input from ACORE

Output to ACORE
DPA ACW HB
Output to ACORE
DPA ACW LB
DPA CCW
4
4
Output to ACORE
Output to ACORE
PR en
PR rst a
1
1
Output to ACORE
Output to ACORE
DCO IN P DS
Output to ACORE
DCO IN P G
Output to ACORE
DCO IN A
DCO IN T
6
64
Output to ACORE
Output to ACORE
DCO IN TF
Output to ACORE
September 18, 2011
156
Signal
DCO tail resistor

Word
length
5
Input/Output/Supply
Comments
Output to ACORE

DCO core
Enable/Disable divide-by-3 divider
after DCO core
which generates input signal of DPA
HB
which generates input signal of DPA
LB
Enable/Disable buer of LB signal
Enable/Disable clock gating signal
in TDC block (expected feature)
SPI standard interface (Master In
Slave Out)
SPI standard interface (Master Out
Slave In)
SPI standard interface (SPI clock)
SPI standard interface (Select signal)
Reset for SPI and associated registers
DIV 1 5 EN
Output to ACORE
HB PA DIV2 EN
Output to ACORE
LB PA DIV2 EN
Output to ACORE
HB BUF EN
LB BUF EN
Test Divider rst a
1
1
1
Output to ACORE
Output to ACORE
Output to ACORE
Clk Gating Enable
Output to ACORE
MISO
Output to pad
MOSI
Input from pad
SCLK
SS n
1
1
Input from pad

Input from pad
RESET
Input from pad
Table B-3: Interface of LSD block

Signal
VDD
VSS
SEL EDGE
Word
length
1
1
1
Input/Output/Supply
Comments
Supply
Supply
DCORE output
VDD for LSD block

VSS for LSD block
SEL EDGE signal for retimer and incrementor
Reference clock used for TDC test
Fractional part of TB tuning word fed
to modulator
Tuning words for drain and source terminal of switches in PVT bank
Tuning words for gate terminal of
switches in PVT bank
Tuning words for Acquisition bank
Tuning words for integer part of tracking bank
Clock for SysMem (CKR)
FREF
DCO TUNE TF
1
5
DCORE input
Output to modulator
DCO IN P DS
DCORE output
DCO IN P G
DCORE output
DCO IN A
DCO IN T
6
64
DCORE output
DCORE output
Mem Clk
Output to SysMem
September 18, 2011
157
Signal
CTL ARSTZ
Word
length
1
Input/Output/Supply
Comments
Input from Sequencer
Asynchronous Reset for OP, OA and

OT block
Enable/Disable PVT path in LSD
Enable/Disable Acquisition path in
LSD
Enable/Disable Tracking path in LSD
Synchronous reset for OP block
Synchronous reset for OA block
Synchronous reset for OT block
Frequency command word
Trigger gear-shifting event for tracking
mode
Ktdc value for TDC normalization
Start command for TDC normalization
Stop command for TDC normalization
Clock for LSD (CKR)
Enable signal for IIR lter bank in
tracking path
Choose proportional coecient after
gear-shifting
Reset signal for LSD block
Coecient for 1st IIR lter
Coecient for 2nd IIR lter
Coecient for 3rd IIR lter
Coecient for 4th IIR lter
Preset tuning word value for PVT bank
Preset tuning word value for Acquisition bank
Preset tuning word value for Tracking
bank
Kdco estimation value for PVT bank
Kdco estimation value for Acquisition
bank
Kdco estimation value for Tracking
bank
Incrementor output
Enable Phase rotation algorithm on
digital side
Integral gain for type-II PLL mode
Trigger the transition from type-I PLL
to type-II PLL
Reset for loop lter of tracking bank
TDC raw output (output of dummy
cells discarded)
Save DCO tuning words to registers
PHE signal for system snapshot
dPHE signal for system snapshot
CTL PLL P
CTL PLL A
1
1

CTL PLL T
CTL SRST P
CTL SRST A
CTL SRST T
FCW
GS trigger
1
1
1
1
32
1

Input from SPI
Input from Sequencere
Ktdc
Ktdc norm start
Ktdc norm stop
LSD CKR
LSD Enable IIR
13
1
1
1
4
Input from SPI

DCORE input
LSD G word
LSD RST
Lambda0
Lambda1
Lambda2
Lambda3
MEM DCO P
MEM DCO A
1
2
2
2
2
7
6

Input from SPI
Input from SPI
Input from SPI
Input from SPI
Input from SPI
Input from SPI
MEM DCO T
6+5
Input from SPI
MEM GAIN P
MEM GAIN A
3
8
Input from SPI

Input from SPI
MEM GAIN T
11
Input from SPI
PHV SMP
PR
10
1
DCORE input
Input from SPI
Rho
SEQ T2 trigger
3
1
Input from SPI

