Copyright

by
Long Chen
2016
 The Dissertation Committee for Long Chen
certifies that this is the approved version of the following dissertation:

Design Techniques for Low-power SAR ADCs in Nano-scale

CMOS Technologies

Committee:

Nan Sun, Supervisor

T. R. Viswanathan

David Z. Pan

Michael Orshansky

Eric Soenen
Design Techniques for Low-power SAR ADCs in Nano-scale
CMOS Technologies

by

Long Chen, B.S.; M.S.E.

DISSERTATION
Presented to the Faculty of the Graduate School of
The University of Texas at Austin
in Partial Fulfillment
of the Requirements
for the Degree of

DOCTOR OF PHILOSOPHY

THE UNIVERSITY OF TEXAS AT AUSTIN

May 2016
Dedicated to my parents and my wife.
 Acknowledgments

First and most importantly, I would like to express my most sincere grati-
tude to my supervisor, Dr. Nan Sun. Without his insightful guidance and continuous
support, I would have never accomplished my research work and thesis. I have been
benefited a lot from his vision and perception. I would also like to thank my com-
mittee members, Dr. T. R. Viswanathan, Dr. David Z. Pan, Dr. Michael Orshansky,
and Dr. Eric Soenen, for their valuable advice and discussions. I am truly grateful to
Dr. Eric Soenen for providing me an opportunity to intern in TSMC Austin. It was
a wonderful experience to work there in 2015. I have been honored and thankful to
many excellent labmates in Sun research group: Kareem Ragab, Arindam Sanyal,
Yeonam Yoon, Wenjuan Guo, Ji Ma, Peijun Wang, Matther Schueler, Xiyuan Tang,
Sungjin Hong, Shaolan Li, Manzur Rahman, Miguel Gandara, Jeonggoo Song, and
Haoyu Zhuang. Special thanks to Arindam Sanyal, Kareem Ragab, Xiyuan Tang, Ji
Ma, Yeonam Yoon, Manzur Rahman and Matther Schueler for their collaboration
in my research. Thanks to Wei-Gi Ho for his insightful technical discussions. I
would also like to thank my friends in Austin for making my life more fun. I am
greatly indebted to my parents for their selfless support and care. Last but not the
least, I am truly fortunate to have my wife Pingshuai Cao in my life. There is no
way I can reach this point without her support, faith and love.

v
 Design Techniques for Low-power SAR ADCs in Nano-scale
CMOS Technologies

Long Chen, Ph.D.

The University of Texas at Austin, 2016

Supervisor: Nan Sun

This thesis presents low power design techniques for successive approxi-
mation register (SAR) analog-to-digital converters (ADCs) in nano-scale CMOS
technologies. Low power SAR ADCs face two major challenges especially at high
resolutions: (1) increased comparator power to suppress the noise, and (2) increased
DAC switching energy due to the large DAC size. To improve the comparator’s
power efficiency, a statistical estimation based comparator noise reduction tech-
nique is presented. It allows a low power and noisy comparator to achieve high
signal-to-noise ratio (SNR) by estimating the conversion residue. A first prototype
ADC in 65nm CMOS has been developed to validate the proposed noise reduc-
tion technique. It achieves 4.5 fJ/conv-step Walden figure of merit and 64.5 dB
signal-to-noise and distortion ratio (SNDR). In addition, a bidirectional single-side
switching technique is developed to reduce the DAC switching power. It can reduce
the DAC switching power and the total number of unit capacitors by 86% and 75%,
respectively. A second prototype ADC with the proposed switching technique is

vi
designed and fabricated in 180nm CMOS technology. It achieves an SNDR of 63.4
dB and consumes only 24 µW at 1MS/s, leading to a Walden figure of merit of 19.9
fJ/conv-step.

This thesis also presents an improved loop-unrolled SAR ADC, which works
at high frequency with reduced SAR logic power and delay. It employs the bidi-
rectional single-side switching technique to reduce the comparator common-mode
voltage variation. In addition, it uses a Vcm -adaptive offset calibration technique
which can accurately calibrate comparator’s offset at its operating Vcm . A proto-
type ADC designed in 40nm CMOS achieves 35 dB at 700 MS/s sampling rate and
consumes only 0.95 mW, leading to a Walden figure of merit of 30 fJ/conv-step.

vii
 Table of Contents

Acknowledgments v

Abstract vi

List of Tables x

List of Figures xi

Chapter 1. Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

Chapter 2. Statistical Estimation Based Noise Reduction 6

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Proposed Statistical Estimation Based Noise Reduction Technique:
Basic Idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Proposed Statistical Estimation Based Noise Reduction Technique:
Mathematical Formulation . . . . . . . . . . . . . . . . . . . . . 15
2.3.1 Estimator based on averaging . . . . . . . . . . . . . . . 17
2.3.2 Maximum likelihood estimator (MLE) . . . . . . . . . . 18
2.3.3 Bayes estimator . . . . . . . . . . . . . . . . . . . . . . 22
2.3.4 Estimator performance comparison . . . . . . . . . . . . 26
2.3.5 Comparison to oversampling and analog scaling . . . . . 29
2.4 Prototype ADC Design . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.1 Detailed circuit schematics . . . . . . . . . . . . . . . . . 31
2.4.2 Measurement Results . . . . . . . . . . . . . . . . . . . 36

viii
Chapter 3. Bidirectional Single-side Switching Technique 43
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.2 Effect of Comparator Common-Mode Variation . . . . . . . . . . 47
3.3 Proposed BSS switching technique . . . . . . . . . . . . . . . . . 54
3.3.1 Comparator input common-mode variation reduction . . . 54
3.3.2 Comparator input common-mode voltage optimization . . 58
3.3.3 Capacitance reduction . . . . . . . . . . . . . . . . . . . 60
3.3.4 Reduced DAC reference energy . . . . . . . . . . . . . . 60
3.4 Prototype ADC Design . . . . . . . . . . . . . . . . . . . . . . . 62
3.4.1 Detailed circuit schematics . . . . . . . . . . . . . . . . . 62
3.4.2 Measurement Results . . . . . . . . . . . . . . . . . . . 66

Chapter 4. Low-power and High-speed Single-channel SAR ADC 73

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.2 Proposed SAR ADC Architecture . . . . . . . . . . . . . . . . . 76
4.3 Prototype ADC Design . . . . . . . . . . . . . . . . . . . . . . . 79
4.3.1 Detailed circuit schematics . . . . . . . . . . . . . . . . . 79
4.3.2 Measurement results . . . . . . . . . . . . . . . . . . . . 85

Chapter 5. Conclusion and Future Directions 91

5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . 92

Appendix 94

Appendix 1. List of publications 95

Bibliography 97

Vita 104

ix
 List of Tables

2.1 Measured performance summary . . . . . . . . . . . . . . . . . . . 42

3.1 Measured performance summary . . . . . . . . . . . . . . . . . . . 71

3.2 Performance comparison . . . . . . . . . . . . . . . . . . . . . . . 72

4.1 Performance comparison . . . . . . . . . . . . . . . . . . . . . . . 90

x
 List of Figures

1.1 FOM versus SNDR plot for recently published SAR ADCs in ISSCC
and VLSI conferences. . . . . . . . . . . . . . . . . . . . . . . . . 2

2.1 Diagram for a b-bit SAR ADC. . . . . . . . . . . . . . . . . . . . . 13

2.2 Simplified SAR ADC model during the LSB comparison. . . . . . . 16
2.3 MSE of V̂res,avg for various N and (a) σ = 0.5 LSB, (b) σ = 2 LSB,
and (c) σ = 1 LSB. . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Value of V̂res,M LE as a function of N and k. . . . . . . . . . . . . . 22
2.5 MSE of V̂res,M LE for different Vres . . . . . . . . . . . . . . . . . . . 23
2.6 Histogram of Vres for (a) a −6-dBFS sinusoidal input, (b) a −20-
dBFS sinusoidal input, and (c) a Gaussian random input with stan-
dard deviation of 10% ADC full range. . . . . . . . . . . . . . . . . 24
2.7 Prior distribution g(Vres ) and posterior distribution g(Vres |k) for
N = 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.8 V̂res,BE as a function of N and k. . . . . . . . . . . . . . . . . . . . 27
2.9 MSE versus x for V̂res,BE . . . . . . . . . . . . . . . . . . . . . . . 27
2.10 MSE versus N for a normal distributed x with σ = 1LSB. . . . . . 28
2.11 Simulated SNR versus N with ±10% variations in the comparator
noise σ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.12 Proposed SAR ADC architecture. . . . . . . . . . . . . . . . . . . 32
2.13 DAC capacitor array floor plan. . . . . . . . . . . . . . . . . . . . . 33
2.14 Schematic of (a) the clock booster; (b) the comparator; and (c) the
timing diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.15 5-bit counter to count k. . . . . . . . . . . . . . . . . . . . . . . . . 35
2.16 Die micrograph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.17 Measured DNL and INL. . . . . . . . . . . . . . . . . . . . . . . . 37
2.18 Dout distribution with and w/o estimation at Vin = 0. . . . . . . . . 38
2.19 Measured 214 -point ADC output spectrum with 96-kHz input. . . . 39
2.20 Measured SNR versus input amplitudes. . . . . . . . . . . . . . . . 40

xi
2.21 Measured SNR improvement versus various comparator noise σ. . . 41
2.22 FOM versus SNDR plot for this work and recently published ADCs
in ISSCC and VLSI conferences. . . . . . . . . . . . . . . . . . . . 42

3.1 Schematic of a strong-arm latch comparator. . . . . . . . . . . . . . 48

3.2 Transient behavior of a strong-arm latch comparator. . . . . . . . . 49
3.3 Comparator schematic for (a) the pre-amplification phase and (b)
the latch regeneration phase. . . . . . . . . . . . . . . . . . . . . . 50
3.4 Simulated comparator performance: (a) G and gm /ID ; (b) noise
and offset; (c) linear fitting for noise and offset with 1/G2 ; (d) re-
solve time; (e) power and (f) F OMcomp . . . . . . . . . . . . . . . . 53
3.5 Proposed bidirectional single-side switching technique. . . . . . . . 55
3.6 Comparator input common-mode variation for different switching
techniques. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.7 Simulated SNDR with comparator offset variation. . . . . . . . . . 58
3.8 Simulated SNDR with both comparator offset variation and a re-
dundant capacitor after the 6th MSB capacitor. . . . . . . . . . . . . 59
3.9 Simulated SNDR with comparator offset variation, noise and a re-
dundant capacitor after the 6th MSB capacitor. . . . . . . . . . . . . 59
3.10 DAC reference energy for an 11-bit SAR ADC with different output
codes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.11 Architecture of the 11-bit prototype SAR ADC. . . . . . . . . . . . 63
3.12 DAC capacitor array floor plan. . . . . . . . . . . . . . . . . . . . . 65
3.13 Schematic for (a) the SAR logic and (b) the DAC switch. . . . . . . 66
3.14 Die micrograph and zoomed view. . . . . . . . . . . . . . . . . . . 67
3.15 Measured DNL and INL. . . . . . . . . . . . . . . . . . . . . . . . 68
3.16 Measured 65536-point FFT spectrum with VDD = 1V. . . . . . . . . 68
3.17 Measured SNDR and SFDR versus input amplitudes. . . . . . . . . 69
3.18 Measured SNDR and SFDR versus input frequencies. . . . . . . . . 69
3.19 Measured Dout distribution with Vcmi = 0.4V and Vcmi = 0.9V . . . 70
3.20 Measured Dout standard deviation with different Vcmi at Vin = 0V. . 71

4.1 (a) Architecture and (b) timing diagram of the proposed SAR ADC. 76
4.2 Dynamic comparator with varactor loading. . . . . . . . . . . . . . 80
4.3 Simulated comparator offset and decision time at different Vcm . . . . 80

xii
4.4 BSS switching scheme. . . . . . . . . . . . . . . . . . . . . . . . . 82
4.5 Simulated time-domain waveforms for comparator inputs Vin+/Vin-
, its common mode voltage Vcm and its differential mode voltage Vdm . 83
4.6 Probability of MSB comparator output being ‘1’ versus its calm
with calp fixed at 1.2V. . . . . . . . . . . . . . . . . . . . . . . . . 84
4.7 Schematic of (a) bootstrapped sampling switch and (b) dynamic OR
gate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.8 Chip micrograph. . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.9 Measured DNL/INL before calibration (dotted line) and after cali-
bration (solid line). . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.10 Measured 1024p FFT spectrum before calibration and after calibra-
tion with 600MS/s sampling rate and 100MHz input. . . . . . . . . 88
4.11 Measured 1024p FFT spectrum with 700MS/s sampling rate and
300MHz input. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.12 Measured SNDR/SFDR under different sampling rates with 50MHz
input (left), under different input frequencies with 700MS/s sam-
pling rate (middle), and under different input amplitudes with 700MS/s
sampling rate and 300MHz input (right). . . . . . . . . . . . . . . . 90

xiii
 Chapter 1

Introduction

1.1 Motivation

The Internet of Things (IoT) is attracting lots of attention from semicon-

ductor industry. It is changing the way how people control their surroundings and
environments. In order to provide multiple functions at low cost, a typical IoT
system-on-a-chip (SoC) commonly integrates sensor interfaces, analog-to-digital
converters (ADCs), local digital processor, embedded memory and multi-protocol
wireless transceivers. The integration of ADCs with adequate performance is criti-
cal in any IoT SoCs. Since many sensor interfaces and ADCs are battery-powered,
low power operation becomes paramount.

Successive approximation register (SAR) ADC is a popular choice due to its

simple architecture and short development cycle. It is more digital friendly and does
not require any opamps compared to pipeline ADCs or sigma-delta ADCs. SAR
ADC can achieve excellent power efficiency of less than one femtojoule (fJ)/conv-
step at low resolution with a target effective number of bits (ENOB) below 10 bits.
This can be visualized from Fig. 1.1, which shows the Walden figure of merit
(FOM) versus the corresponding signal to noise and distortion ration (SNDR) for
recently published SAR ADCs in ISSCC and VLSI conferences. The data for the

1
figure is taken from the survey made available by Dr. Boris Murmann (http://
web.stanford.edu/∼murmann/adcsurvey.html). However, some wire-
less sensor nodes and biomedical devices [Verma and Chandrakasan [2007]; Van Helleputte
et al. [2012]] require low power ADCs with resolution greater than 10 bits. It is
nontrivial to maintain such good power efficiency when extending the ENOB be-
yond 10 bits due to two design challenges. First, designing a low power comparator
with low noise is a main challenge since traditional analog scaling requires four
times the comparator power for every 1-bit reduction in noise. This places a steep
power-noise trade-off. The second challenge for a high-resolution SAR ADC is its
exponentially growing capacitive DAC size, which results in greatly increased DAC
switching power. Thus, it is highly desirable to develop power efficient techniques
to reduce comparator noise and DAC switching power.
1.E+04
10-b ENOB

1.E+03
FOM [fJ/conv-step]

1.E+02

1.E+01

Prototype 1
1.E+00 Prototype 2
ISSCC 2010-2015
VLSI 2010-2015
1.E-01
10 20 30 40 50 60 70 80 90 100 110
SNDR @ f in,hf [dB]

Figure 1.1: FOM versus SNDR plot for recently published SAR ADCs in ISSCC
and VLSI conferences.

2
 In this thesis, a statistical estimation based technique is presented which
can efficiently reduce comparator noise for SAR ADCs. Instead of designing an
accurate comparator with large power consumption, the proposed technique utilizes
the low power and noisy comparator and estimates the conversion residue by firing
the comparator multiple times for LSB bit comparison. The conversion residue can
be estimated by exploring the repeated comparison results. The estimated residue
is then subtracted from the ADC output to accurately represent the input. The key
challenge here is to build an accurate estimator. It has been proved in statistics
that the Bayes estimator achieves lowest estimation error. Thus, Bayes estimator
is chosen for the prototype SAR ADC design. The prototype ADC 1 has been
implemented in 65nm CMOS and measured in the lab. The measured FOM of the
first prototype ADC is 4.5 fJ/conv-step with 64.5dB SNDR, shown in Fig. 1.1. It
suits well for the low power applications with low conversion speed.

Another challenge for low power and high resolution SAR ADC design is
to reduce the DAC switching energy. The DAC is commonly implemented with
binary capacitors in SAR ADCs. To suppress the sampling kT/C noise and provide
good matching accuracy in high resolution designs, the capacitive DAC needs to
be large, leading to a significant power consumption out of total ADC power. This
work proposes a bidirectional single-side switching (BSS) technique, which can
save 86% DAC switching power compared to conventional switching technique.
Moreover, the comparator input common mode variation is reduced in the technique
compared to widely used monotonic switching technique [Liu et al. [2010a]]. The
prototype ADC 2 with BSS switching technique is designed in 180nm CMOS. The

3
measured FOM is 19.9 fJ/conv-step with 63.4dB SNDR. It achieves the state-of-art
power efficiency shown in Fig. 1.1, given its relatively old 180nm process.

Modern high speed serial link transceivers and communication systems em-
ploy high speed and low resolution ADCs. Although SAR ADC is simple and
power efficient, it can not run at a high speed compared to flash ADCs and pipeline
ADCs, mainly due to its serial conversion algorithm. This work investigates a
loop-unrolled architecture with multiple comparators to increase the speed of SAR
ADCs. The loop-unrolled architecture is based on the work of [Jiang et al. [2012]].
However, the linearity in [Jiang et al. [2012]] is not good due to large compara-
tor offset mismatches. This work employs the BSS switching technique and a
Vcm -adaptive calibration technique to effectively reduce the comparator offset mis-
matches and improve the linearity. A third prototype 6-bit loop-unrolled SAR ADC
is implemented in 45nm CMOS and the measured SNDR is 34.8 dB at a sampling
rate of 700 MS/s.

1.2 Organization

Chapter 2 of the thesis presents the low power SAR ADC with statistical
estimation based noise reduction. It also includes the measurement results of the
first prototype designed in 65nm CMOS technology. Chapter 3 presents the bidi-
rectional single-side switching technique to reduce the DAC switching power. The
measurement results for the second prototype designed in 180nm CMOS technol-
ogy are also discussed. Chapter 4 presents the third 40nm high speed SAR ADC
prototype with improved loop-unrolled architecture. The measured results are also

4
shown in Chapter 4. The conclusion is drawn in Chapter 5.

