Ee315b Reader 2013 Stanford Dataconverters

EE315B
VLSI Data Conversion Circuits

- Autumn 2013 Boris Murmann
Stanford University
murmann@stanford.edu
Table of Contents
Chapter 1
Chapter 2
Chapter 3
Chapter 4
Chapter 5
Chapter 6
Chapter 7
Chapter 8
Chapter 9
Chapter 10
Chapter 11
Chapter 12
Introduction
Sampling, Reconstruction, Quantization
Spectral Performance Metrics
Nyquist Rate DACs
Sampling Circuits
Voltage Comparators
Flash ADCs
SAR ADCs
Pipeline ADCs
Time Interleaving
Oversampling ADCs and DACs
Energy Limits in A/D Converters
(This page is intentionally left blank)
Introduction
Boris Murmann
Stanford University
Copyright 2013 by Boris Murmann
B. Murmann
EE315B - Chapter 1
Motivation (1)
This course
Analog
Media and
Transducers
Signal
Conditioning
A/D
Digital
Processing
Signal
Conditioning
D/A
Sensors, Actuators,
Antennas, Storage Media, ...
B. Murmann
EE315B - Chapter 1
Motivation (2)
Benefits of digital signal processing

Reduced sensitivity to "analog" noise
Enhanced functionality and flexibility
Amenable to automated design & test
Direct benefit from the scaling of VLSI technology
"Arbitrary" precision
Issues
Data converters are difficult to design
Especially due to ever-increasing performance requirements
Data converters often present a performance bottleneck

Speed, resolution or power dissipation of the A/D or D/A
converter can limit overall system performance
B. Murmann
EE315B - Chapter 1
A/D Converter ca. 1954
http://www.analog.com/library/analogDialogue/archives/39-06/data_conversion_handbook.html
B. Murmann
EE315B - Chapter 1
Data Converter Applications (1)
Consumer electronics
Audio, TV, Video
Digital Cameras
Automotive control
Appliances
Toys
Communications
Mobile Phones
Personal Data Assistants
Wireless Transceivers
Routers, Modems
EE315B - Chapter 1
B. Murmann
Data Converter Applications (2)
Computing and Control

Storage media
Sound Cards
Data acquisition cards
Instrumentation
Lab bench equipment
Semiconductor test equipment
Scientific equipment
Medical equipment
B. Murmann
EE315B - Chapter 1
Example 1
[Poulton, ISSCC 2003]
High performance digital

oscilloscopes rely on extremely
high performance ADCs
Example
20 GSample/s, 8-bit ADC
10 W Power dissipation
B. Murmann
EE315B - Chapter 1
Example 2
Schvan, ISSCC 2008
Time interleaved architecture using 160 (!) SAR ADCs for

optical networks
B. Murmann
EE315B - Chapter 1
Example 3
A typical cell phone contains:

4 Rx ADCs
Dual Standard, I/Q
4 Tx DACs
Audio, Tx/Rx power
3 Auxiliary ADCs
control, Battery
8 Auxiliary DACs
charge
control, display, ...
A total of 19 data converters!
EE315B - Chapter 1
B. Murmann
Example 4
[Mehta, ISSCC2005]
Low-cost, single chip solutions require embedded data

conversion
Example: 802.11g Wireless LAN chip

2x 11-bit DAC, 176 MSamples/s
2x 9-bit ADC, 80 MSamples/s
B. Murmann
EE315B - Chapter 1
10
Trend Toward Lower ADC Energy

-6
10
-8
P/f s [J]
10
-10
10
-12
10
ISSCC & VLSI 2006-2013

20
30
40
50
60
70
SNDR [dB]
80
90
100
110
B. Murmann, "ADC Performance Survey 1997-2013," [Online]. Available: http://www.stanford.edu/~murmann/adcsurvey.html

B. Murmann
EE315B - Chapter 1
11
Course Objective
Acquire a thorough understanding of the basic principles and

challenges in data converter design
Focus on concepts that are unlikely to expire within the next
decade
Preparation for further study of state-of-the-art "fine-tuned"
realizations
Strategy
Acquire breadth via a complete system walkthrough and a
survey of existing architectures
Acquire depth through a midterm project that entails design
and thorough characterization of a specific circuit example in
modern technology
B. Murmann
EE315B - Chapter 1
12
Staff and Website
Teaching assistant
Vaibhav Tripathi
Administrative support
Ann Guerra, Allen 207
Lecture videos are provided on the web, but please come to

class to keep the discussion intercative
Coursework web page

http://coursework.stanford.edu/homepage/F13/F13-EE-315B-01.html
Please visit the discussion forum regularly

Only enrolled students have full access
EE315B - Chapter 1
B. Murmann
13
Preparation
Course prerequisites
EE214B or equivalent
Device physics and models
Transistor level analog circuits, elementary gain stages
Frequency response, feedback, noise
Prior exposure to Spice, Matlab

Basic signals and systems
Basic probability
Please talk to me if you are not sure whether you have the
required background
B. Murmann
EE315B - Chapter 1
14
Analog Circuit Sequence
B. Murmann
EE315B - Chapter 1
15
Assignments
Homework: (20%)
Handed out on Tue, due following Tue after lecture (1 pm)
Lowest HW score is dropped in final grade calculation
Midterm Project: (40%)

Transistor level design and simulation of a data converter
sub-block (no layout)
Prepare a project report in the format and style of an IEEE
journal paper
Final Exam (40%)
B. Murmann
EE315B - Chapter 1
16
Honor Code
Please remember you are bound by the honor code

I will trust you not to cheat
I will try not to tempt you
But if you are found cheating it is very serious

There is a formal hearing
You can be thrown out of Stanford
Save yourself and me a huge hassle and be honest
For more info

http://www.stanford.edu/dept/vpsa/judicialaffairs/guiding/pdf/
honorcode.pdf
B. Murmann
EE315B - Chapter 1
17
Tools and Technology
Primary tools
Cadence Virtuoso Schematic Editor
Cadence Virtuoso Analog Design Environment
Cadence SpectreRF simulator
You can use your own tools/setups at own risk
Getting started
Read tutorials and setup info provided in the CAD section of the
course website
EE315A/B Technology
0.18-m CMOS
BSIM3v3 models provided under /usr/class/ee315b/models
B. Murmann
EE315B - Chapter 1
18
Reference Books
M. Pelgrom, Analog-to-Digital Conversion, Springer, 2010
Gustavsson, Wikner, Tan, CMOS Data Converters for

Communications, Kluwer, 2000
A. Rodrguez-Vzquez, F. Medeiro, and E. Janssens, CMOS Telecom

Data Converters, Kluwer Academic Publishers, 2003
W. Kester, The Data Conversion Handbook, Newnes, 2005

http://www.analog.com/library/analogdialogue/archives/39-06/data_conversion_handbook.html
B. Razavi, Data Conversion System Design, IEEE Press, 1995
R. Schreier, G. Temes, Understanding Delta-Sigma Data Converters,

Wiley-IEEE Press, 2004
R. v. d. Plassche, CMOS Integrated Analog-to-Digital and Digital-toAnalog Converters, 2nd ed., Kluwer, 2003
J. G. Proakis, and D. G. Manolakis, Digital Signal Processing,

Prentice Hall, 1995
EE315B - Chapter 1
B. Murmann
19
Course Topics
Ideal sampling, reconstruction and quantization
Sampling circuits
Voltage comparators
Nyquist-rate ADCs and DACs
Oversampled ADCs and DACs
Data converter performance trends and limits
Data converter testing, simulation techniques
B. Murmann
EE315B - Chapter 1
20
Sampling, Reconstruction, Quantization
Boris Murmann
Stanford University
EE315B - Chapter 2
B. Murmann
The Data Conversion Problem
Real world signals

Continuous time, continuous amplitude
Digital abstraction
Discrete time, discrete amplitude
Two problems
How to discretize in time and amplitude
A/D conversion
How to "undescretize" in time and amplitude

D/A conversion
B. Murmann
EE315B - Chapter 2
Overview
We'll fist look at these building blocks from a functional, "black

box" perspective
Refine later and look at implementations
B. Murmann
EE315B - Chapter 2
Uniform Sampling and Quantization
Most common way of performing A/D

conversion
Sample signal uniformly in time
Quantize signal uniformly in
amplitude
Key questions
How much "noise" is added due
to amplitude quantization?
How can we reconstruct the
signal back into analog form?
How fast do we need to sample?
Must avoid "aliasing"
B. Murmann
EE315B - Chapter 2
Aliasing Example (1)
Amplitude
fs =
1
= 1000kHz
Ts
fsig = 101kHz
Time
f
v sig ( n ) = cos 2 in n
fs
v sig ( t ) = cos ( 2 fin t )

t n Ts =
n
fs
101
= cos 2
n
1000
EE315B - Chapter 2
B. Murmann
Amplitude
fs =
1
= 1000kHz
Ts
fsig = 899kHz
Time
899
101
899

v sig ( n ) = cos 2
n = cos 2
1 n = cos 2
n
1000
1000
1000
B. Murmann
EE315B - Chapter 2
Amplitude
fs =
1
= 1000kHz
Ts
fsig = 1101kHz
Time
1101
101
1101
v sig ( n ) = cos 2
n = cos 2
1 n = cos 2
n
1000
1000
1000
B. Murmann
EE315B - Chapter 2
Consequence
The frequencies fsig and Nfs fsig (N integer), are

indistinguishable in the discrete time domain
B. Murmann
EE315B - Chapter 2
Sampling Theorem
In order to prevent aliasing, we need

fsig ,max <
fs
2
The sampling rate fs=2fsig,max is called the Nyquist rate
Two possibilities
Sample fast enough to cover all spectral components,
including "parasitic" ones outside band of interest
Limit fsig,max through filtering
B. Murmann
EE315B - Chapter 2
Brick Wall Anti-Alias Filter
B. Murmann
EE315B - Chapter 2
10
Practical Anti-Alias Filter

Desired
Signal
Parasitic
Tone
Attenuation
Continuous
Time
0
fs/2
B/fs
0.5
fs-B
fs
...
Discrete
Time
f/fs
Need to sample faster than Nyquist rate to get good attenuation

"Oversampling"
B. Murmann
EE315B - Chapter 2
11
How much Oversampling?
[v.d. Plassche, p.41]
Alias
Rejection
Filter Order
fs/fsig,max
Can tradeoff sampling speed against filter order
In high speed converters, making fs/fsig,max>10 is usually

impossible or too costly
Means that we need fairly high order filters
B. Murmann
EE315B - Chapter 2
12
Classes of Sampling
Nyquist-rate sampling (fs > 2fsig,max)

Nyquist data converters
In practice always slightly oversampled
Oversampling (fs >> 2fsig,max)

Oversampled data converters
Anti-alias filtering is often trivial
Oversampling also helps reduce "quantization noise"
More later
Undersampling, subsampling (fs < 2fsig,max)

Exploit aliasing to mix RF/IF signals down to baseband
See e.g. Pekau & Haslett, JSSC 11/2005
B. Murmann
EE315B - Chapter 2
13
Subsampling
Aliasing is "non-destructive" if signal is band limited around some

carrier frequency
Downfolding of noise is a severe issue in practical subsampling mixers

Typically achieve noise figure no better than 20 dB (!)
B. Murmann
EE315B - Chapter 2
14
The Reconstruction Problem
As long as we sample fast

enough, x(n) contains all
information about x(t)
fs > 2fsig,max
How to reconstruct x(t) from x(n)?
Ideal interpolation formula
x( t ) =
x( n ) g( t nTs )
n =
g(t ) =
B. Murmann
sin( fs t )
fs t
Very hard to build an analog

circuit that does this
EE315B - Chapter 2
15
Zero-Order Hold Reconstruction
The most practical way of

reconstructing the continuous
time signal is to simply "hold" the
discrete time values
Either for full period Ts or a
fraction thereof
Other schemes exist, e.g.
partial-order hold
See [Jha, TCAS II, 11/2008]
B. Murmann
What does this do to the signal

spectrum?
We'll analyze this in two steps

First look at infinitely narrow
reconstruction pulses
EE315B - Chapter 2
16
Dirac Pulses
xd(t) is zero between pulses

Note that x(n) is undefined at
these times
xd ( t ) = x( t )
( t nTs )
n =
Multiplication in time means

convolution in frequency
Resulting spectrum
Xd ( f ) =
B. Murmann
1
Ts
n
X f
Ts
n =
EE315B - Chapter 2
17
Spectrum
Spectrum of Dirac signal contains replicas of Vin(f) at integer

multiples of the sampling frequency
B. Murmann
EE315B - Chapter 2
18
Finite Hold Pulse
Consider the general case with a

rectangular pulse 0 < Tp Ts
The time domain signal on the left

follows from convolving the Dirac
sequence with a rectangular unit pulse
Spectrum follows from multiplication with

Fourier transform of the pulse
sin( fTp )
Hp (f ) = Tp
Xp (f ) =
fTp
Tp sin( fTp )
Ts
fTp
jfTp
jfTp
n
Xf
Ts
n =
Amplitude Envelope
EE315B - Chapter 2
B. Murmann
19
Envelope with Hold Pulse Tp=Ts

1
0.9
0.8
0.7
0.6
0.5
Ts
0.4
fTp
abs(H(f))
Tp sin( fTp )
0.3
0.2
0.1
0
0
0.5
1.5
f/fs
2.5
f/fs
B. Murmann
EE315B - Chapter 2
20
Envelope with Hold Pulse Tp=0.5Ts

1
0.9
0.8
Tp=Ts
0.7
0.6
0.5
Ts
0.4
fTp
abs(H(f))
Tp sin( fTp )
Tp=0.5Ts
0.3
0.2
0.1
0
0
0.5
1.5
2.5
f/fs
f/f
s
EE315B - Chapter 2
B. Murmann
21
Example
1
Spectrum of
Continuous Time
Pulse Train (Arbitrary
Example)
0.5
0
0
0.5
1.5
2.5
0.5
1.5
2.5
2.5
ZOH Transfer
Function
("Sinc Distortion")
0.5
0
1
ZOH output,
Spectrum of
Staircase
Approximation
original spectrum
0.5
0
0
0.5
1.5
f/fs
B. Murmann
EE315B - Chapter 2
22
Reconstruction Filter
1
Also called
smoothing filter
Same situation
as with anti-alias
filter
A brick wall
filter would be
nice
Oversampling
helps reduce
filter order
0.9
Filter
0.8
Spectrum
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1.5
2.5
f/fs
B. Murmann
EE315B - Chapter 2
23
Summary
Must obey sampling theorem fs > 2fsig,max,

Usually dictates anti-aliasing filter
If sampling theorem is met, continuous time signal can be

recovered from discrete time sequence without loss of
information
A zero order hold in conjunction with a smoothing filter is the

most common way to reconstruct
May need to add pre- or post-emphasis to cancel droop due
to sinc envelope
Oversampling helps reduce order of anti-aliasing and

reconstruction filters
B. Murmann
EE315B - Chapter 2
24
Recap
Next, look at
Transfer functions of quantizer and DAC
Impact of quantization error
EE315B - Chapter 2
B. Murmann
25
Quantization of an Analog Signal

Vq (quantized output)
Transfer Function
Slope=1
Quantization step
Quantization error has

sawtooth shape
Bounded by /2, +/2
Ideally
Infinite input range and
infinite number of
quantization levels
In practice
Finite input range and
finite number of
quantization levels
Output is a digital word
(not an analog voltage)
eq (quantization error)
x (input)
Error eq=q-x
+/2
-/2
x (input)
B. Murmann
EE315B - Chapter 2
26
Conceptual Model of a Quantizer
Encoding block determines how quantized levels are mapped

into digital codes
Note that this model is not meant to represent an actual

hardware implementation
Its purpose is to show that quantization and encoding are
conceptually separate operations
Changing the encoding of a quantizer has no interesting
implications on its function or performance
EE315B - Chapter 2
B. Murmann
27
111
110
101
100
011
010
001
000
Analog Input
eq (quantization error)
Digital Output
Encoding Example for a B-Bit Quantizer
+/2
http://www.analog.com/en/content/0
,2886,760%255F788%255F91285,
00.html
-/2
FSR
x (input)
B. Murmann
Example: B=3
23=8 distinct output codes
Diagram on the left shows
"straight-binary encoding"
See e.g. Analog Devices "MT009: Data Converter Codes" for
other encoding schemes
EE315B - Chapter 2
Quantization error grows out of

bounds beyond code boundaries
We define the full scale range
(FSR) as the maximum input range
that satisfies |eq| /2
Implies that FSR=2B
28
Nomenclature
Overloading - Occurs when an input outside

the FSR is applied
Transition level Input value at the

transition between two codes. By standard
convention, the transition level T(k) lies
between codes k-1 and k
Code width The difference between

adjacent transition levels. By standard
convention, the code width W(k)=T(k+1)-T(k)
Note that the code width of the first and
last code (000 and 111 on previous slide)
is undefined
LSB size (or width) synonymous with

code width
[IEEE Standard 1241-2000]
EE315B - Chapter 2
B. Murmann
29
Implementation Specific Technicalities
So far, we avoided specifying the absolute location of the code

range with respect to "zero" input
The zero input location depends on the particular

implementation of the quantizer
Bipolar input, mid-rise or mid-tread quantizer
Unipolar input
The next slide shows the case with

Bipolar input
The quantizer accepts positive and negative inputs
Represents the common case of a differential circuit
Mid-rise characteristic
The center of the transfer function (zero), coincides with a
transition level
B. Murmann
EE315B - Chapter 2
30
Digital Output
Bipolar Mid-Rise Quantizer
111
110
101
100
011
010
001
000
Analog Input
-FSR/2
+FSR/2
Nothing new here
EE315B - Chapter 2
B. Murmann
31
Bipolar Mid-Tread Quantizer

In theory, less sensitive to infinitesimal disturbance around zero
In practice, offsets larger than /2 (due to device mismatch)
often make this argument irrelevant
Asymmetric full-scale range, unless we use odd number of codes
Digital Output
111
110
101
100
011
010
001
000
Analog Input
FSR/2 + /2
B. Murmann
0 FSR/2 - /2
EE315B - Chapter 2
32
Unipolar Quantizer
Usually define origin where first code and straight line fit intersect
Otherwise, there would be a systematic offset
Usable range is reduced by /2 below zero
Digital Output
111
110
101
100
011
010
001
000
0
B. Murmann
Analog Input
FSR - /2
EE315B - Chapter 2
33
Effect of Quantization Error on Signal
Two aspects
How much noise power does quantization add to samples?
How is this noise power distributed in frequency?
Quantization error is a deterministic function of the signal

Should be able answer above questions using a
deterministic analysis
But, unfortunately, such an analysis strongly depends on the
chosen signal and can be very complex
Strategy
Build basic intuition using simple deterministic signals
Next, abandon idea of deterministic representation and
revert to a "general" statistical model (to be used with
caution!)
B. Murmann
EE315B - Chapter 2
34
Ramp Input
Applying a ramp signal (periodic sawtooth) at the input of the

quantizer gives the following time domain waveform for eq
0.6
0.2
0
e (t) []
0.4
-0.2
-0.4
0
50
100
Time [arbitrary units]
150
What is the average power of this waveform?
Integrate over one period

eq2
T/2
1
=
eq2 ( t )dt
T T/ 2
eq ( t ) =
t
T
eq2
2
=
12
EE315B - Chapter 2
B. Murmann
35
Sine Wave Input

1
1
Vin
V
0.8
0.6
0.6
0.4
0.4
0.2
0.2
eq(t) []
[Volts]
0.8
-0.2
-0.2
-0.4
-0.4
-0.6
-0.6
-0.8
-0.8
-1
0
0.2
0.4
0.6
0.8
-1
0
0.2
0.4
0.6
0.8
Integration is not straightforward
B. Murmann
EE315B - Chapter 2
36
Quantization Error Histogram
Sinusoidal input signal with fsig=101Hz, sampled at fs=1000Hz

8-bit quantizer
Mean=0.000LSB, Var=1.034LSB2/12
120
100
Count
80
60
40
20
0
-0.6 -0.5 -0.4 -0.3 -0.2 -0.1
0 0.1 0.2 0.3 0.4 0.5 0.6

e /
q
Distribution is "almost" uniform

Can approximate average power by integrating uniform
distribution
B. Murmann
EE315B - Chapter 2
37
Statistical Model of Quantization Error
Assumption: eq(x) has a uniform probability density
This approximation holds reasonably well in practice when

Signal spans large number of quantization steps
Signal is "sufficiently active"
Quantizer does not overload
+ / 2
Mean
eq =
/ 2
Variance
eq2 =
+ / 2
/ 2
B. Murmann
EE315B - Chapter 2
eq
deq = 0
eq 2
deq =
2
12
38
Reality Check (1)
Input sequence consists of 1000 samples drawn from Gaussian

distribution, 4=FSR
Mean=-0.004LSB, Var=1.038LSB2/12
150
Count
100
50
0
-0.6 -0.5 -0.4 -0.3 -0.2 -0.1
0 0.1 0.2 0.3 0.4 0.5 0.6

eq/
Error power close to that of uniform approximation

EE315B - Chapter 2
B. Murmann
39
Reality Check (2)
Another sine wave example, but now fsig/fs=100/1000
What's going on here?

Mean=-0.000LSB, Var=0.629LSB2/12
500
Count
400
300
200
100
0
-0.6 -0.5 -0.4 -0.3 -0.2 -0.1
B. Murmann
0 0.1 0.2 0.3 0.4 0.5 0.6

eq/
EE315B - Chapter 2
40
Amplitude
Analysis (1)
fsig/fs=100/1000
Time
Sampled signal is repetitive and has only a few distinct values

This also means that the quantizer generates only a few
distinct values of eq; not a uniform distribution
EE315B - Chapter 2
B. Murmann
41
Analysis (2)
f
v sig ( n ) = cos 2 in n
fs
Signal repeats every m samples, where m is the smallest

integer that satisfies
m
fin
= integer
fs
101
= integer
1000
m = 1000
100
= integer
1000
m = 10
This means that eq(n) has at best 10 distinct values, even if we

take many more samples
B. Murmann
EE315B - Chapter 2
42
Signal-to-Quantization-Noise Ratio
Assuming uniform distribution of eq and a full-scale sinusoidal

input, we have
SQNR =
B. Murmann
Psig
Pqnoise
1 2B
2 2
2
12
= 1.5 22B = 6.02B + 1.76 dB
B (Number of Bits)
SQNR
50 dB
12
74 dB
16
98 dB
20
122 dB
EE315B - Chapter 2
43
Quantization Noise Spectrum (1)
How is the quantization noise power distributed in frequency?

First think about applying a sine wave to a quantizer, without
sampling (output is continuous time)
+ many more harmonics
[Y. Tsividis, ICASSP 2004]
Quantization results in an "infinite" number of harmonics
B. Murmann
EE315B - Chapter 2
44
Now sample the signal at the output

All harmonics (an "infinite" number of them) will alias into
band from 0 to fs/2
Quantization noise spectrum becomes "white"
[Y. Tsividis, ICASSP 2004]
Interchanging sampling and quantization wont change this

situation
EE315B - Chapter 2
B. Murmann
45
Can show that the quantization noise power is indeed

distributed (approximately) uniformly in frequency
Again, this is provided that the quantization error is
"sufficiently random"
2 2
12 fs
References
W. R. Bennett, "Spectra of quantized signals," Bell Syst. Tech. J., pp. 446-72,
July 1948.
B. Widrow, "A study of rough amplitude quantization by means of Nyquist
sampling theory," IRE Trans. Circuit Theory, vol. CT-3, pp. 266-76, 1956.
A. Sripad and D. A. Snyder, "A necessary and sufficient condition for
quantization errors to be uniform and white," IEEE Trans. Acoustics, Speech,
and Signal Processing, pp. 442-448, Oct 1977.
B. Murmann
EE315B - Chapter 2
46
Ideal DAC
Essentially a digitally controlled voltage, current or charge source

Example below is for unipolar DAC
Ideal DAC does not introduce quantization error!
B. Murmann
EE315B - Chapter 2
47
Static Nonidealities
Static deviations of transfer characteristics from ideality

Offset
Gain error
Differential Nonlinearity (DNL)
Integral Nonlinearity (INL)
Useful references
Analog Devices MT-010: The Importance of Data Converter
Static Specifications
http://www.analog.com/en/content/0,2886,761%255F795%255F91286,00.html
"Understanding Data Converters," Texas Instruments

Application Report LAA013, 1995.
http://focus.ti.com/lit/an/slaa013/slaa013.pdf
B. Murmann
EE315B - Chapter 2
48
Offset and Gain Error
Conceptually simple, but lots of (uninteresting) subtleties in how

exactly these errors should be defined
Unipolar versus bipolar, endpoint versus midpoint
specification
Definition in presence of nonlinearities
General idea (neglecting staircase nature of transfer functions):
EE315B - Chapter 2
B. Murmann
49
ADC Offset and Gain Error
Definitions based on bottom and top endpoints of transfer characteristic

LSB before first transition and LSB after last transition
Offset is the deviation of bottom endpoint from its ideal location
Gain error is the deviation of top endpoint from its ideal location with
offset removed
Both quantities are measured in LSB or as percentage of full-scale range

Gain Error
Dout
Dout
Ideal
Vin
Ideal
Vin
Offset
B. Murmann
EE315B - Chapter 2
50
DAC Offset and Gain Error
Same idea, except that endpoints are directly defined by analog

output values at minimum and maximum digital input
Also note that errors are specified along the vertical axis
B. Murmann
EE315B - Chapter 2
51
Comments on Offset and Gain Errors
Definitions on the previous slides are the ones typically used in industry
IEEE Standard suggest somewhat more sophisticated definitions
based on least square curve fitting
Technically more suitable metric when the transfer characteristics are
significantly non-uniform or nonlinear
Generally, it is non-trivial to build a converter with very good gain/offset

specifications
Nevertheless, since gain and offset affect all codes uniformly, these
errors tend to be easy to correct
E.g. using a digital pre- or post-processing operation
Also, many applications are insensitive to a certain level of gain and

offset errors
E.g. audio signals, communication-type signals, ...
More interesting aspect: linearity

DNL and INL
B. Murmann
EE315B - Chapter 2
52
Differential Nonlinearity (DNL)
In an ideal world, all ADC codes would have equal width; all
DAC output increments would have same size
DNL(k) is a vector that quantifies for each code k the deviation

of this width from the "average" width (step size)
DNL(k) is a measure of uniformity, it does not depend on gain

and offset errors
Scaling and shifting a transfer characteristic does not alter its
uniformity and hence DNL(k)
Let's look at an example
B. Murmann
EE315B - Chapter 2
53
ADC DNL Example (1)
B. Murmann
EE315B - Chapter 2
Code (k)
W [V]
undefined
0.5
1.5
1.5
undefined
54
ADC DNL Example (2)
What is the average code width?