Input from sequencer
SRST
TDC Q
1
32

DCORE input
CTL REG DCO

PHE
dPHE
1
32
32

Output to SysMem
Output to SysMem
September 18, 2011
158
Signal
Proportional out

Word
length
32
Input/Output/Supply
Comments
Output to SysMem
Proportional out signal in loop lter for

system snapshot
Filter out signal in loop lter for system
snapshot
TDC decoder output in closed-loop operation for test
TDC decoder output in open-loop operation for test
Registered PVT bank tuning word
Registered Acquisition bank tuning
word
Registered Tracking bank tuning word
Accumulated |TDC FALL-TDC RISE|
value for Ktdc normalization
Overow status in integral path of loop
lter
Overow status in integral path of loop
lter
Overow status of PHE
Filter out
32
Output to SysMem
TDC RISE
Output to TDCMem
TDC RISE test
Output to TDCMem
REG DCO P
REG DCO A
7
6
Output to SPI
Output to SPI
REG DCO T
Tv avg
11
12
Output to SPI
Output to SPI
TypeII OVp
Output to SPI
TypeII OVn
Output to SPI
PHE OVp
PHE OVn
1
1
Output to SPI
Output to SPI
Table B-4: Interface of Sequencer

Signal
VDD
VSS
Enable IIR
Word
length
1
1
4
Input/Output/Supply
Comments
Supply
Supply
Input from SPI
VDD for SPI block

VSS for SPI block
Specify enable signal for IIR
lter bank in tracking path
Specify proportional coecient after gear-shifting
Specify the time to start PVT
mode
Specify the time of mode
switchover from PVT mode
to Acquisition mode
Specify the time to start Ktdc
normalization
switchover from Acquisitioin
mode to Tracking mode
Specify the time of gearshifting event
Specify the time to activate
IIR lter bank
G word
Input from SPI
MEM TIME P
Input from SPI
MEM TIME P2A
Input from SPI
MEM TIME Ktdc Norm
Input from SPI
MEM TIME A2T
Input from SPI
MEM TIME GS
Input from SPI
MEM TIME Lambda Shift
Input from SPI
September 18, 2011
159
Signal
MEM TIME TypeCon
Word
length
7
Input/Output/Supply
Comments
Input from SPI
Specify the time of transition

from Type-I PLL to Type-II
PLL
Specify the time to save DCO
tuning words
reset for other blocks in
DCORE
Asynchronous reset for
modulator
Asynchronous Reset for OP,
OA and OT block
Enable/Disable PVT path in
LSD
Enable/Disable Acquisition
path in LSD
Enable/Disable
Tracking
path in LSD
Synchronous reset for OP
block
Synchronous reset for OA
block
Synchronous reset for OT
block
Save DCO tuning words to
registers
Trigger gear-shifting event for
tracking mode
Start command for TDC normalization
Stop command for TDC normalization
Trigger the transition from
type-I PLL to type-II PLL
Reset for loop lter of tracking bank
Enable signal for IIR lter
bank in tracking path
Choose proportional coecient after gear-shifting
Reset signal for LSD block
MEM TIME REGDCO
Input from SPI
reset
Input from SPI
ARST
Output to modulator
CTL ARSTZ
Output to LSD
CTL PLL P
Output to LSD
CTL PLL A
Output to LSD
CTL PLL T
Output to LSD
CTL SRST P
Output to LSD
CTL SRST A
Output to LSD
CTL SRST T
Output to LSD
CTL REG DCO
Output to LSD
GS trigger
Output to LSD
Ktdc norm start
Output to LSD
Ktdc norm stop
Output to LSD
SEQ T2 trigger
Output to LSD
SRST
Output to LSD
LSD Enable IIR
Output to LSD
LSD G word
Output to LSD
LSD RST
Output to LSD
Table B-5: Interface of SPI block

Signal
VDD
Word
length
1
Input/Output/Supply
Supply
Comments
VDD for SPI block
September 18, 2011
160
Signal
VSS
MISO

Word
length
1
1
Input/Output/Supply
Comments
Supply
DCORE output
VSS for SPI block