5
 Chapter 2

Statistical Estimation Based Noise Reduction

This chapter1 presents a power-efficient noise reduction technique for SAR

ADCs based on the statistical estimation theory. It suppresses both comparator
noise and quantization error by accurately estimating the ADC conversion residue.
It allows a high SNR to be achieved with a noisy low-power comparator and a rel-
atively low resolution DAC. The proposed technique has low hardware complexity,
requiring no change to the standard ADC operation except for repeating the LSB
comparisons. Three estimation schemes are studied and the optimal Bayes estima-
tor is chosen for a prototype 11-bit ADC in 65nm CMOS. The measured SNR is
improved by 7dB with the proposed noise reduction technique.

This chapter is organized as follows: an introduction of existing techniques

is first presented. The basic idea of proposed noise reduction technique is studied
next. Then the formulation is analyzed. Finally, a prototype ADC implementation
is presented, followed by its measurement.

1
This chapter is a partial reprint of the publication: Long Chen, Xiyuan Tang, Arindam Sanyal,
Yeonam Yoon, Jie Cong, and Nan Sun, “A 10.5-b ENOB 645nW 100kS/s SAR ADC with statistical
estimation based noise reduction,” IEEE Custom Integrated Circuit Conference (CICC), 2015, pp.
1–4. I thank all the co-authors for their valuable advice in designing and testing of the prototype.

6
2.1 Introduction

Rapid advances in wireless sensor nodes and biomedical devices place de-
manding requirements on low power and high resolution analog-to-digital convert-
ers (ADCs) [Verma and Chandrakasan [2007]; Van Helleputte et al. [2012]]. Suc-
cessive approximation register (SAR) ADC is a popular choice due to its simple
architecture and short development cycle. It consists of only a capacitive DAC, a
comparator, and a digital SAR logic. Since it is highly digital, it scales well with
technology, and consumes both low power and low area in advanced CMOS pro-
cesses. By using a low power but relatively noisy dynamic comparator, a SAR ADC
does not consume any static current. As a result, it can achieve an excellent power
efficiency of only a few femtojoule (fJ) per conversion step, especially at low resolu-
tion with a target effective number of bits (ENOB) below 10-bit [Tai et al. [2014a];
Harpe et al. [2014]]. Despite many advantages of SAR ADCs, it is nontrivial to
design a high resolution SAR ADC and maintain a high power efficiency when ex-
tending the resolution beyond 10 bits. To reach higher signal-to-noise ratio (SNR),
the comparator noise needs to be reduced. This can be accomplished by brute-force
analog scaling, which is to increase the transistor sizes and power. However, this re-
quires four times the comparator power for every 1-bit reduction in noise, which is
a steep power-noise trade-off. The other challenge for a high-resolution SAR ADC
is its exponentially growing capacitive DAC size, which results in greatly increased
DAC power. In addition, it makes a SAR ADC hard to drive. Facing these chal-
lenges, it is highly desirable to develop a more efficient way to increase SAR ADC
resolution without significantly increasing the comparator power and the DAC size.

7
 There are several prior SAR ADC works that aim to reduce the comparator
power and noise. The technique of [Giannini et al. [2008]] arranges two compara-
tors with different noise and power levels. For a b-bit SAR ADC, it uses a low-
power but high-noise comparator during the first (b − 1) MSB comparison cycles.
To tolerate the resulting comparison errors due to the high noise, a redundant LSB
bit is added. A high-power but low-noise comparator is used only for the LSB and
the redundant bit. As a result, the overall ADC noise is almost the same as the case
where the high-power low-noise comparator is used for every comparison cycle,
but the total comparator power is greatly reduced. The limitation of this approach
is that the offsets of the two comparators need to be tightly matched, which is non-
trivial at high resolution. To address the offset mismatch issue, the majority voting
technique is developed [Harpe et al. [2013]]. It uses only one low-power high-noise
comparator. When low comparator noise is needed at critical decision point with
a small comparator input voltage, the comparator is fired multiple times and the
decision is made via majority voting. This technique does not have the comparator
offset mismatch problem of [Giannini et al. [2008]], but it requires a carefully tuned
metastability detector to sense the comparator input voltage, resulting in increased
design complexity. Also, its conversion speed is reduced due to multiple compar-
isons needed. A similar technique using an optimized vote allocation is reported in
[Ahmadi and Namgoong [2013]]. It obviates the need for a metastability detector,
however, at the cost of further increased number of comparison cycles.

The majority voting technique of [Harpe et al. [2013]; Ahmadi and Nam-
goong [2013]] can effectively reduce the comparator noise and power, but they do

8
not make full use of the information embedded in the voting results. It only cares
about whether there are more ‘1’s or more ‘0’s, and uses it only to make a 1-bit ma-
jority decision. It does not take advantage of the detailed distributions of ‘1’s and
‘0’s, but there is valuable information there that can be exploited. Let us consider an
example that the comparator is fired in total 15 times at a given comparison cycle.
Let us assume that there are two cases: one is that there are eight ‘1’s and seven
‘0’s, and the other is that there are fifteen ‘1’s and no ‘0’s. Since there are more
‘1’s, there is no difference for majority voting, but there is extra information. The
first case indicates that the comparator input is very close to 0, while the second
case of fifteen straight ‘1’s means that its comparator input is greater than zero by
at least several comparator noise standard deviations. As will be shown later, this
extra information can be used to reduce not only the comparator noise but also the
quantization error set by the DAC resolution.

This chapter presents a statistical estimation based technique that can reduce
both the comparator noise and the quantization error for SAR ADCs. Its circuit im-
plementation is simple. It does not require any change to the standard SAR ADC
operation except for repeating the last LSB comparison for multiple times [Chen
et al. [2015]]. It exploits all the information embedded in the comparator output
distribution, not just making a binary majority decision for the LSB bit as in [Harpe
et al. [2013]; Ahmadi and Namgoong [2013]], but to estimate the magnitude of the
comparator input voltage. A useful property of a SAR ADC is that the comparator
input voltage is the ADC conversion residue. If we are able to estimate the residue,
we can subtract it from the ADC output to increase the ADC resolution. Note that

9
this reduces not only the comparator noise, but also the quantization error, which
is impossible with prior works [Giannini et al. [2008]; Harpe et al. [2013]; Ahmadi
and Namgoong [2013]]. Although a ‘1’-bit high-noise comparator cannot provide
an accurate estimation for its input if used only once, we can improve the estima-
tion accuracy by repeating the comparison for multiple times and examining the
number of comparator outputs being ‘1’ or ‘0’. It turns out that the estimation of
an unknown value via multiple noisy binary tests is a classic statistical estimation
problem [Casella and Berger [1990]]. Thus, we can directly borrow the concepts
and theories from statistics to solve our estimation problem. Specifically, this chap-
ter discusses three widely used statistical estimators: the averaging based estimator,
the maximum likelihood estimator (MLE), and the Bayes estimator. Out of them,
the Bayes estimator achieves the lowest estimation error, and thus, is chosen for
our proposed SAR ADC. Note that the estimator is essentially a mapping from the
comparator output distribution to a digital estimate for the comparator input, which
can be easily implemented using a pre-computed look-up table.

This chapter introduces the statistical estimation theory to the field of ADC
design and offers a new perspective. In a broad sense, any form of analog-to-digital
conversion, regardless of its architecture, can be considered as a statistical estima-
tion problem, as its entire operation is to estimate (or convert) an unknown analog
signal by performing a series of comparison tests. The theories from statistics are
helpful when we deal with multiple noisy comparator outputs, as in our case. The
concept of statistical estimation has been exploited in prior studies. For example,
the stochastic flash ADC of [Weaver et al. [2010]] takes advantages of random off-

10
sets in an array of comparators to obtain a 6-bit estimation of its input. This idea
has also been adapted to build a stochastic time-to-digital converter (TDC) [Kratyuk
et al. [2009]]. Recently, the stochastic flash ADC has been used as a back-end of a
SAR ADC [Verbruggen et al. [2015]]. Though independently developed and pub-
lished on close dates, the work of [Verbruggen et al. [2015]] shares a similar big
picture as our work of [Chen et al. [2015]] as it uses multiple comparison results to
estimate the SAR conversion residue. However, there are two key advantages of our
work both in the choice of the estimator and the circuit architecture. First, the work
of [Verbruggen et al. [2015]] uses MLE, which is a sub-optimal choice compared to
the Bayes estimator used in our work. Second, it arranges 16 different comparators
for the LSB estimation. Their offsets need to be very carefully calibrated, which
is a big design and operation burden especially for high resolution applications and
considering process, voltage, and temperature (PVT) variations. By contrast, we
just re-use the original comparator in the SAR ADC, and thus, do not have the off-
set mismatch problem. Our limitation compared to [Verbruggen et al. [2015]] is
reduced conversion speed as it requires a larger number of comparison cycles. Yet,
for the intended low-speed sensor applications, the speed penalty is a minor issue.

To validate the proposed statistical estimation based noise reduction tech-

nique, a prototype 11-bit SAR ADC is implemented in 65nm CMOS. Using the
proposed technique, the measured SNR is improved by 7 dB, which matches well
with the theoretical prediction. Overall, the prototype ADC achieves an ENOB of
10.5-bit at 100kS/s while consuming 0.6µW of power from a 0.7V supply.

11
2.2 Proposed Statistical Estimation Based Noise Reduction Tech-
nique: Basic Idea

Fig. 2.1 shows the simplified block diagram of a single-ended b-bit bottom-
plate sampled SAR ADC. A SAR ADC has a property that its conversion residue
Vres is readily available at the comparator input. We can derive the following rela-
tionship among the ADC input Vin , output Dout , and Vres :

Dout = Vin + ns + Vres (2.1)

where ns represents the kT /C noise directly added to Vin during the sampling
phase. Here, for simplicity of presentation, we have made the following assump-
tions that do not undermine the practicality of the proposed technique: a) we assume
the parasitic capacitor CP = 0. Since its effect is simply attenuating Vres , it can be
easily added in (2.1) by applying a scaling factor to Vres ; b) the comparator offset
is assumed to be zero, as it does not affect the ADC SNR; and c) we ignore the ef-
fect of capacitor mismatch. In practice, if capacitor mismatch is a problem, classic
mismatch calibration technique, such as [Lee et al. [1984]], can be applied jointly
with the proposed technique.

As shown in (2.1), the ADC conversion error, defined as (Dout − Vin ), con-
sists of ns and Vres . To reduce ns , the only option is to increase the DAC capacitance
CDAC , which is not the focus of this work. In a SAR ADC, its conversion error is
typically dominated by Vres , not ns . For example, for a 12-bit SAR ADC with 2V
peak-to-peak differential input swing, 420 fF of single-side DAC capacitance is al-
ready sufficient to suppress the sampling noise ns to be less than the quantization

12
 Figure 2.1: Diagram for a b-bit SAR ADC.

error. Note that 420 fF of total capacitance means that the unit capacitor size is only
0.1 fF, which is smaller than what most SAR ADCs use and what matching requires
[Tripathi and Murmann [2013]]. In practice, ns is usually smaller than Vres , which
is set by the comparator noise and the DAC LSB size. This is especially true for
recent SAR ADCs that use low power dynamic comparators without having a pre-
amplifier [Harpe et al. [2013]; Lee et al. [2014]; Yip and Chandrakasan [2013]]. As
a result, our work focuses on improving ADC SNR by reducing Vres .

Vres consists of three parts: the ADC quantization error (i.e., DAC LSB
step), the comparator noise, and the DAC noise. If the ADC does not have any
comparator noise or DAC noise, Vres is simply the ADC quantization error and
is uniformly distributed between ±1/2 LSB. By contrast, in the presence of large
comparator noise (in a SAR ADC the comparator noise is typically much large
than the DAC noise), Vres is Gaussian distributed with a standard deviation close to

13
the comparator noise. To reduce Vres , a straightforward way is to use a low-noise
comparator and a high-resolution DAC; however, both lead to greatly increased
circuit power.

This chapter proposes a simple and power efficient way to reduce Vres by
using statistical estimation. The core idea is that if we can estimate the value of
Vres , denoted as V̂res , we can increase the ADC SNR by subtracting V̂res from Dout
as:
∗
Dout = Dout − V̂res = Vin + ns + (Vres − V̂res ) (2.2)

∗
which shows that the accuracy of the new ADC output Dout is limited not by Vres
but by the estimation error (Vres − V̂res ). An interesting note is that if the estimation
∗
error can be made small, the resolution of Dout can even surpass the limit set by the
ADC quantization error, because the subtraction of V̂res reduces both comparator
noise and quantization error. This implies that the proposed technique can actually
permit, for example, a SAR ADC with a b-bit DAC array to reach more than b-bit
resolution.

Now with core idea captured in (2.2), the key question to answer is how
we can estimate Vres . We prefer performing estimation without incurring large
hardware and power cost. Since Vres is readily available at the comparator input,
we propose to use the original noisy SAR comparator to estimate Vres . This may
appear counterintuitive because the comparator can only provide a binary decision
and its output is error-prone due to its high noise. Certainly 1-time comparison is
insufficient. What we propose is to simply repeat the LSB comparison for a total of
N times and estimate Vres by examining the number of ‘1’s, denoted as k. This is

14
doable because the comparator output carries information on its input. Qualitatively
speaking, if k = N , we know that Vres is most likely a large positive value; if k = 0,
Vres is most likely negative with a large magnitude; and if k = N/2, Vres is highly
probable to be close to zero.

An intriguing side note is that the estimation of Vres is actually enabled

by the comparator noise. If the comparator does not have any noise, its output
would be straight ‘1’s or ‘0’s, and thus, we cannot extract any information about
the magnitude of Vres , except for its sign. Having noise in the comparator actually
enables us to improve our estimation accuracy on Vres . This phenomenon is actually
an example of stochastic resonance, which is observed in a nonlinear system where
the presence of a small amount of noise can actually improve the overall system
SNR [Harmer et al. [2002]]. Such behavior is impossible in a linear system. In the
SAR ADC, the 1-bit comparator is both nonlinear and noisy, which exactly matches
the requirement for stochastic resonance.

2.3 Proposed Statistical Estimation Based Noise Reduction Tech-

nique: Mathematical Formulation

After presenting the basic idea of our proposed noise reduction technique,
we now quantitatively answer what is the optimum choice of the estimator V̂res
given the number of LSB comparisons N and the number of ‘1’s k. Let us focus
our attention on the repeated LSB comparison, whose model is shown in Fig. 2.2.
d0,i presents the i-th LSB comparison result, where i is from 1 to N . nc represents
the total noise referred to the comparator input. It includes both the comparator

15
noise and the DAC noise. It is typically dominated by the comparator noise. nc is
zero mean, and we denote its standard deviation as σ in the following discussion.

nc
+ { d 0,i} ={0,1,0,...,1}
+
V res -
- Fire N times,
observe k '1'

Figure 2.2: Simplified SAR ADC model during the LSB comparison.

Our goal is to form an estimator V̂res that minimizes the mean square error
(MSE), defined as:

[ ]
MSE = V ar(Vres − V̂res ) = E (Vres − V̂res )2 (2.3)

where V ar and E stands for statistical variance and expectation, respectively [Casella
and Berger [1990]].

It turns out that the estimation of an unknown value out of a series of noisy
binary tests is a classic statistical estimation problem [Casella and Berger [1990]].
Therefore, we directly borrow the concepts and theories from statistics to solve
our estimation problem. Specifically, we discuss three widely used statistical esti-
mators: the simple averaging based estimator, the maximum likelihood estimator
(MLE), and the Bayes estimator. They all can be implemented as digital look-up
tables with similar hardware costs.

16
2.3.1 Estimator based on averaging

One straightforward way to define V̂res is +1 LSB for all straight ‘1’s, −1
LSB for all straight ‘0’s, and performing linear interpolation for other values of k.
Mathematically speaking, this definition is as follows:

2k − N
V̂res,avg = · LSB (2.4)
N

where V̂res,avg linearly increases with k, and is ±1 LSB for k = 0 and N ,

respectively. Although V̂res,avg is easy to construct, it has several drawbacks. First,
because V̂res,avg is bounded by ±1 LSB, it cannot accurately approximate Vres that
is outside of that range due to comparator noise. This can be clearly observed from
Fig. 2.3 that plots the MSE of V̂res,avg as a function of Vres , N , and the comparator
noise σ. If the comparator noise σ is large, there is a high probability for |Vres | > 1
LSB, and V̂res,avg does not work well.

3 3 3
N=2 N=2 N=2
MSE of V̂res,avg [ 2]

MSE of V̂res,avg [ 2]

MSE of V̂res,avg [ 2]

N=4 N=4 N=4

N=8 N=8 N=8
2 2 2
N=16 N=16 N=16
N=32 N=32 N=32
N=64 N=64 N=64
1 1 1

0 0 0
-2 0 2 -2 0 2 -2 0 2
V res [LSB] V res [LSB] V res [LSB]
(a) (b) (c)

Figure 2.3: MSE of V̂res,avg for various N and (a) σ = 0.5 LSB, (b) σ = 2 LSB,
and (c) σ = 1 LSB.

Second, the shape of its MSE curve varies substantially with the comparator
noise σ. The reason is that the comparator noise affects the value of k, but such

17
influence is not captured in (2.4). For example, for a given nonzero Vres , if the
comparator noise is small, the majority of the comparator outputs would be either
all ‘1’s (k = N ) or all ‘0’s (k = 0), leading to V̂res,avg close to ±1 LSB. This results
in a low MSE for Vres = ±1 LSB [see Fig. 2.3(a)], but a large MSE for Vres close
to 0. On the other hand, if the comparator noise is large, it is highly likely that half
of the comparator outputs are ‘1’ (k = N/2), leading to V̂res,avg close to 0. This
yields a small MSE for Vres = 0, but a large MSE elsewhere [see Fig. 2.3(b)]. The
overall best performance for V̂res,avg in terms of a small and relatively flat MSE is
obtained only when the comparator noise σ is close to 1 LSB [see Fig. 2.3(c)]. This
limits its applicability.

Furthermore, although its MSE decreases as N increases, the region with

a small MSE becomes narrower [see Fig. 2.3]. For a nonzero Vres , the MSE of
V̂res,avg does not decrease to 0 even if N goes to infinity. The reason is that V̂res,avg
is a biased estimator of Vres , and the bias does not converge to zero [Casella and
Berger [1990]]. As will be shown later, the aforementioned drawbacks for V̂res,avg
do not exist for the other two estimators.