ADC with perfect uniformity would divide the range between
first and last transition into 6 equal pieces
Hence calculate average code width (i.e. LSB size) as
Wavg =
7.5V 2V
= 0.9167V
6
Now calculate DNL(k) for each code k using
DNL(k) =
W(k) Wavg
Wavg
EE315B - Chapter 2
B. Murmann
55
Result
Code (k)
DNL [LSB]
0.09
-0.45
0.09
0.64
-1.00
0.64
Positive/negative DNL implies wide/narrow code, respectively
DNL = -1 LSB implies missing code
Impossible to have DNL < -1 LSB for an ADC

But possible to have DNL > +1 LSB
Can show that sum over all DNL(k) is equal to zero
B. Murmann
EE315B - Chapter 2
56
A Typical ADC DNL Plot
[Ahmed, JSSC 12/2005]
People speak about DNL often only in terms of min/max number

across all codes
E.g. DNL = +0.63/-0.91 LSB
Might argue in some cases that any code with DNL < -0.9 LSB
is essentially a missing code
Why ?
B. Murmann
EE315B - Chapter 2
57
Impact of Noise
[W. Kester, "ADC Input Noise: The

Good, The Bad, and The Ugly. Is
No Noise Good Noise?" Analogue
Dialogue, Feb. 2006]
In essentially all moderate to high-resolution ADCs, the transition

levels carry noise that is somewhat comparable to the size of an LSB
Noise "smears out" DNL, can hide missing codes
Especially for converters whose input referred (thermal) noise is

larger than an LSB, DNL is a "fairly useless" metric
B. Murmann
EE315B - Chapter 2
58
DAC DNL
Same idea applies

Find output increments for each digital code
Find increment that divides range into equal steps
Calculate DNL for each code k using
DNL(k) =
Step(k) Stepavg
Stepavg
One difference between ADC and DAC is that DAC DNL can be
less than -1 LSB
How ?
B. Murmann
EE315B - Chapter 2
59
Non-Monotonic DAC
DNL(3) =
Step(3) Stepavg
Stepavg
0.5V 1V
= 1.5LSB
1V
In a DAC, DNL < -1LSB implies non-monotinicity
How about a non-monotonic ADC?
B. Murmann
EE315B - Chapter 2
60
Non-Monotonic ADC
Code 2 has two transition levels W(2) is ill defined

DNL is ill-defined!
Not a very big issue, because a non-monotonic ADC is usually

not what we'll design for in practice
B. Murmann
EE315B - Chapter 2
61
Integral Nonlinearity (INL)
General idea
For each "relevant point" of the transfer characteristic,
quantify distance from a straight line drawn through the
endpoints
An alternative, less common definition uses a least square fit
line as a reference
Just as with DNL, the INL of a converter is by definition

independent of gain and offset errors
B. Murmann
EE315B - Chapter 2
62
ADC INL Example (1)
"Straight line" reference

is uniform staircase
between first and last
transition
INL for each code is

INL(k) =
T(k) Tuniform (k)

Wavg
Obviously INL(1) = 0
and INL(7) = 0
INL(0) is undefined
EE315B - Chapter 2
B. Murmann
63
ADC INL Example (2)
Can show that
k 1
INL(k) = DNL(i)
i =1
Means that once we computed DNL, we can easily find INL

using a cumulative sum operation on the DNL vector
Using DNL values from last lecture, we find
B. Murmann
Code (k)
DNL [LSB]
INL [LSB]
0.09
-0.45
0.09
0.09
-0.36
0.64
-0.27
-1.00
0.36
0.64
-0.64
undefined
EE315B - Chapter 2
64
Result
B. Murmann
EE315B - Chapter 2
65
A Typical ADC DNL/INL Plot
[Ishii, Custom
Integrated Circuits
Conference, 2005]
DNL/INL signature often reveals architectural details

E.g. major transitions
We'll see more examples in the context of DACs
Since INL is a cumulative measure, it turns out to be less

sensitive than DNL to thermal noise "smearing"
B. Murmann
EE315B - Chapter 2
66
DAC INL
Same idea applies

Find ideal output values that lie on a straight line between
endpoints
Calculate INL for each code k using
INL(k) =
Vout (k) Voutuniform (k)

Stepavg
Interesting property related to DAC INL

If for all codes |INL| < 0.5 LSB, it
follows that all |DNL| < 1 LSB
A sufficient (but not necessary)
condition for monotonicity
B. Murmann
EE315B - Chapter 2
67
How to Measure DNL/INL in the Lab?
DAC
Apply all input codes, measure output with a precision
voltmeter
ADC
A little more tricky
One option is to build a servo loop that finds the code
transitions
See e.g. Kester, page 5.36
A more popular approach is histogram testing
B. Murmann
EE315B - Chapter 2
68
Basic Histogram Test Setup
Kester, p. 5.39
B. Murmann
EE315B - Chapter 2
69
Histogram Example
Kester, p. 5.40
B. Murmann
EE315B - Chapter 2
70
Sinusoidal Input
Kester, p. 5.42
B. Murmann
Preferred over ramp or

triangular test signals
It is easier much easier

to generate a high
fidelity sinusoid
The histogram now

takes on a bathtub
shape, which can be
mathematically inverted
to find the DNL
EE315B - Chapter 2
71
Correction for Sinusoidal pdf
References
M. V. Bossche, J. Schoukens, and J. Renneboog, Dynamic
Testing and Diagnostics of A/D Converters, IEEE TCAS,
Aug. 1986
IEEE Standard 1057
Kester, Section 5.4
It turns out that is not necessary to know the exact amplitude

and offset of the sine wave input
There exists some confusion about this
The code on the next slide does all the required math to undo
the bathtub shape
B. Murmann
EE315B - Chapter 2
72
DNL/INL Code
function [dnl,inl] = dnl_inl_sin(y);
% transition levels
%DNL_INL_SIN
T = -cos(pi*ch/sum(h));
% dnl and inl ADC output

% input y contains the ADC output
% linearized histogram
% vector obtained from quantizing a
hlin = T(2:end) - T(1:end-1);
% sinusoid
% truncate at least first and last
% Boris Murmann, Aug 2002
% bin, more if input did not clip ADC
% Bernhard Boser, Sept 2002
trunc=2;
hlin_trunc = hlin(1+trunc:end-trunc);
% histogram boundaries
minbin=min(y);
% calculate lsb size and dnl
maxbin=max(y);
lsb= sum(hlin_trunc) / (length(hlin_trunc));

% histogram
dnl= [0 hlin_trunc/lsb-1];
h = hist(y, minbin:maxbin);
misscodes = length(find(dnl<-0.9));
% cumulative histogram
% calculate inl
ch = cumsum(h);
inl= cumsum(dnl);
EE315B - Chapter 2
B. Murmann
73
DNL/INL Code Test

% converter model
B = 6;
DNL = +0.7 / -0.71 LSB, 0 missing codes (DNL<-0.9)
% bits
range = 2^(B-1) - 1;
th = -range:range;
% ideal thresholds
DNL [LSB]
0.5
% thresholds (ideal converter)
-0.5
th(20) = th(20)+0.7; % error

-1
-30
0
code
10
20
30
10
20
30
0.8
% try fs/10!
= round(100 * 2^B / (fs / fx));
t = 0:1/fs:C/fx;
x = (range+1) * sin(2*pi*fx.*t);
0.6
INL [LSB]
-10
INL = +0.76 / -0.063 LSB
fs = 1e6;
fx = 494e3 + pi;
-20
0.4
0.2
0
-0.2
-30
-20
-10
0
code
y = adc(x, th) - 2^(B-1);
hist(y, min(y):max(y));
dnl_inl_sin(y);
B. Murmann
EE315B - Chapter 2
74
Limitations of ADC Histogram Testing
Cannot detect non-monotonicity

The histogram does not capture in which order the codes
occurred
Cannot detect erratic dynamics

E.g. 123, 123, , 123, 0, 124, 124,
Must look directly at ADC output to detect these
Similarly, random noise is not detected and improves DNL

E.g. 9, 9, 9, 10, 9, 9, 9, 10, 9, 10, 10, 10,
Reference
B. Ginetti and P. Jespers, Reliability of Code Density Test
for High Resolution ADCs, Electronics Letters, pp. 22312233, Nov. 1991
B. Murmann
EE315B - Chapter 2
75
Hiding DNL Problems in the Noise
INL suggests that there

may be missing codes
But, DNL is "smeared

out" by noise and does
not show this
Always look at both

DNL/INL
INL usually does not lie

[Source: David Robertson, Analog Devices]
B. Murmann
EE315B - Chapter 2
76
Spectral Performance Metrics
Boris Murmann
Stanford University
B. Murmann
EE315B - Chapter 3
Dynamic Performance Metrics
Time domain
Glitch impulse, aperture uncertainty, settling time,
We'll look at these later, in the context of specific circuits
Frequency domain
Performance metrics follow from looking at converter or
building block output spectrum
"Spectral performance metrics"
Basic idea: Apply one or more tones at converter input

Expect same tone(s) at output, all other frequency components
represent nonidealities
Important to realize that both static (DNL, INL) and dynamic

errors contribute to frequency domain non-ideality
B. Murmann
EE315B - Chapter 3
Alphabet Soup of Spectral Metrics
SNR - Signal-to-noise ratio
SNDR (SINAD) - Signal-to-(noise+distortion) ratio
ENOB - Effective number of bits
DR - Dynamic range
SFDR - Spurious free dynamic range
HD Harmonic distortion
THD - Total harmonic distortion
ERBW - Effective Resolution Bandwidth
IMD - Intermodulation distortion
MTPR - Multi-tone power ratio
B. Murmann
EE315B - Chapter 3
DAC Tone Test/Simulation Setup
B. Murmann
EE315B - Chapter 3
Typical DAC Output Spectrum
[Hendriks, "Specifying Communications DACs, IEEE Spectrum, July 1997]
B. Murmann
EE315B - Chapter 3
ADC Tone Test/Simulation Setup
Basic idea
Apply a clean sinusoid
Compute ADC performance metrics based on output spectrum
B. Murmann
EE315B - Chapter 3
Aside: How to Make a Clean

Sinusoid in the Lab?
BP Filter
Amplitude
Signal generator harmonics
fin
2fin
3fin
4fin
...
...
www.tte.com, or
www.allenavionics.com
EE315B - Chapter 3
B. Murmann
Discrete Fourier Transform Basics
DFT takes a block of N time domain samples (spaced Ts=1/fs)

and yields a set of N frequency bins
N1
X(k) =
x(n)e j2kn/N
n=0
Bin k represents frequency content at kfs/N [Hz]
DFT frequency resolution

Proportional to 1/(NTs) in [Hz/bin]
NTs is total time spent gathering samples
A DFT with N=2integer can be found using a computationally

efficient algorithm
FFT = Fast Fourier Transform
B. Murmann
EE315B - Chapter 3
Matlab Example
clear;
N
50
= 100;
fs = 1000;
40
x = cos(2*pi*fx/fs*[0:N-1]);
s = abs(fft(x));
plot(s, 'linewidth', 2);
DFT Magnitude
fx = 100;
30
20
10
0
0
20
40
60
80
100
Bin #
B. Murmann
EE315B - Chapter 3
Normalized Plot with Frequency Axis

N
= 100;
fs = 1000;
fx = 100;
FS = 1; % full-scale amplitude
s = abs(fft(x));
% remove redundant half of spectrum
s = s(1:end/2);
% normalize magnitudes to dBFS
% dbFS = dB relative to full-scale
s = 20*log10(2*s/N/FS);
DFT Magnitude [dBFS]
x = FS*cos(2*pi*fx/fs*[0:N-1]);
-50
-100
-150
-200
-250
% frequency vector
f = [0:N/2-1]/N;
plot(f, s, 'linewidth', 2);
-300
-350
0
xlabel('Frequency [f/fs]')
0.1
0.2
0.3
Frequency [f/fs]
0.4
0.5
ylabel('DFT Magnitude [dBFS]')
B. Murmann
EE315B - Chapter 3
10
Another Example
0
Same as before, but

now fx=101
This doesn't look the

spectrum of a
sinusoid
-10
What's going on?
-20
-30
-40
-50
-60
-70
-80
-90
0
0.1
0.2
0.3
Frequency [f/fs]
EE315B - Chapter 3
B. Murmann
0.4
0.5
11
Spectral Leakage
The DFT computes the

spectrum of the periodic
repetition of its input
A sequence that contains a noninteger number of sine wave

cycles has discontinuities in its
periodic repetition
Discontinuity looks like a
high frequency signal
component
Power spreads across
spectrum
Two ways to deal with this

Ensure integer number of
periods
Windowing
B. Murmann
EE315B - Chapter 3
12
Integer Number of Cycles

0
N
= 100;
-50
cycles = 9;
fs = 1000;
fx = fs*cycles/N;
Usable test
frequencies are
limited to a multiple
of fs/N
-100
-150
-200
-250
-300
-350
0
0.1
0.2
0.3
Frequency [f/fs]
0.4
0.5
EE315B - Chapter 3
B. Murmann
13
Windowing
Spectral leakage can be attenuated by windowing the time

samples prior to the DFT
Windows taper smoothly down to zero at the beginning and the

end of the observation window
Time domain samples are multiplied by window coefficients on a

sample-by-sample basis
Means convolution in frequency
Sine wave tone and other spectral components smear out
over several bins
Lots of window functions to chose from

Tradeoff: attenuation versus smearing
Example: Hann Window
B. Murmann
EE315B - Chapter 3
14
Hann Window
N=64;
wvtool(hann(N))
Time domain
Frequency domain
50
1
0
Magnitude (dB)
Amplitude
0.8
0.6
0.4
-50
-100
0.2
10
20
30
40
Samples
50
-150
60
0.2
0.4
0.6
0.8
Normalized Frequency ( rad/sample)
EE315B - Chapter 3
B. Murmann
15
Spectrum with Window

N
= 100;
No window
Hann window
fs = 1000;
-20
= 1;
x = A*cos(2*pi*fx/fs*[0:N-1]);
s
= abs(fft(x));
x1 = x.*hann(N);
s1 = abs(fft(x1));
fx = 101;
-40
-60
-80
-100
-120
-140
0
0.1
0.2
0.3
0.4
0.5
f/fs
B. Murmann
EE315B - Chapter 3
16
Integer Cycles versus Windowing
Integer number of cycles

Test signal falls into single DFT bin
Requires careful choice of signal frequency
Ideal for simulations
In lab measurements, can lock sampling and signal
frequency generators (PLL)
"Coherent sampling"
Windowing
No restrictions on signal frequency
Signal and harmonics distributed over several DFT bins
Beware of smeared out nonidealities
Requires more samples for given accuracy
More info
http://www.maxim-ic.com/appnotes.cfm/appnote_number/1040
EE315B - Chapter 3
B. Murmann
17
Example
Now that we've

"calibrated" our test
system, let's look at
some spectra that
involve nonidealities
= 2048;
cycles = 67;
fs = 1000;
fx = fs*cycles/N;
LSB = 2/2^10;
%generate signal, quantize (mid-tread) and take FFT
First look at
quantization noise
introduced by an
ideal quantizer
x = cos(2*pi*fx/fs*[0:N-1]);
x = round(x/LSB)*LSB;
s
= abs(fft(x));
s = s(1:end/2)/N*2;
% calculate SNR
sigbin = 1 + cycles;
noise = [s(1:sigbin-1), s(sigbin+1:end)];
snr = 10*log10( s(sigbin)^2/sum(noise.^2) );
B. Murmann
EE315B - Chapter 3
18
Spectrum with Quantization Noise

2048 point FFT, SNR=61.90dB
0
Spectrum looks fairly uniform
Signal-to-quantization noise
ratio is given by power in
signal bin, divided by sum of
all noise bins
-20
-40
-60
-80
-100
-120
0
0.1
0.2
0.3
Frequency [f/fs]
0.4
EE315B - Chapter 3
B. Murmann
0.5
19
SQNR
Signal Power
Quantization Noise Power
1 VFS
2 2
-20
1 VFS
12 2N
3 2N
2
2
= 6.02 N + 1.76 [dB]

= 6.02 10 + 1.76 [dB]
SQNR =
-40
-60
-80
-100
= 61.9 dB
-120
0
B. Murmann
EE315B - Chapter 3
0.1
0.2
0.3
Frequency [f/fs]
0.4
0.5
20
FFT Noise Floor

0
= 61.9 dBc 30.1 dB

= 92 dBc
Depends on FFT size
Plot is useless if FFT

size is not specified
-20
2048
Nfloor = 61.9 dBc 10log
-40
-60
-80
-100
-120
0
B. Murmann
0.1
0.2
0.3
Frequency [f/fs]
0.4
EE315B - Chapter 3
0.5
21
DFT Plot Annotation
DFT plots are fairly meaningless unless you clearly specifiy the
underlying conditions
Most common annotation

Specify how many DFT points were used (N)
Less common options

Shift DFT noise floor by 10log10(N/2)dB
Normalize with respect to bin width in Hz and express noise
as power spectral density
"Noise power in 1 Hz bandwidth"
B. Murmann
EE315B - Chapter 3
22
Periodic Quantization Noise

Same as before, but cycles =
64 (instead of 67)
fx = fs64/2048 = fs/32
Quantization noise is highly

deterministic and periodic

0
-20
For more random and "white"

quantizion noise, it is best to
make N and cycles mutually
prime
GCD(N,cycles)=1
-40
-60
-80
-100
-120
0
0.1
0.2
0.3
Frequency [f/fs]
0.4
EE315B - Chapter 3
B. Murmann
0.5
23
Typical ADC Output Spectrum

Fairly uniform noise floor due
to additional electronic noise
Harmonics due to
nonlinearities
Definition of SNR
SNR =
-20
Signal Power
Total Noise Power
Total noise power includes all

bins except DC, signal, and
2nd through 7th harmonic
Both quantization noise
and electronic noise affect
SNR
B. Murmann

0
-40
-60
-80
-100
-120
0
EE315B - Chapter 3
0.1
0.2
0.3
0.4
Frequency [f/fs]
24
SNDR and ENOB

Definition
SNDR =
Signal Power
Noise and Distortion Power
Noise and distortion power

includes all bins except DC
and signal
2048 point FFT, SNR=55.9dB, SNDR=47.5dB

0
Effective number of bits

ENOB =
SNDR(dB)-1.76dB
6.02dB
-20
-40
-60
-80
-100
-120
0
0.1
0.2
0.3
0.4
Frequency [f/fs]
EE315B - Chapter 3
B. Murmann
25
Effective Number of Bits
Is a 10-Bit converter with 47.5dB SNDR really a 10-bit

converter?
ENOB =
47.5dB 1.76dB
= 7.6
6.02dB
We get ideal ENOB only for zero electronic noise, perfect

transfer function with zero INL, ...
Low electronic noise is costly

Cutting thermal noise down by 2x, can cost 4x in power
dissipation
Rule of thumb for good power efficiency: ENOB < B-1

B is the "number of wires" coming out of the ADC or the so
called "stated resolution"
B. Murmann
EE315B - Chapter 3
26
Survey Data
SNRBits =
SNR(dB)-1.76dB
6.02dB
R. H. Walden, "Analog-to-digital converter survey and analysis," IEEE J. on

Selected Areas in Communications, pp. 539-50, April 1999
EE315B - Chapter 3
B. Murmann
27
Dynamic Range
DR =
Maximum Signal Power

SNRpeak
Minimum Detectable Signal
SNR
(dB)
PEAK SNR
OVERLOAD
FULL SCALE
0dB
DYNAMIC
RANGE
MDS
B. Murmann
EE315B - Chapter 3
INPUT
Input
Power
AMPLITUDE
[dB]
(dB)
28
SFDR
Definition of "Spurious Free
Dynamic Range"
SFDR =
Signal Power
Largest Spurious Power
Largest spur is often (but not

necessarily) a harmonic of the
input tone
2048 point FFT, SFDR=48.3dB

0
-20
-40
-60
-80
-100
-120
0
0.1
0.2
0.3
0.4
Frequency [f/fs]
EE315B - Chapter 3
B. Murmann
29
SDR and THD
Signal-to-distortion ratio
2048 point FFT, THD=-48.2dB

0
Signal Power
SDR =
Total Distortion Power
Total harmonic distortion

THD =
Total Distortion Power

1
=
Signal Power
SDR
By convention, total distortion

power consists of 2nd through 7th
harmonic
Is there a 6th and 7th harmonic in
the plot to the right?
B. Murmann
-20
-40
-60
-80
-100
-120
0
EE315B - Chapter 3
0.1
0.2
0.3
0.4
Frequency [f/fs]
30
Lowering the Noise Floor

0
Increasing the FFT size let's

us lower the noise floor and
reveal low level harmonics
-20
-40
-60
-80
-100
-120
0
0.1
0.2
0.3
0.4
Frequency [f/fs]
EE315B - Chapter 3
B. Murmann
31
Aliasing
Harmonics can appear at

"arbitrary" frequencies
due to aliasing
f1 = fx = 0.3125 fs
f2 = 2 f1 = 0.6250 fs 0.3750 fs
f3 = 3 f1 = 0.9375 fs 0.0625 fs
f4 = 4 f1 = 1.2500 fs 0.2500 fs
f5 = 5 f1 = 1.5625 fs 0.4375 fs
0
-20
-40
-60
-80
-100
-120
0
B. Murmann
EE315B - Chapter 3
0.1
0.2
0.3
Frequency [f/fs]
0.4
0.5
32
Intermodulation Distortion
Amplitude
f1
f2
SECOND
ORDER
PRODUCTS
f2 - f1
THIRD
ORDER
PRODUCTS
2f1 - f2
2f2 - f1
f1 + f2
Frequency
IMD is important in multi-channel communication systems

Third order products are generally difficult to filter out
EE315B - Chapter 3
B. Murmann
33
MTPR
Amplitude [dB]
MTPR
-20
-40
-60
-80
-100
-120
0.00E+00
4.00E+06
8.00E+06
1.20E+07
1.60E+07
2.00E+07
2.40E+07
2.80E+07
3.20E+07
Frequency(Hz)
Frequency [Hz]
Useful metric in multi-tone transmission systems

E.g. OFDM
B. Murmann
EE315B - Chapter 3
34
Frequency Dependence (1)
All of the above discussed metrics generally depend on frequency

Sampling frequency and input frequency
[Analog Devices, AD9203 Datasheet ]
B. Murmann
EE315B - Chapter 3
35
Frequency Dependence (2)
[Texas Instruments, ADS5541 Datasheet ]
B. Murmann
EE315B - Chapter 3
36
ERBW
Defined as the input frequency at which the SNDR of a

converter has dropped by 3dB
Equivalent to a 0.5-bit loss in ENOB
ERBW > fs/2 is not uncommon, especially in converters

designed for sub-sampling applications
ERBW is only a useful metric when the SNDR is relatively flat

initially and gracefully degrades at high frequencies
EE315B - Chapter 3
B. Murmann
37
Relationship Between INL and

Harmonic Distortion (1)
The INL of an ADC often takes on a quadratic or cubic bow

Meaning that the transfer function can be approximated by
x+a2x2 or x+a3x3
The resulting harmonic distortion usually sets the low frequency

SFDR
Output
Ideal
Quadratic bow
Cubic bow
Input
B. Murmann
EE315B - Chapter 3
38
Relationship Between INL and

Harmonic Distortion (2)
The expression below provides an analytical relationship

between the peak INL due a cubic bow and the resulting HD3
The result provides a reasonable bound even if the INL is not

perfectly cubic
Rule of thumb: HD -20log(2B/INL)
E.g. 1 LSB INL at 10 bits HD -60 dB
20

2
3 3
4
Example of a cubic INL

[Lee, JSSC 12/2007]
B. Murmann
39
SNR Degradation due to DNL (1)
[Source: Ion Opris]
For an ideal quantizer we assumed uniform quatization error

over /2
Let's add uniform DNL over 0.5 LSB and repeat math...
B. Murmann
EE315B - Chapter 3
40
SNR Degradation due to DNL (2)
Integrate triangular pdf

+
e =2
2
e e2
2
1 de = 6
SNR = 6.02 B 1.25 [dB]
3dB
Compare to ideal quantizer

+ /2
e2
2
e =
de =
12
/2
2
SNR = 6.02 B + 1.76 [dB]
Bottom line: non-zero DNL across many codes can easily cost a
few dB in SNR
"DNL noise"
EE315B - Chapter 3
B. Murmann
41
Noise Figure of an ADC
2
Vin,rms
=
Pin =
B. Murmann
1 2 1 VFS
1
V =
= 1V 2
2
2 2
2
2
Vin,rms
1 1V 2
= 10mW = 10dBm
2 50
EE315B - Chapter 3
42
Quantization Noise in 1Hz

0
-20
100MHz
= 10dBm 61.9dB 10log
1Hz
= 131.9
dBm
Hz
NPSD
fs
= Pin SQNR 10log 2

1 Hz
-40
-60
-80
For comparison:
-100
NPSD,Rsource
dBm
= 174
Hz
-120
0
0.1
0.2
0.3
Frequency [f/fs]
EE315B - Chapter 3
B. Murmann
0.4
43
Spot Noise Figure (1)
F=
Total Noise at ADC Input

Noise due to ADC
Noise due to Source Resistor Noise due to Source Resistor
NF = 10log(F) = 174
dBm
dBm
131.9
= 43.1dB
Hz
Hz
Ouch!
B. Murmann
EE315B - Chapter 3
44
0.5
Spot Noise Figure (2)

fs
dBm
+ Pin SQNR 10log 2
NF = 174
Hz
1 Hz
1 V 2
fs
FS
2
dBm
3
2 2
2N
= 174
+ 10log
10log
2
10log
Hz
R
2
1 Hz
Improve NF by increasing R (need a transformer),

resolution (N) and fs
B. Murmann
EE315B - Chapter 3
45
Real World Example
Analog Devices, Tutorial MT-006

B. Murmann
EE315B - Chapter 3
46
Nyquist Rate DACs
Boris Murmann
Stanford University
B. Murmann
EE315B - Chapter 4
Overview
D/A conversion is typically accomplished through the division or

multiplication of a reference voltage, current or charge
Basic configurations
Thermometer
Binary weighted
Segmented
Implementation choices
Resistive, capacitive, current steering
Nyquist rate, oversampling, PWM

Our discussion will focus mainly on high-speed current steering
B. Murmann
EE315B - Chapter 4
Resistor String DAC
VRE F
MSB
LSB
xxxxx
xxxxx
d0
d0
d1
d1
d2
d2
Simple, inherently
monotonic
Small area up to ~8 bits
See e.g. Pelgrom,

JSSC 12/1990
Unsuitable for highresolution, high-speed

designs
OUT
B. Murmann
EE315B - Chapter 4
Thermometer DAC Using Switched Currents
B. Murmann
EE315B - Chapter 4
Inherently monotonic
Need large encoder

with 2B-1 outputs
Impractical for large
B (high resolution)
Thermometer DAC Principle