SPI standard interface (Master
In Slave Out)
SPI standard interface (Master
Out Slave In)
SPI standard interface (SPI
clock)
SPI standard interface (Select
signal)
Reset for SPI and associated registers
Frequency command word
Ktdc value for TDC normalization
Coecient for 1st IIR lter
Coecient for 2nd IIR lter
Coecient for 3rd IIR lter
Coecient for 4th IIR lter
Preset tuning word value for
PVT bank
Preset tuning word value for Acquisition bank
Preset tuning word value for
Tracking bank
Kdco estimation value for PVT
bank
Kdco estimation value for Acquisition bank
Kdco estimation value for Tracking bank
Integral gain for type-II PLL
mode
dithering algorithm
Enable Phase rotation algorithm
on digital side
HB
LB
Control word of capacitor bank
for DPA
MOSI
DCORE input
SCLK
DCORE input
SS n
DCORE input
RESET
DCORE input
FCW
Ktdc
32
13
Output to LSD
Output to LSD
Lambda0
Lambda1
Lambda2
Lambda3
MEM DCO P
2
2
2
2
7
Output
Output
Output
Output
Output
MEM DCO A
Output to LSD
MEM DCO T
6+5
Output to LSD
MEM GAIN P
Output to LSD
MEM GAIN A
Output to LSD
MEM GAIN T
11
Output to LSD
Rho
Output to LSD
CKV SEL
DCORE output
FREF SEL
DCORE output
FREFdither
DCORE output
PR
Output to LSD
PR en
DPA ACW HB
1
4
DCORE output
DCORE output
DPA ACW LB
DCORE output
DPA CCW
DCORE output
September 18, 2011
to
to
to
to
to
LSD
LSD
LSD
LSD
LSD
161
Signal
DIV 1 5 EN
Word
length
1
Input/Output/Supply
Comments
DCORE output
Enable/Disable divide-by-3 divider after DCO core

Enable/Disable divide-by-2 divider which generates input signal of DPA HB
Enable/Disable divide-by-2 divider which generates input signal of DPA LB
Enalbe/Disable buer of LB signal
DCO core
Enable/Disable clock gating signal in TDC block (expected feature)
Specify enable signal for IIR lter bank in tracking path
Specify proportional coecient
after gear-shifting
Specify the time to start PVT
mode
switchover from PVT mode to
Acquisition mode
Specify the time to start Ktdc
normalization
switchover from Acquisitioin
mode to Tracking mode
Specify the time of gear-shifting
event
Specify the time to activate IIR
lter bank
Specify the time of transition
from Type-I PLL to Type-II PLL
Specify the time to save DCO
tuning words
Select the clock for sequencer
Registered PVT bank tuning
word
Registered Acquisition bank tuning word
Registered Tracking bank tuning
word
HB PA DIV2 EN
DCORE output
LB PA DIV2 EN
DCORE output
HB BUF EN
DCORE output
LB BUF EN
DCORE output
DCO tail resistor
DCORE output
Test Divider rst a
DCORE output
Clk Gating Enable
DCORE output
Enable IIR
Output to sequencer
G word
Output to Sequencer
MEM TIME P
Output to Sequencer
MEM TIME P2A
Output to Sequencer
MEM TIME Ktdc Norm
Output to Sequencer
MEM TIME A2T
Output to Sequencer
MEM TIME GS
Output to Sequencer
MEM TIME Lambda Shift
Output to Sequencer
MEM TIME TypeCon
Output to Sequencer
MEM TIME REGDCO
Output to Sequencer
Seq Clk SEL

REG DCO P
1
7
Output to Sequencer
Input from LSD
REG DCO A
Input from LSD
REG DCO T
11
Input from LSD
September 18, 2011
162
Signal
Tv avg

Word
length
12
Input/Output/Supply
Comments
Input from LSD
Accumulated
|TDC FALL-TDC RISE| value
for Ktdc normalization
Choose SysMem input between
PHE, dPHE, Filter output and
Filter output
Word of SysMem with address of
Addr SysMem
Choose TDCMem input between
TDC RISE and TDC RISE test
Word of TDCMem with address
of Addr TDCMem
Choose TDCMem clock between
CKR and FREF
Address of SysMem word to be
read out
Address of TDCMem word to be
read out
Overow status in integral path
of loop lter
Overow status in integral path
of loop lter
reset for other blocks in DCORE
SysMem SEL
Output to SysMem
SysMemOut
32
Input from SysMem
TDCMem SEL
Output to TDCMem
TDCMemOut
Input from TDCMem
TDCMemClk SEL
Output to TDCMem
Addr SysMem
12
Output to SysMem
Addr TDCMem
12
Output to TDCMem
TypeII OVp
Input from LSD
TypeII OVn
Input from LSD
PHE OVp
PHE OVn
Core RESET
1
1
1
Input from LSD

Input from LsD
Output to sequencer
September 18, 2011
Appendix C
Register Map of the SPI Block
Table C-1: The register map of the SPI block.