2.3.2 Maximum likelihood estimator (MLE)

A key reason that V̂res,avg does not achieve a low estimation error is that
it does not assume any prior information on the comparator noise σ. In practice,
σ is chosen by the designer. It can be extracted via SPICE simulations with good
accuracy. Although σ may change due to process, voltage, and temperature (PVT)
variations, we can obtain an accurate value of σ by performing a simple foreground

18
calibration. We can simply set Vin = 0 by shorting the ADC input, and monitor the
standard deviation of Dout . Since we have assumed that comparator noise σ is the
dominant random source over the sampling noise ns , the standard deviation of Dout
simply reflects the value of σ. Since temperature changes slowly, the foreground
calibration does not need to be repeated frequently. In addition, since the statistical
estimation is performed at the LSB level, a 10% change in the value of σ only
causes minor degradation (e.g., a 0.1 LSB error) in the post-estimation ADC SNR.
This greatly relaxes the requirement on the calibration accuracy. Given this, we can
treat the value of σ as a known quantity during the estimation process. We can take
advantage of it to form a much better estimator, which is the maximum likelihood
estimator (MLE).

The definition of MLE is easy to understand. Given the number of compar-

isons N and the number of ‘1’s k, we define the estimator V̂res to be the value that
maximizes the probability of observing k ‘1’s out of N comparisons. MLE has been
thoroughly studied in statistics and has several merits. First, it is consistent. As N
increases, V̂res,M LE converges to Vres and can achieve arbitrary precision [Casella
and Berger [1990]]. Second, it is highly efficient from the information usage point
of view. It achieves the Cramer-Rao lower bound as N goes to infinity, which means
that MLE achieves the lowest asymptotic MSE [Casella and Berger [1990]].

We can derive V̂res,M LE for our problem in the following way. The probabil-
ity of a comparator output being ‘1’ follows Bernoulli distribution with probability:

Vres
P (d0 = 1) = P (Vres + nc > 0) = F ( ) (2.5)
σ

19
where F (x) is the cumulative distribution function of normal distribution with mean
of 0 and variance of 1, given by:
∫ x ∫ x
1 s2
F (x) = f (x)dt = √ e− 2 ds (2.6)
−∞ −∞ 2π

where f (x) is its corresponding probability density function. Assuming the com-
parator hysteresis is negligible, the repeated LSB comparator outputs, {d0 } (i ∈
[1, N ]), can be considered as independently and identically distributed random vari-
ables. Thus, from the probability theory [Casella and Berger [1990]], we know that
∑
k= N i=1 d0,i follows the binomial distribution B(N, F (Vres /σ)) with the proba-

bility given by:

( )
N Vres k ( Vres )N −k
P (k|Vres ) = F( ) 1 − F( ) (2.7)
k σ σ

where P (k|Vres ) means the probability of having k ‘1’s conditioning on Vres . Since
V̂res,M LE maximizes P (k|Vres ) by definition, we have:

dP (k|Vres ) 
V̂res,M LE
=0 (2.8)
dVres

From (2.8), we can derive that:

k
F (V̂res,M LE /σ) = (2.9)
N

This result is intuitive. V̂res,M LE ensures that the probability of the comparator
output being 1 is k/N , and thus, it achieves the highest probability for having in
total k ‘1’s out of N trials. From (2.9), we can solve V̂res,M LE :

k
V̂res,M LE = σ · F −1 ( ) (2.10)
N

20
It shows that V̂res,M LE is linearly proportional to the comparator noise σ. This is
different from V̂res,avg that has no dependence on σ [see (2.4)].

V̂res,M LE defined in (2.10) has one limitation that it does not work for k = 0
and k = N . If we plug k = 0 or k = N into (2.10), V̂res,M LE is ±∞. This is
expected because V̂res,M LE = −∞ achieves the highest probability for k = 0, and
V̂res,M LE = ∞ ensures that k = N . This issue may be minor for a large N , because
the probability of k = 0 and k = N would approach zero. However, for a small
N , k = 0 and k = N do appear, which causes an estimation failure. To solve this
problem, we can re-define:

0.2
V̂res,M LE (k = 0) = σ · F −1 ( ) (2.11)
N
N − 0.2
V̂res,M LE (k = N ) = σ · F −1 ( ) (2.12)
N

For other k lies in [1, N − 1], we still follow the definition of (2.10).

Fig. 2.4 shows V̂res,M LE as a function of k and N . Different from V̂res,M LE ,

the relationship between V̂res,M LE and k is nonlinear. The range of V̂res,M LE ex-
pands with N , from ±1.6σ at N = 3 to ±2.3σ at N = 15. This means that
V̂res,M LE can approximate a wider range Vres as N increases.

To evaluate how accurate V̂res,M LE is, we plot its estimation error as a func-
tion of Vres and N in Fig. 2.5. In general, its MSE decreases as N increases. For
the same N , its MSE is small for a small Vres , but increases as the amplitude of
Vres increases. The reason is that the value of V̂res,M LE is bounded for a given N
(see Fig. 2.4), and thus, does not work well for a very large Vres . However, unlike
V̂res,avg , the range of V̂res,M LE increases with N , and thus, its region with a small

21
 3

1
V̂r es,M L E [ ]

N=3
0 N=5
N=7
-1 N=9
N=11
N=13
-2 N=15

-3
0 5 10 15
k

Figure 2.4: Value of V̂res,M LE as a function of N and k.

MSE is broadened. This is a key advantage of V̂res,M LE compared to V̂res,avg . It

enables V̂res,M LE to accurately estimate a wide range of Vres especially for a large
N . In fact, we can prove that the MSE of V̂res,M LE goes to zero as N approaches
infinity for any Vres [Casella and Berger [1990]].

2.3.3 Bayes estimator

MLE is a significant improvement over the simple averaging based estima-

tor, but it still does not achieve the lowest estimation error. There is one extra piece
of information that MLE does not make use of, which is the distribution of Vres .
This extra information can be exploited to construct a better estimator with a lower
estimation error than MLE.

Before moving ahead, let us first examine the distribution of Vres . Fig. 2.6

22
 10 1
N=2
N=4
N=8
MSE of V̂res,M L E [ 2]

N=16
N=32
10 0
N=64

-1
10

-2
10
-3 -2 -1 0 1 2 3
V res [LSB]

Figure 2.5: MSE of V̂res,M LE for different Vres .

shows the simulated histograms of Vres for an 11-bit SAR ADC assuming the com-
parator noise σ = 1 LSB. Three different input signals are used. Sinusoidal inputs
with −6-dBFS and −20-dBFS are used in Fig. 2.6(a) and (b), respectively. A Gaus-
sian random input with a standard deviation of 10% ADC full swing is used in Fig.
2.6(c). As can be seen, there is negligible difference in the Vres distribution among
the three cases. They are all close to Gaussian distribution with zero mean and a
standard deviation of 1 LSB. This shows that the distribution of Vres has very weak
dependence on the ADC input Vin . This is not hard to understand. By the end of the
11-bit SAR conversion, the conversion residue Vres is almost completely uncorre-
lated with Vin , and is basically set by the comparator noise. Given this observation,
we can confidently approximate Vres as a Gaussian random variable and its proba-

23
bility density function (pdf) g(Vres ) ≡ f (Vres /σ), where f (·) is the pdf of Gaussian
distribution with zero mean and standard deviation of 1.
Probability density function

Probability density function

Probability density function

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
-5 0 5 -5 0 5 -5 0 5
V res [LSB] V res [LSB] V res [LSB]
(a) (b) (c)

Figure 2.6: Histogram of Vres for (a) a −6-dBFS sinusoidal input, (b) a −20-dBFS
sinusoidal input, and (c) a Gaussian random input with standard deviation of 10%
ADC full range.

Now let us derive the optimum estimator for Vres given the prior informa-
tion on its distribution. This problem has been thoroughly studied in the statistical
estimation theory [Casella and Berger [1990]]. It can be rigorously proved that the
best estimator that achieves the minimum MSE given the prior distribution is the
Bayes estimator (BE), which is defined as the mean of the posterior distribution of
Vres after observing k ‘1’s out of N trials. To understand it, let us consider a simple
example of N = 3. We can have 4 different values for k, which is 0, 1, 2, and 3.
For each case, we can calculate the posterior distribution g(Vres |k) using the Bayes
theorem[Casella and Berger [1990]]:

P (k|Vres )g(Vres )
g(Vres |k) = ∫ +∞ (2.13)
−∞
P (k|Vres )g(Vres )dVres

24
where P (k|Vres ) is the probability of observing k ‘1’s conditioning on Vres . Fig.
2.7 plots the Vres prior distribution g(Vres ) together with its posterior distributions
g(Vres |0), g(Vres |1), g(Vres |2), and g(Vres |3). Bayes rule basically allows us to up-
date the distribution of Vres given the observation result k. We can see that the
prior and posterior distributions are different, which is enabled by the knowledge of
k. For example, compared to g(Vres ), the posterior distribution g(Vres |0) is shifted
towards the negative side. This is because after observing all ‘0’s from the compara-
tor outputs, we can update the distribution of Vres , which should be more negatively
biased.

0.7
g(Vres)
0.6 g(Vres|0)
g(Vres|1)
Probability density

0.5
g(Vres|2)
0.4 g(Vres|3)

0.3

0.2

0.1

0
-5 0 5
V res [σ]

Figure 2.7: Prior distribution g(Vres ) and posterior distribution g(Vres |k) for N = 3.

The Bayes estimator is defined as the mean of the posterior distribution.

25
Mathematically, it is given by:
∫ +∞
V̂res,BE (k) ≡ E(Vres |k) = Vres · g(Vres |k)dVres (2.14)
−∞

For the case of N = 3, we can calculate that V̂res,BE (0) = −1σ, V̂res,BE (1) =
−0.3σ, V̂res,BE (2) = +0.3σ, and V̂res,BE (3) = +1σ, respectively.

Note that (2.13) and (2.14) are computationally intensive. Fortunately, we

do not need to solve V̂res,BE for every ADC output. We only need to compute once,
and store the results for all possible k values in a look-up table. This way, once
we know k from the comparator outputs, V̂res,BE can be directly obtained from the
table.

Fig. 2.8 shows V̂res,BE as a function of k and N . Comparing it carefully

with V̂res,M LE shown in Fig. 2.4, we can see that the range of V̂res,BE is smaller
than V̂res,M LE . The reason is that V̂res,BE makes use of the prior distribution of
Vres . Since Vres is concentrated around zero, V̂res,BE is biased more towards zero.

Fig. 2.9 shows MSE for V̂res,BE as a function of Vres and N . Comparing
it with Fig. 2.5 of V̂res,M LE , we see that the MSE of V̂res,BE is smaller than that
of V̂res,M LE for a small Vres in [−2σ, +2σ], but is slightly larger for |Vres | > 2σ.
However, because Vres is known to concentrate around 0, it is expected that the
overall MSE of V̂res,BE is smaller than that of V̂res,M LE .

2.3.4 Estimator performance comparison

To compare the estimation error for the three estimators, we compute the
MSE of V̂res,avg , V̂res,M LE , and V̂res,BE for σ = 1 LSB, and plot them as a function

26
 2

1.5

1
V̂res,B E [ ]

0.5

0 N=3
N=5
-0.5 N=7
N=9
-1 N=11
N=13
-1.5 N=15

-2
0 5 10 15
k

Figure 2.8: V̂res,BE as a function of N and k.

10 1
N=2
N=4
N=8
MSE of V̂res,B E [ 2]

N=16
10 0 N=32
N=64

-1
10

-2
10
-3 -2 -1 0 1 2 3
V res [LSB]

Figure 2.9: MSE versus x for V̂res,BE .

27
of N in Fig. 2.10.

10 0
Vres, avg
Vres, MLE
Vres, BE
MSE [σ2]

10 -1

10 -2
0 1 2
10 10 10
N
Figure 2.10: MSE versus N for a normal distributed x with σ = 1LSB.

As expected, the MSE of V̂res,BE is consistently smaller than that of V̂res,avg

and V̂res,M LE , indicating the highest estimation accuracy. Note that the MSE values
for all three estimators are similar for a small N . However, for a large N , there are
significant differences. For V̂res,avg , its MSE saturates to around 0.2σ 2 . By contrast,
the MSE of V̂res,M LE and V̂res,BE keep decreasing with N at the slope of 10 dB per
decade. This slope is essentially the limit set by Cramer-Rao lower bound [Casella
and Berger [1990]], confirming the high efficiency of both V̂res,M LE and V̂res,BE .
Since V̂res,BE has the lowest MSE, it is chosen for the prototype ADC.

As mentioned earlier, both MLE and Bayes estimators require the knowl-
edge of the comparator noise σ. It is interesting that the extracted comparator

28
noise does not need to be 100% accurate when estimating V̂res,BE . This makes
the proposed technique more robust as the comparator noise can drift with Process-
Voltage-Temperature (PVT) variation. Let us assume a ±10% noise drifting hap-
pens in the prototype ADC. After 1σ=1.0 LSB is extracted, the comparator noise
drifts to 0.9 LSB or 1.1 LSB due to PVT variation. However, the comparator noise
is supposed to be the same and the 1σ comparator noise of 1.0 LSB is used when
performing the Bayes estimation. Fig. 2.11 plots the behavioral simulated SNR
improvement for different σs. As can be seen from Fig. 2.11, the SNR is improved
effectively even though an inaccurate comparator noise 1σ=1.0 LSB is used to es-
timate V̂res,BE for 1σ=0.9 LSB and 1σ=1.1 LSB. Note that the SNR differences
among three cases result from different comparator noise used in the ADC model.
Fig. 2.11 also shows that the proposed technique can be easily extended to achieve
re-configurable resolution by adjusting N . In the prototype design, a 7-dB SNR
improvement is expected which requires N = 17.

2.3.5 Comparison to oversampling and analog scaling

The standard ADC oversampling can also be used to reduce ADC noise
by averaging, and the improvement in SNR is also 10 dB per decade. Thus, it is
meaningful to compare it to the proposed statistical estimation technique. The merit
of oversampling is that it reduces both the sampling noise ns and the comparison
noise nc , while the proposed technique based on statistical estimation only reduces
nc . However, as mentioned earlier, in a SAR ADC, the noise is typically dominated
by nc , and thus, their effect in total noise reduction is similar. The disadvantage of

29
 68
66
SNR [dB]

64
62 σ = 0.9 LSB
σ = 1.0 LSB
60 σ = 1.1 LSB
58
56
0 4 8 12 16 20 24 28 32
N
Figure 2.11: Simulated SNR versus N with ±10% variations in the comparator
noise σ.

oversampling is that it cannot improve the ADC power efficiency. Every doubling
of the oversampling ratio (OSR) leads to twice the ADC power, as it requires repeat-
ing all the sampling, comparison and DAC switching operations. By contrast, the
proposed technique only increases the number of LSB comparisons, so its required
total number of comparator operations is much smaller. Additionally, the DAC is
not switched, and thus, does not dissipate any extra DAC power. As a result, the
power efficiency of the proposed technique is much higher than oversampling.

As in any noise reduction technique, there is always a cost of power. For

the proposed technique, the total power of the comparator increases due to extra
number of LSB comparisons. Take our 11-bit prototype ADC as an example, to
reduce noise by 7 dB, the LSB comparison needs to be fired 17 times, which results

30
in an increase of total comparator power by (11 + 17)/11 ≈ 2.5 times. By contrast,
to obtain the same amount of noise reduction for the same 11-bit ADC, if we choose
the brute-force way to reduce comparator noise by increasing its size and power,
we need to increase the comparator power by 21 times (see Prototype ADC Design
section). Thus, the proposed technique is much more power efficient. The tradeoff
for the proposed technique is reduced conversion rate due to increased number of
LSB comparisons. However, as mentioned earlier, this is only a minor issue for the
intended low-to-medium speed sensor applications.

2.4 Prototype ADC Design

2.4.1 Detailed circuit schematics

To verify the proposed statistical estimation based noise reduction tech-

nique, a 11-bit prototype SAR ADC is designed, whose architecture is shown in
Fig. 2.12. There are only two simple changes made to the standard SAR ADC ar-
chitecture: 1) the SAR logic is modified to repeat the LSB comparison for N = 17
times; and 2) a counter is used to count the number of ‘1’s during LSB compar-
isons to obtain k. A low power supply voltage of 0.7V is chosen to demonstrate the
effectiveness of the proposed technique for low voltage and low power applications.

The DAC is implemented with binary-weighted metal finger (or MoM) ca-
pacitors. Since the DAC power is proportional to the total capacitor value, it is
desired to reduce the unit capacitor Cu for power saving. Considering the noise and
matching requirement, this design chooses Cu = 2 fF. A bidirectional single-side
(BSS) switching technique is adopted to further reduce the DAC reference power by

31
 Vinn Vinp clkin
1.4V 0.7V
booster clk generator

DAC counter k
estimator
Dout sar logic enable
Figure 2.12: Proposed SAR ADC architecture.

86% compared to the conventional switching scheme [Chen et al. [2014a]; Sanyal
and Sun [2014]]. BSS reduces the number of unit capacitors by 4 times, leading
to a small capacitor array of {256, 128, 64, 32, 16, 16, 8, 4, 2, 1, 1}Cu for an 11-
bit ADC. Compared to widely used monotonic switching technique of [Liu et al.
[2010a]], BSS achieves higher SNDR as the comparator input common-mode volt-
age Vcm variation is reduced and Vcm can converge to half Vdd instead of ground. A
redundant capacitor of 16Cu is provided to recover possible errors during the first
several MSB comparisons with large Vcm variation [Chen et al. [2014a]]. The to-
tal capacitance is 528Cu = 1056fF, leading to 88-µV differential sampling kT /C
noise. Since the unit capacitor is only 2 fF, the routing parasitic capacitors have a
considerable influence on the capacitor matching accuracy. A segmented common-
centroid layout technique is used for better matching. The floor plan for a single
side capacitor array is shown in Fig. 2.13. The capacitors are separated into two
groups {C8 , C7 , ..., C4 } and {C3 , C2 , C1 , C0 }. The common-centroid rule is applied
horizontally for the first group and vertically for the second. Dummy cells, repre-

32
sented as D, are added for better matching.

vertical common centroid

horizontal common centroid
Figure 2.13: DAC capacitor array floor plan.