15
Output
14
14
13
13
13
12
12
12
12
11
11
11
11
11
10
10
10
10
10
10
10
11
12
13
14
15
Input Code
B. Murmann
EE315B - Chapter 4
Binary Weighted DAC
B. Murmann
EE315B - Chapter 4
No encoder needed
Monotonicity is not
guaranteed
Consider transition
100000. to
011111.
2B-1 source must
match sum of others
to within 1 LSB to
make transition
monotonic
Binary Weighted DAC Principle
MSB Transition
Problem
B. Murmann
EE315B - Chapter 4
Implementation of Weighted Elements
B. Murmann
EE315B - Chapter 4
Segmented DAC
B. Murmann
Binary weighted
section with Bb bits
Thermometer section
with Bt = B-Bb bits
Typically Bt ~ 48
Reasonably small
encoder
Easier to achieve
monotonicity
EE315B - Chapter 4
Segmented DAC (2-2)

Transition
Problem Limited
to LSB Section
B. Murmann
EE315B - Chapter 4
10
Static Errors (1)
DNL and INL due to unit element mismatch
Systematic Errors
Contact and wiring resistance (IR drop)
Edge effects in unit element arrays
Process gradients
Finite current source output resistance
Systematic errors can be mitigated by proper layout and

switching sequence design
See e.g. [Lin, JSSC 12/98], [Van der Plas, JSSC 12/99]
B. Murmann
EE315B - Chapter 4
11
Static Errors (2)
In the best possible scenario, we are then limited by random errors,

which are due to material roughness, randomness in etching, etc.
The distribution of random errors is usually well approximated by a

Gaussian PDF (central limit theorem)
References
C. Conroy et al., Statistical Design Techniques for D/A
Converters, IEEE J. Solid-State Ckts., pp. 1118-28, Aug. 1989.
P. Crippa, et al., "A statistical methodology for the design of highperformance CMOS current-steering digital-to-analog converters,"
IEEE Trans. CAD of ICs and Syst. pp. 377-394, Apr. 2002.
B. Murmann
EE315B - Chapter 4
12
Gaussian Distribution
f(x) =
1
2
( x )2
2 2
X=
0.4
f(x)
0.3
0.2
0.1
0
-3
-2
-1
0
x/X
EE315B - Chapter 4
B. Murmann
13
Yield (1)
2
P ( C X +C ) =
1
2
+C X
e 2
C
dX = erf
P(-C
C)
P(-X xX +X)
0.8
0.6
0.4
0.2
0
0
0.5
1.5
X/
2.5
B. Murmann
EE315B - Chapter 4
14
Yield (2)
C
P(-C X C) [%]
P(-C X C) [%]
0.2000
15.8519
2.2000
97.2193
0.4000
31.0843
2.4000
98.3605
0.6000
45.1494
2.6000
99.0678
0.8000
57.6289
2.8000
99.4890
1.0000
68.2689
3.0000
99.7300
1.2000
76.9861
3.2000
99.8626
1.4000
83.8487
3.4000
99.9326
1.6000
89.0401
3.6000
99.9682
1.8000
92.8139
3.8000
99.9855
2.0000
95.4500
4.0000
99.9937
EE315B - Chapter 4
B. Murmann
15
Example
Measurements show that the current in a production lot of

current sources follows a Gaussian distribution with = 0.1 mA
and = 10 mA
What fraction of current sources is within 3% (or 1%) of
the mean?
Relative matching ("coefficient of variation")

u =
I 0.1mA
= stdev =
= 1%
I 10mA
Fraction of current sources within 3%

C = 3 99.73%
Fraction of current sources within 1%

C = 1 68.27%
B. Murmann
EE315B - Chapter 4
16
Mismatch in MOS Current Sources

I = I1 I2 gmVt + I1
g
I
m Vt +
I1
I1
1
W
Cox
2
L
Vt =
A Vt
WL
A
WL
Example
W=500m, L=0.2m, gm/ID=10S/A, AVt=5mV-m, A=1%-m
2
S 5mV 1%
I = 10
+
=
A 10 10
( 0.5%)2 + ( 0.1%)2
= 0.51%
EE315B - Chapter 4
B. Murmann
17
DNL of Thermometer DAC
DNL(k) =
Step(k) Stepavg
Stepavg
Ik
=
1 N
I
N j =1 j
N
1
I
N j =1 j
Ik I I
=
I
I
I
stdev (DNL(k) ) = stdev = u
I
Standard deviation of DNL for each code is simply equal to

relative matching (u) of unit elements
Example
Say we have unit elements with u = 1% and want 99.73% of
all converters to meet the spec
Which DNL specification value should go into the datasheet?
B. Murmann
EE315B - Chapter 4
18
DNL Yield Example (1)
First cut solution

For 99.73% yield, need C = 3
DNL = u = 1%
3 DNL = 3%
DNL specification for a yield of 99.73% is 0.03 LSB
Independent of target resolution (?)
Not quite right

Must keep in mind that a converter will meet specs only if all
codes meet DNL spec, i.e. DNL(k) < DNLspec for all k
A converter with more codes is less likely to have all codes
meet the specification
Let's see if this is significant
B. Murmann
EE315B - Chapter 4
19
Let's say there are N codes, and assume that all DNL(k) values
are independent, then
P(all codes meet spec) = P(single code meets spec)N
P(all codes meet spec)1/N = P(single code meets spec)
Lets look at two examples N=63 (6 bits) and N=4095 (12 bits)
0.99731/63 = 0.99995708
0.99731/4095 = 0. 99999929929
Can calculate modified confidence intervals using Matlab

For N=63, C = sqrt(2)*erfinv(0.99731/63) = 4.09
For N=4095, C = sqrt(2)*erfinv(0.99731/4095) = 4.97
Refined result for 99.97% yield

N=63: DNL spec should be 0.0409 LSB
N=4095: DNL spec should be 0.0497 LSB
B. Murmann
EE315B - Chapter 4
20
Getting a more accurate yield estimate for the preceding

example wasn't all that hard
Unfortunately things won't always be that simple
E.g. in a segmented DAC, DNL(k) are no longer independent
The "typical" DAC designer tends to rely on simulations rather

than trying to formulate "exact" yield equations
Get rough estimate using simple (often optimistic)
expressions
Run "Monte Carlo" simulations in Matlab to find actual yield
or to center specs
Still important to have a qualitative feel for what may cause
discrepancies
EE315B - Chapter 4
B. Murmann
21
INL (1)
INL(k) =
Iout (k) Iout,uniform (k)

Stepavg
k
Ij
=
j =1
k N
I
N j=1 j
1 N
I
N j =1 j
N Ij
=
j =1
k
Ij +
j =1
Ij
j = k +1
A N
k
A +B
A N
A
X
var (INL(k) ) = var
k = N2 var
= N2 var
A +B
A +B
Y
B. Murmann
EE315B - Chapter 4
22
INL (1)
For a quotient of random variables

X
var X
Y Y
2X 2Y
cov ( X, Y )
2 + 2 2
X
Y
Y
X
[Dennis E. Blumenfeld, Operations Research Calculations Handbook, Online:

http://www.engnetbase.com/ejournals/books/book_summary/toc.asp?id=701]
After identifying the means (), variances (2) and covariance

(cov) needed in the above approximation, it follows that
k
var (INL(k)) k 1 u2
N
k
INL (k) u k 1
N
EE315B - Chapter 4
B. Murmann
23
INL (2)
6
INL (k)/u
N=64
N=128
4
20
40
60
80
100
120
140
Standard deviation of INL is maximum at mid-scale (k=N/2)

INL u
N N/ 2 1
1
= u N u 2B
1
2
N 2
2
For a more elaborate derivation of this result see

[Kuboki et al., IEEE Trans. Circuits & Systems, 6/1982]
B. Murmann
EE315B - Chapter 4
24
Achievable Resolution
2

B log2 4 INL = 2 + 2log2 INL
u
u
Example: INL= 0.1 LSB (at mid-scale code)

u
1%
8.6
0.5%
10.6
0.2%
13.3
0.1%
15.3
EE315B - Chapter 4
B. Murmann
25
INL Yield
Again, we should ask how many DACs will meet the spec for a
given INL (worst code)
It turns out that this is a very difficult math problem
Two solutions
Do the math
G. I. Radulov et al., "Brownian-Bridge-Based Statistical
Analysis of the DAC INL Caused by Current Mismatch," IEEE
TCAS II, pp. 146-150, Feb. 2007.
Yield simulations
Good rule of thumb

For high target yield (>95%), the probability of "all codes
meet INL spec" is very close to "worst code meets INL spec"
B. Murmann
EE315B - Chapter 4
26
DNL/INL of Binary Weighted DAC
INL same as for thermometer DAC

Why?
DNL is not same for all codes, but depends on transition
Consider worst case: 0111 1000

Turning on MSB and turning off all LSBs
2
DNL
= 2B 1 1 u2 + 2B 1 u2 = 2B 1 u2
14
4244
3 1424
3
0111...
1000...
8I
4I
2I
Example
B = 12, u = 1% DNL = 0.64 LSB
Much worse than thermometer DAC
EE315B - Chapter 4
B. Murmann
27
DNL (4-bit Example)

15
2 )22
( DNL
D NL //
u
10
6
8
DAC input code
10
12
14
code
B. Murmann
EE315B - Chapter 4
28
Simulation Example
DNL and INL of 12 Bit converter (from converter decision thresholds)
DNL [in LSB]
2
-1 / +0.1 LSB, avg=-9.3e-005, std.dev=0.035, range=1.4
-1.3
1
0
-1
-2
500
1000
1500
2000
bin
code
2500
3000
3500
4000
INL [in LSB]
2
-0.8 / +0.8 LSB, avg=-1.1e-013, std.dev=0.37, range=1.6
1
0
-1
500
1000
1500
2000
bin
code
2500
3000
3500
4000
EE315B - Chapter 4
B. Murmann
29
Another Random Run

DNL and INL of 12 Bit converter (from converter decision thresholds)
DNL [in LSB]
2
1
-0.9 / +0.4 LSB, avg=-7.5e-005, std.dev=0.039, range=1.3
0
-1
500
1000
1500
2000
2500
3000
3500
4000
bin
code
INL [in LSB]
2
1
-0.7 / +0.7 LSB, avg=3.3e-014, std.dev=0.33, range=1.3
0
-1
500
1000
1500
2000
2500
3000
3500
4000
bin
code
Peak DNL not at mid-scale!

Important to realize that this is just one single statistical
outcome
B. Murmann
EE315B - Chapter 4
30
Multiple Simulation Runs (100)

Overlay Plot
RMS DNL and INL
[Lin & Bult, JSSC 12/1998]

EE315B - Chapter 4
B. Murmann
31
DNL/INL of Segmented DAC
INL
Same as in thermometer
DAC
Example: B=Bb+Bt=4+4=8
DNL
Worst case occurs when
LSB DAC turns off and
one more MSB DAC
element turns on
Essentially same DNL as
a binary weighted DAC
with Bb+1 bits
B. Murmann
16I
8I
EE315B - Chapter 4
16I
4I
16I
2I
32
Comparison
Thermometer
INL
(worst case)
DNL
(worst case)
Number of
Switched
Elements
Binary
Weighted
Segmented
1
u 2B
2
u 2Bb +1 1
u 2B 1
2B 1
Bb + 2Bt 1
EE315B - Chapter 4
B. Murmann
33
Example (B=12, u=1%)
INL
DNL
(worst)
(worst)
Number of
Switched
Elements
Thermometer
0.32
0.01
4095
Binary Weighted
0.32
0.64
12
Segmented (Bb=7, Bt=5)
0.32
0.16
38
DAC Architecture
B. Murmann
EE315B - Chapter 4
34
Dynamic DAC Errors (1)
Finite settling time and slewing

Finite RC time constant
Signal dependent slewing
Feedthrough
Coupling from switch signals to DAC output
Clock feedthrough
Glitches due to timing errors

Current sources wont switch simultaneously
Dynamic DAC errors are generally hard to model!
B. Murmann
EE315B - Chapter 4
35
Dynamic DAC Errors (2)
References
Gustavsson, Chapter 12
M. Albiol, J.L. Gonzalez, E. Alarcon, "Mismatch and dynamic
modeling of current sources in current-steering CMOS D/A
converters," IEEE TCAS I, pp. 159-169, Jan. 2004
Doris, van Roermund, Leenaerts, Wide-Bandwidth High
Dynamic Range D/A Converters, Springer 2006.
T. Chen and G.G.E. Gielen, "The analysis and improvement
of a current-steering DAC's dynamic SFDR," IEEE Trans.
Ckts. Syst. I, pp. 3-15, Jan. 2006.
B. Murmann
EE315B - Chapter 4
36
Glitch Impulse (1)
DAC output waveform depends on timing

Consider binary weighted DAC transition 0111 1000
ideal
10
Ideal
0
1
1.5
2.5
early
10
LSBs early, MSB late
0
1
1.5
2.5
late
10
LSBs late, MSB early
0
1
1.5
2
Time
2.5
EE315B - Chapter 4
B. Murmann
37
Glitch Impulse (2)
Worst case glitch impulse (area): t 2B-1
LSB area: T
Need t 2B-1 << T which implies t << T/2B-1
B. Murmann
fs [MHz]
t [ps]
12
<< 488
20
16
<< 1.5
1000
10
<< 2
EE315B - Chapter 4
38
Commercial Example
B. Murmann
EE315B - Chapter 4
39
Implementation Example
[T. Miki, Y. Nakamura, M. Nakaya, S. Asai, Y. Akasaka, and Y. Horiba, An 80-MHz 8bit CMOS D/A Converter, IEEE J. of Solid-State Circuits, pp. 983-988, Dec. 1986.]
B. Murmann
EE315B - Chapter 4
40
Mitigating IR Drop
B. Murmann
EE315B - Chapter 4
41
Basic Differential Pair Switch
B. Murmann
EE315B - Chapter 4
42
Commonly Used Techniques
Retiming
Latches in (or close to) each current cell
Latch controlled by global clock to ensure that current cells
switch simultaneously (independent of decoder delays)
Make before break

Ensure uninterrupted current flow, so that tail current source
remains active
Low swing driver

Drive differential pair with low swing to minimize coupling
from control signals to output
Cascoded tail current source for high output impedance

Ensures that overall impedance at output nodes is code
independent (necessary for good INL)
B. Murmann
EE315B - Chapter 4
43
Example Current Cell Implementation
[Barkin & Wooley, JSSC 4/2004]
B. Murmann
EE315B - Chapter 4
44
Constant Clock Load Latch

Q
D
MN1
INV1
INV2
QB
DB
MN2
INV6
Capacitive load seen

at CLK the same for
all possible cases,
H-L, L-H, H-H or L-L
MN3
INV3
INV7
CLKB
INV5
MN4
INV4
CLK
Mercer, US patent ,7,023,255 4/4/2006
EE315B - Chapter 4
B. Murmann
45
High Performance DAC Examples (1)

100MHz
1GHz
[Van den Bosch, JSSC 3/2001]
B. Murmann
EE315B - Chapter 4
46
[Schafferer, ISSCC 2004]
B. Murmann
EE315B - Chapter 4
47
[Schafferer, ISSCC 2004]
B. Murmann
EE315B - Chapter 4
48
[Lin, ISSCC 2009]
EE315B - Chapter 4
B. Murmann
49
Binary Weighted Charge Redistribution DAC

CP (top plate parasitic)
+
2N1 C
8C
4C
2C
C VOUT
VRE F
b1
(msb)
bN3
bN2
Vout =
bN1
2B C
B
2 C + Cp
bN
(lsb)
Vref
bi
i
i =1 2
Can redistribute charge onto OTA + feedback capacitor to

mitigate gain error due to Cp
B. Murmann
EE315B - Chapter 4
50
Charge-based Pipeline DAC (1)
[Manganaro et al., "A dual 10-b 200-MSPS pipelined D/A converter with DLL-based
clock synthesizer," IEEE JSSC 11/2004]
B. Murmann
EE315B - Chapter 4
51
Charge-based Pipeline DAC (2)
(fclk=200MHz)
B. Murmann
EE315B - Chapter 4
52
Sampling Circuits
Boris Murmann
Stanford University
B. Murmann
EE315B - Chapter 5
Recap
How to build circuits that "sample"?
Ideal Dirac sampling is impractical

Need a switch that opens, closes and acquires signal within
an infinitely small time
Practical solution
"Track and hold"
B. Murmann
EE315B - Chapter 5
Outline
Elementary track-and-hold circuit and its nonidealities
First order improvements to elementary track-and-hold
Advanced techniques
Clock bootstrapping
Bottom plate sampling
Settling and noise analysis in charge-redistribution

track-and-hold circuit
Noise simulation example
B. Murmann
EE315B - Chapter 5
Ideal Track-and-Hold Circuit
B. Murmann
EE315B - Chapter 5
Signal Nomenclature
time
Continuous Time Signal
T/H Signal
("Sampled Data Signal")
Clock
Discrete Time Signal
B. Murmann
EE315B - Chapter 5
Circuit with MOS Switch
Pedestal Error
B. Murmann
EE315B - Chapter 5
Nonidealities
Finite acquisition time
Tracking nonlinearity
Signal dependent hold instant
Thermal noise
Clock jitter
Hold mode feedthrough and leakage
Charge injection and clock feedthrough
EE315B - Chapter 5
B. Murmann
Finite Acquisition Time
Consider various input signal scenarios
1. Input is a sampled data signal, i.e. the

output of another switched capacitor stage
2. Input is a slowly varying continuous time
signal, e.g. the input of an oversampling
ADC
3. Input is a rapidly varying continuous time
signal, e.g. the input of a Nyquist or subsampling ADC
B. Murmann
EE315B - Chapter 5
Finite Acquisition Time Case 1
For simplicity, neglect finite rise time of the input signal

Consider worst case the output is required to settle from 0 to
the full-scale voltage of the system (VFS)
First order MOS switch model
= RC
Vout (t) = VFS 1 e t /
EE315B - Chapter 5
B. Murmann
)
9
Finite Acquisition Time Case 1
T
Vout,err s = VFSe
2
Ts
/
2
= VFSeN
Typically think about settling error in terms of the number of settling

time constants (N) required for LSB settling in a B-bit system
VFS e
N=
B. Murmann
1 VFS
2 2B
Ts / 2
ln 2 2B
)
EE315B - Chapter 5
>4.9
10
>7.6
14
>10.4
18
>13.2
10
Finite Acquisition Time Case 2 & 3
Consider a sinusoidal input around 0, and 0 also as the initial

condition for Vout (for notational simplicity)
Vin (t) = A cos( t + )

Vout (t) =
A cos( )
2 2
1 +4
144
2444
3
initial transient
A cos( t + )
14
+2444
2 2 3
144
steady-state response
= atan()
B. Murmann
EE315B - Chapter 5
11
At t=Ts/2, the error in the held signal consists of two parts
Residual error due to initial exponentially decaying initial

transient term
In order to minimize this error, we need to chose N
appropriately, as calculated for the step input scenario
Error due to magnitude attenuation and phase shift in the steady

state term
This error depends only on the RC time constant and the
input frequency; it cannot be reduced by extending the
length of the track phase
How significant is the error due to the steady-state term?
B. Murmann
EE315B - Chapter 5
12
As an example, lets compute the percent amplitude error for the

N values derived previously ( LSB, B-bit settling to a step)
A
A err =
B. Murmann
1 + ( )
A
= 1
1
1 + ( )
= 1
T /2
1 + 2fin s
N
= 1
f
1 + in
N fs
Aerr (fin = fs/20)
Aerr (fin = fs/2)
4.9
0.052%
4.9%
10
7.6
0.021%
2.1%
14
10.4
0.011%
1.1%
18
13.2
0.007%
0.7%
EE315B - Chapter 5
13
Summary Finite Acquisition Time
Precise settling to an input step is accomplished within 513

RC time constants (depending on precision)
Precise tracking of a high-frequency continuous time input

signal tends to impose more stringent requirements
Number to remember: ~1% attenuation error at Nyquist
(fin=fs/2) for N ~10
In applications where attenuation is tolerable, the RC time

constant requirements then tend to follow from the distortion
specs
The larger the attenuation, the larger the instantaneous
voltage drop across the (weakly nonlinear) MOSFET
undesired harmonics
Lets look at that next
B. Murmann
EE315B - Chapter 5
14
Voltage Dependence of Switch

ID(triode) = Cox
RON
dI
D(triode)
dVDS
RON =
V
W
VGS Vt DS VDS
L
2
Cox
VDS 0
=
Cox
W
( V Vt )
L GS
1
W
( Vin Vt )
L
Two problems
RON is modulated by Vin (assuming e.g. =VDD=const.)
Transistor turn off is signal dependent, occurs when =Vin+Vt
B. Murmann
EE315B - Chapter 5
15
Track Mode Nonlinearity
[Razavi, Data Conversion System Design, p.16]
Output tracks well when input voltage is low

Gets distorted when voltage is high due to increase in RON
B. Murmann
EE315B - Chapter 5
16
Analysis
ID K ( VGS Vt ) VDS
C
K 2
VDS
2
dVout
K
2
= K ( Vout Vt )( Vin Vout ) ( Vin Vout )
dt
2
"All" we need to do is solve the above differential equation
Can use Volterra Series analysis

General method that allows us to calculate the frequency
domain response of nonlinear circuits with memory
See e.g. Stanford EE214
Luckily someone has already done this for us

See [Yu, TCAS II, 2/1999]
EE315B - Chapter 5
B. Murmann
17
Result
HD3 =
Amplitude of third harmonic

Amplitude of fundamental
2
fin
1
A
1
A

4 VGS Vt
4 VGS Vt fs N
VGS is the "quiescent point" value of the gate-source voltage; i.e.

in the zero crossing of the sine input
For low distortion

Make amplitude smaller than VGS-Vt
Low swing bad for SNR
Make 1/ much larger than (input frequency)

Big switch may cost lots of power to drive, comes with large
parasitic capacitances
B. Murmann
EE315B - Chapter 5
18
Numerical Example
Parameters
VDD = VCK = 1.8V
Signal is centered about VDD/2 = 0.9V
VGS-Vt = 1.8V-0.9V-0.45V = 0.45V
A = 0.2V
N = 0.5Ts/ = 10
fin = fs/2
2
1 0.2 1
HD3
= 42dB
4 0.45 2 10
Not all that great
EE315B - Chapter 5
B. Murmann
19
Signal Dependent Sampling Instant (1)

()
Tf
Must make fall time of sampling clock (Tf) much faster than
maximum dVin/dt
B. Murmann
EE315B - Chapter 5
20
Signal Dependent Sampling Instant (2)
Distortion analysis result (see Yu, TCAS II, 2/1999]

HD3 =
Amplitude of third harmonic

Amplitude of fundamental
3 A
Tf
8 VCK
Example: VCK = 1.8V, A = 0.5V, Tf = 100ps, = 2100MHz

2
HD3
3 0.5
2100 106 100 1012 = 79dB
8 1.8
EE315B - Chapter 5
B. Murmann
21
Hold Mode Feedthrough

CDS
Want to make Rout as small as possible
Consider using a T-switch when hold-mode

feedthrough is a problem (for low-speed
applications)
Consider using half circuit cross coupling for

high-speed applications
B. Murmann
EE315B - Chapter 5
22
Hold Mode Leakage

Example:
B. Murmann
EE315B - Chapter 5
23
Gate Leakage Data
A. Annema, et al., Analog circuits in ultra-deep-submicron CMOS, IEEE J. Solid-State Circuits, pp. 132-143, Jan. 2005.
In 65nm CMOS, gate capacitance droop rate is ~1V/s (!)

Issue is solved with high-k dielectrics in post-65nm technologies
B. Murmann
EE315B - Chapter 5
24
Thermal Noise (1)
Questions
What is the noise variance of the Vout samples in hold mode?
What is the spectrum of the discrete time sequence
representing these samples?
Nearly white, provided that the number of settling time
constants (N) is large (see EE315A for a derivation)
EE315B - Chapter 5
B. Murmann
25
Thermal Noise (2)
Sample values Vout(n) correspond to instantaneous values of the

track mode noise process
From Parseval's theorem, we know that the time domain power

(or variance) of this process is equal to its power spectral
density integrated over all frequencies
Further, given that the process is ergodic, this number must also be equal to the "ensemble" variance,
i.e. the variance of a sample taken at a particular time
2
v out
1
= 4kTR
1 + sRC
f
var [ Vout (n)] =
B. Murmann
2
v out,tot
1
kT
= 4kTR
df =
1 + j2f RC
C
0
EE315B - Chapter 5
26
Alternative Derivation
The equipartition theorem says that each degree of freedom"

(typically a quadratic energy variable) of a system in thermal
equilibrium holds an average energy of kT/2
In our system, the degree of freedom is the energy stored on the

capacitor
1
1
Cv out 2 = kT
2
2
v out 2 =
kT
C
EE315B - Chapter 5
B. Murmann
27
Implications of kT/C Noise
Example: suppose we make the kT/C noise equal to the

quantization noise of a B-bit ADC
kT 2
=
,
C 12
VFS
2B
2B
C = 12kT
V
FS
For a given B, both C and R (via N on slide 19) are fully determined
Example numbers for VFS=1V and fs=100MHz:
B
8
10
12
14
16
18
B. Murmann
C [pF]
0.003
0.052
0.834
13.3
213
3,416
EE315B - Chapter 5
R []
246,057
12,582
665
36
1.99
0.11
28
Commercial Example
B. Murmann
EE315B - Chapter 5
29
How to Estimate Thermal Noise

in the Lab?
One option is to look at the measured SNR and back-calculate

the thermal component by subtraction of quantization noise and
DNL noise
Another (potentially better) option is to take a code histogram

with grounded input
S. Ruscak and L. Singer, Using Histogram

Techniques to Measure A/D Converter Noise
http://www.analog.com/library/analogDialogue/
archives/29-2/cnvtrnoise.html
B. Murmann
EE315B - Chapter 5
30
Sampling Jitter
In any sampling circuit, electronic noise causes random timing

variations in the actual sampling clock edge
Adds "noise" to samples, especially if dVin/dt is large
Vin = Change in V in during t
Vin
dVin
t
dt
t = Sampling jitter
Analysis
Consider sine wave input signal
Assume t is random with zero mean and standard deviation t
EE315B - Chapter 5
B. Murmann
31
Analysis
{ }
Vin2
dV 2
dV 2
2
2
in
in
E
t = E dt E t
dt
{ }
2
d
1

2
E A cos [ 2 fin t ] 2t ( 2 A fin ) 2t
2

dt
=
B. Murmann
1

2
1

2
EE315B - Chapter 5
1

32
Result
1.E+11
1.E+10
fin [Hz]
1.E+09
t = 0.1ps
1.E+08
t = 1ps
1.E+07
t = 10ps
1.E+06
1.E+05
10
20
30
40
50
60
70
80
90
100
110
120
SNR
[dB]
SNR
aperture
jitter [dB]
B. Murmann
EE315B - Chapter 5
33
ADC Performance Survey (ISSCC & VLSI 97-12)

Data: http://www.stanford.edu/~murmann/adcsurvey.html
B. Murmann
EE315B - Chapter 5
34
Why is it Hard to Achieve Low Jitter?