Physical
ter
Regis-
Address
Corresponding
Internal Register
control
status
Meaning(Internal control signal)

Control word of SPI
Status word of SPI (2b10: Idle;
2b00: command input; 2b01: data
access)
8h00
8h00
fcw reg
Frequency command word (32 bits)
219815571
CONTROL
STATUS
6h00
6h01
FCW0 0
FCW0 1
FCW0 2
FCW0 3
SysMem 0
SysMem 1
TDCMem 0
TDCMem 1
DPAACW
6h04
6h05
6h06
6h07
6h09
6h0A
6h0B
6h0C
6h10
DPACCW
DCOdivctrl
6h11
6h12
dpa ccw reg

dco divctrl reg
DCOtailres
ClkGating
6h13
6h14
dco tailres reg

clkgating reg
ckvfrefsel
6h16
ckvfref sel reg
MEMDCOP
MEMDCOA
6h19
6h1A
memdcop reg
memdcoa reg
Default value
addr sysmem reg Read address for SysMem (12 bits)
addr tdcmem reg Read address for TDCMem(16 bits)
dpa acw reg
{DPA ACW HB[3:0],

DPA ACW LB[3:0]}
DPA CCW[3:0]
{HB BUF EN,
LB BUF EN,
HB PA DIV2 EN,
LB PA DIV2 EN, DIV 1 5 EN}
DCO tail resistor[4:0]
{Test Divider rst a, Seq Clk SEL,
Clk Gating Enable}
{SysMem SEL[1:0],
FREF SEL,
FREFdither,
PR,
CKV SEL,
PR en}
MEM DCO P[6:0]
MEM DCO A[5:0]
8b0000 0000
4b0010
5b00000
5b00000
3b100
6b11 1000
0
0
September 18, 2011
164
Physical
ter
Register Map of the SPI Block

Regis-
Address
MEMDCOT 0
MEMDCOT 1
MEMGAINP
MEMGAINA
MEMGAINT 0
MEMGAINT 1
RhoGword
Lambda
6h1B
6h1C
6h1D
6h1E
6h1F
6h20
6h21
6h22
EnableIIR
Ktdc 0
Ktdc 1
TIMEP
TIMEP2A
TIMEKtdc
TIMEA2T
TIMEGS
TIMELambda
TIMEType
TIMEREGDCO
6h23
6h24
6h25
6h28
6h29
6h2A
6h2B
6h2C
6h2D
6h2E
6h2F
September 18, 2011
Corresponding Meaning(Internal control sigInternal Reg- nal)

ister
memdcot reg
MEM DCO T[10:0]
memgainp reg
memgaina reg
MEM GAIN P[2:0]

MEM GAIN A[7:0]
memgaint reg
MEM GAIN T[10:0]
rhogword reg
lambda reg
enableiir reg
{Rho[1:0],G word[2:0]}
{Lambda3[1:0],
Lambda2[1:0],
Lambda1[1:0], Lambda0[1:0]}
Enable IIR[3:0]
Ktdc reg
Ktdc[12:0]
timep reg
timep2a reg
timektdc reg
timea2t reg
timegs reg
timelambda reg
timetype reg
timeregdco reg
MEM
MEM
MEM
MEM
MEM
MEM
MEM
MEM
TIME
TIME
TIME
TIME
TIME
TIME
TIME
TIME
P[6:0]
P2A[6:0]
Ktdc Norm[6:0]
A2T[6:0]
GS[6:0]
Lambda Shift[6:0]
TypeCon[6:0]
REGDCO[6:0]
Default value
0
4
51
655
5b11111
8b00 10 10 01
4b0011
2846
0
0
0
0
0
0
0
0
Appendix D
Figures for TDC Open-Loop Test
3.5
x 10
Bin Count: 3.26e+04
Bin Count: 3.29e+04
Bin Center: 16
Bin Edges: [15.5, 16.5]
Bin Center: 17
Bin Edges: [16.5, 17.5]
2.5
1.5
0.5
0
5
5
10
15
20
25
30
35
TDC static test. External clock frequency is 100 MHz203*0.5 KHz
40
45
Figure D-1: TDC static test histogram (fext is 100 MHz-203*0.5 kHz).
September 18, 2011
166
3.5
x 10
Bin Count: 3.33e+04

3
Bin Center: 16
Bin Edges: [15.5, 16.5]
Bin Count: 3.22e+04

Bin Center: 17
Bin Edges: [16.5, 17.5]
2.5
1.5
0.5
0
5
5
10
15
20
25
30
35
40
TDC static test. External clock frequency is 100 MHz202*0.5 KHz
Figure D-2: TDC static test histogram (fext is 100 MHz-202*0.5 kHz).
September 18, 2011
45
167
3.5
x 10
Bin Count: 3.27e+04
Bin Count: 3.28e+04
Bin Center: 15
Bin Edges: [14.5, 15.5]
Bin Center: 16
Bin Edges: [15.5, 16.5]
2.5
1.5
0.5
0
5
5
10
15
20
25
30
35
TDC static test. External clock frequency is 100 MHz+38*0.5 KHz
40
45
Figure D-3: TDC static test histogram (fext is 100 MHz+38*0.5 kHz).
September 18, 2011
168
3.5
x 10
Bin Count: 3.3e+04
Bin Count: 3.25e+04
Bin Center: 15
Bin Edges: [14.5, 15.5]
Bin Center: 16
Bin Edges: [15.5, 16.5]
2.5
1.5
0.5
0
5
0
5
10
15
20
25
30
35
40
September 18, 2011
45
169
3.5
x 10
Bin Count: 3.23e+04
Bin Center: 14
Bin Edges: [13.5, 14.5]
Bin Count: 3.32e+04