To ensure good sampling linearity, bottom-plate sampling is used. In addi-

tion, a clock booster shown in Fig. 2.14(a) is employed to boost the sampling clock
voltage from 0.7V to 1.4V. As a result, simple small NMOS transistors can be used
to sample Vin instead of an array of bootstrapped switches or large CMOS switches
to reduce the design complexity and the switch driving power.

Fig. 2.14(b) shows the dynamic strong-arm latch comparator. It uses a

PMOS input pair to minimize the flicker noise and the substrate coupling. Thanks to
the proposed noise reduction technique, the comparator thermal noise requirement
is relaxed, leading to substantially reduced comparator power. In this design, the
SPICE simulated total input referred noise at the comparator input is 480µV or 0.7
LSB. Note that this is much larger than the sampling kT /C noise, and thus, the

33
 VDD

CB
EN

clkin clkout

(a)
VDD

clkc
Vinn Vinp

clkc clkc

clkc clkc

(b)

clk

sample
normal sar operation
repeated LSB comparisons

(c)

Figure 2.14: Schematic of (a) the clock booster; (b) the comparator; and (c) the
timing diagram.

34
overall ADC noise is dominated by the comparator noise.

The ADC timing diagram is shown in Fig. 2.14(c). It uses a synchronous

clocking scheme. The frequency of the master clock is 32 times faster than the
sampling rate. The first 4 clock cycles are used for input sampling to ensure high
sampling accuracy; the subsequent 11 cycles are used for normal SAR operation;
and the final 17 cycles are used for repeated LSB comparisons. This clock allo-
cation scheme can be easily implemented using a ripple counter based clock di-
vider and several AND gates. The SAR logic is built by standard shift registers.
When the normal SAR operation finishes, the last shift register makes transition
and ready = 1 is generated. A 5-bit counter shown in Fig. 2.15 is enabled, which
records the number of ‘1’s during the LSB comparisons and obtains k for statistical
estimation. The counter consumes no power during normal SAR operation. The
counted k is given by:

k = 16 × D4 + 8 × D3 + 4 × D2 + 2 × D1 + D0 (2.15)

D4
D Q D Q D Q D Q D Q
ready
Q Q Q Q Q
Voutp R D0 R D1 R D2 R D3 R
clks
Figure 2.15: 5-bit counter to count k.

35
 Figure 2.16: Die micrograph.

2.4.2 Measurement Results

The prototype ADC is implemented in 65nm CMOS process. Fig. 2.16

shows the die photo. The ADC occupies an active area of 0.03 mm2 , which is
dominated by the DAC. The power supply is 0.7V and the sampling rate is 100kS/s.
Fig. 2.17 shows the measured DNL and INL, which are +1.04/ − 1 LSB and
+1.57/ − 1.23 LSB. According to the INL plot, there exists a 1-LSB systematic
mismatch between the 6 MSB capacitors and the 6 LSB capacitors, which arises
from the unmatched surrounding environment due to the segmented layout strategy
in Fig. 2.13 and inaccurate parasitic capacitor extraction. A simple foreground
calibration similar to [Chen and Brodersen [2006]] is performed and the appreciable

36
periodic INL transition pattern reduces.

Before calibration
2

1
INL [LSB]

-1

-2
0 500 1000 1500 2000

After calibration
2

1
INL [LSB]

-1

-2
0 500 1000 1500 2000
Output code

Figure 2.17: Measured DNL and INL.

To verify the proposed noise reduction technique, we first measure the ADC
noise (e.g., the variation of Dout ) at Vin = 0. The measured probability densities
for Dout before and after noise reduction are shown in Fig. 2.18 together with fitted
normal distributions. Before noise reduction, the standard deviation of Dout is 0.73
LSB. It indicates the comparator input referred noise is about 500µV, which is in
agreement with SPICE simulation. After noise reduction, the standard deviation
∗
of Dout is reduced by 7 dB to 0.33 LSB, which matches well with the estimation
theory. Note that if the conventional SAR ADC design approach is used, the com-
parator noise needs to be reduced to 0.16 LSB in order for the total ADC noise to

37
be 0.33 LSB, which also includes the 0.29 LSB quantization error. This means that
the total comparator power needs to be increased by 21 times. By contrast, in our
proposed noise reduction technique, the total comparator power is only increased
by 2.5 times, which firmly proves its higher power efficiency compared to brute-
forth analog scaling. Once the comparator noise is extracted, V̂res,BE (k) can be
computed using (2.13) and (2.14).

1.2
w/o estimation
Probability density function

1 with estimation

0.8

0.6

0.4

0.2

0
1042 1043 1044 1045 1046 1047 1048 1049
Output code

Figure 2.18: Dout distribution with and w/o estimation at Vin = 0.

Fig. 2.19 shows the measured spectrum for a 96-kHz full-scale input sam-
pled at 100kS/s. The reason choosing the 96-kHz frequency input is our high-
quality low-distortion band-pass filter has a cut-off frequency at 90 kHz. The mea-
sured SNDR and SNR are 59.4 and 59.7 dB for 96-kHz input, respectively. After
applying the proposed noise reduction technique, the noise floor is clearly lowered.
SNDR and SNR are improved to 64.5 dB and 65 dB, respectively. The correspond-

38
ing ENOB is 10.5-bit.

0
w/o estimation
-20 with estimation
Spectrum [dBFS]

-40

-60

-80

-100

-120
0 0.1 0.2 0.3 0.4 0.5
Normalized frequency f/fs

Figure 2.19: Measured 214 -point ADC output spectrum with 96-kHz input.

Fig. 2.20 shows the SNR with varying input amplitudes. The SNR im-
provement using simple averaging based estimator V̂res,avg is limited to only 2.2dB.
Maximum likelihood estimator (MLE) achieves 5.8dB SNR improvement, which
is better than averaging based estimator. Using the Bayes estimator V̂res,BE , the
SNR can be improved by 7dB, which is 4.8dB better than that of averaging based
estimator and 1.2dB better than that of MLE. This matches well with the analysis.
When the input is very large, the SNR improvement decreases slightly to 5.3 dB,
which is caused by the unwanted capacitive coupling from the ADC input to the
reference lines, discovered during measurements. Such SNR loss can be recovered
by layout optimization to reduce the coupling. To evaluate the robustness of SNR
improvement using Bayes estimator, various extracted comparator noise σs, which

39
can be caused by PVT variation, are used to estimate V̂res,BE . Fig. 2.21 shows the
measured SNR improvement versus σs. The peak SNR improvement is large in the
middle and small at two sides, just like a parabolic distribution. The peak SNR im-
provement is 6.9 dB at σ = 0.73 LSB. As long as the extracted σ ∈ [0.6, 0.86] LSB,
the SNR improvement is greater than 6 dB and one more effective bit is achieved
with Bayes estimation.

70
SNR w/o estimation
SNR with avaraging
60 SNR with MLE
SNR with BE
SNR [dB]

50

40

30

20
-35 -30 -25 -20 -15 -10 -5 0
Input amplitude [dBFS]
Figure 2.20: Measured SNR versus input amplitudes.

The ADC consumes 0.6 µW from a 0.7 V power supply. The compara-
tor, DAC, clock generator, and SAR logic consume 70 nW, 102 nW, 193 nW and
280 nW, respectively. With the noise reduction technique, the comparator power
accounts for only 10% of the total power at the ENOB of 10.5-bit. The digital
power, including both clock generator and SAR logic, dominates the overall ADC
power. It can be substantially reduced via optimization and/or going to a more

40
 8

Measured SNR improvement [dB]

6

0
0.2 0.4 0.6 0.8 1 1.2 1.4
Comparator input referred noise [LSB]
Figure 2.21: Measured SNR improvement versus various comparator noise σ.

advanced technology node, without affecting SNR. The measured figure-of-merit

(FOM) for the prototype ADC is 4.5 fJ/conversion-step. The performance of the
proposed ADC is summarized in Table 2.1. Fig. 2.22 shows Walden FoM versus
SNDR for this work and recently published ADCs in ISSCC and VLSI conferences.
The data for the figure is taken from the survey made available by Dr. Boris Mur-
mann (http://web.stanford.edu/∼murmann/adcsurvey.html). As
can be seen, this work achieves the state-of-the-art power efficiency, especially
among ADCs with SNDR greater than 64 dB. Note that there is still large space
for further performance improvements especially on the design of the clock gener-
ator and the SAR logic.

41
 Table 2.1: Measured performance summary

Process [nm] 65
Sampling rate [kS/s] 100
Resolution [bit] 11
Active area [mm2 ] 0.03
Power supply [V] 0.7
Total power [µW] 0.6
DNL [LSB] +1.04 / −1.00
INL [LSB] +1.57 / −1.23
Noise Reduction? No Yes
Dynamic range [dB] 60 67
Peak SNR [dB] 59.7 65
Peak SNDR [dB] 59.4 64.5
ENOB [bit] 9.6 10.5
Walden FoM [fJ/conv-step] 9 4.5
1.E+04

1.E+03
FOM [fJ/conv-step]

1.E+02

1.E+01

This work
1.E+00
ISSCC 2010-2015
VLSI 2010-2015
1.E-01
10 20 30 40 50 60 70 80 90 100 110
SNDR @ f in,hf [dB]

Figure 2.22: FOM versus SNDR plot for this work and recently published ADCs in
ISSCC and VLSI conferences.

42
 Chapter 3

Bidirectional Single-side Switching Technique

This chapter1 presents a low-power SAR ADC with a bidirectional single-

side (BSS) switching technique. Compared to the conventional SAR switching
scheme, it reduces the DAC reference power and the total number of unit capacitors
by 86% and 75%, respectively. It also minimizes the power dissipated in driving
DAC switches as it has only one single-side switching event every comparison cy-
cle. Compared to the monotonic switching technique [Liu et al. [2010a]] that also
has only one switching event, it reduces the comparator input common-mode volt-
age variation by 2 times. Moreover, its comparator input common-mode voltage
does not converge to ground but to Vcm . This greatly reduces the comparator off-
set and noise. It obviates the need for a specially designed comparator and allows
the use of a low-power strong-arm latch. A prototype with proposed technique has
been developed in 180nm CMOS.

This chapter is organized as follows: an introduction of existing switching

techniques is first presented. The effect of comparator common-mode variation is

1
This chapter is a partial reprint of the publication: Long Chen, Arindam Sanyal, Ji Ma and
Nan Sun, “A 24-uW 11-bit 1-MS/s SAR ADC with a bidirectional single-side switching technique”,
IEEE ESSCIRC, pp. 219–222, 2014. I thank all the co-authors for their valuable advice in designing
and testing of the prototype.

43
analyzed next. Finally, the proposed BSS switching technique is presented, fol-
lowed by the prototype implementation and measurement.

3.1 Introduction

Successive approximation register (SAR) analog-to-digital converter (ADC)

is scaling friendly and power efficient. It is suitable for low power and low volt-
age applications, such as wireless sensor networks, implantable bio-sensors, and
portable medical electronics [Harpe et al. [2013]; Van Elzakker et al. [2010]; Verma
and Chandrakasan [2006]; Zhang et al. [2012]]. It consists of a capacitive DAC, a
comparator, and a SAR logic block. The DAC can take up a large portion of the
total ADC power especially at high resolution with large capacitors. The conven-
tional DAC switching technique based on trial and error is not power efficient. To
address this issue, several low-power DAC switching techniques have been devel-
oped, including the split capacitor technique [Ginsburg and Chandrakasan [2005]],
the Vcm -based switching technique [Zhu et al. [2010]], and the monotonic switching
technique [Liu et al. [2010a]]. The split capacitor technique reduces DAC reference
power by 37% compared to the conventional switching scheme. The Vcm -based
switching technique reduces DAC reference power by 88%, but it requires addi-
tional switches to pass Vcm . Due to the reduced overdrive voltage, these switches
need to be large, leading to increased power in driving these switches. This lim-
itation is more pronounced at low power supply voltage and high conversion rate
[Sanyal and Sun [2014]].

The monotonic switching technique of [Liu et al. [2010a]] has attracted in-

44
creasing attention as it reduces both the DAC reference power and power dissipated
in driving switches. It achieves 81% reduction in DAC reference power. It does
not need capacitor splitting or additional large switches. Moreover, it requires only
one switching event every comparison cycle. Thus, its switching activity is only
33% of the conventional switching technique and 50% of the Vcm -based switching
technique. In addition, it reduces the total number of unit capacitors by a factor
of 2. Despite the merits mentioned above, monotonic switching has one key lim-
itation. Its comparator input common-mode voltage decreases every comparison
cycle and eventually converges to ground. Although a comparator with a PMOS
input pair can be used to maintain functionality, the large change in the common-
mode voltage causes varying comparator offset, resulting in ADC nonlinearity. In
addition, the large overdrive voltage for the input pair leads to increased compara-
tor noise and degrades the SNR. To address these issues, the authors of [Liu et al.
[2010a]] designed a special comparator by stacking an extra tail transistor acting
like a dynamic current source, so that the input pairs overdrive voltage can be kept
relatively constant. However, this solution increases design effort and requires ad-
ditional bias circuit. In addition, the proposed comparator consumes static power
during the evaluation phase. Another approach to address the common-mode varia-
tion problem is to split MSB capacitors and switch two half capacitors on both sides
of the DAC capacitor array every comparison [Liu et al. [2010b]]. This technique
requires additional inverter for each split capacitor, which increases digital power
and slows down speed. In addition, the number of switches also needs to be doubled
for split bits, resulting in more layout efforts and more power dissipated on driving

45
the parasitic capacitors.

This chapter presents a bidirectional single-side (BSS) switching technique

that maintains all the aforementioned merits of the monotonic switching technique
and solves the large common-mode variation issue [Chen et al. [2014a]]. Instead
of switching DAC capacitors from Vref to ground monotonically as in [Liu et al.
[2010b]], the proposed technique first switches the MSB capacitor from ground to
Vref , and then, switches other capacitors from Vref to ground. Consequently, the
range of comparator input common-mode variation is reduced by 2 times. More
importantly, the common-mode voltage converges to Vcm instead of ground, which
greatly reduces the comparator noise and offset variation. As a result, the proposed
technique obviates the need for the specially designed comparator and permits the
use of a low-power strong-arm latch as a comparator. Furthermore, the proposed
switching method achieves an additional bit over monotonic switching by using
Vcm as reference for the LSB capacitor. For the same resolution, the proposed tech-
nique reduces the total number of unit capacitors by 2 times compared to monotonic
switching and 4 times compared to conventional switching. It reduces the DAC ref-
erence power by 86%, which is also more than that for monotonic switching.

An 11-bit prototype ADC using the proposed BSS switching technique is

fabricated in 180nm CMOS process. To minimize power and area, the smallest
available MIM capacitor of 2 fF is used as the unit capacitor. To ensure good
matching, a segmented common-centroid layout with surrounding dummy capaci-
tors is developed, which allows the ADC to achieve 77 dB of SFDR without any
calibration. Operating at 1 MS/s, it consumes 24 µW of power under 1 V power

46
supply. It achieves 10.3-b ENOB and an FOM of 19.9 fJ per conversion step.

3.2 Effect of Comparator Common-Mode Variation

The strong-arm latch comparator in Fig. 3.1 is widely used in ADCs and
memory read-out circuits [Kobayashi et al. [1993]; Van der Plas et al. [2006]].
Its input common-mode voltage strongly affects the comparator offset, noise, and
speed [Wicht et al. [2004]; Nuzzo et al. [2008]]. Since the common-mode voltage
variation is a key issue in both monotonic switching and the proposed BSS switch-
ing techniques, let us first analyze its effects on the comparator performance.

Fig. 3.2 shows a typical transient behavior of the comparator with a 1mV in-
put differential voltage. When clkc is low, the comparator is in reset with both VX1,2
and Voutp,n pulled to VDD . When clkc is high, the comparator is in evaluation mode.
Here its operation can be divided into two phases, the pre-amplification phase and
the latch regeneration phase as shown in Fig. 3.3, with the turn-on of the PMOS
cross-coupled pair separating two phases. During the pre-amplification phase, the
PMOS cross-coupled pair is in cut-off and the comparator works as a dynamic inte-
grator. From differential-mode point of view, the comparator input voltage induces
a differential drain current, which is integrated on the output capacitive load CO
and produces a differential output voltage Vout that grows linearly with time. From
the common-mode point of view, VX1,2 decrease after clkc goes high. Once they
reach VDD − VT n3,4 , M3 and M4 are turned on and Voutp,n start to decrease. When
Voutp,n decrease to VDD − VT p5,6 , the PMOS crossed coupled pair is turned on and
the comparator enters the latch phase. Since the PMOS crossed coupled pair is in

47
 VDD

clkc clkc
M9 M5 M6 M10
Voutp Voutn

M7 M3 M4 M8

VX1 VX2
Vinn Vinp
M1 M2

clkc
Mb

Figure 3.1: Schematic of a strong-arm latch comparator.

positive feedback and provides exponentially growing gain, the PMOS pair domi-
nates the behavior of the comparator in the latch phase, while other transistors can
be considered to the first order as a current source before Vout grows very large that
cuts off a transistor in the PMOS pair or the current source.

A critical parameter that links the two phases is the pre-amplifier gain G,
defined as the differential voltage gain of the comparator at the end of the pre-
amplification phase. It affects the comparator offset and noise. We can model

48
 τcomp
clkc
1.8
1.6
1.4

[mV]
τ preamp
1.2 Voutp
Voltage [V]

5 Vout
1
0.8 0 Vx
Voutn
0.6 −5

0.4
0.2 Vx2

0 τpreamp τlatch Vx1

−0.2
0 100 200 300
Time [ps]

Figure 3.2: Transient behavior of a strong-arm latch comparator.

the input referred offset of the pre-amplification phase as Vos,preamp , whose main
contributor is VT n1,2 mismatch in the input pair. For the latch phase, we can model
its offset referred to the output nodes at the beginning of the latch generation phase
as Vos,latch , whose main contributor is VT p5,6 mismatch in the PMOS cross coupled
pair. As a result, the input referred offset of the entire comparator, Vos , is given by:
Vos,latch
Vos = Vos,preamp + (3.1)
G
In terms of root-mean-square (rms) offset, we have:
√
2
σos,latch
2
σos = σos,preamp + (3.2)
G2

49
 Voutp Voutn
VDD
CO CO
M3 M4

VX1 VX2 M5 M6
CX CX
Vinn Vinp Voutp Voutn
M1 M2

clkc
Mb

(a) (b)

Figure 3.3: Comparator schematic for (a) the pre-amplification phase and (b) the
latch regeneration phase.