Inverters have little to no
supply rejection
Example:
1V supply, 20ps rise time
100mV noise on the supply
corresponds to about 2ps
of jitter
Noisy VDD
Jitter
Free
Clock
Jittered
Output
Clock
In addition, there is thermal

amplitude noise (kT/C) that
also translates into jitter
B. Murmann
EE315B - Chapter 5
35
How Well Can We Do?
In todays optimized designs, jitter numbers as low as 50 fsrms have

been achieved, see e.g. Ali, JSSC 8/2006
However, several hundred psrms are more common, especially
in noisy SoC environments
Differential signaling can help (particularly at the chip boundary),

but comes with diminishing returns due to reduced amplitude
and signal slope (relative to inverters)
B. Murmann
EE315B - Chapter 5
36
Significance of Jitter
In light of the above reasoning, sampling jitter has become one

of the main showstoppers for further improvements in the ADC
speed-resolution product
Example
fin = 10GHz, t = 300ps SNRjitter = 34.5dB (ouch!)
What saves us in many application is that the signal is not a

sinusoid, but instead spread in some way across frequency
Proper analysis for applications that push the boundaries must

therefore take the signal properties into account
Reference: Da Dalt, TCAS1, 9/2002
EE315B - Chapter 5
B. Murmann
37
Generalized Calculation
Da Dalt showed that in general, the jitter SNR can be computed

using the curvature of the signals autocorrelation function
=
1

=
0
0
Side note
In this result, the signal is assumed to be stationary
An extension for cyclostationary signals was provided by ElChammas, TCAS1, May 2009
Useful for systems with simple signal constellations, such as NRZ
B. Murmann
EE315B - Chapter 5
38
Intuition
Gradual decay in
autocorrelation
Slow
signal
Sharp decay in
autocorrelation
Fast
signal
EE315B - Chapter 5
B. Murmann
39
Example: High-Speed Link
White Info
Source
TX
Channel
h(t)
RX
Rx is the autocorrelation function of the input signal (assumed to

be white and stationary in this example)
Ry is the autocorrelation function of the signal at the sampler
=
= = [ ]

B. Murmann
EE315B - Chapter 5
40
Example
Chanel Frequency Response
Chanel Impulse Response
0
5
-10
-30
h(t) [V/ns]
S21 [dB]
-20
-40
fs/2
-50
3
2
-60
1
-70
-80
0
10
f [GHz]
15
20
0
4.4
4.6
4.8
Time [ns]
EE315B - Chapter 5
B. Murmann
5.2
41
Example
Convolved Impulse Response
Second Derivative
x 10
0.12
-5
1
0.5
0.1
0
w''() [1/ps2]
w()
0.08
0.06
0.04
-0.5
-1
-1.5
-2
0.02
0
-1
B. Murmann
-2.5
-0.5
0
[ns]
0.5
-3
-0.5
EE315B - Chapter 5
0
[ns]
0.5
42
Example
=
0
1
2.8 10 1
=
= 2.3 10
0.12

0
1
= 46.8
300
= 20 log
Another way to think about this is to compute the sinusoid

frequency that causes the same jitter noise
, =
(Better!)
1
= 2.4
2
Bottom line: the fact that the signal is wideband and filtered by
the channel helps, but the jitter spec will still be non-trivial
B. Murmann
EE315B - Chapter 5
43
How to Measure/Quantify the Sampling

Jitter of an ADC in the Lab?
Use sinusoids as a test signal
Method 1: Leverage the known proportionalities and extract jitter

estimate using an amplitude and frequency sweep
RMS jitter is noise proportional to input frequency
RMS is noise proportional to input amplitude
This works well when jitter is the dominant issue, and not masked
by other nonidealities (like thermal noise)
Method 2: Leverage the fact that jitter causes cyclostationary noise

Jitter causes almost no noise near peaks of sinusoid, but lots of
noise near zero crossing
Noise beat at twice the signal frequency
B. Murmann
EE315B - Chapter 5
44
Jitter Noise Beat

[Aaron Buchwald]
Can back calculate jitter noise by comparing
variance at zero crossing with variance near peak
(which is set by thermal and quantization noise)
B. Murmann
EE315B - Chapter 5
45
Jitter Estimation Based on Spectrum
Reference
D.M. Hummels, W. Ahmed, W., F.H. Irons, "Measurement of random
sample time jitter for ADCs," Proc. ISCAS, pp.708-711, May 1995.
Spectrum of squared sequence
contains a tone proportional to jitter
After removal of harmonics:
B. Murmann
EE315B - Chapter 5
46
Matlab Code
% Jitter estimation technique proposed by Hummels, ISCAS 1995
% x is a vector that contains an integer number of cycles of a sampled sine wave
function [sigma_jitter_est, sigma_noise_est, fsin] = estimate_jitter(x, show_plot)
% take fft, remove DC, signal and harmonics
N = length(x);
s = fft(x);
[sigamp, sigbin]=max(abs(s));
A_est = sigamp/N*2;
cycles = sigbin-1;
fsin = cycles/N;
harmbins = 1 + abs([2:8]*cycles - N*round([2:8]*cycles/N));
sn=s;
sn(1)=eps;
sn(sigbin)=eps;
sn(N+2-sigbin)=eps;
sn(harmbins) =eps;
sn(N+2-harmbins)=eps;
% inverse fft, compute jitter and noise estimates
xinv = ifft(sn, 'symmetric');
s_inv = abs(fft(xinv.^2));
jitterbin = 1 + abs(2*cycles - N*round(2*cycles/N));
sigma_jitter_est = sqrt(2*2*s_inv(jitterbin)/N/(2*pi*fsin)^2/A_est^2);
sigma_noise_est = sqrt(s_inv(1)/N-2*s_inv(jitterbin)/N);
B. Murmann
EE315B - Chapter 5
47
Charge Injection and Clock Feedthrough
Pedestal Error
Analyze two extreme cases

Very large Tf (slow gating)
Very small Tf (fast gating)
B. Murmann
EE315B - Chapter 5
48
Slow Gating
H
VIN + V T
V IN
t
HOLD
L
VO
VIN
V
V
t
toff
All channel charge has disappeared by toff without introducing

error; it is absorbed by the input source
EE315B - Chapter 5
B. Murmann
49
Slow Gating Model for t > toff

Vout = Vin Vout
Vout = Vin
Col
( Vin + Vt L ) = Vin (1 + ) + Vos
Col + C
Col
Col + C
Vos =
Gain Error
Col
( Vt L )
Col + C
Offset Error
Example: C=1pF, L=0V, Vt=0.45V, W=20m, Col=0.1fF/m, Col=2fF

= 0.2%
B. Murmann
Vos = 0.9mV
EE315B - Chapter 5
50
Fast Gating
Channel charge
cannot change
instantaneously
Resulting surface
potential decays via
charge flow to source
and drain
Charge divides
between source and
drain depending on
impedances loading
these nodes
H
VIN + VT
Surface
Potential
Qch
S
t < to
VO
V IN
V
V
t > to
(1-)Q
chch
2Q
B. Murmann
Q
ch
2 Qch
EE315B - Chapter 5
51
Charge Split Ratio Data
Tf
RonC2
G. Wegmann et al., "Charge injection in analog MOS switches," IEEE J. SolidState Circuits, pp. 1091-1097, June 1987.
Y. Ding and R. Harjani, "A universal analytic charge injection model," Proc.
ISCAS, pp. 144-147, May 2000.
B. Murmann
EE315B - Chapter 5
52
Interpretation
RonC2 and Tf are usually comparable, or at least not more than

an order of magnitude apart
This brings us into the range of 0.11 on the chart by
Wegmann
This means that the charge split will in practice have some
dependence on the impedances seen on the two sides of the
transistor
Remember: Slightly more charge will go to the side with lower

impedance
EE315B - Chapter 5
B. Murmann
53
Fast Gating Model for t > toff

Vout = Vin Vout = Vin (1 + ) + Vos
Vout = Vin
Col
1Q
( H L ) + ch
Col + C
2 C
Assuming 50/50
charge split
Qch = WLCox [ H Vin Vt ]

=
1 WLCox
2
C
Vos =
Col
1 WLCox
( H L )
( H Vt )
Col + C
2
C
Example: C=1pF, H-L=1.8V, Vt=0.45V, W=20m, LCox=2fF/m

Col=0.1fF/m, Col=2fF
= +2%
B. Murmann
Vos = 30.6mV
EE315B - Chapter 5
54
Transition Fast/Slow Gating

Fast gating
Slow gating
Fast gating
| |
Slow gating
|Vos|
TtFf
TtfF
|| and |Vos| decrease as the fall time of (Tf) increases and

approach the limit case of slow gating
Unfortunately, high-speed switched capacitor circuits tend to

operate in fast gating regime
EE315B - Chapter 5
B. Murmann
55
Impact of Technology Scaling
1 Qch
2 C
R
Cox
T
1
= s = N RC
2fs
2
W
( V Vt )
L GS
L2
Qch
L2
V
N
fs
Charge injection error to speed ratio benefits from shorter

channels and increases in mobility (e.g. due to strain)
B. Murmann
EE315B - Chapter 5
56
Outline
Advanced techniques
Clock bootstrapping

B. Murmann
EE315B - Chapter 5
57
Improvements
Charge cancelation
Try to cancel channel charge by injecting a charge packet
with opposite sign
Differential sampling
Use a differential circuit to suppress offset
CMOS switch
Try to balance the nonidealities of NMOS device with a
parallel PMOS
B. Murmann
EE315B - Chapter 5
58
Charge Cancellation
Q1 = 0.5Qch1 + Qol1
L =L
W =0.5W
1
Q2 Qch2 + 2Qol2 0.5Qch1 + Qol1
Q1 Q2 0
[Eichenberger and Guggenbhl, JSSC 8/89]
Cancellation is never perfect, since channel charge of M1 will not

exactly split 50/50
E.g. if Rs is very small, most of M1s channel charge will flow
toward the input voltage source
Not a precision technique, just an attempt to do a partial clean-up
EE315B - Chapter 5
B. Murmann
59
Differential Sampling (1)
VO1
+
VI1
C
CH
VIC =
VO2
+
VI2
VID = VI1 VI2
C
CH
VI1 + VI2
2
VOD = VO1 VO2

VOC =
VO1 + VO2
2
VO1 = (1 + 1 ) VI1 + VOS1

VO2 = (1 + 2 ) VI2 + VOS2
+
+
VOD = 1 + 1 2 VID + ( 1 2 ) VIC + ( VOS1 VOS2 ) 1 + 1 2 VID
2
2
+ VOS2 1 + 2
+ VOS2
V
V

+
VOC = 1 2 VID + 1 + 1 2 VIC + OS1
VIC + OS1
1+
2
2
2
2
4
B. Murmann
EE315B - Chapter 5
60
Differential Sampling (2)
Assuming good matching between the two half circuits, we have

Small residual offset in VOD
Good rejection of coupling noise, supply noise,
Small common-mode to differential-mode gain
Unfortunately, VOD has essentially same gain error as the basic

single ended half circuit
This also means that there will be nonlinear terms

Out simplistic analysis assumed that the channel charge is
linearly related to Vin
This is true only to first order (consider e.g. backgate effect)
Expect to see nonlinear distortion along with gain error
EE315B - Chapter 5
B. Murmann
61
CMOS Switch
VO
+
VIN
Qchp WpLpCox VIN L Vtp
CH
Assuming fast gating, 50/50 charge split and WnLn = WpLp

1
1
Qchn + Qchp C
2
Vo 2
= ox
C
C
Qchn WnLnCox ( H VIN Vtn )
V Vtp

VIN H L + tn
2
2
Charges fully cancel e.g. for VIN = (H-L)/2 = VDD/2, and Vtn=|Vtp|,
but there is still signal dependent residual injection
B. Murmann
EE315B - Chapter 5
62
On Resistance of CMOS Switch
At least in principle, adding a PMOS can also help with the

problem of signal dependent Ron in track mode
For increasing VIN, NMOS resistance goes up, PMOS
resistance goes down
VO
+
VIN
CCH
W
W
nCox ( VGSn Vtn ) pCox VGSp Vtp
L n
L p
EE315B - Chapter 5
B. Murmann
63
Analysis
R
W
W
nCox ( VGSn Vtn ) pCox VGSp Vtp
L n
L p
W
W
W
W
nCox ( VDD Vtn ) nCox pCox v in pCox Vtp
L
L
L p
L n
n
p
1
W
nCox VDD Vtn Vtp
L n
W
W
if n = p
L n
L p
Independent of Vin too good to be true!

Missing factors
Backgate effect
Short channel effects
B. Murmann
EE315B - Chapter 5
64
Real CMOS Switch

100
NMOS
PMOS
NMOS || PMOS
R []
80
60
40
20
0
0
0.5
1
Vin [V]
1.5
Design
Size P/N ratio to minimize change in R over input range
Size P and N simultaneously to meet distortion specs
PMOS brings limited benefit unless the input signal range is

large or centered near VDD
B. Murmann
EE315B - Chapter 5
65
Outline
Advanced techniques
Clock bootstrapping

B. Murmann
EE315B - Chapter 5
66
Clock Bootstrapping
+
VDD
-
+
VDD
-
Cboot
Cboot
VGS=VDD=const.
Vin
A. Abo, "Design for Reliability of Low-voltage, Switched-capacitor Circuits," PhD Thesis, UC Berkeley, 1999.
Phase 1
Cboot is precharged to VDD
Sampling switch is off
Phase 2
Sampling switch is on with VGS=VDD=const.
To first order, both Ron and channel charge are signal
independent
B. Murmann
EE315B - Chapter 5
67
Waveforms
A. Abo, "Design for Reliability of Low-voltage, Switched-capacitor Circuits," PhD Thesis, UC Berkeley, 1999.
B. Murmann
EE315B - Chapter 5
68
Circuit Implementation
Switch
A. Abo et al., A 1.5-V, 10-bit, 14.3-MS/s CMOS Pipeline Analog-toDigital Converter, IEEE J. Solid-State Circuits, pp. 599, May 1999
B. Murmann
EE315B - Chapter 5
69
Limitations
Efficacy of bootstrap circuit is

reduced by
Backgate effect
Parasitic capacitance
between at top plate of C3
B. Murmann
Cpar
Cpar
Cboot
W
nCox
VDD
Vin Vtn [ Vin ]
1
424
3
Cboot + Cpar
L n Cboot + Cpar
Backgate effect
EE315B - Chapter 5
70
Alternative Implementation
Dessouky and Kaiser, "Input switch configuration suitable for rail-to-rail operation of
switched opamp circuits," Electronics Letters, Jan. 1999.
Less complex, but Cpar tends to be larger due to two parasitic

well capacitances
EE315B - Chapter 5
B. Murmann
71
Performance of Bootstrapped Samplers
Bootstrapped top plate sampling (as opposed to bottom plate

more later) tends to work very well up to ~10bit resolution
Example
[Louwsma, JSSC 4/2008]
B. Murmann
EE315B - Chapter 5
72
High-Speed Example without Bootstrap
[Choi and Abidi, JSSC 12/2001]
For lower resolution applications, it can be OK to drop the

bootstrap circuit
B. Murmann
EE315B - Chapter 5
73
Bottom Plate Sampling
What if we want to do much better, e.g. 16 bits?
Basic idea
Sample signal at the "grounded" side of the capacitor to
achieve signal independent sampling
References
D. G. Haigh and B. Singh, A switching scheme for SC filters
which reduces the effect of parasitic capacitances
associated with switch control terminals, in Proc. IEEE Int.
Symp. Circuits and Systems, 1983, pp. 586589.
K.-L. Lee and R. G. Meyer, Low-Distortion SwitchedCapacitor Filter Design Techniques, IEEE J. Solid-State
Circuits, pp. 1103-1113, Dec. 1985.
First look at single ended half circuit for simplicity
B. Murmann
EE315B - Chapter 5
74
Bottom Plate Sampling Analysis (1)
Turn M2 off "slightly" before M1

Typically a few hundred ps
delay between falling edges
of e and
During turn off, M2 injects charge

Q2
Q2
1
WLCox ( H Vtn )
2
To first order, the charge injected

by M2 is signal independent
Voltage across C
VC = Vin +
B. Murmann
Q 2
C
EE315B - Chapter 5
75
Q1
B. Murmann
Next, turn off M1
M1 will inject signal dependent

charge onto the series
combination of C and the
parasitic capacitance at its
bottom plate (Cpar)
Looks like, this is not much

different from the conventional
top-plate sampling?
But wait
1
WLCox ( H Vin Vtn )
2
EE315B - Chapter 5
76
Q X = CVin Q2
Charge injected by M2
(Signal independent)
Interesting observation
Even though M1 injects
some charge, the total
charge at node X cannot
change!
Idea
Process total charge at
node X instead of looking at
voltage across C
The charge can be processed

in two ways
Open-loop
Closed-loop (charge
redistribution)
EE315B - Chapter 5
B. Murmann
77
Open-Loop Charge Processing

Q X = CVin Q2
1
Vin
VX =
M1
QX
Q2
C
= Vin
C + Cp
C + Cp C + Cp
(no term due to signal

dependent charge!)
Vx
1e
Cpar
M2
1e
Remaining drawback
Cpar (and buffer input capacitance) is usually weakly nonlinear
and will introduce some harmonic distortion
B. Murmann
EE315B - Chapter 5
78
Closed-Loop Charge Processing
Amplifier forces voltage at node X to zero

Means that charge at node X must redistribute onto
feedback capacitor Cf
EE315B - Chapter 5
B. Murmann
79
Charge Conservation Analysis

Charge at node X during 1:
Q X1 = CVin Q2 + 0 Cf
Charge at node X during 2:
Q X2 = Cf Vout
Q X1 = Q x2
Charge Conservation:
CVin Q2 = Cf Vout
Vout =
Q2
C
Vin +
Cf
Cf
Offset term due to signal independent injection from M2 can be

easily removed using a differential architecture
B. Murmann
EE315B - Chapter 5
80
Clock Generation
[A. Abo, "Design for Reliability of Low-voltage, Switched-capacitor Circuits," PhD

Thesis, UC Berkeley, 1999]
B. Murmann
EE315B - Chapter 5
81
Fully Differential Circuit
B. Murmann
EE315B - Chapter 5
82
Analysis (1)
Q1m = CVinp + Q
Q2m = CVxm Cf Vop Vxm
Q1p = CVinm + Q
Q2p = CVxp Cf Vom Vxp
1)
Q1m = Q2m
2)
Q1p = Q2p
Vxm = Vxp
Vop + Vom
2
= Voc
EE315B - Chapter 5
B. Murmann
83
Analysis (2)
Subtracting 1) and 2) yields

Vop Vom =
C
( Vinp Vinm )
Cf
Adding 1) and 2) yields
C Vinp + Vinm + 2Q = ( C + Cf ) Vxp + Vxm Cf Vop + Vop

Vxc =
Cf
Q
C
+
V oc
V ic
C + Cf C + C f
C + Cf
Variations in Vic show up as common mode variations at the

amplifier input
Need amplifier with good CMRR
B. Murmann
EE315B - Chapter 5
84
T/H with Common Mode Cancellation
S.H. Lewis & P.R. Gray, "A Pipelined 5 MSample/s 9-bit Analog-to-Digital Converter", IEEE
J. Solid-State Circuits, pp. 954-961, Dec. 1987
Shorting switch allows to re-distribute only differential charge on

sampling capacitors
Common mode at OPAMP input becomes independent of common

mode at circuit input terminals (IN+/IN-)
Original idea: Yen & Gray, JSSC 12/1982

EE315B - Chapter 5
B. Murmann
85
Analysis (1)
During 1
During 2
C
C
V
V
V
V
C
C
V
C
Charge conservation at Vip,Vim and Vfloat
( Vip + Vim ) C = ( Vfloat Vxp ) C + ( Vfloat Vxm ) C

Vic = Vfloat Vxc
Vfloat = Vic + Vxc
B. Murmann
EE315B - Chapter 5
86
Analysis (2)
Common mode charge conservation at amplifier inputs

Vic C Voc Cf = ( Vfloat Vxc ) C ( Voc Vxc ) Cf
Vic C = ([ Vic + Vxc ] Vxc ) C + Vxc Cf
0 = Vxc
Amplifier input common mode (Vxc) is independent of

Input common mode (Vic)
Output common mode (Voc)
B. Murmann
EE315B - Chapter 5
87
Flip-Around T/H
[W. Yang et al., "A 3-V 340-mW 14-b 75-MSample/s CMOS ADC With 85-dB SFDR at
Nyquist Input", IEEE J. Solid-State Circuits, pp. 1931-1936, Dec. 2001]
Sampling caps are "flipped around" OTA and used as feedback

capacitors during 2
Main advantage: improved feedback factor (lower noise, higher speed)
Main disadvantage: OTA is subjected to input common mode variations
B. Murmann
EE315B - Chapter 5
88
Sampling Network Design Considerations

1-
M1-
1+
1
M1
M1- switches only needed to

set common mode; M1 is
actual sampling switch
Make M1 larger than M1-
Ideally turn off M1- before M1

In practice, usually OK to
turn off simultaneously
In track mode, the total path

resistance is R(M3) plus
bottom plate switch resistance
Since R(M3) is signal
dependent, make its
resistance small compared
to that of bottom plate
network
1+
1-
M1[Lin, Kim and Gray, JSSC 4/1991]
11
1+
2
B. Murmann
EE315B - Chapter 5
89
Schematic Entry and Layout of M1
Use antiparallel devices to implement M1

Needed in simulation to guarantee circuit symmetry
E.g. BSIM model is not necessarily perfectly symmetric with
respect to drain/source!
Needed in layout to ensure symmetry in presence of

drain/source asymmetry due to processing artifacts
B. Murmann
EE315B - Chapter 5
90
What Ultimately Limits Linearity?
Track mode nonlinearity due to R=f(Vin)

Mitigate using clock bootstrapping and proper partitioning of
total path resistance
Eventually, bootstrapping falls apart high frequencies, due to
parasitics capacitances inside the bootstrap circuit
Mismatch in half-circuit charge injection due to R=f(Vin)

Bottom plate switches in the two half circuits see input
dependent impedance; this creates input dependent charge
injection mismatch
Bootstrapping helps; ultimately limited by backgate effect
This effect is often fairly independent of frequency
(somewhat dependent on realization of top plate switch)
In high performance designs, can achieve ~80-100dB linearity

up to a few hundreds of MHz
EE315B - Chapter 5
B. Murmann
91
Capacitors
Metal-Insulator-Metal (MIM)
Vertical Parallel Plate (VPP)
[Ng, Trans. Electron Dev., 7/2005]
[Aparicio, JSSC 3/2002]
Typically 1-2 fF/m2 (10-20 fF/m2 for advanced structures)

For 1 fF/m2, a 10 pF capacitor occupies ~100m x 100m
Both MIM and VPP capacitors have good electrical properties

Mostly worry about parasitic caps
Series and parallel resistances are often not a concern
B. Murmann
EE315B - Chapter 5
92
Plate Parasitics
Ideal Capacitor
Symbol
Typical Integrated
Circuit Capacitor
Node n1 is usually the "physical" top plate of the capacitor

Makes nomenclature very confusing, since this plate is
typically used as the "electrical" bottom plate in a sampling
circuit (in the context of "bottom plate sampling")
Typical values for a MIM capacitor

=1%, =10%
B. Murmann
EE315B - Chapter 5
93
Proper Connection of Capacitors
Fat plate is oriented away from virtual ground nodes to avoid

reduction of feedback factor and reduce potential noise coupling
B. Murmann
EE315B - Chapter 5
94
Outline
Advanced techniques
Clock bootstrapping

B. Murmann
EE315B - Chapter 5
95
Settling and Noise Analysis
Cf
1
Cs
Gm
CL
B. Murmann
EE315B - Chapter 5
96
First Order Amplifier Model
Piecewise linear half-circuit

x
xd
od
2
x
gm v x
for
io =
ID sign ( v x )
gm v x < ID
else
EE315B - Chapter 5
B. Murmann
97
Linear Settling (Small Input Step)
Vofinal
t=0
-Vistep
v o (t) = Vofinal 1 e t /
(ignoring feedforward zero)
Important parameter: Return factor or "feedback factor"

=
B. Murmann
Cf
Cf + Cs + Cx
EE315B - Chapter 5
98
Waveform Detail
Static
Error 0
Dynamic
Error d(t)
1
V
V o/V
/V
out o,ideal
out,ideal
0.8
0.6
0.4
0.2
0
0
10
t/
EE315B - Chapter 5
B. Murmann
99
Static Settling Error
Ideal output voltage for t

Vofinal,ideal = Vistep
Actual output voltage (from detailed analysis)

Vofinal = Vistep
Cs T0
Cf 1 + T0
T0 = gmro = avo
Define static settling error

0 =
Cs
Cf
Vofinal Vofinal,ideal
Vofinal,ideal
T
1
1
1
1
T
+
=
=
1
1+ T
T
Example: T0=1000 0.1% static settling error
B. Murmann
EE315B - Chapter 5
100
Dynamic Settling Error
t /
Vofinal
v o (t) Vofinal Vofinal 1 e
dynamic (t) =
=
= e t /
Vofinal
Vofinal
N=
ts
= ln ( d )
dynamic
1%
4.6
0.1%
6.9
0.01%
9.2
EE315B - Chapter 5
B. Murmann
101
Time Constant
Cf
C f + Cs + C x
R=
1
gm
CLtot = CL + (1 ) Cf
B. Murmann
1 CLtot
gm
EE315B - Chapter 5
102
Transconductor Current
0
Vofinal
t=0
-Vistep
v o (t) = Vofinal 1 e t /
During linear settling, the current delivered by the transconductor is

io CLtot
dv o (t)
V
= CLtot ofinal e t /
dt
Peak current occurs at t=0

io max = CLtot
Vofinal
EE315B - Chapter 5
B. Murmann
103
Slewing
The amplifier can deliver a maximum current of ID

If |io|max > ID, slewing occurs
io max = CLtot
CLeff
Vofinal
>I
1 CLtot D
gm
Vofinal
> ID
gm
1
>
Vofinal
ID
Example: =0.5, Vofinal=0.5V gm/ID > 4 S/A will result in slewing
Very hard to avoid slewing, unless

We are willing to bias at very low gm/ID (power inefficient)
Feedback factor is small (large closed-loop gain, CS/Cf)
Output voltage swing is small
B. Murmann
EE315B - Chapter 5
104
Output Waveform with Initial Slewing
v o (t) = Voslew + Volin 1 e (t tslew )/
v o (t) =
ID
t = SR t
CLtot
Continuous derivative in the transition slewinglinear requires