Bin Center: 15
Bin Edges: [14.5, 15.5]
2.5
1.5
0.5
0
5
5
10
15
20
25
30
35
40
45
September 18, 2011
170
3.5
x 10
Bin Count: 3.29e+04
Bin Count: 3.26e+04
Bin Center: 14
Bin Edges: [13.5, 14.5]
Bin Center: 15
Bin Edges: [14.5, 15.5]
2.5
1.5
0.5
0
5
5
10
15
20
25
30
35
40
September 18, 2011
45
Appendix E
Figures for ADPLL Top Level Simulation
September 18, 2011
172

3KDVHQRLVHRIRXWSXWVLJQDO)5()LV0+]/RFNHGIUHTXHQF\LV0+]
)&:LVUEZLVH
3KDVH1RLVH#N+]RIIVHWLVG%F+]#N+]RIIVHWLVG%F+]#N+]RIIVHWLVG%F+]#0+]RIIVHWLVG%F+]
,QWHJUDWHG'6%3KDVH1RLVHIURP.+]WR0+]LVG%F

Figure E-1: Phase noise result.fv =3800 MHz,Tinv =12.5 ps,fR =33.8688 MHz.RBW=10 kHz.
)&:LVUEZLVH

September 18, 2011
173
)&:LVUEZLVH

)&:LVUEZLVH

September 18, 2011
174
)&:LVUEZLVH

)&:LVUEZLVH

September 18, 2011
175
)&:LVUEZLVH

)&:LVUEZLVH

September 18, 2011
176
)&:LVUEZLVH

)&:LVUEZLVH

Figure E-10: Phase noise result.fv =3556.214 MHz,Tinv =12.5 ps,fR =33.8688 MHz.RBW=10 kHz.
September 18, 2011
177
)&:LVUEZLVH

)&:LVUEZLVH

September 18, 2011
178
)&:LVUEZLVH

Figure E-13: Phase noise result.fv =3803 MHz,Tinv =12.45 ps,fR =40 MHz.RBW=10 kHz.
)&:LVUEZLVH

September 18, 2011
179
)&:LVUEZLVH

)&:LVUEZLVH

September 18, 2011
180

)&:LVUEZLVH

)&:LVUEZLVH

September 18, 2011
181
)&:LVUEZLVH

)&:LVUEZLVH

September 18, 2011
182
)&:LVUEZLVH

)&:LVUEZLVH

Figure E-22: Phase noise result.fv =3560.01 MHz,Tinv =12.45 ps,fR =40 MHz.RBW=10 kHz.
September 18, 2011
183
)&:LVUEZLVH

)&:LVUEZLVH

September 18, 2011
184

[
$FTXLVLWLRQPRGHHQGV
7UDFNLQJPRGHEHJLQV
397PRGHHQGV
$FTXLVLWLRQPRGHEHJLQV

G7IV
;
<
;
<

397PRGHEHJLQV
;
<H

7LPHXV

G7IV
;
<

;
<

7LPHXV

Figure E-25: ADPLL settling transient in PVT mode and acquisiton mode(fv = 3300 MHz)
September 18, 2011

185

G7IV

;
<

7LPHXV

)UHTXHQF\
VHWWOHV

G7IV
;
<

;
<
)DVWWUDFNLQJHQGV
*HDUVKLIWLQJHYHQW

$FTXLVLWLRQPRGHHQGV
7UDFNLQJPRGHEHJLQV
;
<

7LPHXV
Figure E-26: ADPLL settling transient in tracking mode(fv = 3300 MHz)
September 18, 2011
186
[

397PRGHEHJLQV
;
<H
G7IV

$FTXLVLWLRQPRGHHQGV
7UDFNLQJPRGHEHJLQV
397PRGHHQGV

;
<
;
<

7LPHXV

G7IV

;
<
;
<

7LPHXV

Figure E-27: ADPLL settling transient in PVT mode and acquisiton mode(fv = 4050 MHz)
September 18, 2011

187

G7IV

;
<

7LPHXV

G7IV

;
<

;
<
)UHTXHQF\
VHWWOHV
;
<
)DVWWUDFNLQJHQGV
*HDUVKLIWLQJHYHQW
$FTXLVLWLRQPRGHHQGV
7UDFNLQJPRGHEHJLQV

7LPHXV
Figure E-28: ADPLL settling transient in tracking mode(fv = 4050 MHz)
September 18, 2011
188

$FTXLVLWLRQPRGHHQGV
7UDFNLQJPRGHEHJLQV
397PRGHHQGV

G7IV
;
<
;
<

;
<
397PRGHEHJLQV

7LPHXV

G7IV

;
<
;
<

7LPHXV

Figure E-29: ADPLL settling transient in PVT mode and acquisiton mode(fv
3556.214 MHz)
September 18, 2011