Similarly, we can also write down the expression for the comparator noise:
√
2
σn,latch
2
σn = σn,preamp + (3.3)
G2

From (3.1)-(3.3), we can see that the offset and noise depend strongly on G, and
a large G is preferred as it attenuates the contribution from the latch phase. The
value of G can be estimated to the first order in the following way. We first cal-
culate the time duration τpreamp of the pre-amplification phase by examining the
common-mode voltage change. At the beginning of the pre-amplification phase,
the common-mode voltages at Voutp,n and VX1,2 are both VDD . By the end of the
pre-amplification phase, the common-mode voltage at Voutp,n and VX1,2 are approx-
imately VDD − VT p5,6 and VDD − VT p5,6 − VT n3,4 , respectively. For simplicity,
assuming that the common-mode drain current of the comparator input pair ID is

50
unchanged during the integration, we have:

CO VT p5,6 + CX (VT p5,6 + VT n3,4 )

τpreamp ≈ (3.4)
ID

where CX and CO are the total capacitive load at VX1,2 and Voutp,n (see Fig. 3.3).
Similarly, assuming gm of the input pair is unchanged during the integration, the
total amount of differential charge ∆Q produced by the input pair with small input
voltage ∆Vin is:
∆Q = ∆I · τpreamp ≈ gm ∆V · τpreamp (3.5)

By the end of the integration phase, the majority of the differential charge ∆Q is at
the output nodes Voutp,n (the amount of differential charge at VX1,2 is very small as
shown in Fig. 3.2), and thus, we can derive G:

∆Vout ∆Q/CO
G ≈ ≈
∆Vin ∆V
{ }
gm CX
≈ VT p5,6 + (VT p5,6 + VT p3,4 ) (3.6)
ID CO

Since the comparator input common-mode Vcmi determines gm /ID of the

input transistor, it has a strong influence on G. To increase G and reduce offset
and noise, we prefer a small Vcmi . To verify the analysis above, SPICE simulation
is performed in 0.18-µm CMOS process with VDD = 1.8V . G is extracted by
examining the voltage difference at Voutp,n when the output common-mode drops to
VDD − VT p5,6 . Fig. 3.4(a) shows that G depends strongly on Vcmi . It decreases from
12.6 at Vcmi = 0.6V to 0.6 at Vcmi = 1.8V . Also, gm /ID proportionally decreases
as Vcmi increases. These results match (3.6). σos and σn are extracted from Monte
Carlo and transient noise simulations, respectively. Fig. 3.4(b) shows that they both

51
increase with Vcmi due to the reduction in G. Fig. 3.4(c) plots their square as a
function of 1/G2 . The close linear fitting with a fitting coefficient r > 0.99 clearly
validates the models of (3.1)-(3.3).

Vcmi also strongly affects the comparator resolve time τcomp defined here
as the time it takes for the comparator output differential voltage to reach 0.7VDD .
Fig. 3.4(d) shows τcomp as a function of Vcmi for a fixed 1mV input. τcomp decreases
as Vcmi increases, which can be explained by (3.4) as a larger ID shortens the pre-
amplification phase τcomp . An interesting observation is that the minimum τcomp
is not obtained at Vcmi = VDD . The reason is that if Vcmi is too large, the time
duration of the latch regeneration phase is longer due to the reduction in G [Wicht
et al. [2004]]. In addition, the comparator power Pcomp depends mildly on Vcmi , as
shown in Fig. 3.4(e). The reason is that a large Vcmi leads to increased short circuit
current.

Overall, a small Vcmi is preferred for small offset, low noise, and low power,
but it leads to slow speed. This represents a trade-off in the choice of Vcmi . To
provide a holistic evaluation of the comparator performance, we can define a com-
parator figure-of-merit F OMcomp shown below:

1
F OMcomp = (3.7)
σnα × β
τcomp × Pcomp
γ

where α, β and γ are weighting factors to be determined by the designer. For

example, for noise optimized design, we can give more weight to σn by having a
larger α. Or if speed is the top priority, we can emphasize τcomp in F OMcomp by
enlarging β. For a balanced design with α = β = γ = 1, F OMcomp is plotted

52
 20 10 1.5 15
Gain σn
gm/ID σos

g /I [V−1]
1 10

σos [mV ]
σn [mV ]
G

10 5

m D
0.5 5

0 0 0 0
0.6 0.9 1.2 1.5 1.8 0.6 0.9 1.2 1.5 1.8
Vcmi [V] Vcmi [V]
(a) (b)
2 200 500
σ2n
1.5 2 150
σos 400
τcomp [ps]
σos [mV ]
σ2n [mV2]

1 100
300
2

0.5 50
200
0 0
0 1 2 3 0.6 0.9 1.2 1.5 1.8
2 Vcmi [V]
1/G
(c) (d)
35 1
FOM [(µWpsV)−1]

0.8
30
Pcomp [µW]

0.6
25 0.4

20 0.2

0
0.6 0.9 1.2 1.5 1.8 0.6 0.9 1.2 1.5 1.8
Vcmi [V] Vcmi [V]
(e) (f)

Figure 3.4: Simulated comparator performance: (a) G and gm /ID ; (b) noise and
offset; (c) linear fitting for noise and offset with 1/G2 ; (d) resolve time; (e) power
and (f) F OMcomp .

53
shown in Fig. 3.4(f). Its optimum is at around Vcmi = 0.8V , which is more than 6
times larger than that at Vcmi = 1.8V .

3.3 Proposed BSS switching technique

Fig. 3.5 shows an example of the proposed bidirectional single-side (BSS)

switching technique applied to a 4-bit top-plate sampled SAR ADC. During the
sampling phase, the input is sampled onto the DAC array with the 2C MSB capac-
itor connected to ground and other capacitors connected to Vref . Since the input is
available at the comparator input, the MSB decision can be made immediately after
the sampling phase. Depending on the comparison result, one side of the 2C MSB
capacitors is switched from ground to Vref . Following the second comparator de-
cision, one side of the second MSB capacitor (the middle capacitor C) is switched
from Vref to ground. After the third comparison, one side of the LSB capacitor (the
rightmost capacitor C) is switched from Vref to Vcm . Finally, the fourth comparison
is launched. This operation scheme can be easily generalized for more number of
bits and also bottom-plate sampling.

3.3.1 Comparator input common-mode variation reduction

The key limitation for monotonic switching is that its comparator input
common-mode voltage Vcmi monotonically decreases from Vcm to ground, which
requires the use of a specially designed comparator with a PMOS input pair. Note
that there exists a complementary version for the original monotonic switching by
connecting all DAC capacitors initially to ground. This way, Vcmi monotonically

54
 (1/4)CVref2 A
VrefVref
2C C C
Yes
Vin > (1/2)Vref?
0 2C C C No
VrefVref VrefVref Vref VrefVref
2C C C 0 2C C C Yes (3/4)CVref2 B
Vinp
Vinn Vin < 0?

2C C C 2C C C No Vref VrefVref (3/4)CVref2 C

VrefVref VrefVref 2C C C
0 Yes
Vin = Vinp - Vinn Vin > (-1/2)Vref?
No
2C C C
VrefVref (1/4)CVref2 D

Vcmi = (1/2)Vref Vcmi = (3/4)Vref

A Vcm C Vref Vcm

2C C C 2C C C

-(3/16)CVref2 (1/16)CVref2
Vin > (7/8)Vref? Vin > (-1/8)Vref?
Vref 2C C C Vref Vref 2C C C
2C C C Yes VrefVrefVref 2C C C Yes Vref Vref
Vin > (3/4)Vref? Vin > (-1/4)Vref?
Vref Vref Vref
2C C C No 2C C C No
2C C C 2C C C
VrefVrefVref VrefVref
(3/16)CVref2 -(1/16)CVref2
Vin > (5/8)Vref? Vin > (-3/8)Vref?
2C C C 2C C C
VrefVrefVcm VrefVcm

B VrefVcm D VrefVrefVcm
2C C C 2C C C

2 2
-(1/16)CVref (3/16)CVref
Vin > (3/8)Vref? Vin > (-5/8)Vref?
VrefVref 2C C C VrefVrefVref 2C C C
2C C C Yes Vref Vref 2C C C Yes Vref
Vin > (1/4)Vref? Vin > (-3/4)Vref?
VrefVref Vref VrefVref
2C C C No 2C C C No
2C C C 2C C C
Vref Vref Vref
2 2
(1/16)CVref -(3/16)CVref
Vin > (1/8)Vref? Vin > (-7/8)Vref?
2C C C 2C C C
Vref Vcm Vcm

Vcmi = (5/8)Vref Vcmi = (9/16)Vref Vcmi = (5/8)Vref Vcmi = (9/16)Vref

Figure 3.5: Proposed bidirectional single-side switching technique.

increases from Vcm to VDD , which permits the use of the strong-arm latch compara-
tor with an NMOS input pair. However, as shown in Fig. 3.4, having Vcmi = VDD
results in large comparator offset and noise.

55
 Fig. 3.6 shows the variation of Vcmi for an 11-bit SAR ADC with different
switching schemes. Note that Vcmi depends on the comparison cycle but is inde-
pendent from Vin . For the proposed bidirectional up-down switching of Fig. 3.5,
Vcmi first increases, then decreases, and finally converges to Vcm . As a result, com-
pared to monotonic switching, the amount of Vcmi variation is reduced by a factor
of 2. Furthermore, because Vcmi converges to Vcm , the comparator noise and offset
are significantly reduced. To compare the ADC performance with monotonic up
switching and the proposed switching technique, a behavioral model for an 11-bit
SAR ADC is built in MATLAB using the strong-arm latch comparator parameters
extracted via SPICE simulation (see Fig. 3.4). For simplicity, other components in
the SAR ADC are assumed to be ideal.

In a conventional SAR ADC, Vcmi is fixed, and thus, the comparator offset
is a constant and does not affect the ADC linearity. By contrast, in both monotonic
switching and the proposed switching schemes, Vcmi changes every comparison cy-
cle, which leads to varying comparator offset (see Fig. 3.4) and degrades the ADC
linearity. To examine the influence of offset variation, we first perform 1000-time
Monte Carlo simulations for the 11-bit ADC with Vcmi dependent offset variation
but no noise. The SNDR histograms are shown in Fig. 3.7. Since Vcmi in mono-
tonic switching changes from Vcm to VDD , it leads to large varying offset, which
significantly degrades the ADC SNDR. For the proposed switching scheme, since
its Vcmi variation is only VDD /4 and it starts from Vcm and ends at Vcm , its SNDR
is much higher. However, because Vcmi still undergoes large changes in the first
several MSB comparisons, there are still appreciable SNDR degradations in cases

56
 1
Monotonic up

0.8 Optimized speed

0.6
Vcmi / Vref

Bidirectional up down

0.4

Optimized noise

0.2

Monotonic down
0
1 2 3 4 5 6 7 8 9 10 11
Comparison cycle

Figure 3.6: Comparator input common-mode variation for different switching tech-
niques.

where the comparator offsets are large and have considerable variations. To solve
this problem, a small redundant capacitor can be added after the 6th MSB capac-
itor, which corrects the errors due to offset variations in the first 6 comparisons.
For the comparisons afterwards, the change in Vcmi is within Vref /128, and thus,
the comparator offset variation is negligible and does not degrade SNDR. Fig. 3.8
shows the simulation results with redundancy added. The performance for the ADC
with the proposed switching scheme is fully restored. The ADC performance for
the monotonic switching is also improved, but there is still about 20% probability
for having less than 67 dB SNDR. Although adding redundancy can effectively ad-

57
dress the offset variation problem, it cannot solve the SNDR loss due to increased
comparator noise. For monotonic up switching, because Vcmi converges to VDD , its
comparator noise is much larger than that for the proposed technique especially for
the last several noise sensitive LSB comparisons. This leads to SNDR degradation.
Fig. 3.9 shows the simulated SNDR histograms with offset, noise, and redundancy.
The average SNDR for monotonic switching is 60 dB, while that for the proposed
switching is 67 dB. Note that this 7 dB SNDR improvement comes without any
penalty in the comparator power. Moreover, assuming a target SNDR of 63 dB
as in our prototype, the proposed switching technique can achieve close to 100%
yield, while that for the monotonic switching is less than 10%.

Monotonic switching Proposed switching

200 300

150
200
Count

Count

100
100
50

0 0
40 60 80 50 60 70
SNDR SNDR

Figure 3.7: Simulated SNDR with comparator offset variation.

3.3.2 Comparator input common-mode voltage optimization

The proposed switching technique can be generalized to allow the compara-

tor common-mode voltage to converge to any desired voltage. It provides designers
the freedom to optimize the comparator for different design specifications. For ex-
ample, for low-speed high-resolution applications, it is desirable to reduce Vcmi to

58
 Monotonic switching Proposed switching
800 400

600 300

Count

Count
400 200

200 100

0 0
40 60 80 67 68 69
SNDR SNDR

Figure 3.8: Simulated SNDR with both comparator offset variation and a redundant
capacitor after the 6th MSB capacitor.

Monotonic switching Proposed switching

400 300

300
200
Count

Count

200
100
100

0 0
40 60 80 60 65 70
SNDR SNDR

Figure 3.9: Simulated SNDR with comparator offset variation, noise and a redun-
dant capacitor after the 6th MSB capacitor.

minimize the comparator noise (see Fig. 3.4). To this end, we can initialize the
DAC array in such a way that the second MSB capacitor is connected to ground
while all other capacitors are connected to Vref . As a result, during DAC switching,
Vcmi first goes down, then goes up, and finally goes down to 0.25 × 1.8 = 0.45V
(see Fig. 3.6). In this configuration, the simulated comparator input referred noise
is only 100 µV, which is 70% less than that for Vcmi = 0.9V . For high-speed
medium-resolution applications, it is preferred to place Vcmi at higher voltages (see

59
Fig. 3.4). This can be achieved by letting Vcmi go up during the first two MSB
comparisons and then go down, as shown in Fig. 3.6. This way, Vcmi converges to
0.75VDD , leading to a 10 ps shorter comparator resolve time. Note that the speed
improvement by adjusting Vcmi can be more significant in advanced technology
nodes with low power supply voltage.

3.3.3 Capacitance reduction

For the 4-bit SAR ADC shown in Fig. 3.5, the proposed BSS switching
technique requires a total capacitance of only 8C. By contrast, the conventional
switching technique requires in total 32C and the monotonic switching requires
16C. Thus, for a fixed unit capacitor size which is often determined by matching
accuracy and/or the fabrication technology, the proposed technique reduces the to-
tal DAC capacitance by 4 times, leading to reduced chip area and power. When de-
signed for the same kT /C noise with a fixed total DAC capacitance, the proposed
technique allows a 4 times bigger unit capacitor, which provides better matching
and simplifies layout design.

3.3.4 Reduced DAC reference energy

The energy required from Vref during each DAC switching is shown in blue
boxes along the arrows in Fig. 3.5. The proposed technique consumes zero energy
during the switching of the MSB capacitor, as it involves only charge redistribu-
tion [Sanyal and Sun [2014]]. For the 4-bit case using the proposed technique, the
2
average reference energy during the comparison phase is (0.5CVref . Once all con-

60
versions finish, the DAC array needs to be reset to initial condition for sampling,
2
which dissipates an additional energy of 1.75CVref . Thus, the total reference en-
2
ergy is 2.25CVref . For an n-bit SAR ADC, the total reference energy including
both comparison and reset phases is:

∑
n−1
1
Eref = (2 n−3
+ 2n−3−i − )CVref
2
(3.8)
i=2
2

For an 11-bit prototype ADC, the reference energy for different output codes
is plotted in Fig. 3.10. On average, the conventional switching technique con-
2 2
sumes 2729CVref . The monotonic switching technique consumes 512CVref . The
2
proposed technique consumes 383CVref , which is 86% and 25% smaller than the
conventional and monotonic switching schemes, respectively.
Reference energy [CVref]

1000 Monotonic switching

2

Proposed switching
800

600

400

200

0
0 500 1000 1500 2000
Output code
Figure 3.10: DAC reference energy for an 11-bit SAR ADC with different output
codes.

Compared to monotonic switching, the proposed BSS switching technique

requires an additional reference voltage Vcm , but the Vcm generator does not need

61
to deliver a large current as it is only connected to the small LSB capacitor (see
Fig. 3.5). It is interesting to note that the averaged energy drawn from Vcm is 0.
The reason is that Vcm sends energy (charge) to the DAC for 50% of the time while
it receives energy (charge) from the DAC for the rest of the time. Thus, a large
capacitor can be placed at Vcm to average out the charges, and output resistance for
the Vcm generator can be large. This reduces the power and the design effort for the
Vcm generator. Moreover, since Vcm is used only for the LSB comparison, it does
not need to be accurate. Simulation shows that 5% deviation in Vcm only results in
a DNL error of 0.05 LSB. This further simplifies Vcm generation.

3.4 Prototype ADC Design

A prototype implementing the proposed bidirectional single-side switching

technique was designed in 180nm CMOS process. The following subsections pro-
vide detailed description of the prototype.

3.4.1 Detailed circuit schematics

Fig. 3.11 shows an 11-bit SAR ADC using the proposed BSS switching
technique.

The capacitive DAC is implemented using binary-weighted capacitors. A

redundant capacitor C3 = 8C is added to recover conversion errors during the first
several MSB comparisons with large comparator common-mode variations. The
total capacitance including both sides of the DAC array is 1040C for the proposed
technique, which is 4 times smaller than that for the conventional switching tech-

62
 Vinp Vcm
Vref
S8 S7 S6 S5 S4 S3 S3 S2 S1 S0 S0
C8 C7 C6 C5 C4 C3 C3 C2 C1 C0 C0

256C 128C 64C 32C 16C 8C 8C 4C 2C C C

Vcm
Vcm
256C 128C 64C 32C 16C 8C 8C 4C 2C C C
C8 C7 C6 C5 C4 C3 C3 C2 C1 C0 C0
SAR
S8 S7 S6 S5 S4 S3 S3 S2 S1 S0 S0 Logic
Vinn D0~D11
Vref Vcm clk
Figure 3.11: Architecture of the 11-bit prototype SAR ADC.

nique that requires 4096C. In our design, the size of the unit capacitor is limited by
matching requirement. For our proposed BSS switching technique, the equivalent
LSB capacitor size is 0.5C due to the use of Vcm . To ensure that the worst case
DNL is within 1 LSB, we have [Lin and Bult [1998]; Harpe et al. [2011]]:
√
1040
3σDN L = 3 σu < 1 (3.9)
0.5

where σu is the unit capacitor mismatch in percentage. From (9), has to be smaller
than 0.5%. This translates to the minimum MIM capacitor size of 2 fF based on
foundry provided mismatch data. Out of serendipity, the smallest available MIM
capacitor is also 2 fF. Thus, it is chosen as the unit capacitor size. The total DAC ca-
pacitance is 2 pF. The kT /C noise of about 100 µV does not limit the performance
of the ADC with an LSB size of 1.8 mV under VDD = 1.8V .