Volin
ID
=
CLtot
Volin =
ID
CLtot
EE315B - Chapter 5
B. Murmann
105
Dynamic Error with Slewing

t slew = ( Vofinal Volin )
Voslew = Vofinal Volin
CLtot
ID
Using the above result, we can now calculate the dynamic error
during the final linear settling portion
For t > t slew :
t t slew ) /
v o (t) = Voslew + Volin 1 e (
( t t slew ) /
Vofinal
v o (t) Vfinal Voslew + Volin 1 e
d (t) =
=
Vfinal
Vofinal
d (t) =
B. Murmann
Volin ( t t slew ) /
e
Vofinal
EE315B - Chapter 5
106
Noise Analysis
Cf
1
Cs
Gm
CL
Noise due to switches
Noise due to amplifier and switches
EE315B - Chapter 5
B. Murmann
107
Tracking Phase (1)
Variable of interest is total integrated

"noise charge" at node X, qx2
Cumbersome to compute using

standard analysis
Find transfer function from each
noise source (3 resistors) to qx
Integrate magnitude squared
expressions from zero to infinity
and add
Much easier
Use equipartition theorem
C
V
B. Murmann
EE315B - Chapter 5
108
Tracking Phase Noise Charge
Energy stored at node X is

1 q2x
1 q2x
=
2 Ceff 2 Cs + Cf
C
V
Apply equipartition theorem

1 q2x
1
= kT
2 Cs + C f 2
q2x = kT ( Cs + Cf )
Note that any additional parasitic

capacitance at node X will
increase the sampled noise
charge!
EE315B - Chapter 5
B. Murmann
109
Redistribution Phase Noise

4kTRon f
>1 excess noise factor
1
R
Gm
4kTRon1f
4kT
f
Gm
In a proper design, Ron1 and Ron2 will be much smaller than

1/Gm, else the switches would significantly affect the dynamics,
which would be very wasteful
It is much easier to design switches with low on-resistance
than an amplifier with very large Gm
B. Murmann
EE315B - Chapter 5
110
Output Referred Noise Comparison

2
Ron1 noise referred to vo
C
2
N1 = 4kTRon1f s H( j)
Cf
Ron2 noise referred to vo
N2 = 4kTRon2 f H( j)
Amplifier noise referred to vo
Na =
C
4kT
2
f 1 + s H( j)
Gm
Cf
Cs
1 +
Cf
Na

=
>> 1
N1 GmRon1 C 2
s
Cf
Na
Cs
=
1+
>> 1
N2 GmRon2
Cf
Amplifier noise dominates over noise due to Ron1, Ron2

EE315B - Chapter 5
B. Murmann
111
Total Integrated Amplifier Noise
1
Gm
4kTGm f
v o2
1
1
= 4kT
R
f
R
j CLtot
v o2
B. Murmann
1
R
= 4kT
f
R
1 + j RCLtot
0
EE315B - Chapter 5
df =
1 kT
CLtot
112
Adding up the Noise Contributions

2
1
q2x = kT ( Cs + Cf )
2
v o,1
=
Cs + C f
kT
=
C2
C2f
f
q2x
kT Cs
=
1 +
C
f
Cf
2
v o,tot
=
2
v o,2

1 kT
CLtot
kT Cs
1 kT
1 +
+
Cf
Cf
CLtot
EE315B - Chapter 5
B. Murmann
113
Noise in Differential Circuits
In differential circuits, the noise power is doubled (because there

are two half circuits contributing to the noise)
But, the signal power increases by 4x

Looks like a 3dB win?
DRsingle
V 2
o
kT
C
DRdiff
2V )
(
kT
C
=2
V o2
kT
C
Yes, theres a 3dB win in DR, but it comes at twice the power
dissipation (due to two half circuits)
Can get the same DR/power in a single ended circuit by

doubling all cap sizes and gm
B. Murmann
EE315B - Chapter 5
114
Outline
Advanced techniques
Clock bootstrapping

B. Murmann
EE315B - Chapter 5
115
SC Noise Simulation
There are at least three ways to simulate noise in switched

capacitor circuits
Basic .ac/.noise Spice simulations

Simulate noise in each clock phase separately
Activate 1 switches, run .noise and integrate noise
charge at relevant node over all frequencies and refer to
output
Activate 2 switches, run .noise and integrate noise over
all frequencies at the output
Sum integrated noise from the two phases
This is analogous to the way we carried out the hand
analysis
B. Murmann
EE315B - Chapter 5
116
SC Noise Simulation
Periodic Steady State Simulation

First run PSS analysis to find the periodic operating point
Analogous to .op for .ac/noise
Next run PNOISE analysis
Computes total noise, taking all clock phases, noise
aliasing, noise correlations, etc. into account
Transient Noise
Direct simulation of all noise sources using a transient
simulation
Most physical way of simulating noise
EE315B - Chapter 5
B. Murmann
117
Example T/H Circuit
Switches sized 5 times faster, i.e. N = 510 = 50

Clp
cl
Clm
cl
vic
vic
Cfm
cf
p1
cs
Csp
cs
Csm
p1
voc
vdd
vocs
Cfp
cf
vic
OTA Gm chosen such than (Ts/2) / OTA = 10
vic
fs = 100 MHz, = 2, = 1
Cs = Cf = 100 fF, CL = 500 fF, Cpar 0
B. Murmann
EE315B - Chapter 5
118
.Noise Simulation (f1)
*** Compute noise charge at node X and refer to output via Cf
en vno 0 vcvs vol =( cs*v(x,s) + cf*v(x,f) )/cf
.ac dec 100 100 1000Gig
.noise v(vno) vdummy
(showing half circuit for simplicity)
EE315B - Chapter 5
B. Murmann
119
PSD [V2/Hz]
.Noise Simulation (1)

10
10
10
-15
-20
-25
Integral [uVrms]
10
B. Murmann
x 10
10
10
10
Frequency [Hz]
10
10
10
12
-4
406Vrms
4
2
0 2
10
10
10
10
Frequency [Hz]
EE315B - Chapter 5
10
10
10
12
120
PSD [V2/Hz]
.Noise Simulation (2)

10
10
10
-15
-20
-25
Integral [uVrms]
10
x 10
10
10
10
Frequency [Hz]
10
10
10
12
-4
266Vrms
2
0 2
10
10
2
v out,tot
=
B. Murmann
10
10
Frequency [Hz]
10
10
( 406Vrms )2 + ( 266Vrms )2
10
12
= 485Vrms
EE315B - Chapter 5
121
PSS Simulation Setup
Set tstab if your circuit needs time

to reach steady state
(e.g. clock bootstrap circuits)
Under options set maxacfreq to
the highest frequency from which
you expect noise to fold down
B. Murmann
EE315B - Chapter 5
122
PSS Waveforms (Clocks)

4.75ns
B. Murmann
EE315B - Chapter 5
123
PNOISE Simulation Setup

Numsidebands fmax/fs, where fmax is the
maximum frequency from which you expect
significant noise folding
timedomain means simulator computes

spectrum of discrete time noise samples at
the specified sampling instant
Sampling instant (4.75ns in this example)
B. Murmann
EE315B - Chapter 5
124
How to Chose Parameters

Maxacfreq must be set commensurate with the speed
of the switches
Common pitfall: Use nice 1m switches must
consider noise from DC to daylight
In our example
Maxacfreq 10
1
10
=
N fs
2RonC
Maxacfreq 3 50 100MHz = 15GHz
EE315B - Chapter 5
B. Murmann
125
How to Chose Parameters

In traditional PSS/PNOISE simulators (such as
SpectreRF), simulation time increases rapidly for large
values of Numsidebands
Berkeley Design Automation (BDA) Analog FastSPICE
(AFS) automatically includes an infinite number of
sidebands at significantly reduced simulation times
Numsidebands
B. Murmann
Maxacfreq 15GHz
=
= 150
fs
100MHz
EE315B - Chapter 5
126
PNOISE Result (SpectreRF)

Sampled Noise PSD
Integrated Noise
EE315B - Chapter 5
B. Murmann
127
Transient Noise Using BDA AFS
Simulated 1000 samples
Only one critical setting: noisefmax = 15 GHz
B. Murmann
EE315B - Chapter 5
128
Comparison
Very good agreement between hand calculation and all three

simulation approaches
Method
1 Noise 2 Noise Total

Comment
[Vrms] [Vrms] [Vrms]
Calculation
406
245
474
.NOISE
406
266
485
PNOISE
475
Somewhat smaller than .NOISE

due to finite Maxacfreq and
Numsidebands
TRAN
NOISE
477
Somewhat smaller than .NOISE

due to statistical fluctuations (can
use more samples)
B. Murmann
Neglected switch noise during 2

smaller value than .NOISE
EE315B - Chapter 5
129
Advantages of TRAN NOISE

Takes advantage of rapid advancements in very fast,
spice-accurate transient simulators (such as BDA AFS)
Can handle much larger circuits compared to
PSS/PNOISE (PSS tends to have convergence issues
for large circuits)
Intuitive inspection of waveforms
No need to combine noises manually
Compared to .ac/.noise simulation flow
Applicable also to non-periodic circuits
B. Murmann
EE315B - Chapter 5
130
Further Reading on Noise
Basics
B. Murmann, "Thermal Noise in Track-and-Hold Circuits: Analysis and
Simulation Techniques," IEEE Solid-State Circuits Magazine, vol. 4, no.
2, pp. 46-54, June 2012.
SC integrator noise analysis

R. Schreier, J. Silva, J. Steensgaard, and G.C. Temes, "Designoriented estimation of thermal noise in switched-capacitor circuits,"
IEEE TCAS1, vol.52, no.11, pp. 2358-2368, Nov. 2005.
Total integrated noise expressions for more complex OTAs

(two-stage, etc.)
A. Dastgheib and B. Murmann, Calculation of Total Integrated Noise in
Analog Circuits, IEEE TCAS1, vol. 55, no. 10, pp. 2988-2993, Oct.
2008.
Transient noise simulation example for a complete ADC

Lennart Mathe and David C. Lee, "Analog-to-Digital Converter
Performance Signoff with Analog FastSPICE Transient Noise at
Qualcomm," Berkeley Design Automation, www.berkeley-da.com.
B. Murmann
EE315B - Chapter 5
131
Summary Sampling Circuits
Three predominant implementation styles

Purely passive
Source follower T/H, up to ~9-10bit accuracy
Charge redistribution or flip-around architecture
In a typical, properly designed circuit only the most fundamental

issues are significant
Jitter, kT/C noise
Charge injection is not a problem if properly handled

E.g., use bottom plate sampling
B. Murmann
EE315B - Chapter 5
132
Voltage Comparators
Boris Murmann
Stanford University
B. Murmann
EE315B - Chapter 6
Recap
Ultimately, building a quantizer requires circuit elements that

"make decisions"
The most widely used "decision circuit" is a voltage comparator
B. Murmann
EE315B - Chapter 6
Preview - Flash ADC
2B-1
Dout
Vin
2B-1 Decision Levels
B. Murmann
EE315B - Chapter 6
Ideal Voltage Comparator
Function
Compare the instantaneous values of two analog voltages
(e.g. an input signal and a reference voltage) and generate a
digital 1 or 0 indicating the polarity of that difference
We essentially want to implement infinite gain
B. Murmann
EE315B - Chapter 6
How to Implement High Gain?
Considerations
Amplification need not be linear
Amplification need not be continuous time if comparator is
used in a sampled data system
Clock signal will tell comparator when to make a decision
We will focus on this case, since it is the predominant scenario
Implementation options to be looked at

Multi-stage amplification
E.g. cascade of resistively loaded differential pairs
Regenerative latch using positive feedback

E.g. cross coupled inverters
EE315B - Chapter 6
B. Murmann
Input Signal Scenarios

Sampled Data Input
Our focus
Continuous Time Input
B. Murmann
EE315B - Chapter 6
Cascade of Open-Loop Amplifiers
In each stage:
u =
gm
const.
Cgs
A 0 = gmR =
u
0
0 =
1
= u
RC A 0
What is the best choice for a given gain requirement?

Single stage with lots of voltage gain (OTA)?
Only a few stages with moderate gain?
Lots of stages with low gain?
EE315B - Chapter 6
B. Murmann
Step Response (1)
When the input is a sampled data signal, we want to achieve the

fastest possible amplification in response to an input step
Vout (s) = Vin (s)A(s) =
Vistep
s
Av
N
(1+ s A )
1/N
u =
1
u
A v = AN
0
1/N

1/N N 1 A
Vout (t) = Vistep A v 1 e u A v u v
i!
i=0
A(t)
B. Murmann
EE315B - Chapter 6
Step Response (2)

10
Step response A(t)
d(N=3)
10(1-e-1)
6
4
N=1
N=3
N=5
Av*(1-e-1)
0
0
10
15
Time t/u
Three stage amplifier wins! (for Av=10)

EE315B - Chapter 6
B. Murmann
Delay versus Number of Stages
A( d ) = A v 1 e1
(numerically)
50
Av=10
Av=100
Av=1000
Delay time d/u
40
30
20
10
6
8
10
N (Number of stages)
12
14
Using a single stage is a non-starter

The optimum number of stages show a shallow minimum
B. Murmann
EE315B - Chapter 6
10
Optimum Number of Stages
Optimum number of stages
12
Numerical result
ln(Av)
10
8
6
4
2
0 0
10
10
10
10
Av (Total gain)
10
10
Nopt ln(A v )
EE315B - Chapter 6
B. Murmann
11
Optimum Gain per Stage

A0,opt (Optimum gain per stage)
4
3.5
3
2.5
2
1.5
1 0
10
Nopt ln(A v )
B. Murmann
Numerical result
e
4.5
10
10
10
Av (Total gain)
Nopt
A v = A 0,opt
EE315B - Chapter 6
10
Nopt
10
A 0,opt e
12
Cascade of "Integrators" (1)
Intuition
Load resistors (in a cascade of open-loop amplifiers) shunt
current away from load capacitance; this slows down
amplification
Analysis using integrator stages
g
v o1 = m vin = u vin
sC
s
v oN =
N
u
sN
vin
EE315B - Chapter 6
B. Murmann
13
Cascade of Integrators (2)

Vistep N
u
N
s s
Vout (t) = Vistep N

u
10
10
8
Step response A(t)
Step response A(t)
Vout (s) =
4
N=1 (with R)
N=1 (integ)
4
N=3 (with R)
N=3 (integ)
Av*(1-e-1)
0
0
B. Murmann
10
Time t/u
tN
N!
Av*(1-e-1)
15
0
0
EE315B - Chapter 6
10
Time t/u
15
14
Cascade of Integrators (3)
Cascade of integrators achieves faster amplification than

cascade of resistively loaded stages
Delay time
1/N
d = u [(N! A( d ))]
A( d ) =
Vout ( d )
Vinstep
Optimum number of stages approximately given by

Nopt = 1.1ln [ A( d )] + 0.79
[Wu, JSSC 12/1988]
Effective gain per stage is still relatively close to e=2.7183
EE315B - Chapter 6
B. Murmann
15
Regenerative Sense Amplifier (Latch)

Conceptual Circuit

t=0
Inverter
Transconductance

[Figueiredo]

/

1: Set up initial condition (vOD0)

2: Enable positive feedback
B. Murmann
EE315B - Chapter 6

/

16

Simulation Example
I4
I7
vdm
vp
ideal_balun
vcm
vm
vic
vip
vop
vim
vom
vp
vdm
ideal_balun
vm
vcm
p2!
vid
M0
10u/0.18u
vod
voc
M4
10u/0.18u
p1!
vip
C3
cl
vid
vdc:vdd vdc:vic
M1
5u/0.18u
M3
5u/0.18u
C0
cl
V5
vdc:vid
B. Murmann
CL=100 fF
p2!
vic
V2
vop
p1
p1b
p1e
p1eb
p2
p2b
p2e
clock_gen
p1!
vdd
V0
vim
vom
p1!
p2!
EE315B - Chapter 6
17
Transient Response
Nodes vOP and vON for differential inputs of 1mV, 1V, 1nV and 1pV
2 goes high
B. Murmann
EE315B - Chapter 6
18
Transient Response
Differential Mode
Common Mode
EE315B - Chapter 6
B. Murmann
19
Transient Response (Log Scale)
=
/
=
/
B. Murmann
600ps
600ps

=
=
= 43

ln 10
6 2.3
ln

EE315B - Chapter 6
20
Comparison
10
Amplification A(t)
4
N=3 (with R)
N=3 (integ)
Latch
Av*(1-e-1)
0
0
t/ t/
Time
10
15
Latch is much faster than cascade of amplifiers/integrators
EE315B - Chapter 6
B. Murmann
21
Latch Gain
B. Murmann
A(d)
d/
10
2.3
100
4.6
1,000
6.9
10,000
9.2
EE315B - Chapter 6
22
"The" Architecture
Why bother using pre-amplification (Av)?
Pre-amplification may be used for several reasons

Lower offset (latch offset tends to be large)
Lower thermal noise
Attenuation of "kickback noise
Separate input from output
Common mode rejection
B. Murmann
EE315B - Chapter 6
23
Basic Topology Examples (1)
Class-A (constant current)
Figure from [Figueiredo, TCAS2, July 2006]

B. Murmann
EE315B - Chapter 6
24
Class-AB
Input pair has constant bias, but latch draws current only
while making a decision
B. Murmann
EE315B - Chapter 6
25
Fully dynamic
Draws (significant) current only while making a decision
Original paper:
[Kobayashi, JSSC, April 1993]
B. Murmann
EE315B - Chapter 6
26
Offset in Latches
vOP
vON
VOS,static
Inverter switching voltage: VS

Initial common mode: VOC0
CLP
CLN
Load capacitor mismatch: CL
Static offset (VOS) due to transistor mismatch (primarily Vt)
Dynamic offset due to load capacitor mismatch

For an analysis see Nikoozadeh, TCAS II, Dec. 2006
,
1

2
Example:
1 10
0.3 0.9 = 30
2 100
Minimize, if possible
EE315B - Chapter 6
B. Murmann
27
Input Referred Offset with Preamplification
2VOS = 2VOS1 +
1
A 2v
2VOS2
Example: VOS1=3mV, VOS2=30mV, Av=10

VOS =
B. Murmann
( 3mV )2 +
1
10
( 30mV )2
EE315B - Chapter 6
= 4.2mV
28
Dealing With Offset
Options depend strongly on the application context
May use an ADC architecture that inherently tolerates offset

More later
May use a relatively coarse tuning mechanism

E.g. tune offset to lie within LSB at the 6-bit level
May have to apply precision techniques

Autozeroing, etc. (see EE315A), for high-end instrumentation
Simply use large transistors

Usually not power efficient
B. Murmann
EE315B - Chapter 6
29
Example: Coarse Offset Tuning
The circuit below uses capacitve imbalance to tune the offset

An example of using a "bug" as a feature
See e.g. Van der Plas, ISSCC 2006
Watch out for PSRR degradation due to circuit imbalance
B. Murmann
EE315B - Chapter 6
30
Example: Precision Comparator

[http://www.elecdesign.com/Articles/Index.cfm?ArticleID=3956]
Used in the AD7671, 16-bit, 1-MS/s successive approximation ADC

Uses cascaded output series offset cancellation (see EE315A)
B. Murmann
EE315B - Chapter 6
31
Electronic Noise in Latch Comparators
Class-A topology
Integrate noise at regenerative node over all frequencies and
refer to the input kT/CL
See Opris, Electronics Letters, July 1997
Dynamic topology
Very hard to analyze due to the time variant behavior
See Nuzzo, IEEE TCAS1, July 2008
Fortunately, the end result is not too far from class-A-like
noise estimation in the decision point kT/CL
B. Murmann
EE315B - Chapter 6
32
Kickback
E.g. Flash ADC

resistor ladder
EE315B - Chapter 6
B. Murmann
33
Kickback Mitigation Techniques

Neutralization
Neutralization and Switch Isolation
Figures from [Figueiredo, TCAS2, July 2006]
B. Murmann
EE315B - Chapter 6
34
Kickback Mitigation Techniques
Extra switches act

similar to cascodes
[Sundstrom, MIXDES 2007]
B. Murmann
EE315B - Chapter 6
35
Other Design Considerations
Input capacitance and linearity of input capacitance

Consider e.g. loading of the input network in a flash ADC
Overdrive recovery
Consider a very large input, followed by a very small input of
opposite polarity
Metastability
What if the input is so small that the comparator cannot
decide in the given amount of time?
B. Murmann
EE315B - Chapter 6
36
Overdrive Recovery Test

Latch
activation
1
Comparator
input
Preamp
output
[Razavi, p. 183]
The test is passed when the latch output changes polarity in

case 1, but not in case 2
EE315B - Chapter 6
B. Murmann
37
Metastability (1)
The plot below shows the time needed to regenerate a given

differential latch input to full logic swing
35
30
treg/
25
= ln
20
15
10
5
0 -15
10
10
-10
10
-5
10
vOD0/V/VDD
B. Murmann
EE315B - Chapter 6
38
Metastability (2)
What if the input is so small that we run out of time?
This is called a metastable state, i.e. at the end of the clock

cycle there is no clear logic decision
The following logic gates will go into an erratic state that is hard
to predict, and so we classify this outcome as an error
In the following analysis we will try to estimate the probability at

which metastable states will occur
Application requirements vary widely

Error correction coding in certain communication sytems can
absorb probabilities of up to 10-3
High-speed wireline links (without coding) or instrumentation
equipment may require probabilities better than 10-18
EE315B - Chapter 6
B. Murmann
39
Metastability (3)
Assume that the maximum available time is tmax
The minum latch input we can regenerate is then

, =
In case, we have a preamp, we can do slightly better

, =

/
1

/
Now assume that the input is uniformly distributed over some

range, and compute the probabily of seeing a metastable state
PDF

,

/
=
=
, ,
-vID0,max
+vID0,max
-vID0,min +vID0,min
B. Murmann
EE315B - Chapter 6
40
Mean Time to Failure
In some applications it makes sense to think about metastability

in terms of the mean time between metastable events
Mean time to failure (MTF)
MTF [days]
10
=
1

10
fs=10GHz
fs=100MHz
-2
10
-18
10
-16
10
-14
10
-12
10
-10
10
Pmeta
EE315B - Chapter 6
B. Murmann
41
Example: Flash ADC
In a flash ADC, there is always (exactly) one comparator that

sees an input within /2
The probability of seeing a metastable state within the array is

therefore found using vID0,max = /2

=

/

/
=
1 V

2 2
2
The graph on the next slide plots an example

Sampling rate fs = 1GHz, tmax 1/2fs = 500ps
B = 6, Av =1, VFS = VDD
B. Murmann
EE315B - Chapter 6
42
Example: Flash ADC

0
10
-5
Pmeta
10
-10
10
-15
10
-20
10
10
20
30
40
50
[ps]
60
70
80
The curve is incredibly steep in the region of Pmeta < 10-5

Very strong incentive to minimize as much as possible
Ultimitaley bounded by Cgg/gm = 1/T
B. Murmann
EE315B - Chapter 6
43
Metastability Detectors?
Why not build a clever logic circuit that detects the metastable
state and overrides it with a valid logic state?
This sounds good, but is difficult to do in reality
Logic gates look like "low pass filters" compared to a latch
A latch provides the maximum possible gain per unit of time;
any gate delay added after the latch (that eats into the
regeneration time) is counterproductive
Most (if not all) metastability detector ideas
proposed in literature tend to create new
metastable states
Related article: R. Ginosar, "Fourteen ways to
fool your synchronizer," 2003
B. Murmann
EE315B - Chapter 6
44
Comparator Examples (1)
Mehr & Dalton, JSSC 7/1999
B. Murmann
EE315B - Chapter 6
45
Mehr & Dalton, JSSC 7/1999
B. Murmann
EE315B - Chapter 6
46
Schinkel, ISSCC 2007: "Double tail sense amplifier"
EE315B - Chapter 6
B. Murmann
47
Miyahara, ASSCC 2009
B. Murmann
EE315B - Chapter 6
48
Lee, JSSC, April 2011
EE315B - Chapter 6
B. Murmann
49
Tripathi, ESSCIRC 2013
6ps
B. Murmann
EE315B - Chapter 6
50
Selected References (1)

1. Y. S. Yee, L. M. Terman and L. G. Heller, A 1mV MOS Comparator, IEEE J. of
Solid-State Circuits, vol. SC-13, pp. 294-297, June 1978.
2. A. Yukawa, A CMOS 8-Bit High-Speed A/D Converter IC, IEEE J. of Solid-State
Circuits, vol. SC-20, pp. 775-779, June 1985.
3. B. J. McCarroll, C. G. Sodini, and H.-S. Lee, A High-Speed CMOS Comparator for
Use in an ADC, IEEE J. of Solid-State Circuits, vol. 23, pp. 159-165, Feb. 1988.
4. J.-T. Wu and B. A. Wooley, A 100-MHz Pipelined CMOS Comparator, IEEE J. SolidState Circuits, vol. 23, pp. 1379-1385, Dec. 1988.
5. B. Razavi and B. A. Wooley, A 12-b 5-MSample/s Two-Step CMOS A/D Converter,
IEEE J. Solid-State Circuits, vol. 27, pp. 1667-1678, Dec. 1992.
6. B. Razavi and B. A. Wooley, Design Techniques for High-Speed, High-Resolution
Comparators, IEEE J. Solid-State Circuits, vol. 27, pp. 1916-1926, Dec. 1992.
9.
M. Choi and A. A. Abidi, "A 6-b 1.3-GSample/s A/D converter in 0.35-m CMOS,"
IEEE J. Solid-State Circuits, pp. 1847-1858, Dec. 2001.
10. I. Mehr and D. Dalton, "A 500-MSample/s, 6-bit Nyquist-rate ADC for disk-drive readchannel applications," IEEE J. Solid-State Circuits, pp. 912-920, July 1999.
11. I. Mehr and L. Singer, A 55-mW, 10-bit, 40-MSample/s Nyquist-Rate CMOS ADC,
IEEE J. Solid-State Circuits, pp. 318-25, March 2000.
12. G.R. Couranz, and D.F. Wann, "Theoretical and Experimental Behavior of
Synchronizers Operating in the Metastable Region," IEEE Trans. Computers, vol. C24, no.6, pp. 604-616, June 1975.
B. Murmann
EE315B - Chapter 6
51