189

G7IV

;
<

7LPHXV

G7IV

;
<
;
<

;
<
)DVWWUDFNLQJHQGV
*HDUVKLIWLQJHYHQW
)UHTXHQF\
VHWWOHV
$FTXLVLWLRQPRGHHQGV
7UDFNLQJPRGHEHJLQV

7LPHXV

Figure E-30: ADPLL settling transient in tracking mode(fv = 3556.214 MHz)
September 18, 2011
190
September 18, 2011
Appendix F
DPA Layout
September 18, 2011
192
DPA Layout
Figure F-1: Layout of the module catip adpll dpaswitcharray
September 18, 2011
193
Figure F-2: Layout of the module catip adpll dpacaparray
September 18, 2011
194
September 18, 2011
DPA Layout
Bibliography
[1] R. Staszewski, C. Hung, K. Maggio, J. Wallberg, D. Leipold, and P. Balsara,

All-digital phase-domain TX frequency synthesizer for Bluetooth radios in 0.13m
CMOS, in Solid-State Circuits Conference, 2004. Digest of Technical Papers. ISSCC.
2004 IEEE International. IEEE, 2004, pp. 272527.
[2] P. Colantonio, F. Giannini, R. Giofr`e, M. Medina, D. Schreurs, and B. Nauwelaers,
High frequency Class E design methodologies, in Gallium Arsenide and Other Semiconductor Application Symposium, 2005. EGAAS 2005. European. IEEE, 2005, pp.
329332.
[3] R. Staszewski, J. Wallberg, S. Rezeq, C. Hung, O. Eliezer, S. Vemulapalli, C. Fernando, K. Maggio, R. Staszewski, N. Barton et al., All-digital PLL and GSM/EDGE
transmitter in 90nm CMOS, in Solid-State Circuits Conference, 2005. Digest of Technical Papers. ISSCC. 2005 IEEE International. IEEE, 2005, pp. 316600.
[4] T. Lee, The design of CMOS radio-frequency integrated circuits.
Pr, 2004.
Cambridge Univ
[5] F. Gardner, Phaselock techniques. Wiley-Blackwell, 2005.