Since the unit capacitor size is only 2 fF, the routing parasitic capacitors

63
have a considerable influence on the capacitor matching accuracy. Although digital
calibration techniques can be used to compensate capacitor mismatches [Chen et al.
[2014b]; Ragab et al. [2015]; Chang et al. [2013]], they increase design complexity
and digital power. In this design, a segmented common-centroid layout technique is
used to simplify the routing, thus minimizing the parasitic capacitors of the routing
wires. Fig. 3.12 shows the floor plan for a single side capacitor array. The ca-
pacitors are separated into two groups {C8 , C7 , ..., C3 } and {C3 , C2 , C1 , C0 }. The
common-centroid rule is applied horizontally for the first group and vertically for
the second. The second group is placed at the right side of capacitor array instead
of at the center [Liu et al. [2010a]] to simplify routing. Dummy cells, represented
as D, are added for better matching. The capacitors’ top plates are connected using
the highest metal M6 while the bottom plates are connected using a lower metal
M2. This minimizes the parasitic capacitors between routing wires. Post-layout
simulation shows that with the proposed layout, the ADC is able to achieve an INL
within ±0.5 LSB and an SFDR of 79 dB.

Bottom plate sampling is used to ensure high linearity. The sampling switch
is implemented as a transmission gate, and an NMOS switch is used to pass Vcm .
Both the transmission gate and NMOS switch are sized large enough to reduce
resistance and ensure linearity.

The comparator is a strong-arm latch (see Fig. 3.1). No pre-amplifier is

used. Based on simulation, the comparator rms input referred noise is 300 µV at
Vcmi = 0.9V under 1.8 V power supply. The comparator outputs are connected to
two nearby inverters that isolate the comparator output nodes from other loading

64
 vertical common centroid
horizontal common centroid
Figure 3.12: DAC capacitor array floor plan.

components, which reduces offset due to load mismatch. An XOR gate follows the
inverters to generate a valid signal to trigger the SAR logic.

Since the target ADC sampling rate is low, a synchronous SAR logic is used
for simplicity and robustness. Similar to [Zhang et al. [2012]], a 4-bit counter and
a decoder are used to generate the sampling clock clks and comparator clock clkc.
To perform the proposed switching technique, a modified SAR logic is developed
as shown in Fig. 3.13(a). It is built with a ring counter and a data register. It pro-
duces the initial control sequence of {0, 1, 1, ..., 1} for the DAC at the end of the
sampling phase. Once the comparator resolves after the first comparison, a valid
signal is generated and triggers the ring counter. It then triggers the correspond-
ing DFF in the data register to store the comparator output, which also serves as
the DAC control signal to perform the proposed switching technique until the last

65
comparison.

As shown in Fig. 3.13(b), a gated inverter is used as the DAC control switch.
During the sampling phase with clks being high, the DAC switch is disabled, and
thus, it does not disturb the bottom plate sampling. During the comparison phase
with clks being low, the capacitor is switched between Vref and ground depending
on the control signal. The sizes of the top three MSB switches are scaled according
to their connected capacitor sizes. The LSB switch is sized 4 times the minimum
size due to its reduced overdrive voltage when passing Vcm . Other switches have
the same minimum size since the settling requirements are relaxed with small ca-
pacitors. All digital gates are optimized based on the method of logical efforts.
clks Vref
Valid
S R R R R R R
DFF DFF DFF DFF DFF DFF DFF
D Q D Q D Q D Q D Q D Q D Q clks

dac<i>

R S S S S S Ci
DFF DFF DFF DFF DFF DFF out<0>
D Q D Q D Q D Q D Q D Q Vin,comp
clksb
Comp

dac<11> dac<10> dac<9> dac<2> dac<1>

(a) (b)

Figure 3.13: Schematic for (a) the SAR logic and (b) the DAC switch.

3.4.2 Measurement Results

The prototype SAR ADC is fabricated in 0.18-µm CMOS process. Fig.

3.14 shows the die photo. The active area is 0.1 mm2 . The rest of the chip area
is filled with de-coupling capacitors and metal fills. The power supply is 1.8V
and the sampling rate is 1.05 MS/s. Fig. 3.15 shows the measured DNL and INL,

66
which are +0.56/ − 0.83 LSB and +0.64/ − 0.69 LSB. The major INL jumps occur
at first and second MSB transitions, which are due to capacitor mismatches. The
measured SNDR and SFDR are 64.3 dB and 78.7 dB, respectively. No calibration
is performed. To evaluate the influence of process variation, 4 chips are randomly
picked and tested. Their measured SNDR and SFDR are consistently beyond 64 dB
and 78 dB. The measured total power consumption is 60.4 µW at VDD = 1.8V.

Switches
320 µm Capacitor Array
330 µm

ADC Comp
Core
Switches

Capacitor Array

Figure 3.14: Die micrograph and zoomed view.

To reduce the power consumption, the measurements are repeated with a

reduced power supply of 1V while maintaining the same sampling rate of 1.05
MS/s. Fig. 3.16 shows the measured spectrum with 500 kHz full-scale input at
VDD = 1V . SNDR and SFDR are 63.4 dB and 76.6 dB, respectively.

Fig. 3.17 shows the measured SNDR and SFDR versus the input amplitude.
The SNDR is limited by noise rather than distortion. Fig. 3.18 shows the SNDR
and SFDR versus the input frequency. The SNDR stays almost constant at 63 dB.

Comparator noise plays a key role in the resolving accuracy. To verify the

67
 1

0.5

DNL [LSB] 0

0.5

1
0 500 1000 1500 2000
Output code
1

0.5
INL [LSB]

0.5

1
0 500 1000 1500 2000
Output code

Figure 3.15: Measured DNL and INL.

0
SNDR = 63.4 dB
Spectrum [dBFS]

SFDR = 76.6 dB
−40

−80

−120

0 0.1 0.2 0.3 0.4 0.5

Normalized frequency [f/fs]

Figure 3.16: Measured 65536-point FFT spectrum with VDD = 1V.

analysis, we vary the comparator input common-mode voltage Vcmi and measure
ADC output Dout at Vin = 0V. The measured probability densities for Dout at

68
 80
SNDR

SNDR and SFDR [dB]

SFDR
60

40

20
−35 −30 −25 −20 −15 −10 −5 01
Input amplitude [dBFS]

Figure 3.17: Measured SNDR and SFDR versus input amplitudes.

80
SNDR and SFDR [dB]

70

60
SNDR
SFDR
50
100 200 300 400 500 600 700 800 900
Input frequency [kHz]

Figure 3.18: Measured SNDR and SFDR versus input frequencies.

VDD = 1V with Vcmi = 0.4V and Vcmi = 0.9V are shown in Fig. 3.19 together with
the fitted normal distributions. When Vcmi = 0.4V, Dout is centered at code 1051
and spread over only 3 bins with a standard deviation of 0.36 LSB. By contrast,
when Vcmi is increased to 0.9V, Dout is centered at code 1049 and spread over 8
bins with a standard deviation of 0.70 LSB. The shift of Dout center indicates the

69
comparator offset variation, which is due to Vcmi induced preamplifier gain change
as explained in (3.1). The increase in the ADC noise, which is dominated by the
comparator, is also the result of preamplifier gain reduction as explained in (3.3).
Fig. 3.20 shows the measured ADC rms noise with different Vcmi . It clearly shows
that a larger Vcmi leads to increased comparator noise, which matches well with the
analysis.

1.2
Probability density function

data for Vcm=0.4V

1 fitted curve for Vcm=0.4V
0.8 data for Vcm=0.9V
fitted curve for Vcm=0.9V
0.6
0.4
0.2
0
1046 1048 1050 1052
Dout

Figure 3.19: Measured Dout distribution with Vcmi = 0.4V and Vcmi = 0.9V .

The measured total ADC power is 24 µW at VDD = 1V. The DAC, the
comparator, and the SAR logic consume 5.8 µW, 8 µW, and 10.2 µW, respectively.
The DAC reference power accounts for 24% of the total power. This compares
favorably to other works with conventional switching technique, such as [Zhang
et al. [2012]] whose DAC power is 62% of the total power. The figure-of-merit
(FOM) is 19.9 fJ/conversion-step.

Table 3.1 summarizes the ADC performance under both 1.8V and 1V power
supplies. Table 3.2 compares it with recent works with similar power supply volt-

70
 0.8

[LSB] 0.7

0.6

0.5
n

0.4

0.4 0.5 0.6 0.7 0.8 0.9

Vcmi [V]

Figure 3.20: Measured Dout standard deviation with different Vcmi at Vin = 0V.

ages and sampling rates. This prototype ADC achieves the state-of-the-art perfor-
mance, especially considering its relatively old 180 nm process.

Table 3.1: Measured performance summary

Process [nm] 180

Sampling rate [MS/s] 1
Resolution [bit] 11
Input capacitance [pF] 2.1
2
Active area [mm ] 0.1
Power supply [V] 1.8 1.0
DNL [LSB] +0.56/-0.83 +0.53/-0.85
INL [LSB] +0.64/-0.69 +0.68/-0.91
SNDR [dB] 64.3 63.4
SFDR [dB] 78.7 76.6
ENOB [bit] 10.4 10.3
Power breakdown
DAC power [µW] 11.5 5.8
Comparator power [µW] 20.1 8.0
SAR logic power [µW] 28.8 10.2
Total power [µW] 60.4 24.0

71
 Table 3.2: Performance comparison

Liu Liu Huang Harpe Chang Kuo This work

et al. et al. et al. et al. et al. and
[2010a] [2010b] [2013] [2011] [2013] Hsieh
[2011]
Process [nm] 130 180 90 90 65 180 180
Supply voltage [V] 1.2 1.0 1.0 1.0 1.2 0.9 1.0
Sampling rate [MS/s] 50 10 30 10 50 1 1
Resolution [bit] 10 10 10 8 12 10 11
ENOB [bit] 9.2 9.8 9.2 7.8 10.9 8.4 10.3
Power [µW] 826 98 980 26.3 2090 7.2 24
FOM [fJ/conv-step] 29 11 57 12 21.9 21.6 19.9

72
 Chapter 4

Low-power and High-speed Single-channel SAR ADC

This chapter presents a low-power and high-speed single-channel SAR ADC.

It uses a loop-unrolled architecture with multiple comparators. Each comparator is
used not only to make a comparison but also to store its output and generate an
asynchronous clock to trigger the next comparator. The SAR logic is significantly
simplified to increase speed and reduce power. The comparator offset and decision
time are optimized with a bidirectional single-side switching technique by control-
ling the input common-mode voltage Vcm . To remove the nonlinearity due to com-
parators’ offset mismatch, a simple and effective Vcm -adaptive offset calibration
technique is proposed. A 40nm prototype ADC achieves 35dB SNDR and 48dB
SFDR at 700MS/s sampling rate. It consumes 0.95mW, leading to a Walden FOM
of 30fJ/conv-step.

This chapter is organized as follows: an introduction of existing high speed

SAR ADC design techniques is first presented. The proposed SAR ADC archi-
tecture is shown next. Finally, the detailed prototype ADC implementation and

0
This chapter is a partial reprint of the publication: Long Chen, Kareem Ragab, Xiyuan Tang,
Jeonggoo Song, Arindam Sanyal, and Nan Sun, “A 0.95-mW 6-b 700-MS/s single-channel loop-
unrolled SAR ADC in 40-nm CMOS,” to appear in IEEE Transactions on Circuits and Systems
II (TCAS–II). I thank all the co-authors for their valuable advice in designing and testing of the
prototype.

73
measurement are presented.

4.1 Introduction

High-speed low-resolution analog-to-digital converters (ADCs) are required

by many demanding applications, such as high speed serial link transceivers and
communication systems. Compared with pipelined and ∆Σ ADCs, SAR ADCs are
more power efficient and scaling friendly due to their mostly digital architecture.
Several techniques have been developed to increase the speed of SAR ADCs [Jiang
et al. [2012]; Chen and Brodersen [2006]; Yang et al. [2010]; Wei et al. [2012];
Hong et al. [2015]; Kull et al. [2013]; Tai et al. [2014b]; Lin et al. [2010]; Chen et al.
[2013]; Rahman et al. [2014]; Verbruggen et al. [2012]]. The first asynchronous
SAR ADC was proposed in [Chen and Brodersen [2006]] to shorten the time du-
ration of each comparison cycle. The SAR ADC speed can also be improved by
using multi-bit-per-cycle architectures to reduce the number of comparisons, how-
ever at the cost of increased hardware complexity [Wei et al. [2012]; Hong et al.
[2015]]. Other effective high-speed techniques include using alternate comparators
to save the comparator reset time [Kull et al. [2013]] or pipelining two-stage SAR
ADCs [Tai et al. [2014b]]. Recently, several works arrange multiple comparators to
further increase the speed [Jiang et al. [2012]; Lin et al. [2010]; Chen et al. [2013];
Rahman et al. [2014]; Verbruggen et al. [2012]]. A binary-search ADC was pro-
posed in [Lin et al. [2010]], which describes a transitional structure between flash
and SAR ADCs. However, the hardware cost is high in [Lin et al. [2010]] as this
technique requires additional switching networks and 2N −1 comparators for an N -

74
bit design. The loop-unrolled architecture of [Jiang et al. [2012]; Verbruggen et al.
[2012]] employs a dedicated comparator for each comparison cycle. The compari-
son result is stored directly at the comparator output. As a result, the SAR logic is
greatly simplified, leading to reduced power and delay. Although more compara-
tors are used compared to the conventional SAR architecture, the total comparator
power does not increase since each of them is fired only once during the whole
conversion. Nevertheless, the comparator common-mode voltage Vcm varies signif-
icantly and eventually goes to VDD in [Jiang et al. [2012]; Verbruggen et al. [2012]],
resulting in large comparator offsets and reduced linearity. Both work [Verbruggen
et al. [2012]] and [Jiang et al. [2012]] require complicated calibrations for compara-
tors’ offset mismatches, which increase both the power consumption and the design
complexity.

This work proposes a novel loop-unrolled SAR ADC with two new key
techniques to improve the linearity and the power efficiency. First, in order to ad-
dress the large Vcm variation issue, a bidirectional single-side switching technique
is employed. It reduces the comparator offset by appropriately controlling Vcm .
The comparator decision time is also optimized. In addition, it allows a reduced
number of DAC unit capacitors, which reduces the DAC area and the routing par-
asitics. Second, to further improve the linearity, a novel Vcm -adaptive offset cali-
bration technique is proposed to calibrate the comparators’ offset mismatch. The
proposed calibration technique has very low hardware complexity. It can calibrate
the comparator offset at its operating Vcm following the proposed switching proce-
dure. A prototype ADC is implemented in 40nm CMOS. It achieves 34.8dB SNDR

75
and 47.8dB SFDR at a sampling rate of 700MS/s, while consuming only 0.95mW
power from a 1.2V supply.

4.2 Proposed SAR ADC Architecture

Figure 4.1: (a) Architecture and (b) timing diagram of the proposed SAR ADC.

The proposed 6-bit SAR ADC architecture is shown in Fig. 4.1(a). It con-
sists of a clock generator, a sampling network, 2 capacitive DACs, 6 comparators,

76
and a calibration unit. The clock generator generates the required timing phases.
The sampling switches S1 and S2 are bootstrapped to ensure high sampling lin-
earity at high input frequencies. The DACs are implemented using binary weighted
capacitors. Small dynamic comparators with offset calibration are used to minimize
power consumption. Dynamic OR gates with reset and controlled delay are used to
generate the asynchronous clocks. The comparator offset mismatch is foreground
calibrated by the calibration unit. The LSB comparator outputs are combined by a
NOR gate to generate a ready signal, which indicates the end of the whole ADC
conversion.

The timing diagram is shown in Fig. 4.1(b). When clks is low, the input
voltages are top-plate sampled on the DACs through S1 and S2 . All comparators’
outputs are reset to low through dynamic OR gates controlled by clkse. The ready
signal is reset to high. When the sampling phase ends, both clkse and clks go to
high. The MSB comparator is triggered by clk[5] which is two gates delay of clkse.
Depending on the sampled input voltages, the MSB comparator makes its decision
dp[5] and dn[5], which directly control the capacitive DACs to perform the proposed
switching technique without the need for any shift register based SAR logic. The
dynamic OR gate delay is controlled to provide adequate time for DAC settling. It
generates clk[4] to trigger the second MSB comparator. This procedure propagates
in a ‘domino’ fashion until the LSB comparison finishes. The ready signal goes to
low. The next sampling phases starts after the comparator outputs are latched by
the falling edge of clkse.

Compared to the conventional SAR ADC, the SAR logic is greatly sim-

77
plified in this loop-unrolled architecture. There is no need for any shift register
based sequencer or DFF based data storage because all comparator results are di-
rectly stored at the comparator outputs. The comparators are not reset during the
ADC conversion process. The asynchronous clock is generated easily by ORing
the comparators’ outputs. The reduced logic complexity reduces the circuit power,
minimizes the chip area, and increases the speed. In the proposed architecture,
the conversion time is reduced in three ways compared to the conventional asyn-
chronous SAR ADC [Chen and Brodersen [2006]]. First, no DFF or latch delay
is needed to store the comparator output in our critical path. Second, comparators
are reset simultaneously, and thus, no comparator reset time is needed for every
comparison cycle. The comparator reset time can be a speed bottleneck especially
in most advanced technology node where both logic delay and DAC settling time
are small [Kull et al. [2013]]. Third, the proposed design allows independent op-
timization for each comparison cycle. In other words, each OR gate delay can be
adjusted based on the corresponding DAC settling time. The comparator power can
also be optimized using the technique in [Ahmadi and Namgoong [2015]]. Overall,
the optimized critical path for each comparison cycle in the proposed design can be
represented as:
T = tcomp,decision + max{tDAC , tOR } (4.1)

This design optimizes tcomp,decision by optimizing comparator input common-

mode voltage with a bidirectional single-side switching technique. tDAC and tOR
are optimized by using small unit capacitor and specially designed dynamic OR
gate, respectively. Based on SPICE simulation, tDAC is greater than tOR for the

78
first 2 MSB bits where the DAC capacitors are large. tOR dominates over tDAC for
the last 4 LSB bits.