9.
H. J. M. Veendrick, The Behavior of Flip-Flops Used as Synchronizers and

Prediction of Their Failure Rate, IEEE J. Solid-State Circuits, April 1980.
10. B. Zojer, et al., A 6-bit/200-MHz full Nyquist A/D converter, IEEE J. Solid-State
Circuits, vol. SC-20, pp. 780-786, June 1985.
11. K.-L.J. Wong and C.-K.K. Yang, "Offset compensation in comparators with minimum
input-referred supply noise," IEEE J. Solid-State Circuits, pp. 837-840, May 2004.
12. A. Graupner, "A Methodology for the Offset-Simulation of Comparators,"
http://www.designers-guide.org/Analysis/comparator.pdf.
13. D. Schinkel et al., "A Double-Tail Latch-Type Voltage Sense Amplifier with 18ps
Setup+Hold Time," ISSCC Dig. Techn. Papers, pp. 314-315, 2007.
14. P.P. Nuzzo, et al., "Noise Analysis of Regenerative Comparators for Reconfigurable
ADC Architectures, IEEE Trans. Circuits Syst. I, pp.1441-1454, July 2008.
15. B.S. Leibowitz, et al., "Characterization of Random Decision Errors in Clocked
Comparators," Proc. IEEE CICC, pp.691-694, Sep. 2008.
16. Analog Devices, "Find Those Elusive ADC Sparkle Codes and Metastable States"
http://www.analog.com/en/content/0,2886,760%255F788%255F91218,00.html
19. M. Miyahara and A. Matsuzawa, "A low-offset latched comparator using zero-static
power dynamic offset cancellation technique," ASSC, pp. 233-236, Nov. 2008.
20. P.M. Figueiredo, "Comparator Metastability in the Presence of Noise," IEEE TCAS1,
vol. 60, no. 5, pp.1286-1299, May 2013.
B. Murmann
EE315B - Chapter 6
52
Flash ADCs
Boris Murmann
Stanford University
EE315B - Chapter 7
B. Murmann
Nyquist ADC Architectures
Nyquist rate
Word-at-a-time
Flash ADC
Instantaneous comparison with 2B-1 reference levels
Multi-step
E.g. pipeline ADCs
Coarse conversion, followed by fine conversion of residuum
Bit-at-a-time
E.g. successive approximation ADCs
Conversion via a binary search algorithm
SPEED
Level-at-a-time
E.g. single or dual slope ADCs
Input is converted by measuring the time it takes to
charge/discharge a capacitor from/to input voltage
B. Murmann
EE315B - Chapter 7
ADC Survey (ISSCC & VLSI 1997-2013)

10
Flash
Folding
Two-Step
Pipeline
10
10
SAR
Other
1000fs
in,hf
[Hz]
100fs
10
10
rms
rms
Jitter
Jitter
20
30
40
50
60
70
80
SNDR [dB]
90
100
110
120
hf
EE315B - Chapter 7
B. Murmann
Flash ADC Speed

7
Clock Speed [GHz]
Flash ADCs
Microprocessors
5
4
3
2
1
0
1996
1998
2000
2002
2004
2006
2008
Year
B. Murmann
EE315B - Chapter 7
Modern Flash ADC Architecture

Vadc
Vadc
Dout
T/H
Fast
Speed limited only by comparator
decision
High complexity, large input capacitance
Resolution tends to limited to 6 bits
B. Murmann
Wallace Encoder
CLK
Vin
Usually implemented as a fully

differential circuit
EE315B - Chapter 7
Why Use a T/H?
One reason: Timing offsets between comparators may cause

bubbles in the thermometer code
Single bubble
Double
bubble
Fast
moving
input
K. Uyttenhove, M.S.J. Steyaert, "A 1.8-V 6-bit 1.3-GHz flash ADC in 0.25-m CMOS," JSSC, pp. 1115-1122, July 2003.
B. Murmann
EE315B - Chapter 7
Bubble Problem in a Poor Mans Encoder
e.g. due to
timing offset
between two
comparators
Correct output: 1000, actual output: 1110 (!)
B. Murmann
EE315B - Chapter 7
Bubble Tolerant Encoder
Protects against single bubbles
Reference: C. W. Mangelsdorf, A 400-MHz Flash Converter with Error

Correction, IEEE J. Solid-State Circuits, pp. 184-191, Feb. 1990.
B. Murmann
EE315B - Chapter 7
Modern Solution: Wallace Encoder
B. Murmann
Simply a ones adder
F. Kaess, et al., New encoding

scheme for high-speed flash
ADCs, ISCAS 1997, pp. 58
Complexity used to be an issue in

older technology; not a problem in
sub-100nm CMOS
Number of adders needed for

N bits:
EE315B - Chapter 7
Alternative (or Additional) Solution
Still want to use a T/H, even if bubbles are not a problem

Simplified comparator design
Well-behaved input impedance
T/H need not be all that complex
Source follower T/H
Y. Tamba, and K. Yamakido, "A CMOS 6 b 500 MS/s ADC for a hard disk drive read channel, ISSCC 1999, pp. 324-325
B. Murmann
EE315B - Chapter 7
10
How about Metastability?
Different gates interpret metastable output X differently

Correct output: 0111 or 1000, actual output: 1111 (!)
Again, a simple encoder does not handle this very well
B. Murmann
EE315B - Chapter 7
11
Solution 1: Wallace Encoder
B. Murmann
Similar reasoning as with

bubbles
Error due to metastable

input is bounded by 1LSB
EE315B - Chapter 7
12
Solution 2: Latch Pipelining
Use additional latches to create extra gain before generating

decoder signals
Usually too power hungry and area inefficient
EE315B - Chapter 7
B. Murmann
13
Solution 2: Gray Encoding

Thermometer Code
Gray
T1
T2
T3
T4
T5
T6
T7
G3
G2
G1
B3
B2
B1
G3
B. Murmann
Binary
G2
G1
EE315B - Chapter 7
Each Ti affects only one Gi

Avoids disagreement
of interpretation by
multiple gates
14
Efficient Implementation
Reference
C. Portmann and T. Meng, Power-Efficient Metastability Error
Reduction in CMOS Flash A/D Converters, IEEE J. Solid-State
Circuits, pp. 1132-40 , Aug. 1996.
EE315B - Chapter 7
B. Murmann
15
Offset
Typically want offset of each comparator <1/4LSB with high

probability (3 sigma)
If we budget half of the input referred offset for the latch, the
other half for the pre-amp, this means pre-amp offset must
be <1/4LSB / sqrt(2)
3VOS = 3
AVT
WL
<
1 FSR
4 2 2B
E.g. 6-bit flash ADC, FSR=1V

3
AVT
1 1V
<
= 2.8mV
WL 4 2 26
Huge!
2
2
2
A
3
18.4m
3 4mV m
2
VT
WL >
= 102m
= 18.4m W >
=
0.18m
2.8mV
2.8mV
B. Murmann
EE315B - Chapter 7
16
Re-cap of Options
Simply use large devices

For each extra bit, need to increase width by 4x, also need
to double number of comparators
Assuming constant current density, this means each
additional bit costs 8x in power!
Dynamic offset cancellation (autozeroing, etc.)

Tends to cost speed
More commonly used

Offset averaging
Offset calibration
Incorporation of redundant comparators
B. Murmann
EE315B - Chapter 7
17
Offset Averaging (1)
R2/R1=1.3
[Kattmann & Barrow, ISSCC 1991]

B. Murmann
EE315B - Chapter 7
18
DNL/
Offset Averaging (2)
[Bult & Buchwald, JSSC 12/1997]
[Scholtens & Vertregt, JSSC 12/2002]

B. Murmann
EE315B - Chapter 7
19
6-bit Flash ADC with Averaging

[Choi & Abidi, JSSC 12/2001]
Averaging networks designed to

reduce input referred offset by 3x
B. Murmann
EE315B - Chapter 7
20
Offset Calibration Using Global DAC
S. Sutardja, "360 Mb/s (400 MHz) 1.1 W 0.35m CMOS PRML read channels with 6
burst 8-20 over-sampling digital servo," ISSCC Dig. Techn. Papers, Feb. 1999.
EE315B - Chapter 7
B. Murmann
21
Calibration Using one DAC per Comparator
(Reset switches not shown)

8
Voutn
Voutp
Dcalp
CAL
Vlo Vhi
Vrefn
MUX/Encoder
Vinp
Vinn
Vrefp
CAL
Dcaln
Vlo Vhi
[M. El-Chammas, VLSI 2010]
B. Murmann
EE315B - Chapter 7
22
Offset Calibration at Start-Up
Out
RefIn-
INC/
DEC
Start-up calibration is
reasonably stable at flash
ADC resolutions
The correction circuit shown
on the previous slide has a
relatively low temperature
dependence
AVG
Control Bits
Out
Output oscillates between one

and zero as the input-referred
offset converges to zero
Input-referred
Offset
In+
Ref+
Offset = 0
t
B. Murmann
EE315B - Chapter 7
23
Tuning Range Issue (1)
Ideally, the trim-DAC must bring the offset well within one LSB
of the flash ADC
The required trim-DAC resolution can become unacceptably

large if near minimum-size transistors (with large Vt mismatch)
in the latest technology are used
A clever workaround is to utilize the error correction capability of

the Wallace tree encoder (ones counter)
B. Murmann
EE315B - Chapter 7
24
Tuning Range Issue (2)

[Verma, ISSCC 2013]
+
Vref0
Vref0 (eff)
+
Vref1
Vref2
Vref2 (eff)
+
Vref3
Vref1 (eff)
3
Vref3 (eff)
EE315B - Chapter 7
B. Murmann
25
Ones Counting is Equivalent to Reordering

+
Vref0
[Verma, ISSCC 2013]
2
Vref0 (eff)
+
Vref1
Vref2
Vref3
Vref1 (eff)
1
Vref2 (eff)
3
Vref3 (eff)
Reduced tuning range

B. Murmann
EE315B - Chapter 7
26
Comparator Redundancy (1)
Interesting idea: Do not worry about offsets at all, just provide

enough levels that properly add up on average
90mW
900mW
Wallace Encoder
Paulus et al., "A 4GS/s 6b flash ADC in 0.13um CMOS," VLSI Circuits Symposium, 2004
B. Murmann
EE315B - Chapter 7
27
Comparator Redundancy (2)
Another idea: Build a "sea of imprecise comparators", then

determine which ones to use
C. Donovan, M. P Flynn, "A 'digital' 6-bit ADC in 0.25-m CMOS," IEEE J.

Solid-State Circuits, pp. 432-437, March 2002.
B. Murmann
EE315B - Chapter 7
28
Differential Pair Redundancy
Idea: Have multiple input pairs and use the one you like
[Chen, VLSI 2013]
B. Murmann
EE315B - Chapter 7
29
Details
B. Murmann
EE315B - Chapter 7
30
Performance Comparison
[Chen, VLSI 2013]
B. Murmann
EE315B - Chapter 7
31
Calibration Using Reference String
[Park & Flynn, "A 3.5 GS/s 5-b Flash ADC in 90 nm CMOS," CICC 2006]
B. Murmann
EE315B - Chapter 7
32
Details
B. Murmann
EE315B - Chapter 7
33
Measured Results
B. Murmann
EE315B - Chapter 7
34
Reducing Complexity
Even with calibration, flash ADCs tend to hit a wall beyond 6 bits
of resolution
Large calibration range
Large input capacitance
Complex encoder
Several techniques have been developed to ease the pain and

potentially allow going beyond 6 bits (at the cost of some speed)
Interpolation
Folding, folding & interpolation
Subranging
EE315B - Chapter 7
B. Murmann
35
Interpolation
Idea
Interpolation between preamp outputs
Reduces number of preamps

Reduced input capacitance
Reduced area, power dissipation
Same number of latches
Important side-benefit
Decreased sensitivity to preamp offset
Improved DNL
B. Murmann
EE315B - Chapter 7
36
Concept
[van de Plassche, p.118]
B. Murmann
EE315B - Chapter 7
37
Differential Implementation
VA + VB
=0
2
B. Murmann
EE315B - Chapter 7
VA = VB
38
Higher Order Interpolation

Resistors produce
additional levels
Define interpolation factor
as ratio ratio of latches
and preamps
The example shown on
this slide has M=8
[H. Kimura et al, A 10-b 300-MHz

Interpolated-Parallel A/D Converter,
IEEE J. of Solid-State Circuits, pp.
438-446, April 1993.
B. Murmann
EE315B - Chapter 7
39
Potential Issues with Interpolation
Must ensure that "linear range" of adjacent preamplifiers

overlaps
Sets upper bound on preamp gain
Resistor string reduces signal path bandwidth

Sets upper bound on interpolation factor, typically around 4
For interpolation factors >2, amplifier nonlinearity can limit the

precision of zero crossings
See e.g. Van de Plassche, p.121
B. Murmann
EE315B - Chapter 7
40
Folding
MSB
ADC
VIN
Digital
Output
LSB
ADC
Folding Circuit
Main idea: Re-use comparator several times across

the full scale range
B. Murmann
EE315B - Chapter 7
41
Simplest Variant
[Verbruggen, JSSC, March 2009]
Folding circuit
B. Murmann
EE315B - Chapter 7
42
Creating Multiple Folds
Parameter K sets output common mode

EE315B - Chapter 7
B. Murmann
43
Folder Output
Folder Output
0.5
Accurate only at
zero-crossings
Ideal Folder
CMOS Folder
-0.5
0
0.5
1.5
2.5
3.5
Lowdown
0.1
Most folding ADCs

do not actually use
the folds, but only the
zero-crossings!
Error
0.05
0
-0.05
-0.1
0
0.5
B. Murmann
1.5
2
Vin /
2.5
3.5
EE315B - Chapter 7
44
Multiple Folds Using Only Zero Crossings
B. Murmann
Way too complex, need

one folder per decision
in the fine ADC
Idea
Use interpolation to
eliminate some of
the folders
EE315B - Chapter 7
45
Interpolation
Same idea as discussed previously
B. Murmann
EE315B - Chapter 7
46
Complete Folding & Interpolating ADC
B. Murmann
EE315B - Chapter 7
3-bit coarse ADC
3-bit fine ADC

2 folders with 4x
interpolation
8 levels
We have therefore
built a 6-bit ADC
with only 16
comparators
(instead of 63)
47
8-bit, 70 MS/s Folding and Interpolating ADC
[B. Nauta and G. Venes, JSSC. Dec. 1985]
B. Murmann
EE315B - Chapter 7
48
8-bit, 1.6 GS/s Folding and Interpolating ADC
[Taft et al., JSSC 12/2004]
B. Murmann
EE315B - Chapter 7
49
8b, 10GS/s Folding and Interpolating ADC
[Landolt, PRIME 2012]

For Rohde & Schwarz Scope
0.25um SiGe BiCMOS
9 Watts
B. Murmann
EE315B - Chapter 7
50
Performance
EE315B - Chapter 7
B. Murmann
51
Subranging
4-bit Flash ADC
B. Murmann
2+2-bit Subranging ADC
EE315B - Chapter 7
52
8-bit, 1 GS/s Subranging ADC
Coarse ADC selects tap range for fine ADC

[Chung, VLSI 2011]
B. Murmann
EE315B - Chapter 7
53
Summary Flash ADCs & Extensions
Today, flash ADC design boils down to an effective offset

management approach
Calibration and/or redundancy
Extensions such as averaging, folding, interpolation and

subranging can be used to alleviate the tradeoffs
But these techniques havent been popular in recent years
The reason may be the strong competition from pipelined
and time interleaved architectures
More later
B. Murmann
EE315B - Chapter 7
54

Flash ADCs, Offset Averaging
1. K. Kattmann and J. Barrow, "A Technique for Reducing Differential Non-Linearity
Errors in Flash A/D Converters," ISSCC Digest of Technical Papers, pp. 170-171,
Feb. 1991.
2. K. Bult and A. Buchwald, "An embedded 240-mW 10-b 50-MS/s CMOS ADC in 1mm2," IEEE J. of Solid-State Ckts., pp. 1887-1895, Dec. 1997.
3. M. Choi and A. A. Abidi, A 6 b 1.3 Gsample/s A/D converter in 0.35m CMOS, IEEE
J. Solid-State Circuits, pp. 18471858, Dec. 2001.
4. C. S. Scholtens and M. Vertregt, A 6-b 1.6-Gsample/s Flash ADC in 0.18-m CMOS
Using Averaging Termination, IEEE J. of Solid-State Ckts., vol. 37, pp. 1599-1609,
Dec. 2002.
5. X. Jiang, M-C. F. Chang, "A 1-GHz signal bandwidth 6-bit CMOS ADC with powerefficient averaging," IEEE J. of Solid-State Circuits, pp. 532- 535, Feb. 2005.
Folding and Interpolating A/D Converters
6. R.C. Taft et al., "A 1.8-V 1.6-GSample/s 8-b self-calibrating folding ADC with 7.26
ENOB at Nyquist frequency," IEEE J. Solid-State Ckts., pp. 2107-2115 Dec. 2004.
7. B. Nauta and A. G. W. Venes, A 70-MS/s 110-mW 8-b CMOS Folding and
Interpolating A/D Converter, IEEE J. of Solid-State Circuits, pp. 1302-1308, Dec.
1995.
B. Murmann
EE315B - Chapter 7
55

8.
9.
10.
11.
12.
13.
14.
15.
16.
M. P. Flynn and B. Sheahan, A 400-MSample/s, 6-b CMOS Folding and

Interpolating ADC, IEEE J. of Solid-State Circuits, pp. 1932-1938, Dec. 1998.
H. Pan, M. Segami, M. Choi, J. Cao, F. Hatori and A. Abidi, A 3.3V, 12b,
50MSample/s A/D converter in 0.6mm CMOS with over 80dB SFDR, ISSCC Digest
of Technical Papers, pp. 40-41, Feb. 2000.
M.-J. Choe, B.-S. Song and K. Bacrania, A 13b 40MSample/s CMOS pipelined
folding ADC with background offset trimming, ISSCC Digest of Technical Papers,
pp. 36-37, Feb. 2000.
G. Hoogzaad and R. Roovers, A 65-mW, 10-bit, 40-Msample/s BiCMOS Nyquist
ADC in 0.8 mm2, IEEE J. Solid-State Circuits, pp. 1796-1802, Dec. 1999.
K. Nagaraj, F. Chen, T. Le and T. R. Viswanathan, Efficient 6-bit A/D converter using
a 1-bit folding front end, IEEE J. Solid-State Circuits, pp. 1056-1062, Aug. 1999.
M.-J. Choe and B.-S. Song, An 8b 100MSample/s CMOS pipelined folding ADC,
VLSI Symposium Digest of Technical Papers, pp. 81-82, Jun. 1999.
K. Bult and A. Buchwald, An embedded 240-mW 10-b 50-MS/s CMOS ADC in 1
mm2, IEEE J. Solid-State Circuits, vol. 32, pp. 1887-1895, Dec. 1997.
P. Vorenkamp and R. Roovers, A 12-b, 60-MSample/s cascaded folding and
interpolating ADC, IEEE J. Solid-State Circuits, vol. 32, pp. 1876-1886, Dec. 1997.
A. G. W. Venes and R. J. van de Plassche, An 80-MHz, 8-b CMOS folding A/D
converter with distributed track-and-hold preprocessing, IEEE J. Solid-State Circuits,
vol. 31, pp. 1846-1853, Dec. 1996.
B. Murmann
EE315B - Chapter 7
56

17. M. P. Flynn and D. J. Allstot, CMOS folding A/D converters with current-mode
interpolation, IEEE J. Solid-State Circuits, vol. 31, pp. 1248-1257, Dec. 1996.
18. R. Roovers and M. S. J. Steyaert, A 175 Ms/s, 6 b, 160 mW, 3.3 V CMOS A/D
converter, IEEE J. Solid-State Circuits, vol. 31, pp. 938-944, Jul. 1996.
19. W. T. Colleran and A. A. Abidi, A 10-b, 75-MHz two-stage pipelined bipolar A/D
converter, IEEE J. Solid-State Circuits, vol. 28, pp. 1187-1199, Dec. 1993.
20. J. van Valburg and R. J. van de Plassche, An 8-b 650-MHz folding ADC, IEEE J.
Solid-State Circuits, vol. 27, pp. 1662-1666, Dec. 1992.
21. R. J. van de Plassche and P. Baltus, An 8-bit 100-MHz full-Nyquist analog-to-digital
converter, IEEE J. Solid-State Circuits, vol. 23, pp. 1334-1344, Dec. 1988.
22. R. E. J. Van de Grift, I. W. J. M. Rutten and M. van der Veen, An 8-bit video ADC
incorporating folding and interpolation techniques, IEEE J. Solid-State Circuits, vol.
SC-22, pp. 944-953, Dec. 1987.
23. R. E. J. van de Grift and R. J. van de Plassche, A monolithic 8-bit video A/D
converter, IEEE J. Solid-State Circuits, vol. SC-19, pp. 374-378, Jun. 1984.
24. R.C. Taft, et al., "A 1.8V 1.0GS/s 10b self-calibrating unified-folding-interpolating
ADC with 9.1 ENOB at Nyquist frequency," ISSCC Digest of Technical Papers,
pp.78-79, 79a, Feb. 2009.
B. Murmann
EE315B - Chapter 7
57
SAR ADCs
Boris Murmann
Stanford University
EE315B - Chapter 8
B. Murmann
Similarity Between Non-Flash ADC

Architectures
Vin
Minimize
D/A
Dout
Most ADC architectures (other than flash) are based on minimizing

the error between input and a D/A signal approximation
SAR ADC uses comparator to sense
Sigma-delta ADC minimizes via integration and feedback
Pipeline uses distributed DAC
B. Murmann
Successive Approximation Register ADC
Binary search over DAC output
High accuracy achievable (16+ Bits)

Relies on highly accurate comparator
Moderate speed (1+ MHz)
B. Murmann
EE315B - Chapter 8
High Performance Example
B. Murmann
EE315B - Chapter 8
Low Power Example
M.D. Scott, B.E. Boser, K.S.J. Pister, "An ultralow-energy ADC for Smart
Dust," IEEE J. Solid-State Circuits, pp. 1123 -1129, July 2003.
EE315B - Chapter 8
B. Murmann
Classical Implementation
See e.g. [McCreary, JSSC 12/1975]
C1A = C1B = C
B. Murmann
C2 = 2C
C3 = 4C
EE315B - Chapter 8
CB = 2B 1C
6
Sampling Phase (5-bit Example)
Total charge at node Vx after opening Sx

Q = Vin 32C = Vin Ctotal
B. Murmann
EE315B - Chapter 8
Bit5 Test (MSB)
Q = Vin Ctotal = (Vx Vref ) 16C + Vx 16C + Cp
Ctotal
1
Vx = Vref Vin
2
Ctotal + Cp
Vx<0 Vin>0.5Vref Bit5=1
Vx>0 Vin<0.5Vref Bit5=0
B. Murmann
EE315B - Chapter 8
Bit4 Test (Assuming Bit5=0)
Q = Vin Ctotal = (Vx Vref ) 8C + Vx 24C + Cp
Ctotal
1
Vx = Vref Vin
4
Ctotal + Cp
Vx<0 Vin>0.25Vref Bit4=1
Vx>0 Vin<0.25Vref Bit4=0
B. Murmann
EE315B - Chapter 8
Challenges at High Resolution
Capacitor mismatch
Use self calibration to obtain precision beyond raw
technology matching [Lee, JSSC 12/84]
Minimum capacitor size, e.g. Cmin=10fF 216Cmin = 655pF

Solution: Implement "two-stage" or "multi-stage" capacitor
network to reduce array size [Yee, JSSC 8/79]
Typically mandates some form of calibration
Ultimately, even with calibration, there is still a lower bound

based on kT/C in the sampling phase
E.g. 16-bit resolution, Ctotal ~ 100pF required
How to drive such an ADC at high speeds?
B. Murmann
EE315B - Chapter 8
10
Recent Publication Trend

12
SAR ADC Papers at ISSCC and VLSI Symposium

10
Count
0
1996
1998
B. Murmann
2000
2002
2004
2006
Year
2008
2010
2012
EE315B - Chapter 8
2014
11
Recent Innovations in SAR ADCs
Mostly focused on low-to-moderate resolution!
Use of extremely small unit capacitors (<1fF)

E.g. [Shikata, VLSI 2011]
Energy efficient switching techniques
E.g. [Liu, VLSI 2010]
Asynchronous operation
E.g. [Chen, CICC 2006]
Incomplete DAC settling enabled by redundancy
E.g. [Liu, VLSI 2010] 100 MS/s, ~ 9.5 ENOB
SAR ADC pipelining
E.g. [Lee, VLSI 2010] 50MS/s, ~10.5 ENOB
Massive time interleaving
E.g. [Doris, ISSCC 2011] 2.6 GS/s, ~ 8 ENOB
B. Murmann
EE315B - Chapter 8
12
Small Metal Fringe Capacitors
Want to minimize unit caps as much as possible for low-tomoderate resolution SAR ADCs
0.5fF unit capacitors
[Shikata, VLSI 2011]
13
13
B. Murmann
Asynchronous Timing
[Chen & Brodersen, JSSC 12/2006]
Key insight: At most one of the SAR decisions can be close to

metastability; all others will be very fast
Asynchronous timing exploits this by allocating the most time to

the hardest decision
B. Murmann
EE315B - Chapter 8
14
Energy Efficient Switching (1)
Conventional switching scheme is

wasteful in terms of energy
[Liu, JSSC 4/2010]

[Liu, JSSC 4/2010]
EE315B - Chapter 8
B. Murmann
15
Top plate sampling

First decision free
[Liu, JSSC 4/2010]
Total of 81% switching energy reduction
Caveat: Need comparator with good common-rejection

B. Murmann
EE315B - Chapter 8
16
Constant Common Mode CDAC
[Kull, ISSCC 2013]
B. Murmann
EE315B - Chapter 8
17

MSB cap split into two
[Ginsburg, ISCAS 2005]
Key insight: Do not have to switch entire MSB cap for first DAC transition
Total energy savings of 37%
B. Murmann
EE315B - Chapter 8
18
Resistive DAC
[Wei, ISSCC 2011]
Multi-bit/cycle architecture yielding 8 bits at 400 MS/s

EE315B - Chapter 8
B. Murmann
19
DAC Settling Errors
0(t)
t
tdecision
bk
Input
-1 x +1
(conceptual model)
2-1
1
b0
1
b1
2-1
b2
2-2
B. Murmann
Register
EE315B - Chapter 8
Output x
20
Error in First Decision

1
k=0
x
0
x
-1
EE315B - Chapter 8
B. Murmann
21
Redundancy (Radix < 2)

x
bk
-1
1
b0
1
-2
b1
b2
-1
-2
b3
Register
-3
Output x
<
>
B. Murmann
EE315B - Chapter 8
22
Error in First Decision ( = 22/3)

1
k=0
x
0
Error /2
Note that
-1
EE315B - Chapter 8
B. Murmann
23
Tolerable Error ( = 22/3)

1
k=0
Tolerable
Error
x
/2
-1
B. Murmann
EE315B - Chapter 8
24
First Hardware Implementation
Z. Boyacigiller, B. Weir, and P. Bradshaw, An error-correcting 14b/20s CMOS A/D

converter, in ISSCC Dig. Techn. Papers, Feb. 1981, pp. 6263.
B. Murmann
EE315B - Chapter 8
25
Other Ways to Introduce Redundancy
Radix = 2, but extra comparators

T. C. Verster, A Method to Increase the Accuracy of FastSerial-Parallel Analog-to-Digital Converters, IEEE Trans. on
Electronic Computers, vol. EC-13, no. 4, pp. 471473, Aug.
1964.
Radix = 2, but extra cycles

V. Giannini et al., An 820W 9b 40MS/s Noise-Tolerant
Dynamic-SAR ADC in 90nm Digital CMOS, in ISSCC Dig.
Techn. Papers, Feb. 2008, pp. 238239.
C.-C. Liu et al., A 10b 100MS/s 1.13mW SAR ADC with
binary-scaled error compensation, in ISSCC Dig. Techn.
Papers, 2010, pp. 386387.
B. Murmann
EE315B - Chapter 8
26
SAR ADC with Redundant Comparator