[6] H. Darabi, H. Jensen, and A. Zolfaghari, Analysis and Design of Small-Signal Polar
Transmitters for Cellular Applications, Solid-State Circuits, IEEE Journal of, vol. 46,
no. 6, pp. 12371249, 2011.
[7] S. Drago, D. Leenaerts, B. Nauta, F. Sebastiano, K. Makinwa, and L. Breems, A
200 A Duty-Cycled PLL for Wireless Sensor Nodes in 65 nm CMOS, Solid-State
Circuits, IEEE Journal of, vol. 45, no. 7, pp. 13051315, 2010.
September 18, 2011
196
Bibliography
[8] R. Staszewski, I. Bashir, and O. Eliezer, RF built-in self test of a wireless transmitter, Circuits and Systems II: Express Briefs, IEEE Transactions on, vol. 54, no. 2,
pp. 186190, 2007.
[9] Y. Park and D. Wentzlo, An All-Digital 12 pJ/Pulse IR-UWB Transmitter Synthesized From a Standard Cell Library, Solid-State Circuits, IEEE Journal of, vol. 46,
no. 5, pp. 11471157, 2011.
[10] J. Borremans, K. Vengattaramane, V. Giannini, B. Debaillie, W. Van Thillo, and
J. Craninckx, A 86 MHz12 GHz Digital-Intensive PLL for Software-Dened Radios,
Using a 6 fJ/Step TDC in 40 nm Digital CMOS, Solid-State Circuits, IEEE Journal
of, vol. 45, no. 10, pp. 21162129, 2010.
[11] M. Zanuso, S. Levantino, C. Samori, and A. Lacaita, A Wideband 3.6 GHz Digital
Fractional-N PLL With Phase Interpolation Divider and Digital Spur Cancellation, IEEE journal of solid-state circuits, vol. 46, no. 3, pp. 627638, 2011.
[12] T. Tokairin, M. Okada, M. Kitsunezuka, T. Maeda, and M. Fukaishi, A 2.1-to2.8-GHz Low-Phase-Noise All-Digital Frequency Synthesizer With a Time-Windowed
Time-to-Digital Converter, Solid-State Circuits, IEEE Journal of, vol. 45, no. 12, pp.
25822590, 2010.
[13] E. Temporiti, C. Weltin-Wu, D. Baldi, M. Cusmai, and F. Svelto, A 3.5 GHz Wideband ADPLL With Fractional Spur Suppression Through TDC Dithering and Feedforward Compensation, Solid-State Circuits, IEEE Journal of, vol. 45, no. 12, pp.
27232736, 2010.
[14] R. Staszewski, P. Balsara, and I. ebrary, All-digital frequency synthesizer in deepsubmicron CMOS. Wiley Online Library, 2006.
[15] P. Eendrik, Time-to-Digital Converter (TDC) for WiMAX ADPLL in State-of-TheArt 40-nm CMOS, Masters thesis, Delft University of Technology, 2011.
[16] M. Lee, M. Heidari, and A. Abidi, A low-noise wideband digital phase-locked loop
based on a coarsene time-to-digital converter with subpicosecond resolution, SolidState Circuits, IEEE Journal of, vol. 44, no. 10, pp. 28082816, 2009.
[17] L. Xu, S. Lindfors, K. Stadius, and J. Ryynanen, A 2.4-GHz low-power all-digital
phase-locked loop, Solid-State Circuits, IEEE Journal of, vol. 45, no. 8, pp. 1513
1521, 2010.
[18] K. Kundert, Predicting the phase noise and jitter of PLL-based frequency synthesizers, pp. 4669, 2003.
[19] M. Perrott, CppSim System Simulator.
[20] V.-A. committee, SystemVerilogAMS.
September 18, 2011
Bibliography
197
[21] D. Leeson, A simple model of feedback oscillator noise spectrum, Proceedings of the
IEEE, vol. 54, no. 2, pp. 329330, 1966.
[22] E. Temporiti, C. Weltin-Wu, D. Baldi, R. Tonietto, and F. Svelto, A 3 GHz fractional
all-digital PLL with a 1.8 MHz bandwidth implementing spur reduction techniques,
Solid-State Circuits, IEEE Journal of, vol. 44, no. 3, pp. 824834, 2009.
[23] R. Staszewski, K. Waheed, S. Vemulapalli, P. Vallur, M. Entezari, and O. Eliezer,
Elimination of spurious noise due to time-to-digital converter, in Circuits and Systems Workshop,(DCAS), 2009 IEEE Dallas. IEEE, 2009, pp. 14.
[24] R. Staszewski, S. Vemulapalli, and K. Waheed, An all-digital oset PLL architecture, in Radio Frequency Integrated Circuits Symposium (RFIC), 2010 IEEE. IEEE,
2010, pp. 1720.
[25] K. Waheed, M. Sheba, R. Staszewski, F. Dulger, and S. Vamvakos, Spurious free
time-to-digital conversion in an ADPLL using short dithering sequences, in Custom
Integrated Circuits Conference (CICC), 2010 IEEE. IEEE, 2010, pp. 14.
[26] R. Staszewski, K. Waheed, S. Vemulapalli, F. Dulger, J. Wallberg, C. Hung, and
O. Eliezer, Spur-free all-digital PLL in 65nm for mobile phones, in Solid-State
Circuits Conference Digest of Technical Papers (ISSCC), 2011 IEEE International.
IEEE, 2011, pp. 5254.
[27] WiMAX Forum Mobile Radio Specication, WiMAX FORUM, 2011.
[28] SPI Block Guide V03.06, Motorola, Inc., 2003.
[29] Wikipedia, Serial Peripheral Interface Bus.
[30] A. Ba, Ultra-low-power Digitally-controlled Oscillator for Event-driven Transmitter,
Masters thesis, Delft University of Technology, 2011.
[31] P. Eendrik, W. Jiang, M. van de Gevel, F. Verwaal, and R. Staszewski, Timeto-digital converter (TDC) for WiMAX ADPLL in 40-nm CMOS, in Proc. of 20th
European Conference on Circuit Theory and Design (ECCTD11). IEEE, 2011, pp.
370373.
[32] N. Sokal and A. Sokal, Class E - A new class of high-eciency tuned single-ended
switching power ampliers, Solid-State Circuits, IEEE Journal of, vol. 10, no. 3, pp.
168176, 1975.
[33] F. Raab, Idealized operation of the class E tuned power amplier, Circuits and
Systems, IEEE Transactions on, vol. 24, no. 12, pp. 725735, 1977.
September 18, 2011
198
Bibliography
[34] R. Staszewski, K. Muhammad, D. Leipold, C. Hung, Y. Ho, J. Wallberg, C. Fernando,

K. Maggio, R. Staszewski, T. Jung et al., All-digital TX frequency synthesizer and
discrete-time receiver for Bluetooth radio in 130-nm CMOS, Solid-State Circuits,
IEEE Journal of, vol. 39, no. 12, pp. 22782291, 2004.
[35] P. Cruise, C. Hung, R. Staszewski, O. Eliezer, S. Rezeq, K. Maggio, and D. Leipold, A
digital-to-RF-amplitude converter for GSM/GPRS/EDGE in 90-nm digital CMOS,
in Radio Frequency integrated Circuits (RFIC) Symposium, 2005. Digest of Papers.
2005 IEEE. IEEE, 2005, pp. 2124.
[36] R. Staszewski, J. Wallberg, C. Hung, G. Feygin, M. Entezari, and D. Leipold, LMSbased calibration of an RF digitally controlled oscillator for mobile phones, Circuits
and Systems II: Express Briefs, IEEE Transactions on, vol. 53, no. 3, pp. 225229,
2006.
[37] D. Chowdhury, L. Ye, E. Alon, and A. Niknejad, An Ecient Mixed-Signal 2.4GHz Polar Power Amplier in 65-nm CMOS Technology, Solid-State Circuits, IEEE
Journal of, vol. 46, no. 8, pp. 17961809, 2011.
September 18, 2011