4.3 Prototype ADC Design

4.3.1 Detailed circuit schematics

Fig. 4.2 shows the dynamic comparator with offset calibration. Two vari-
able MOS capacitors are added at the drain of input transistors M 1 and M 2 for
calibration purpose. Comparator offset and decision time are two key parameters
in this design. Unlike the conventional SAR ADC with one single comparator, the
loop-unrolled architecture employs 6 comparators. All comparators have offsets
and their offset mismatches degrade the ADC linearity. Comparator decision time
is in the critical timing path as shown in (4.1). Thus, it is desirable to design com-
parator with small offset and fast decision time. One important factor that influences
both offset and decision time is the comparator input common-mode voltage Vcm
[Wicht et al. [2004]]. A small Vcm is preferred to reduce the offset. The reason is
that the pre-amplification gain is larger at small Vcm , which suppresses the offset
contribution from the latch. A large Vcm helps reduce the pre-amplification time
but the time duration of the latch regeneration phase is longer due to the reduction
in the pre-amplification gain [Wicht et al. [2004]]. There exists an optimized Vcm
for decision time. The simulated 1σ offset and decision time are shown in Fig.
4.3. Here the decision time is defined as the time it takes for the comparator output
differential voltage to reach 0.7VDD given 1mV differential input. It suggests the
optimal value of Vcm is around 0.8V, where the decision time is minimized and the

79
offset is also small.

M7 M5 M6 M8

clk clk

M9 M3 M4 M10

Vin+ Vin-
M1 M2
dn dp

M11 M12
clk
M0
calp calm

Figure 4.2: Dynamic comparator with varactor loading.

40 50
1 offset
Decision time

30 45
Decision time [ps]
1 offset [mV]

20 40

10 35

0 30
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2
Vcm [V]

Figure 4.3: Simulated comparator offset and decision time at different Vcm .

80
 A modified bidirectional single-side (BSS) switching technique based on
[Chen et al. [2014a]; Sanyal and Sun [2014]] is employed in the design. The BSS
technique can reduce the number of unit capacitors by 4 times compared to conven-
tional switching technique and 2 times compared to monotonic switching technique
[Jiang et al. [2012]; Verbruggen et al. [2012]]. The binary capacitors used in the
DAC are [8CU , 4CU , 2CU , CU , CU ] for 6-bit implementation, with a unit capacitor
CU of 1.9fF. Thanks to the small unit capacitor, the DAC settling time tDAC is short
and it is easy to guarantee the settling error to be within 1/2 LSB. No redundancy
is provided to avoid additional comparison cycles. The reduced number of unit
capacitors reduces the layout complexity since fewer interconnections are needed.
The measured result shows the capacitors are matched well for 6-bit accuracy with
a similar layout strategy in [Chen et al. [2014a]]. Unlike prior loop-unrolled works
[Jiang et al. [2012]; Verbruggen et al. [2012]], no capacitor calibration is needed
here.

The proposed BSS switching scheme is shown in Fig. 4.4. DAC is con-
nected to [0, 0, 0, 1, 1/2] when sampling the inputs, where 0, 1 and 1/2 indicate
ground, VDD and VDD /2, respectively. Instead of switching DAC capacitors from
ground to VDD monotonically, the proposed technique switches the first 3 MSB ca-
pacitors [8CU , 4CU , 2CU ] from ground to VDD , and then, switches the left side CU
capacitors from VDD to ground.

The simulated Vcm variation is shown in Fig. 4.5. In the proposed switching
scheme, Vcm stays close to 0.8V after the first comparison cycle, where the opti-
mized decision time is achieved and comparator offset is kept small [Chen et al.

81
 0 0 0 1 1/2 dn[5] dn[4] dn[3] dp[2] dp[1]

Vinp 8C 4C 2C C C 8C 4C 2C C C

Vinm 8C 4C 2C C C 8C 4C 2C C C

Figure 4.4: BSS switching scheme.

[2016]]. Compared with monotonic switching scheme in [Jiang et al. [2012]; Ver-
bruggen et al. [2012]], where Vcm can converge to VDD , the 1σ offset can be reduced
from 32mV to 10mV and decision time can be reduced by 10% in the proposed
switching technique. Note that Vcm decreases slightly when the comparators are
fired due to the comparator kick-back noise. The kick-back noise is a common-
mode noise to the first order, which does not disturb the settling of the comparator
differential mode voltage Vdm (see Fig. 4.5). Thus, the kick-back noise does not
degrade the linearity for this 6-bit ADC.

In the prototype ADC, the full-scale differential input swing is 1.4V and 1
LSB size is about 20mV. The optimized comparator 1σ offset is 10mV, which is
comparable to the LSB size. To avoid linearity degradation due to offset mismatch,
a Vcm -adaptive offset calibration technique is proposed. The calibration technique
works as follows. When the ADC is in the calibration mode, the calibration control
switch S0 is on and a zero differential input voltage is sampled through S1 and S2
(see Fig. 4.1). A low-frequency external clock clkext is provided. Each OR gate’s
delay is set large enough such that before each comparator is triggered, its differen-
tial inputs settle well and stay close to zero. If there is no offset, each comparator’s

82
 1.2 Vin+
Vin-
Vcm
Vdm
0.8 clks
Voltages [V]

0.4

0
clk[5]

clk[4]
clk[3]
clk[2]
clk[1]
clk[0]

0.4
0 0.5 1 1.5 2
Time [ns]

Figure 4.5: Simulated time-domain waveforms for comparator inputs Vin+/Vin-, its
common mode voltage Vcm and its differential mode voltage Vdm .

output jumps between “1” and “0” due to thermal noise. Comparator thermal noise
in the design is about 400µV, which is much smaller than the offset. With a large
offset in presence, the comparator’s output keeps staying at either “1” or “0”. The
offset can be calibrated by tuning the MOSFET based varactors shown in Fig. 4.2,
whose values are controlled by its gate voltage calp/calm. The calibration range
is designed to be 30mV which is 3 times the simulated 1σ offset. By observing the
comparator’s output, we can tell whether the comparator offset has been calibrated
or not. The calibration is finished when the comparator output is evenly distributed
between “1” and “0”. 1024 ADC outputs are captured by a logic analyzer dur-
ing the calibration. The comparator offset is removed when the probability that its

83
output code equals to “1” is around 50%. Note that the probability does not need
to be exactly 50% since the calibration accuracy is relaxed for a 6-b design. The
measured probability of comparator outputs being “1” versus its calibration voltage
calm for the MSB comparator is plotted in Fig. 4.6. The proposed calibration is
simple as it only requires one additional switch S0. It does not require special DAC
patterns to generate the operating Vcm in [Verbruggen et al. [2012]] or special input
voltages that cause metastability at different comparators in [Jiang et al. [2012]].
Furthermore, since the same DAC switching procedure happens during calibration,
each comparator offset is calibrated at the same Vcm as that of the normal ADC
operation. This is necessary in BSS or monotonic switching technique since Vcm
varies in each comparison cycle and the comparator offset depends strongly on Vcm
[Chen et al. [2014a]; Wicht et al. [2004]].
Probability of output = 1

0.8

0.6

0.4

0.2

0
0 0.1 0.2 0.3 0.4 0.5
calm voltage [V]

Figure 4.6: Probability of MSB comparator output being ‘1’ versus its calm with
calp fixed at 1.2V.

A clock generator similar to [Chen and Brodersen [2006]] is used. The input
sampling switches S1 and S2 are bootstrapped with the circuit shown in Fig. 4.7(a)
[Siragusa and Galton [2004]]. It is simple and does not require two complementary

84
clock phases [Jiang et al. [2012]]. A thick-oxide device provided in the process
is used to tolerate a potential voltage higher than 1.2V. OR gate needs to have a
reset function together with controlled delay with the aim to provide enough time
for DAC settling. Since the unit capacitor employed in this design is quite small
(1.9fF), the DAC settling time is comparable with the CMOS logic gate delay. To
minimize the number of logic gates and their delay, a dynamic OR gate shown in
Fig.4.7(b) is used instead of traditional CMOS gates. V B is used to control the
current flowing through OR gate, thus tuning the gate delay.

Figure 4.7: Schematic of (a) bootstrapped sampling switch and (b) dynamic OR
gate.

4.3.2 Measurement results

The prototype ADC is fabricated in 40-nm CMOS and occupies an active

area of only 0.004mm2 , as shown in Fig. 4.8. Unlike the floor plan in [Jiang
et al. [2012]], where the DAC and SAR logic are placed side by side, SAR logic

85
is placed between two CDACs in our design. By doing this, long routing wires are
avoided, which saves both area and power. The outputs are decimated by 32 using
5 divide-by-2 toggle flip-flops to facilitate ADC measurements. All measurements
are performed under 1.2V power supply.

Figure 4.8: Chip micrograph.

Fig. 4.9 shows the measured DNL and INL before and after perform-
ing the proposed comparator offset calibration. Before calibration, large DNL
and INL jumps happen due to comparator offset mismatches. After performing
the Vcm -adaptive offset calibration, DNL and INL are reduced significantly to -
0.4LSB/0.9LSB and -0.3/0.6LSB, respectively. The effectiveness of the calibration
can also be observed from the measured spectrum shown in Fig. 4.10, where 4.3dB
SNDR and 8.5dB SFDR improvements are achieved after calibration.

86
 3
Before cal
2 After cal
DNL [LSB]

1

2
0 10 20 30 40 50 60
Output code

3
Before cal
2 After cal
INL [LSB]

1

2
0 10 20 30 40 50 60
Output code

Figure 4.9: Measured DNL/INL before calibration (dotted line) and after calibration
(solid line).

Fig. 4.11 shows the measured FFT spectrum with 700MS/s sampling rate
and 300MHz input. The measured SNDR and SFDR are 34.8dB and 47.8dB,
respectively, leading to 5.5-bit ENOB. When the input frequency is reduced to
10MHz, the measured SNDR and SFDR are 35.1dB and 49.5dB, respectively. Fig.
4.12 shows SNDR and SFDR at different sampling frequencies with a 50MHz in-

87
 0 0
Before cal After cal

10 SNDR=30.5dB
SFDR=41.8dB

10 SNDR=34.8dB
SFDR=50.3dB

20
20
Spectrum [dBFS]

30
30

40
40

50
50

60
60

70
70

80
80
0 3.75 7.5 0 3.75 7.5
Frequency [MHz] Frequency [MHz]

Figure 4.10: Measured 1024p FFT spectrum before calibration and after calibration
with 600MS/s sampling rate and 100MHz input.

put, at different input frequencies with 700MS/s sampling rate, and at different input
amplitudes with 300MHz input and 700MS/s sampling rate. The measured SNDR
is above 34.7dB across the whole Nyquist band. The maximum sampling rate of
700MS/s is limited by insufficient sampling time discovered during chip measure-
ments. When clkse goes low, it takes longer time to reset comparators in mea-
surements than that in simulation, which reduces the effective sampling time. The
total power consumption at 700MS/s is 0.95mW, whose breakdown is: 0.84mW
used by SAR logic and comparator, 0.09mW used by DAC, and 0.02mW used by
clock generator. The measured Walden figure-of-merit (FOM) [Walden [1999]]
at Nyquist rate is 30fJ/conversion-step and Schreier FOM [Schreier and Temes
[2005]] is 153.4dB. Table 4.1 compares the performance of the prototype ADC

88
with other recently published high-speed 6-bit SAR ADCs. As shown in Table 4.1,
this work achieves the highest SNDR and the best Walden FOM with the smallest
chip area among them.

0
fs=700MS/s

10 fin=300MHz
SNDR=34.8dB

20 SFDR=47.8dB
Spectrum [dBFS]

30

40

50

60

70

80
0 2.1875 4.375 6.5625 8.75 10.9375
Frequency [MHz]

Figure 4.11: Measured 1024p FFT spectrum with 700MS/s sampling rate and
300MHz input.

89
 55 50
55
SNDR/SFDR [dB]
50
50 40
45
45 SNDR SNDR 30
SFDR SFDR
40 40
20
35 35
SNDR
SFDR
30 30 10
200 400 600 0 100 200 300 -20 -15 -10 -5 0
fs [MS/s] fin [MHz] Input amplitude [dBFS]

Figure 4.12: Measured SNDR/SFDR under different sampling rates with 50MHz
input (left), under different input frequencies with 700MS/s sampling rate (middle),
and under different input amplitudes with 700MS/s sampling rate and 300MHz
input (right).

Table 4.1: Performance comparison

Chen Yang Tai Jiang This work

and et al. et al. et al.
Broder- [2010] [2014b] [2012]
sen
[2006]
Resolution [bit] 6 6 6 6 6
Process [nm] 130 65 40 40 40
Supply voltage [V] 1.2 1.2 1.2 1.0 1.2
Active area [mm2 ] 0.060 0.110 0.009 0.014 0.004
Number of channels 2 2 1 1 1
Sampling rate [GS/s] 0.6 1 1 1.25 0.7
SNDR(Nyq) [dB] 32 29 31.2 26.8 34.8
ENOB [bit] 5.0 4.5 4.9 4.1 5.5
Power [mW] 5.3 6.7 5.3 6.08 0.95
FOM [fJ/conv-step] 276 290 180 272 30

90
 Chapter 5

Conclusion and Future Directions

5.1 Conclusion

This thesis focuses on improving the power efficiency of SAR ADCs. Three
prototype ADCs have been developed to validate the proposed techniques. Chap-
ter 2 presents a novel noise reduction technique for SAR ADC based on statistical
estimation. To the authors’ best knowledge, this is the first work that comprehen-
sively introduces the statistical estimation theory to the field of ADC design. The
proposed technique requires minimum change to the original SAR ADC design. It
can significantly reduce both the comparator noise and the quantization error. It is
suitable for applications that require low-power high-resolution SAR ADCs.

Chapter 3 presents a low-power ADC with a bidirectional single-side switch-

ing technique. It solves the linearity degradation problem due to comparator input
common-mode voltage variation. The proposed switching technique provides de-
signers the freedom to choose where the comparator common-mode voltage con-
verges to. It is suitable for applications that require both low power and high reso-
lution.

Chapter 4 presents a high-speed and low-power single-channel loop-unrolled

SAR ADC. It proposed a simple method to calibrate the comparator offsets. The

91
proposed switching technique allows the designers to control comparator input
common-mode voltage for comparator offset and speed optimization. It can be
easily time-interleaved (TI) for even higher speed applications.

5.2 Future Directions

The proposed statistical estimation based noise reduction technique requires

the knowledge of the comparator noise. Since comparator noise is sensitive to PVT
variation, an interesting direction is to build a comparator noise extraction circuit
on chip. Thus, the look up table of the estimator can be updated when the noise
changes. It is also desirable to implement a fully reconfigurable SAR ADC by
providing the ability to change the estimation times N .

In the second prototype ADC, a unit capacitor of 2fF is used, which is the
minimum MIM capacitor provided in the PDK. Since the DAC switching power
is proportional to the unit capacitor size, a smaller unit capacitor is desired. This
can be achieved by custom designing the unit capacitor with metal wires. It is also
interesting to investigate low power calibration techniques to calibrate the capacitor
mismatch with aim to use ultra small unit capacitor. Another interesting path to
reduce the DAC size is to use hybrid architecture, such as [Sanyal et al. [2014]].
By doing this, the first stage SAR ADC only needs to resolve less number of bits,
leading to reduced DAC size and power. During the measurement, we learned that
the SAR logic consumes a significant power out of total ADC power. Although the
power of logic circuits reduces in more advanced technologies, it is still desirable
to explore circuits design techniques to optimize the logic power.

92
 Although the power and linearity of the loop-unrolled SAR ADC have been
improved by using optimized switching technique and Vcm -adaptive calibration
technique, there is still some work which can be done in the future. One pos-
sible direction is to design the background calibration technique to calibrate the
comparator offset mismatches. The other direction could be time-interleaving the
single channel ADC to achieve higher speed.

93
Appendix

94
 Appendix 1

List of publications

1. Long Chen, Kareem Ragab, Xiyuan Tang, Jeonggoo Song, Arindam Sanyal,
and Nan Sun, “A 0.95-mW 6-b 700-MS/s single-channel loop-unrolled SAR
ADC in 40-nm CMOS,” accepted in IEEE Transactions on Circuits and Sys-
tems II (TCAS–II).

2. Long Chen, Arindam Sanyal, Ji Ma, Xiyuan Tang, and Nan Sun, “Com-
parator common-mode variation effects analysis and its application in SAR
ADCs,” accepted in 2016 IEEE International Symposium on Circuits and
Systems (ISCAS).

3. Long Chen, Xiyuan Tang, Arindam Sanyal, Yeonam Yoon, Jie Cong, and Nan
Sun, “A 10.5-b ENOB 645nW 100kS/s SAR ADC with statistical estimation
based noise reduction,” IEEE Custom Integrated Circuit Conference (CICC),
2015, pp. 1–4.

4. Yeonam Yoon, Koungtae Lee, Sungjin Hong, Xiyuan Tang, Long Chen, and
Nan Sun, “A 0.04-mm2 modular Σ∆ ADC with VCO-based integrator and
0.9-mW 71-dB SNDR distributed digital DAC calibration,” IEEE Custom
Integrated Circuit Conference (CICC), 2015, pp. 1–4.

5. Arindam Sanyal, Long Chen, and Nan Sun, “Dynamic element matching with
signal-independent element transition rates for multibit delta sigma modula-
tors,” IEEE Transactions on Circuits and Systems I (TCAS–I), vol. 62, no. 5,
pp. 1325–1334, 2015.

6. Kareem Ragab, Long Chen, Arindam Sanyal, and Nan Sun, “Digital back-
ground calibration for pipelined ADCs based on comparator decision time
quantization”, IEEE Transactions on Circuits and Systems II (TCAS–II), vol.
62, no. 5, pp. 456–460, 2015.