+ok
+. > +
= .
2-1
-ok
b0
1
b1
b2
2-1
Register
2-2
<
R. Vitek et al., A 0.015mm2 63fJ/conversion-step 10-bit 220MS/s SAR ADC with 1.5b/step
redundancy and digital metastability correction, in Proc. IEEE Custom Integrated Circuits
Conference, Sep. 2012, pp. 14.
EE315B - Chapter 8
B. Murmann
27
Normal Operation
1
k=0
+o1
x
+o0
-o1
0
-o0
-1
B. Murmann
EE315B - Chapter 8
28

k=0
+o0
+o1
0
-o1
-o0
-1
EE315B - Chapter 8
B. Murmann
29
SAR ADC with Redundant Step

x
2-1
1
b0
2-2
b1
b2
2-1
bk
b3
2-2
Register
= +

= = +.
= = .
V. Giannini et al., An 820W 9b 40MS/s Noise-Tolerant Dynamic-SAR ADC in 90nm Digital

CMOS, in ISSCC Dig. Techn. Papers, Feb. 2008, pp. 238239.
B. Murmann
EE315B - Chapter 8
30
Normal Operation
1
k=0
x
0
Characteristic
crossing
-1
EE315B - Chapter 8
B. Murmann
31

1
k=0
x
0
-1
B. Murmann
EE315B - Chapter 8
32
Redundant Step Alternate Method

x
2-1
1
b0
1
bk
-
2-1
b1
b1R
2-1
2-1
b2
Register
2-2
C.-C. Liu et al., A 10b 100MS/s 1.13mW SAR ADC with binary-scaled error compensation, in
ISSCC Dig. Techn. Papers, Feb. 2010, pp. 386387.
EE315B - Chapter 8
B. Murmann
33
Extra Levels
The addition of the redundant step provides extra levels that help
counter errors in previous decisions
b1R + b2
b2
+0.25+0.125 = 0.375
+0.125
+0.25-0.125 = +0.125
-0.125
-0.25+0.125 = -0.125
-0.25-0.125 = -0.375
B. Murmann
EE315B - Chapter 8
34
Normal Operation
1
k=0
2R
x
0
-1
EE315B - Chapter 8
B. Murmann
35

1
k=0
x
0
2R
-1
B. Murmann
EE315B - Chapter 8
36
Potential Speed Benefit of Using

Redundancy
T. Ogawa et al., IEICE Trans. on Fundamentals of

Electronics, Communications and Computer Sciences,
Feb. 2010.
Benefit is clearly significant at high resolution

However, one must carefully evaluate if redundancy helps
significantly at low resolutions (e.g. 6-8 bits)
B. Murmann
EE315B - Chapter 8
37
Metastability in an Asynchronous SAR
Worst case input is near 1/6 VFS
[Chen, JSSC 12/2006]
B. Murmann
EE315B - Chapter 8
38
Example: 8-bit SAR (1)

There are seven easy decisions, and one hard decision
It can be shown that the sum of the easy decisions will take
anywhere between 25-33, depending on the input
For the hard decision, the input is within 1LSB, and we can
assume that it is uniformly distributed
PDF
=
-LSB
+LSB
vod0,min
2
= ln
= ln
ln
EE315B - Chapter 8
B. Murmann
39

Number of time constants to be allocated for the hard
decision, given a desired metastability rate (B=8)
60
50
thard/
40
30
20
10
-20
10
-15
-10
10
10
10
-5
P meta
B. Murmann
EE315B - Chapter 8
40
Continue using the following assumptions (design and technology

dependent)
Total cycle time Ts = 1ns (1GS/s)
Sampling window = Ts/8
Sum of DAC settling times = Ts/3
Sum of logic delays = Ts/6
The time available for regeneration is treg,tot = (3/8)Ts = 375ps
The metastability rate estimate as a function of is
2 e
=2 e
2 e
, ()
EE315B - Chapter 8
B. Murmann
41

0
10
teasy=33
teasy=25
-5
10
Pmeta
-10
10
-15
10
-20
10
[ps]
Incredibly steep tradeoff fast comparator is key
B. Murmann
EE315B - Chapter 8
42
Comparator Reset
Fast reset is just as as important as fast regeneration

The design below ping-pongs between two comparators to
increase the available reset time
[Kull, ISSCC 2013]
B. Murmann
EE315B - Chapter 8
43
Aside: Counter/Single-Slope ADC

[Harpe, ASSCC 2010]
One of the oldest ADC architectures
Simply measure time it takes for a ramp to reach input voltage
Ramp can be generated using a counter + DAC or current

source+capacitor
Slow, because counter range increases exponentially with bitresolution
B. Murmann
EE315B - Chapter 8
44
Pipeline ADCs
Boris Murmann
Stanford University
B. Murmann
EE315B - Chapter 9
Outline
Background
History and state-of the art performance
Derivation from SAR architecture
Pipeline ADC basics

Ideal block diagram and operation, impact of block
nonidealities
Ways to deal with nonidealities

Redundancy, calibration
CMOS implementation details

Stage scaling, MDAC design
Architectural options
OTA sharing, SHA-less front-end
Research topics
B. Murmann
EE315B - Chapter 9
Pipeline ADCs An Old Idea (1956)
B.D. Smith, An Unusual Electronic Analog-Digital Conversion Method, IRE Transactions on

Instrumentation, vol. PGI-5, pp. 155-160, June 1956.
EE315B - Chapter 9
B. Murmann
Todays Pipeline ADC Performance Range

10
Flash
Folding
Two-Step
Pipeline
10
10
SAR
Other
1000fs
in,hf
[Hz]
100fs
10
10
rms
rms
Jitter
Jitter
20
30
40
50
60
70
80
SNDR [dB]
90
100
110
120
hf
B. Murmann
EE315B - Chapter 9
SAR ADC
bk
-1
-2
b0
b1
b2
-1
b3
-2
Register
-3
Output x
EE315B - Chapter 9
B. Murmann
SAR Pipeline (1)
b0
1
-2
b1
b2
-1
-2
b3
Register
-3
Output x
B. Murmann
bk
EE315B - Chapter 9
SAR Pipeline (2)
b0
bk
b1
b2
-1
b3
-2
Register
-3
Output x
EE315B - Chapter 9
B. Murmann
SAR Pipeline (3)
b1
b0
1
b2
-1
-2
B. Murmann
EE315B - Chapter 9
b3
-3
Output x
Observation
Pipeline and SAR ADCs are mathematically equivalent

(considering ideal components)
Therefore, not surprisingly, all variants of redundancy can also

be incorporated in a pipeline ADC
It is just less obvious, and usually explained quite differently
More later
EE315B - Chapter 9
B. Murmann
General Pipeline ADC Block Diagram

Align & Combine Bits
Dout
Vin
SHA
Stage 1
Stage n-1
Vin1
G11
G
ADC
Vres1
Stage n
Each stage contains a T/H

(not shown)
DAC
D1
B. Murmann
EE315B - Chapter 9
10
Concurrent Stage Operation

CLK
Align & Combine Bits

D
V
SHA
Stage 1
Stage 2
Stage 3
ACQUIRE
CONVERT
ACQUIRE
CONVERT
BUFFER
ACQUIRE
CONVERT
ACQUIRE
Stages operate on the input signal like a shift register
New output data every clock cycle, but each stage introduces
clock cycle latency
B. Murmann
EE315B - Chapter 9
11
Pipelining An Old Idea
B. Murmann
EE315B - Chapter 9
12
Data Alignment
Digital shift register aligns sub-conversion results in time
Digital output is taken as weighted sum of stage bits
B. Murmann
EE315B - Chapter 9
13
Latency
[Analog Devices, AD9226 Data Sheet]
B. Murmann
EE315B - Chapter 9
14
Pipeline ADC Characteristics
Number of components grows linearly with resolution

Unlike flash ADC, where components ~ 2B
Pipeline ADC trades latency for conversion speed

Throughput limited by speed of one stage
Enables high-speed operation
Latency can be an issue in some applications

E.g. in feedback control loops
Pipelining only possible with good analog "memory elements"

Calls for implementation in CMOS using switched-capacitor
circuits
B. Murmann
EE315B - Chapter 9
15
Stage Analysis
Ignore timing/clock delays for simplicity
D = Q(Vin )
B. Murmann
EE315B - Chapter 9
Vres = G [Vin Vdac ]
16
Unity Gain Quantizer Model
B. Murmann
1
1 2
LSB =
2
2 2B
EE315B - Chapter 9
17
Stage Model with Ideal DAC
D = Vin + q
Vres = G q
Residue of pipeline stage (Vres) is equal to (-gain) times subADC quantization error
B. Murmann
EE315B - Chapter 9
18
"Residue Plot" (2-bit Sub-ADC)
{-3/4, -1/4, 1/4, 3/4}
Sub-ADC
Decision
Levels
B. Murmann
EE315B - Chapter 9
19
Pipeline Decomposition
Often convenient to look at pipeline as single stage plus

backend ADC
B. Murmann
EE315B - Chapter 9
20
Resulting Model
res
in
1
qb
q
out
G
Dout = Vin + q 1
Gd
qb
+
Gd
With Gd=G
B. Murmann
EE315B - Chapter 9
21
Canonical Extension
q( n 1)
qn
G( n 1)
G q 2
G2
Dout = Vin + q1 1 1 +
1
+ n 1
1
+ ... + n 2
Gd( n 1)
Gd 1 Gd 1 Gd 2
G
G
dj
j =1
First stage has most stringent precision requirements
Note that above model assumes that all stages use same reference
voltage (same full scale range)
This is true for most designs, one exception is [Limotyrakis 2005]
B. Murmann
EE315B - Chapter 9
dj
j =1
22
General Result Ideal Pipeline ADC

With ideal DACs and ideal digital weights (Gdj=Gj)
qn
Dout = Vin + n 1
n 1
BADC = Bn + log 2G j
j =1
G j
j =1
The only error in Dout is that of last quantizer, divided by

aggregate gain
Aggregate ADC resolution is independent of sub-ADC
resolutions in stage 1...n-1 (!)
Makes sense to define "effective" resolution of jth stage as
Rj=log2(Gj)
B. Murmann
EE315B - Chapter 9
23
Questions
How to pick stage gain G for a given sub-ADC resolution?
Impact and compensation of nonidealities?

Sub-ADC errors
Amplifier offset
Amplifier gain error
Sub-DAC error
Begin to explore these questions using a simple example

First stage with 2-bit sub-ADC, followed by 2-bit backend
B. Murmann
EE315B - Chapter 9
24
Upper Bound for Stage Gain

Vin
Vres
G1
G
qb
Db
q
B
{-1/2 1/2B}
+1
G/2B
Vres
2G/2B
-G/2B
-1
-1
Dout = Vin +
0
Vin
qb
G
+1
qb
Grows out of LSB bounds for G>2B

EE315B - Chapter 9
B. Murmann
25
Issue with G=2B

in
res
1
qb
b
q
B
Overrange!
res
qb
in
Any error in sub-ADC decision levels will overload backend ADC and
thereby deteriorate ADC transfer function
B. Murmann
EE315B - Chapter 9
26
Idea #1: G slightly less than 2B
Effective stage resolution can be non-integer (R=log2G)

E.g. R = log23.2 = 1.68 bits
See e.g. [Karanicolas 1993]
B. Murmann
EE315B - Chapter 9
27
Idea #2: G < 2B, but Power of Two
Effective stage resolution is an integer

E.g. R = log22 = 1 = B-1
Digital hardware requires only a few adders, no need to
implement fractional weights (see appendix)
See e.g. [Mehr 2000]
B. Murmann
EE315B - Chapter 9
28
Idea #3: G=2B, Extended Backend Range
No redundancy in stage with errors

Extra decision levels in succeeding stage used to bring
residue "back into the box"
See e.g. [Opris 1998]

EE315B - Chapter 9
B. Murmann
29
Variant of Idea #2: "1.5-bit stage"
Sub-ADC decision levels placed to minimize comparator count
Can accommodate errors up to
B = log2(2+1) = 1.589 (sub-ADC resolution)
R = log22 = 1 (effective stage resolution)
See e.g. [Lewis 1992]
B. Murmann
EE315B - Chapter 9
30
Summary on Sub-ADC Redundancy
We can tolerate sub-ADC errors as long as

The residue stays "inside the box", or
Another stage downstream returns the residue "into the box"
before it reaches last quantizer
This result applies to any stage in an n-stage pipeline

Can always decompose pipeline into single stage + backend
ADC
In literature, sub-ADC redundancy schemes are often called

"digital correction" a misnomer in my opinion
There is no explicit error correction!

Sub-ADC errors are absorbed in the same way as their
inherent quantization error
As long as there is no overranging
EE315B - Chapter 9
B. Murmann
31
Amplifier Offset
Push
Amplifier offset can be referred toward stage input and results in

Global offset
Usually no problem, unless "absolute ADC accuracy" is required
Sub-ADC offset
Easily accommodated through redundancy
B. Murmann
EE315B - Chapter 9
32
Gain Errors
qn
G +
Dout = Vin + q1 1 1
+ ... + n 1
Gd 1
dj
j =1
Want to make Gd1 = G1+

EE315B - Chapter 9
B. Murmann
33
Digital Gain Calibration (1)
Error in analog gain is not a problem as long as "digital gain

term" is adjusted appropriately
Problem
Need to measure analog gain precisely
Example
Digital calibration of a 1-bit first stage with 1-bit redundancy
(R=1, B=2)
Note
Even if all Gdj are perfectly adjusted to reflect the analog gains, the ADC will have nonzero DNL and INL, bounded by 0.5LSB. This can be explained by the fact that the
residue transitions may not correspond to integer multiples of the backend-LSB. This
can cause non-uniformity in the ADC transfer function (DNL, INL) and also nonmonotonicity (see [Markus, 2005]).
In case this cannot be tolerated
B. Murmann
Add redundant bits to ADC backend (after combining all bits, final result can be truncated back)
Calibrate analog gain terms
EE315B - Chapter 9
34
Vres = G [Vin Vdac ]

Db = Vres + qb
EE315B - Chapter 9
B. Murmann
35
Step1:
Step2:
1)
Db( 1) = G [Vin + 0.25] + (qb
2)
Db( 2 ) = G [Vin 0.25] + (qb
1)
2)
Db( 1) Db( 2 ) = 0.5 G + (qb
(qb
Can minimize impact of quantization error using

Averaging (thermal noise dither)
Extra backend resolution
B. Murmann
EE315B - Chapter 9
36
DAC Calibration
Essentially same concept as gain calibration

Step through DAC codes and use backend to measure errors
Store coefficients for each DAC transition in a look-up table
B. Murmann
EE315B - Chapter 9
37
Recursive Stage Calibration
First few stages have most stringent accuracy requirements

Errors of later stages are attenuated by aggregate gain
Commonly used algorithm [Karanicolas 1993]

Take ADC offline
Measure least significant stage that needs calibration first
Move to next significant stage and continue toward stage 1
B. Murmann
EE315B - Chapter 9
38
Calibration Hardware Example
[Chuang 2002]
B. Murmann
EE315B - Chapter 9
39
Alternative Schemes
Other foreground calibration schemes

Calibrate ADC starting from first stage [Singer 2000]
Connect stages in a circular loop [Soenen 1995]
Background calibration
See e.g. [Ming 2001]
Makes sense primarily when calibration parameters are
expected to drift
Capacitor ratios do not drift!
Background calibration is justifiable e.g. when drift in OTA
open-loop gain is an issue
B. Murmann
EE315B - Chapter 9
40
Nonlinearity Compensation with

Background Calibration
[Murmann 2003]
B. Murmann
EE315B - Chapter 9
41
Combining the Bits (1)
Example1: Three 2-bit stages, no redundancy
Dout = D1 +
B. Murmann
1
1
D2 +
D3
4
16
EE315B - Chapter 9
42

D1
XX
D2
XX
D3
XX
-----------Dout DDDDDD
Only bit shifts

No arithmetic circuits needed
EE315B - Chapter 9
B. Murmann
43
Example2: Three 2-bit stages, one bit redundancy in stages 1

and 2 (6-bit aggregate ADC resolution)
Vin
B1=3
R1=2
B2=3
R2=2
Stage 1
Stage 2
B3=2
Stage 3
8 Wires
???
6 Wires
Dout[5:0]
B. Murmann
EE315B - Chapter 9
44

Dout
1
1
= D1 + D2 +
D3
4
16
B. Murmann
D1
XXX
D2
XXX
D3
XX
-----------Dout DDDDDD
Bits overlap
Need adders (Still, no
good reason for calling
this "digital correction"...)
EE315B - Chapter 9
45
For fractional weights (e.g. radix <2), there is no need to

implement complex multipliers
Can still use simple bit shifts; push actual multiplication into lowresolution output
E.g. a 1x10 bit multiplication needs only one adder
See e.g. [Karanicolas 1993]
B. Murmann
EE315B - Chapter 9
46
Outline
Background
History and state-of the art performance
General idea of multi-step A/D conversion
Pipeline ADC basics

Ideal block diagram and operation, impact of block
nonidealities
Ways to deal with nonidealities

Redundancy, calibration
CMOS implementation details

Stage scaling, MDAC design
Architectural options
OTA sharing, SHA-less front-end
Research topics
EE315B - Chapter 9
B. Murmann
47
Stage Implementation
Flash ADC
"MDAC"
SwitchedCapacitor
Circuit
B. Murmann
EE315B - Chapter 9
48
Generic Circuit
1:
Q = Vin kCs
2 :
Q = Vres Cf + Vrefp mCs + Vrefn ( k m ) Cs
Vres =
( k m ) Cs V
kCs
mCs
Vin
Vrefp +
refn
Cf
Cf
Cf
= G ( Vin Vdac )
B. Murmann
EE315B - Chapter 9
49
Endless List of Design Parameters
Stage resolution, stage scaling factor
Stage redundancy
Thermal noise/quantization noise ratio
OTA architecture
OTA sharing?
Switch topologies
Comparator architecture
Front-end SHA vs. SHA-less design
Calibration approach (if needed)
Time interleaving?
Technology and technology options (e.g. capacitors)
A very complex optimization problem!
B. Murmann
EE315B - Chapter 9
50
Thermal Noise Considerations
Total input referred noise

Thermal noise + quantization noise
Costly to make thermal noise smaller than quantization noise
Example: VFS=1V, 10-bit ADC

Nquant=LSB2/12=(1V/210)2/12=(280Vrms)2
Design for total input referred thermal noise ~280Vrms or
larger, if SNR target allows
Total input referred thermal noise is the sum of noise in all

stages
How should we distribute the total thermal noise budget
among the stages?
Let's look at an example
B. Murmann
EE315B - Chapter 9
51
Stage Scaling (1)
Example: Pipeline using 1-bit (effective) stages (G=2)
Total input referred noise power

1
1
1
Ntot kT +
+
+ ...
C1 4C2 16C3
B. Murmann
EE315B - Chapter 9
52
Stage Scaling (2)

C1/2
C1
Vin
C2/2
C2
Gm
C3/2
C3
Gm
Gm
1
1
Ntot kT +
+
+ ...
C1 4C2 16C3
If we make all caps the same size, backend stages contribute

very little noise
Wasteful, because Power ~ Gm ~ C
EE315B - Chapter 9
B. Murmann
53
Stage Scaling (3)

C1/2
C1
Vin
C2/2
C2
Gm
C3/2
C3
Gm
Gm
1
1
Ntot kT +
+
+ ...
C1 4C2 16C3
How about scaling caps down by 22=4x per stage?

Same amount of noise from every stage
All stages contribute significant noise
Noise from first few stages must be reduced
Power ~ Gm ~ C goes up!
B. Murmann
EE315B - Chapter 9
54
Stage Scaling (4)
[Cline 1996]
Optimum capacitior scaling lies approximately midway between

these two extremes
EE315B - Chapter 9
B. Murmann
55
Shallow Optimum
[Chiu 2004]
Capacitor scaling factor = 2Rx
B. Murmann
x=1 scaling exactly by stage gain
EE315B - Chapter 9
56
Practical Approach to Stage Scaling
Start by assuming caps are scaled precisely by stage gain

E.g. for 1-bit effective stages:
C/2
C
Vin
C/4
C/2
Gm
C/8
C/4
Gm
Gm
Refine using first pass circuit information & Excel spreadsheet

Use estimates of OTA power, parasitics, minimum feasible
sampling capacitance etc.
Or, buy a circuit optimization tool
B. Murmann
EE315B - Chapter 9
57
Stage Scaling Examples (1)
[Cline 1996]
B. Murmann
EE315B - Chapter 9
58
Stage Scaling Examples (2)

[Ishii 2005]
EE315B - Chapter 9
B. Murmann
59
How Many Bits Per Stage?
Low per-stage resolution (e.g. 1-bit effective)

Need many stages
+ OTAs have small closed loop gain, large feedback factor
High speed
High per-stage resolution (e.g. 3-bit effective)

+ Fewer stages
OTAs can be power hungry, especially at high speed
Significant loading from flash-ADC
Qualitative conclusion
Use low per-stage resolution for very high speed designs
Try higher resolution stages when power efficiency is most
important constraint
B. Murmann
EE315B - Chapter 9
60
Power Tradeoff is Fairly Flat!
[Chiu 2004]
= parasitic cap at output/total
sampling cap in each stage
(junctions, wires, switches, )
ADC power varies by only ~2x across different stage resolutions!

EE315B - Chapter 9
B. Murmann
61
Examples
Reference
[Yoshioka, 2007] [Jeon, 2007]
[Loloee 2002]
[Bogner 2006]
Technology
90nm
90nm
0.18um
0.13um
Bits
10
10
12
14
Bits/Stage
1-1-1-1-1-1-1-3
2-2-2-4
1-1-1-1-1-1-1-1-1-1-2
3-3-2-2-4
SNDR [dB]
~56
~54
~65
~64
Speed [MS/s]
80
30
80
100
Power [mW]
13.3
4.7
260
224
mW/MS/s
0.17
0.16
3.25
2.24
Low power is possible for a wide range of architectures!
B. Murmann
EE315B - Chapter 9
62
Re-Cap
Choosing the "optimum" per-stage resolution and stage scaling

scheme is a non-trivial task
But optima are shallow!
Quality of transistor level design and optimization is at least as

important (if not more important than) architectural
optimization
Next, look at circuit design details

Assume we're trying to build a 10-bit pipeline
Recent technology, feature size ~0.18m or smaller

Moderate to high-speed ~100MS/s
1-bit effective/stage, using "1.5-bit" stage topology
Dedicated front-end SHA
EE315B - Chapter 9
B. Murmann
63
1.5-Bit Stage Implementation

[Abo 1999] ([Lewis 1992])
Cf is used as sampling cap during acquisition phase, as

feedback cap in redistribution phase
Helps improve feedback factor (max. 1/3 max. 1/2)
B. Murmann
EE315B - Chapter 9
64
Residue Plot
[Abo 1999]
B. Murmann
EE315B - Chapter 9
65
Stage 1 Matching Requirements
Cs
C
1+
Cf
C
Error in residue transition must be accurate to within a fraction of

9-bit backend LSB
Typically want C/C ~0.1% or better
B. Murmann
EE315B - Chapter 9
66
Capacitor Matching
0.1% "easily" achievable in current technologies

Even with metal sandwich caps, see e.g. [Verma 2006]
Beware of metal density related issues, "copper dishing"
For MIMCap matching data see e.g. [Diaz 2003]
What if we needed much higher resolution than 10 bits?