Thesis Verilog AMS Model

Uploaded by

Document Informationclick to expand document informationverilog ams modeling, pll, adpll, dpll, verilog-A, behavioral modeling, VLSI, semiconductor, Analog, Digital, mixed signal behavioral modeling

Document Informationclick to expand document information

Copyright:

Available Formats

Thesis Verilog AMS Model

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Thesis Verilog AMS Model

Uploaded by

Copyright:

Available Formats

MSc Thesis

ADPLL Design for WiMAX

Delft University of Technology

c 2011 by Wenlong Jiang

September 18, 2011

DELFT UNIVERSITY OF TECHNOLOGY

Dated: September 18, 2011

Dr. Ir. Nick van der Meijs

Ir. Frank Verwaal

Ir. Marcel van de Gevel

Dr. Xuefei Bai

Mr. Sejed Amir Reza Ahmadi Mehr

September 18, 2011

September 18, 2011

2 ADPLL Architecture and Building Blocks

2-3 TDC and Incrementor in ADPLL System . . . . . . . . . . . . . . . . . . . .

September 18, 2011

3 Modeling, Simulation and Analysis of ADPLL System

ADPLL Top Level

September 18, 2011

6 DPA Design for ADPLL

A-1 Verilog-AMS Code for DCO . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

C Register Map of the SPI Block

D Figures for TDC Open-Loop Test

E Figures for ADPLL Top Level Simulation

September 18, 2011

September 18, 2011

1-1 Architecture of direct conversion transmitter. . . . . . . . . . . . . . . . . . .

1-4 Simplied block diagram of a spectrum analyzer. . . . . . . . . . . . . . . . .

1-6 Eect of LO phase noise in a receiver (reciprocal mixing). . . . . . . . . . . .

1-7 Eect of LO phase noise in a transmitter. . . . . . . . . . . . . . . . . . . . .

1-10 Simplied schematic of ADPLL in [1](modulation part is not drawn). . . . . .

Bird view of ADPLL in this project. . . . . . . . . . . . . . . . . . . . . . . . 18

Construction of icker noise with single sampling clock and oversampling. . . .

3-8 TDC Transfer function for type II PLL ( = 26 , = 215 ). . . . . . . . . . .

3-13 S-domain result for ADPLL PN (fv =3800 MHz). . . . . . . . . . . . . . . . .

3-14 S-domain result for ADPLL PN (fv =3300 MHz). . . . . . . . . . . . . . . . .

Transient for zero phase restart. . . . . . . . . . . . . . . . . . .

September 18, 2011

4-6 PHE transient for CKV frequency of 3793.3156 MHz. . . . . . . . . . . . . .

4-8 ADPLL Spectrum with phase rotation (RBW=1 kHz) . . . . . . . . . . . . .

4-9 ADPLL Spectrum with phase rotation (RBW=10 kHz) . . . . . . . . . . . . .

4-10 Schematic of modulator for FREF dithering. . . . . . . . . . . . . . . . .

4-11 ADPLL Spectrum with FREF dithering (RBW=1 kHz) . . . . . . . . . . . . .

4-12 ADPLL Spectrum with FREF dithering (RBW=10 kHz) . . . . . . . . . . . .

Top Level View of ACORE in ADPLL. . . . . . .

5-5 TDC open-loop static test. . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

5-9 ADPLL settling transient in tracking mode(fv = 3800 MHz) . . . . . . . . . . 114

Top level schematic of DPA . . . . . . . . . . .

6-13 DPA eciency curve with respect to ACW (HB). . . . . . . . . . . . . . . . . 127

6-16 Phase noise performance of DPA LB. . . . . . . . . . . . . . . . . . . . . . . 129

result.fv =3800 MHz,Tinv =12.5 ps,fR =33.8688 MHz.RBW=10 kHz.172

E-26 ADPLL settling transient in tracking mode(fv = 3300 MHz) . . . . . . . . . . 185

E-28 ADPLL settling transient in tracking mode(fv = 4050 MHz) . . . . . . . . . . 187

September 18, 2011

September 18, 2011

1-1 Specication for the WiMAX ADPLL System. . . . . . . . . . . . . . . . . .

2-1 Frequency resolution for PB/AB/TB in DCO core. . . . . . . . . . . . . . . .