95
 7. Arindam Sanyal, Kareem Ragab, Long Chen, T. R. Viswanathan, Shouli Yan
and Nan Sun, “A hybrid SAR-VCO ∆Σ ADC with first-order noise shaping”,
IEEE Custom Integrated Circuit Conference (CICC), pp. 1–4, 2014.

8. Long Chen, Arindam Sanyal, Ji Ma and Nan Sun, “A 24-uW 11-bit 1-MS/s
SAR ADC with a bidirectional single-side switching technique”, IEEE Euro-
pean Solid-State Circuit Conference (ESSCIRC), pp. 219–222, 2014.

9. Long Chen, Ji Ma, and Nan Sun, “Capacitor mismatch calibration for SAR
ADCs based on comparator metastability detection,” IEEE International Sym-
posium on Circuits and Systems (ISCAS), pp. 2357–2360, 2014.

10. Manzur Rahman, Long Chen, and Nan Sun, “Algorithm and implementation
of digital calibration of fast converging radix-3 SAR ADC,” IEEE Interna-
tional Symposium on Circuits and Systems (ISCAS), pp. 1336–1339, 2014.

11. Long Chen, Manzur Rahman, Sha Liu, and Nan Sun, “A fast radix-3 SAR
analog-to-digital converter,” IEEE International Midwest Symposium on Cir-
cuits and Systems (MWSCAS), pp. 1148–1151, 2013.

Future publications
1. Long Chen, Xiyuan Tang, Arindam Sanyal, Yeonam Yoon, Jie Cong, and Nan
Sun, “A 0.7V 0.6µW 100kS/s low-power SAR ADC with statistical estima-
tion based noise reduction,” submitted to IEEE Journal of Solid-State Circuits
(JSSC).

96
 Bibliography

Naveen Verma and Anantha P. Chandrakasan. An Ultra Low Energy 12-bit Rate-
Resolution Scalable SAR ADC for Wireless Sensor Nodes. Solid-State Circuits,
IEEE Journal of, 42(6):1196–1205, 2007.

Nick Van Helleputte, Sunyoung Kim, Hyejung Kim, Jong Pal Kim, Chris Van Hoof,
and Refet Firat Yazicioglu. A 160µW biopotential acquisition ASIC with fully
integrated IA and motion-artifact suppression. In Solid-State Circuits Conference
Digest of Technical Papers (ISSCC), 2012 IEEE International, pages 118–120,
2012.

Chun-Chen Liu, Soon-Jyh Chang, Guan-Ying Huang, and Ying-Zu Lin. A 10-bit
50-MS/s SAR ADC With a Monotonic Capacitor Switching Procedure. Solid-
State Circuits, IEEE Journal of, 45(4):731–740, 2010a.

Tao Jiang, Wing Liu, Freeman Y. Zhong, Charlie Zhong, Kangmin Hu, and
Patrick Yin Chiang. A Single-Channel, 1.25-GS/s, 6-bit, 6.08-mW Asyn-
chronous Successive-Approximation ADC With Improved Feedback Delay in
40-nm CMOS. Solid-State Circuits, IEEE Journal of, 47(10):2444–2453, 2012.

Hung-Yen Tai, Yao-Sheng Hu, Hung-Wei Chen, and Hsin-Shu Chen. A

0.85fJ/conversion-step 10b 200kS/s subranging SAR ADC in 40nm CMOS. In
Solid-State Circuits Conference, 2014. Digest of Technical Papers. ISSCC. 2014
IEEE International, pages 196–197, 2014a.

Pieter Harpe, Eugenio Cantatore, and Arthur Van Roermund. An oversampled

12/14b SAR ADC with noise reduction and linearity enhancements achieving up
to 79.1dB SNDR. In Solid-State Circuits Conference, 2014. Digest of Technical
Papers. ISSCC. 2014 IEEE International, pages 194–195, 2014.

Vito Giannini, Pierluigi Nuzzo, Vincenzo Chironi, Andrea Baschirotto, Geert

Van der Plas, and Jan Craninckx. An 820µW 9b 40MS/s Noise-Tolerant

97
 Dynamic-SAR ADC in 90nm Digital CMOS. In Solid-State Circuits Confer-
ence, 2008. Digest of Technical Papers. ISSCC. 2008 IEEE International, pages
238–610, 2008.

Pieter Harpe, Eugenio Cantatore, and Arthur Van Roermund. A 10b/12b 40 kS/s
SAR ADC with data-driven noise reduction achieving up to 10.1b ENOB at 2.2
fJ/Conversion-Step. Solid-State Circuits, IEEE Journal of, 48(12):3011–3018,
2013.

Mahdi Ahmadi and Won Namgoong. A 3.3fJ/conversion-step 250kS/s 10b SAR

ADC using optimized vote allocation. In Custom Integrated Circuits Conference
(CICC), 2013 IEEE, pages 1–4, 2013.

Long Chen, Xiyuan Tang, Arindam Sanyal, Yeonam Yoon, Jie Cong, and Nan Sun.
A 10.5-b ENOB 645 nW 100kS/s SAR ADC with statistical estimation based
noise reduction. In Custom Integrated Circuits Conference (CICC), 2015 IEEE,
pages 1–4, 2015.

George Casella and Roger L. Berger. Statistical inference. Duxbury Press, 1990.

Skyler Weaver, Benjamin Hershberg, Peter Kurahashi, Daniel Knierim, and Un-Ku
Moon. Stochastic Flash Analog-to-Digital Conversion. IEEE Transactions on
Circuits and Systems I: Regular Papers, 57(11):2825–2833, 2010.

Volodymyr Kratyuk, Pavan Kumar Hanumolu, Kerem Ok, Un-Ku Moon, and Kar-
tikeya Mayaram. A Digital PLL With a Stochastic Time-to-Digital Converter.
IEEE Transactions on Circuits and Systems I: Regular Papers, 56(8):1612–1621,
2009.

Bob Verbruggen, Jorgo Tsouhlarakis, Takaya Yamamoto, Masao Iriguchi, Ewout

Martens, and Jan Craninckx. A 60 dB SNDR 35 MS/s SAR ADC With
Comparator-Noise-Based Stochastic Residue Estimation. Solid-State Circuits,
IEEE Journal of, 50(9):2002–2011, 2015.

Hau-Seung Lee, Hodges David A., and Paul R. Gray. A self-calibrating 15 bit
CMOS A/D converter. Solid-State Circuits, IEEE Journal of, 19(6):813–819,
1984.

98
Vaibhav Tripathi and Boris Murmann. Mismatch characterization of small metal
fringe capacitors. In Custom Integrated Circuits Conference (CICC), 2013 IEEE,
pages 1–4, 2013.

Sunghyuk Lee, Anantha P. Chandrakasan, and Hae-Seung Lee. A 1GS/s 10b

18.9mW time-interleaved SAR ADC with background timing-skew calibration.
In Solid-State Circuits Conference, 2014. Digest of Technical Papers. ISSCC.
2014 IEEE International, pages 384–385, 2014.

Marcus Yip and Anantha P. Chandrakasan. A Resolution-Reconfigurable 5-to-10-

Bit 0.4-to-1 V Power Scalable SAR ADC for Sensor Applications. Solid-State
Circuits, IEEE Journal of, 48(6):1453–1464, 2013.

Gregory P. Harmer, Bruce R. Davis, and Derek Abbott. A review of stochastic

resonance: circuits and measurement. IEEE Transactions on Instrumentation
and Measurement, 51(2):299–309, 2002.

Long Chen, Arindam Sanyal, Ji Ma, and Nan Sun. A 24-µW 11-bit 1-MS/s SAR
ADC with a bidirectional single-side switching technique. In Solid-State Circuits
Conference, 2014. ESSCIRC 2014. Proceedings of the 40th European, pages
219–222. IEEE, 2014a.

Arindam Sanyal and Nan Sun. An Energy-Efficient Low Frequency-Dependence

Switching Technique for SAR ADCs. Circuits and Systems II: Express Briefs,
IEEE Transactions on, 61(5):294–298, 2014.

Shuo-Wei Michael Chen and Robert W. Brodersen. A 6-bit 600-MS/s 5.3-mW

Asynchronous ADC in 0.13-µm CMOS. Solid-State Circuits, IEEE Journal of,
41(12):2669–2680, 2006.

Michiel Van Elzakker, Ed Van Tuijl, Paul Geraedts, Danil Schinkel, Eric
A. M. Klumperink, and Bram Nauta. A 10-bit charge-redistribution ADC con-
suming 1.9 µW at 1 MS/s. Solid-State Circuits, IEEE Journal of, 45(5):1007–
1015, 2010.

Naveen Verma and Anantha P. Chandrakasan. A 25 µW 100kS/s 12 bit ADC for

wireless micro-sensor applications. pages 822–831, Feb 2006.

99
Dai Zhang, Ameya Bhide, and Atila Alvandpour. A 53-nW 9.1-ENOB 1-kS/s SAR
ADC in 0.13- µm CMOS for medical implant devices. Solid-State Circuits, IEEE
Journal of, 47(7):1585–1593, 2012.

Brian P. Ginsburg and Anantha P. Chandrakasan. An energy-efficient charge re-

cycling approach for a SAR converter with capacitive DAC. In Circuits and
Systems, 2005. ISCAS 2005. IEEE International Symposium on, pages 184–187.
IEEE, 2005.

Yan Zhu, Chi-Hang Chan, U-Fat Chio, Sai-Weng Sin, Seng-Pan U Ben, Rui Paulo
Martins, and Franco Maloberti. A 10-bit 100-MS/s Reference-Free SAR ADC
in 90 nm CMOS. Solid-State Circuits, IEEE Journal of, 45(6):1111–1121, 2010.

Chun-Cheng Liu, Soon-Jyh Chang, Guan-Ying Huang, Ying-Zu Lin, and Chung-
Ming Huang. A 1V 11fJ/conversion-step 10bit 10MS/s asynchronous SAR ADC
in 0.18µm CMOS. In VLSI Circuits (VLSIC), 2010 IEEE Symposium on, pages
241–242. IEEE, 2010b.

Tsuguo Kobayashi, Kazataka Nogami, Tsukasa Shirotori, and Yuihiro Fujimoto. A

current-controlled latch sense amplifier and a static power-saving input buffer for
low-power architecture. Solid-State Circuits, IEEE Journal of, 28(4):523–527,
1993.

Geert Van der Plas, Stefaan Decoutere, and Stphane Donnay. A 0.16pJ/Conversion-
Step 2.5mW 1.25GS/s 4b ADC in a 90nm Digital CMOS Process. In Solid-
State Circuits Conference, 2006. ISSCC 2006. Digest of Technical Papers. IEEE
International, pages 2310–2312. IEEE, 2006.

Bernhard Wicht, Thomas Nirschl, and Doris Schmitt-Landsiedel. Yield and speed
optimization of a latch-type voltage sense amplifier. Solid-State Circuits, IEEE
Journal of, 39(7):1148–1158, 2004.

Pierluigi Nuzzo, Fernando De Bernardinis, Pierangelo Terreni, and Geert Van der
Plas. Noise Analysis of Regenerative Comparators for Reconfigurable ADC Ar-
chitectures. IEEE Transactions on Circuits and Systems I: Regular Papers, 55
(6):1441–1454, 2008.

100
Chi-Hung Lin Lin and Klaas Bult. A 10-b, 500-MSample/s CMOS DAC in 0.6
mm2 . Solid-State Circuits, IEEE Journal of, 33(12):1948–1958, 1998.

Pieter J. A. Harpe, Cui Zhou, Yu Bi, Nick P. Van der Meijs, Xiaoyan Wang, Kath-
leen Philips, Guido Delmans, and Harmke De Groot. A 26 µW 8 bit 10 MS/s
Asynchronous SAR ADC for Low Energy Radios. Solid-State Circuits, IEEE
Journal of, 46(7):1585–1595, 2011.

Long Chen, Ji Ma, and Nan Sun. Capacitor mismatch calibration for SAR ADCs
based on comparator metastability detection. In Circuits and Systems (ISCAS),
2014 IEEE International Symposium on, pages 2357–2360, 2014b.

Kareem Ragab, Long Chen, Arindam Sanyal, and Nan Sun. Digital Background
Calibration for Pipelined ADCs Based on Comparator Decision Time Quanti-
zation. IEEE Transactions on Circuits and Systems II: Express Briefs, 62(5):
456–460, 2015.

Albert H. Chang, Hae-Seung Lee, and Duane Boning. A 12b 50MS/s 2.1mW
SAR ADC with redundancy and digital background calibration. In Solid-State
Circuits Conference, 2013. ESSCIRC 2013. Proceedings of the 39th European,
pages 109–112, 2013.

Guan-Ying Huang, Soon-Jyh Chang, Chun-Cheng Liu, and Ying-Zu Lin. 10-bit
30-MS/s SAR ADC Using a Switchback Switching Method. IEEE Transactions
on Very Large Scale Integration (VLSI) Systems, 21(3):584–588, 2013.

Chien-Hung Kuo and Cheng-En Hsieh. A high energy-efficiency SAR ADC based
on partial floating capacitor switching technique. In Solid-State Circuits Confer-
ence, 2011. ESSCIRC 2011. Proceedings of the 37th European, pages 475–478,
2011.

Jing Yang, Thura Lin Naing, and Robert W. Brodersen. A 1 GS/s 6 Bit 6.7 mW
Successive Approximation ADC Using Asynchronous Processing. Solid-State
Circuits, IEEE Journal of, 45(8):1469–1478, 2010.

Hegong Wei, Chi-Hang Chan, U-Fat Chio, Sai-Weng Sin, Seng-Pan U, Rui Paulo
Martins, and Franco Maloberti. An 8-b 400-MS/s 2-b-Per-Cycle SAR ADC With
Resistive DAC. Solid-State Circuits, IEEE Journal of, 47(11):2763–2772, 2012.

101
Hyeok-Ki Hong, Hyun-Wook Kang, Dong-Shin Jo, Dong-Suk Lee, Yong-Sang
You, Yong-Hee Lee, Ho-Jin Park, and Seung-Tak Ryu. A 2.6b/cycle-architecture-
based 10b 1 JGS/s 15.4mW 4x-time-interleaved SAR ADC with a multistep
hardware-retirement technique. In Solid-State Circuits Conference, 2015. Digest
of Technical Papers. ISSCC. 2015 IEEE International, pages 1–3, 2015.

Lukas Kull, Thomas Toifl, Martin Schmatz, Pier Andrea Francese, Christian
Menolfi, Matthias Brandli, Marcel Kossel, Thomas Morf, Toke Meyer Ander-
sen, and Yusuf Leblebici. A 3.1 mW 8b 1.2 GS/s Single-Channel Asynchronous
SAR ADC With Alternate Comparators for Enhanced Speed in 32 nm Digital
SOI CMOS. Solid-State Circuits, IEEE Journal of, 48(12):3049–3058, 2013.

Hung-Yen Tai, Cheng-Hsueh Tsai, Pao-Yang Tsai, Hung-Wei Chen, and Hsin-Shu
Chen. A 6-bit 1-GS/s Two-Step SAR ADC in 40-nm CMOS. Circuits and Sys-
tems II: Express Briefs, IEEE Transactions on, 61(5):339–343, 2014b.

Ying-Zu Lin, Soon-Jyh Chang, Yen-Ting Liu, Chun-Cheng Liu, and Guan-Ying
Huang. An Asynchronous Binary-Search ADC Architecture With a Reduced
Comparator Count. IEEE Transactions on Circuits and Systems I: Regular Pa-
pers, 57(8):1829–1837, 2010.

Long Chen, Manzur Rahman, Sha Liu, and Nan Sun. A fast radix-3 SAR analog-
to-digital converter. In Circuits and Systems (MWSCAS), 2013 IEEE 56th Inter-
national Midwest Symposium on, pages 1148–1151, 2013.

Manzur Rahman, Long Chen, and Nan Sun. Algorithm and implementation of
digital calibration of fast converging Radix-3 SAR ADC. In Circuits and Systems
(ISCAS), 2014 IEEE International Symposium on, pages 1336–1339, 2014.

Bob Verbruggen, Masao Iriguchi, and Craninckx Jan. A 1.7 mW 11b 250 MS/s 2-
Times Interleaved Fully Dynamic Pipelined SAR ADC in 40 nm Digital CMOS.
Solid-State Circuits, IEEE Journal of, 47(12):2880–2887, 2012.

Muhammad Ahmadi and Won Namgoong. Comparator Power Minimization Anal-

ysis for SAR ADC Using Multiple Comparators. IEEE Transactions on Circuits
and Systems I: Regular Papers, 62(10):2369–2379, 2015.

102
Long Chen, Arindam Sanyal, Ji Ma, Xiyuan Tang, and Nan Sun. Comparator
common-mode variation effects analysis and its application in SAR ADCs. In
Circuits and Systems (ISCAS), 2016 IEEE International Symposium on, 2016.

Eric Siragusa and Ian Galton. A digitally enhanced 1.8-V 15-bit 40-MSample/s
CMOS pipelined ADC. Solid-State Circuits, IEEE Journal of, 39(12):2126–
2138, 2004.

Robert H. Walden. Analog-to-digital converter survey and analysis. IEEE Journal

on Selected Areas in Communications, 17(4):539–550, 1999.

Richard Schreier and Gabor C. Temes. Understanding delta-sigma data converters.

Wiley-IEEE Press, 2005.

Arindam Sanyal, Kareem Ragab, Long Chen, T. R. Viswanathan, Shouli Yan, and
Nan Sun. A hybrid SAR-VCO ∆Σ ADC with first-order noise shaping. In
Custom Integrated Circuits Conference (CICC), 2014 IEEE Proceedings of the,
pages 1–4, 2014.

103
 Vita

Long Chen received the B.S. degree from the Institute of Microelectronics,
Tsinghua University, Beijing, China in 2011 and his M.S.E. degree from the Elec-
trical and Computer Engineering Department, the University of Texas at Austin,
Austin, TX, USA in 2014. He has interned in Silicon Laboratories, Austin and Tai-
wan Semiconductor Manufacturing Company, Austin. He is currently working as
a Staff II Design Engineer in Broadcom Limited, Austin. His research interests are
in analog and mixed-signal circuits design.

email address: [email protected]

This dissertation was typeset with LATEX† by the author.

† A
LTEX is a document preparation system developed by Leslie Lamport as a special version of
Donald Knuth’s TEX Program.

104