Digital calibration
Multi-bit first stage
Each extra bit resolved in the first stage alleviates precision
requirements on residue transition by 2x
For fixed capacitor matching, can show that each (effective) bit
moved into the first stage
Improves DNL by 2x
Improves INL by sqrt(2)x
Multi-bit examples: [Singer 1996] [Kelly 2001] [Lee 2007]
EE315B - Chapter 9
B. Murmann
67
Typical Reference Generator

[Brooks 1994]
External decoupling caps provide dynamic currents

Low power reference buffer
B. Murmann
EE315B - Chapter 9
68
Comparators
Can tolerate large offsets and large noise with appropriate

redundancy
Consume negligible power in a good design

50-100W or less per comparator
Lots of implementation options

Resistive/capacitive reference generation
Different pre-amp/latch topologies

EE315B - Chapter 9
B. Murmann
69
Comparator Examples
Vin
[Chiu 2004]
[Mehr 2000]
B. Murmann
EE315B - Chapter 9
70
OTA Design Considerations
Static amplifier error = 1/(DC Loop Gain)

E.g. for 0.1% accuracy in first stage of 10-bit ADC, need loop
gain > 60dB
Dynamic settling error

Typically want to settle outputs to ~1/8 LSB accuracy within
1/2 clock cycle
Thermal noise
Size capacitors to satisfy kT/C noise requirement
Start by picking an OTA topology that will deliver sufficient gain

Or think about ways to compensate finite gain error
General references on OTA design

[Boser 2005], [Murmann, EE315A]
EE315B - Chapter 9
B. Murmann
71
Two-Stage Folded Cascode OTA

1V
[Ishii 2005]
~0.5V
(1Vpp,diff)
Works down to VDD=1V with reasonable output swing
Gain ~ (gmro)3 ~ 103 = 60dB
Use gain boosting to achieve larger gain
B. Murmann
EE315B - Chapter 9
72
How Fast Can We Go? (1)
Non-dominant pole in two-stage amplifier hard to move past fT/5
For 73 degrees phase margin (optimum for fast settling), loop

crossover frequency is 1/3 of non-dominant pole frequency
Settling linearly to 0.1% precision takes 7 loop time constants;

typically budget ~10 time constants
Ideally, we'd have 1/2 clock cycle to settle linearly, but there is
some time needed for slewing and non-overlap clock timing
Assume 60% of half cycle is available for linear settling
In summary
fCLK ,max =
fT 1
f
1
2 0.5 0.6 = T
5 3
10
80
EE315B - Chapter 9
B. Murmann
73
How Fast Can We Go? (2)
Technology
NMOS fT
(at moderate VGS-Vt ~150mV)
fCLK,max = fT/80
0.35um
10GHz
125 MHz
0.18um
30GHz
375 MHz
90nm
90GHz
1.125GHz (?)
Sampling speeds of 200-300MHz are "easily" achievable in

today's technologies
fT is no longer a showstopper
Speed ultimately constrained by power, power efficiency
and/or clock jitter
B. Murmann
EE315B - Chapter 9
74
Switches
[Ishii 2005]
Make switch RC ~ 10 times

faster than OTA
Avoids speed degradation
Minimizes switch noise
contribution
See e.g. [Schreier 2005]
Avoids stability issues due to

poles in feedback network
Three choices for switches

Single N or P device
Transmission gate
Bootstrapped NMOS
For high swing nodes that
require constant Ron
EE315B - Chapter 9
B. Murmann
75
Front-End SHA
Need constant RON here to
minimize signal dependent
charge injection from S1N, S1P
Minimize
Jitter!
S1P
S1N
[Ishii 2005]
B. Murmann
EE315B - Chapter 9
76
Total Integrated OTA Noise (1)

Cf
N1 = 1 +
g m11 + g m 31
2...4
g m1
N2 = 1 +
g m 61
2
g m51
Cs
Cc
1
kT
kT
2
Vod
= 2 N1
+ 2( N2 + 1)
Cc
CLtot
Stage 1
Cf
Cf + Cs + Cgs1
Stage 2
ignore in first
cut design
CLtot = CL + (1 ) Cf + Cparasitic
EE315B - Chapter 9
B. Murmann
77
Total Integrated OTA Noise (2)
Assuming =1, N1=N2=2

1 kT
kT
2
Vod
=4
+6
Cc
CLtot
OTA noise partitioning problem

How should we split noise between stage1 and stage2
terms?
In this design example we'll use a 2/3, 1/3 split

This is yet another design/optimization parameter
With this assumption, we have

2
Vod
= 18
B. Murmann
kT
CLtot
Cc =
EE315B - Chapter 9
CLtot
3
78
Stage 1 Noise
Cs2
Cs1=C1P+C2P
Cs1 / 2
1
Cs1 + Cgs1 3
2
Vod
,1 = 18
1C
CLtot = Cs 2 + 1 s1
3 2
Vid2 ,1 =
kT
Cs 2 + Cs1 / 3
kT
2 Cs 2 + Cs1 / 3
18
2
EE315B - Chapter 9
B. Murmann
79
SHA Noise
=
Cs 0
1
Cs 0 + Cgs1 2
Cs1
1C
CLtot = Cs1 + 1 s 0
2 2
Cs0
Design choice: Cs0 = Cs1

From sample phase (1)
2
2
Vod
,0 = Vid ,0 = 18
B. Murmann
kT
kT
kT
+
16
Cs1 + Cs 0 / 4 Cs 0
Cs1
EE315B - Chapter 9
80
Noise Budgeting
Total input referred noise budget, assuming VFS,diff=1V

Nthermal = Nquant=LSB2/12=(1V/210)2/12=(280Vrms)2
Reasonable "first cut" partitioning of input referred noise

SHA 1/2
Stage 1 1/4
All remaining stages 1/4
Vid2 ,0 = 16
Vid2 ,1 =
kT 1
2
= ( 280Vrms ) Cs1 = 1.66 pF
Cs1 2
9
kT
1
2
= ( 280Vrms ) Cs 2 = 0.38 pF
2 Cs 2 + Cs1 / 3 4
EE315B - Chapter 9
B. Murmann
81
Capacitor Sizes
Cs0
1.66pF
Cs1
1.66pF
Cs2
0.38pF
Cs3
190fF
Cs4
85fF
Cs5
42.fF (minimum)
Cs10
42.fF (minimum)
/2
/2
/2
Now refine these numbers using simulation and Excel spreadsheet

Iterate over assumptions/design choices to optimize design
B. Murmann
EE315B - Chapter 9
82
Reality Check
[Honda 2007]
Not too far off from a practical design
B. Murmann
EE315B - Chapter 9
83
Amplifier Sharing (1)

[Min 2003]
B. Murmann
EE315B - Chapter 9
Limited power savings

because amplifiers have
different specs
84
Amplifier Sharing (2)

[Kurose 2006]
Sharing of amplifiers is most efficiently done in a pair of

converters that process I/Q signals
B. Murmann
EE315B - Chapter 9
85
SHA-Less Architectures (1)
Motivation
SHA can burn up to 1/3 of total ADC power
Removing front-end SHA creates acquisition timing mismatch

issue between first stage MDAC & Flash
Sampler
(MDAC)
[Chiu 2004]
B. Murmann
EE315B - Chapter 9
86
SHA-Less Architectures (2)
Strategies
Use first stage with large redundancy; this can help absorb
fairly large skew errors
Try to match sampling sub-ADC/MDAC networks
Bandwidth and clock timing
[Mehr 2000]
B. Murmann
EE315B - Chapter 9
87
Pipelined SAR
[Lee, JSSC, April 2011]
B. Murmann
EE315B - Chapter 9
88
Pipeline Cyclic ADC (1)
b1
b0
b2
-1
b3
-2
-3
Output x
EE315B - Chapter 9
B. Murmann
89
Pipeline Cyclic ADC (2)
Register
b0
1
b1
b2
-1
-2
B. Murmann
EE315B - Chapter 9
b3
-3
Output x
90
Implementation Example
1
Vo1
C
Vo1 = 1 + 5 Vo 2
C6
[Erdogan et al. JSSC 12/99]
Vo2
Vo 2 = Vip Vref
) CC4
3
EE315B - Chapter 9
B. Murmann
91
Discussion
Advantages
Area efficient
Typically only one or two switched capacitor stages plus
comparator
Easy to calibrate
Need to measure only one coefficient (capacitor ratio)
Disadvantages
Slow
Need many clock cycles for a single conversion
Sub-optimal power efficiency

Cannot scale stages like in a pipeline ADC
Noise and accuracy requirements decrease from MSB to LSB
cycle, but invested circuit energy per cycle is (usually) constant
B. Murmann
EE315B - Chapter 9
92
General
S. Kawahito, "Low-Power Design of Pipeline A/D Converters," Proc. CICC, pp.
505-512, Sep. 2006.
Redundancy & Calibration

S. H. Lewis et al., "A 10-b 20-Msample/s analog-to-digital converter," IEEE
JSSC, pp. 351-358, Mar. 1992.
A. N. Karanicolas et al. "A 15-b 1-Msample/s digitally self-calibrated pipeline
ADC," IEEE J. Of Solid-State Circuits, pp. 1207-1215, Dec. 1993.
E. G. Soenen et al., "An architecture and an algorithm for fully digital correction
of monolithic pipelined ADCs," IEEE TCAS II, pp. 143-153, Mar. 1995.
I. E. Opris et al., "A single-ended 12-bit 20 MSample/s self-calibrating pipeline
A/D converter," IEEE JSSC, pp. 1898-1903, Dec. 1998.
L. Singer et al., "A 12 b 65 MSample/s CMOS ADC with 82 dB SFDR at 120
MHz," ISSCC Dig. Techn. Papers, pp. 38-39, Feb. 2000.
I. Mehr et al., "A 55-mW, 10-bit, 40-Msample/s Nyquist-rate CMOS ADC," IEEE
JSSC, pp. 318-325, Mar. 2000.
J. Ming et al., "An 8-bit 80-Msample/s pipelined analog-to-digital converter with
background calibration," IEEE JSSC, pp. 1489-1497, Oct. 2001.
S.-Y. Chuang et al., "A Digitally Self-Calibrating 14-bit 10-MHz CMOS Pipelined
A/D Converter," IEEE JSSC, pp. 674-683, Jun. 2002.
J. Markus et al., "On the monotonicity and linearity of ideal radix-based A/D
converters," IEEE Trans. Instr. and Measurement, pp. 2454-2457, Dec. 2005.
EE315B - Chapter 9
B. Murmann
93
Implementation
T. Cho, "Low-Power Low-Voltage Analog-to-Digital Conversion Techniques
using Pipelined Architecures, PhD Dissertation, UC Berkeley, 1995,
http://kabuki.eecs.berkeley.edu/~tcho/Thesis1.pdf.
L. A. Singer at al., "A 14-bit 10-MHz calibration-free CMOS pipelined A/D
converter," VLSI Circuit Symposium, pp. 94-95, Jun. 1996.
A. Abo, "Design for Reliability of Low-voltage, Switched-capacitor Circuits,"
PhD Dissertation, UC Berkeley, 1999,
http://kabuki.eecs.berkeley.edu/~abo/abothesis.pdf.
D. Kelly et al., "A 3V 340mW 14b 75MSPS CMOS ADC with 85dB SFDR at
Nyquist," ISSCC Dig. Techn. Papers, pp. 134-135, Feb. 2001.
A. Loloee, et. al, A 12b 80-MSs Pipelined ADC Core with 190 mW
Consumption from 3 V in 0.18-um, Digital CMOS, Proc. ESSCIRC, pp.
467-469, 2002
B.-M. Min et al., "A 69-mW 10-bit 80-MSample/s Pipelined CMOS ADC,"
IEEE JSSC, pp. 2031-2039, Dec. 2003.
Y. Chiu, et al., "A 14-b 12-MS/s CMOS pipeline ADC with over 100-dB
SFDR, IEEE JSSC, pp. 2139-2151, Dec. 2004.
S. Limotyrakis et al., "A 150-MS/s 8-b 71-mW CMOS time-interleaved
ADC," IEEE JSSC, pp. 1057-1067, May 2005.
T. N. Andersen et al., "A Cost-Efficient High-Speed 12-bit Pipeline ADC in
0.18-um Digital CMOS," IEEE JSSC, pp. 1506-1513, Jul 2005.
B. Murmann
EE315B - Chapter 9
94

P. Bogner et al., "A 14b 100MS/s digitally self-calibrated pipelined ADC in
0.13um CMOS," ISSCC Dig. Techn. Papers, pp. 832-833, Feb. 2006.
D. Kurose et al., "55-mW 200-MSPS 10-bit Pipeline ADCs for Wireless
Receivers," IEEE JSSC, pp. 1589-1595, Jul. 2006.
S. Bardsley et al., "A 100-dB SFDR 80-MSPS 14-Bit 0.35um BiCMOS
Pipeline ADC," IEEE JSSC, pp. 2144-2153, Sep. 2006.
A. M. A. Ali et al, "A 14-bit 125 MS/s IF/RF Sampling Pipelined ADC With
100 dB SFDR and 50 fs Jitter," IEEE JSSC, pp. 1846-1855, Aug. 2006.
S. K. Gupta, "A 1-GS/s 11-bit ADC With 55-dB SNDR, 250-mW Power
Realized by a High Bandwidth Scalable Time-Interleaved Architecture,"
IEEE JSSC, 2650-2657, Dec. 2006.
M. Yoshioka et al., "A 0.8V 10b 80MS/s 6.5mW Pipelined ADC with
Regulated Overdrive Voltage Biasing," ISSCC Dig. Techn. Papers, pp. 452453, Feb. 2007.
Y.-D. Jeon et al., "A 4.7mW 0.32mm2 10b 30MS/s Pipelined ADC Without a
Front-End S/H in 90nm CMOS," ISSCC Dig. Techn. Papers, pp. 456-457,
Feb. 2007.
K.-H. Lee et al., "Calibration-free 14b 70MS/s 0.13um CMOS pipeline A/D
converters based on highmatching 3D symmetric capacitors," Electronics
Letters, pp. 35-36, Mar. 15, 2007.
K. Honda et al., "A Low-Power Low-Voltage 10-bit 100-MSample/s Pipeline
A/D Converter Using Capacitance Coupling Techniques," IEEE JSSC, pp.
757-765, Apr. 2007.
EE315B - Chapter 9
B. Murmann
95
Per-Stage Resolution and Stage Scaling

D. W. Cline, "Noise, Speed, and Power Tradeoffs in Pipelined Analog to
Digital Converters", PhD Dissertation, UC Berkeley, November, 1995,
http://kabuki.eecs.berkeley.edu/~cline/thesis/thesis.pdf.
D. W. Cline et al., "A power optimized 13-b 5 MSamples/s pipelined analogto-digital converter in 1.2um CMOS," IEEE JSSC, Mar. 1996
Y. Chiu, "High-Performance Pipeline A/D Converter Design in DeepSubmicron CMOS," PhD Dissertation, UC Berkeley, 2004.
H. Ishii et al., "A 1.0 V 40mW 10b 100MS/s pipeline ADC in 90nm CMOS,"
Proc. CICC, pp. 395-398, Sep. 2005.
OTA Design, Noise

B. E. Boser, "Analog Circuit Design with Submicron Transistors,"
Presentation at IEEE Santa Clara Valley, May 19, 2005,
http://www.ewh.ieee.org/r6/scv/ssc/May1905.htm
R. Schreier et al., "Design-oriented estimation of thermal noise in switchedcapacitor circuits," IEEE TCAS I, pp. 2358-2368, Nov. 2005.
B. Murmann, "Thermal Noise in Track-and-Hold Circuits: Analysis and
Simulation Techniques," IEEE Solid-State Circuits Magazine, vol.4, no.2,
pp. 46-54, June 2012.
B. Murmann
EE315B - Chapter 9
96
Capacitor Matching Data

C. H. Diaz et al., "CMOS technology for MS/RF SoC," IEEE Trans. Electron
Devices, pp. 557-566, Mar. 2003.
A. Verma et al., "Frequency-Based Measurement of Mismatches Between
Small Capacitors," Proc. CICC, pp. 481-484, Sep. 2006.
V. Tripathi and B. Murmann, "Mismatch Characterization of Small Metal
Fringe Capacitors," CICC 2013.
Reference Generator
T. L. Brooks et al., "A low-power differential CMOS bandgap reference,"
ISSCC Dig. Techn. Papers, pp. 248-249, Feb. 1994.
B. Murmann
EE315B - Chapter 9
97
Time Interleaving
Boris Murmann
Stanford University
EE315B - Chapter 10
B. Murmann
Time-Interleaved ADC
M times increased throughput!
[Ken Poulton]
[W. Black and D. Hodges, JSSC, Dec. 1980]
B. Murmann
EE315B - Chapter 10
Typical Timing
Sampling
clocks of each
ADC
Each ADC samples for 1/Mth of the period
Still need very high acquisition bandwidth in each ADC slice

No free lunch
EE315B - Chapter 10
B. Murmann
Why Time Interleaving?

70
TI SAR
TI Pipeline
TI Flash
Other
SNDRHF [dB]
60
50
40
30
20
10 8
10
10
10
10
11
10
Sampling Rate [Hz]

B. Murmann
EE315B - Chapter 10
How Many Channels?

[El-Chammas, PhD Thesis, Stanford University, 2010]
Feasible
Solutions
Secondary factors
dominate the power
(e.g. flash resistor ladder)
Shallow power minimum
A complex optimization problem!

Luckily, first order analyses show that the optimum is shallow
EE315B - Chapter 10
B. Murmann
Time Interleaving Errors

0
o0
G0
ADC0
Ts+1
o1
x(t)
ADC1
G1
y[n]
(N-1)Ts+N-1
oN-1
GN-1
Gain
B. Murmann
ADCN-1
Offset
EE315B - Chapter 10
Timing Skew
Offset Errors
[Ken Poulton]
Signal independent noise pattern

B. Murmann
EE315B - Chapter 10
Impact of Offset Errors
[Gustavsson, p.262]
OS/FS
E.g. FS=1V, OS=1mV ENOB~9bits!
B. Murmann
EE315B - Chapter 10
Gain Errors
[Ken Poulton]
Very similar to amplitude modulation

B. Murmann
EE315B - Chapter 10
Impact of Gain Errors
[Gustavsson, p.266]
Gain
E.g. Gain=0.1% ENOB~10bits
B. Murmann
EE315B - Chapter 10
10
Timing Errors
Very similar to PM, jitter
[Ken Poulton]
Can be analyzed like jitter, using curvature of signal

autocorrelation (see earlier discussion)
Only difference is that a factor of N/N-1 shows up, since only
N-1 channels are skewed
See [El-Chammas & Murmann, IEEE TCAS 5/2009]
B. Murmann
EE315B - Chapter 10
11
Impact of Timing Errors

[Gustavsson, p.267]
Above chart is for M=4 channels
E.g. fin=100MHz, phase skew=3ps ENOB~9bits!
B. Murmann
EE315B - Chapter 10
12
Compensation of Interleaving Errors
Gain and offset

Several relatively simple techniques exist
Can compensate in analog and/or digital domain
Timing skew
Much more difficult to handle, fully digital compensation
often too complex for practical designs
Popular solutions
Measure errors in digital domain, compensate via
adjustable delay lines typically incur a jitter penalty
Measure errors in digital domain, compensate by
skewing equalizer taps
See overview paper by B. Razavi, CICC 2012
EE315B - Chapter 10
B. Murmann
13
Digital Detection and Correction
ADC0
y0[n]
Detection
Correction
~y1[n]
y1[n]
Sub-ADC
output
B. Murmann
Digital Backend
x(t)
ADC1
~y0[n]
Correction
EE315B - Chapter 10
Digitally corrected
sub-ADC output
14
Digital Detection and Analog Correction
B. Murmann
EE315B - Chapter 10
15
Foreground vs. Background

Calibration
Foreground Calibration
Background Calibration
B. Murmann
EE315B - Chapter 10
16
Example: Fully Digital Gain Error

Background Cal
[Tsai, TCAS1, 2/2009]
A similar approach can be used for digital offset

compensation
B. Murmann
EE315B - Chapter 10
17
Example: Foreground Analog Gain &

Offset Cal
gain
cntrl
comp
offset
cntrl
G2(0)
63
comp
C Calibration
Logic
Interleave 0
Decoder
5-bit
8-bit
Calibration
DAC
vga
offset
cntrl
G1(0)
Flash
other
interleave
(not shown)
[Verma, ISSCC 2013]
B. Murmann
EE315B - Chapter 10
Coarse
nonlinearity
compensation
in G2(0)
18
Clock Generator
Sources of Timing Skew
All traces and all transistors have mismatch, which can easily
result in ~10ps total timing skew for a complex clock tree
[A. Agarwal, CICC 2008]
EE315B - Chapter 10
B. Murmann
19
Minimizing Skew by Design
[Louwsma, JSSC 4/2008]
Global CML master clock times the sampling instants; local phase
gating for each ADC
Skew ~0.5ps (!)
B. Murmann
EE315B - Chapter 10
20
Hiding the Timing Skew Using a Global T/H
Typically not feasible in ultra-high-speed converters
Successfully used in moderately high-speed designs

S. Gupta et al., A 1-GS/s 11-bit ADC With 55-dB SNDR, 250-mW
Power Realized by a High Bandwidth Scalable Time-Interleaved
Architecture , JSSC 12/2006.
B. Setterberg et al., A 14b 2.5GS/s 8-Way-Interleaved Pipelined ADC
with Background Calibration and Digital Dynamic Linearity
Correction, ISSCC 2013.
EE315B - Chapter 10
B. Murmann
21
Global Bottom Plate Sampling
[Gustavsson, TCAS1, 9/2000]

X
B. Murmann
EE315B - Chapter 10
Big issue: parasitics

at node X
22
Better: Local Re-Timing & Fine Tuning

3.5ps tuning range, 250fs steps
[Kull, VLSI 2013]
B. Murmann
EE315B - Chapter 10
23
Typical Delay Cell
Causes a penalty in supply induced

jitter manageable if delay range is
small
K. Poulton et al., A 4GSample/s 8b ADC in 0.35um CMOS, ISSCC 2002

B. Murmann
EE315B - Chapter 10
24
Delay Control via Varactor DAC
Encoder
Control Bits
EE315B - Chapter 10
B. Murmann
25
Measuring the Skew Using Aux ADC

Auxiliary ADC
B. Murmann
MUX
...
..
[El-Chammas, VLSI 2010]
EE315B - Chapter 10
26
Clocking the Aux ADC

1
t
2
t
3
CAL
3
Period of CAL is e.g. 17/8 period of 1 - 8

B. Murmann
EE315B - Chapter 10
27
Blind Estimation of the Skew
[Haftbaradaran, CICC 2007]
Assuming well-behaved input signal statistics, it also possible to

detect the skew blindly
Key idea: pairwise signal correlations along the array must be

equal, unless there is skew
B. Murmann
EE315B - Chapter 10
28
System-Level Absorption of Skew

[Tsai, TCAS1, 2/2009]
Split equalizer into M paths and adapt coefficients separately
For complex equalizers typically found in ADC-based links, one

must carefully evaluate if the power overhead is justifiable
Can also combine this idea with analog time skew control
B. Murmann
EE315B - Chapter 10
29
Hierarchical Interleaving
[Doris, JSSC 12/2011]

B. Murmann
EE315B - Chapter 10
30
Quadrature Clocking
[Doris, JSSC 12/2011]
Hierarchical interleaving with 4x split in first rank

Drive the samplers directly with QVCO signal
Low jitter, small phase skew (need only small tuning range, if any)
B. Murmann
EE315B - Chapter 10
31
Examples (1)
[Greshishchev, ISSCC 2010]
160 SAR ADCs interleaved to resolve 6 bits at 40GS/s (!)
B. Murmann
EE315B - Chapter 10
32
Examples (2)
[Kull, VLSI 2013]
8 Interleaved SARs,
each running at
1.1GS/s
8.8GS/s, 50mW
B. Murmann
EE315B - Chapter 10
33
Examples (2)
[El-Chammas & Murmann, VLSI 2010]
8x Interleaved 5-bit Flash
12GS/s, 81mW
65nm CMOS
B. Murmann
EE315B - Chapter 10
34
Examples (3)
[Setterberg, ISSCC 2013]
14b, 2.5 GS/s ADC, 80-dB linear up to 1 GHz input

Employs 2 million logic gates for digital correction
Power dissipation = 24 Watts
B. Murmann
EE315B - Chapter 10
35
Examples (4)
[Danesh, VLSI 2011]
128 single slope ADCs interleaved to resolve 7 bits at 1GS/s

Power efficient (~26mW)
B. Murmann
EE315B - Chapter 10
36
Oversampling ADCs and DACs
Boris Murmann
Stanford University
B. Murmann
EE315B - Chapter 11
Overview
Oversampling A/D conversion
Decimation filters
Oversampling D/A conversion
B. Murmann
EE315B - Chapter 11
Recap
Sampling theorem
fs > 2fsig ,max
One good reason for sampling faster ("oversampling")

Can use lower order anti-alias filter
B. Murmann
EE315B - Chapter 11
Anti-Alias Filtering
B. Murmann
EE315B - Chapter 11
Quantization Noise
Ne(f)
2/12
fB
fs/2
Recall that the "noise" introduced by quantizer is evenly

distributed across all frequencies
Provided that quantization error sequence is "sufficiently
random"
Idea: Let's filter out the noise beyond f=fB!
B. Murmann
EE315B - Chapter 11
Digital Noise Filter (1)
f B 2
f s / 2 12
2
12
Total quantization noise at digital output is reduced proportional

to "oversampling ratio" M=(fs/2)/fB
B. Murmann
EE315B - Chapter 11
Digital Noise Filter (2)
Increasing M by 2x, means 3-dB reduction in quantization noise

power, and thus 1/2 bit increase in resolution
"1/2 bit per octave"
Is this useful?
Reality check
Want 16-bit ADC, fB=1MHz
Use oversampled 8-bit ADC with digital lowpass filter
8-bit increase in resolution necessitates oversampling by 16
octaves
fs 2 fB M = 2 1MHz 216
131GHz
B. Murmann
EE315B - Chapter 11
Noise Shaping
Idea: "Somehow" build an ADC that has most of its quantization

noise at high frequencies
Key: Feedback
B. Murmann
EE315B - Chapter 11
Noise Shaping Using Feedback (1)
Y ( z ) = E ( z ) + A ( z ) X ( z ) A ( z )Y ( z )
= E (z)
A(z)
1
+ X (z)
1+ A ( z )
1+ A ( z )
= E ( z ) HE ( z ) + X ( z ) H X ( z )
123
123
Noise
Transfer
Function
Signal
Transfer
Function
EE315B - Chapter 11
B. Murmann
Noise Shaping Using Feedback (2)

Y (z) = E (z)
A(z)
1
+ X (z)
1+ A ( z )
1+ A ( z )
1
424
3
1
424
3
Noise
Transfer
Function
Signal
Transfer
Function
Objective
Want to make STF unity in the signal frequency band
Want to make NTF "small" in the signal frequency band
If the frequency band of interest is around DC (0fB) we

achieve this by making |A(z)| >>1 at low frequencies
Means that NTF is <<1
Mans that STF 1
B. Murmann
EE315B - Chapter 11
10
Discrete Time Integrator
V ( z ) = z 1U ( z ) + z 1V ( z )
v ( k ) = u ( k 1) + v ( k 1)
z 1
V (z)
U (z)
1 z
1
z 1
z = e j T
"Infinite gain" at DC (=0, z=1)
EE315B - Chapter 11
B. Murmann
11
First Order Delta-Sigma Modulator
E(z)
X(z)
+
-
z-1
1-z-1
Y(z)
+
DAC
1
1
Y (z) = E (z)
+ X ( z ) z 1 = E ( z ) 1 z 1 + X ( z ) z 1
1
1
1+
1+
z 1
z 1
Output is equal to delayed input plus filtered quantization noise
B. Murmann
EE315B - Chapter 11
12
NTF Frequency Domain Analysis

He ( z ) = 1 z 1
e j T / 2 e j T / 2
He ( j ) = 1 e j T = 2e j T / 2
= 2e
T
2
T
j sin
2
T
= 2 sin
j
e
T
2
f
He ( f ) = 2 sin ( fT ) = 2 sin
fs
"First order noise Shaping"
Quantization noise is attenuated at low frequencies, amplified at

high frequencies
EE315B - Chapter 11
B. Murmann
13
In-Band Quantization Noise (1)
Question: If we had an ideal digital lowpass, what would be the

achieved SQNR as a function of oversampling ratio?
Can integrate shaped quantization noise spectrum up to fB and

compare to full-scale signal
Pqnoise =
fB
fB
f
2 2
2 sin
12 fs
fs
2 2
12 fs
df

f
2 df
fs
3
2 2 2fB
2 2 1
12 3 fs
12 3 M 3
B. Murmann
EE315B - Chapter 11
14
In-Band Quantization Noise (2)
Assuming a full-scale sinusoidal signal, we have
SQNR
Psig
Pqnoise
1 2 1
2
2
= 1.5 2B 1 2 3 M 3
=
2
2
2
1
4
1
24
3
Due to noise
12 3 M 3
shaping &
digital filter
1.76 + 6.02B 5.2 + 30 log( M )
[dB ]
(for large B)
Each 2x increase in M results in 8x SQNR improvement

9dB (1.5bits) per octave oversampling
EE315B - Chapter 11
B. Murmann
15
SQNR Improvement
Example revisited
Want 16-bit ADC, fB=1MHz
Use oversampled 8-bit ADC, first order noise shaping and
(ideal) digital lowpass filter
SQNR improvement compared to case without oversampling is
-5.2dB+30log(M)
8-bit increase in resolution (48dB SQNR improvement)

would necessitate M60
Not all that bad!
B. Murmann
SQNR improvement
16
31dB (~5 bits)
256
67dB (~11 bits)
1024
85dB (~14 bits)
EE315B - Chapter 11
16
DAC Requirements
Y (z) = E (z)
A(z)
1
+ X ( z ) DAC ( z )
1+ A ( z )
1+ A ( z )
DAC error is indistinguishable from signal

Means that DAC must be precise to within target resolution
For the previous example, this means that we need an 8-bit

DAC whose output levels have 16-bit precision
EE315B - Chapter 11
B. Murmann
17
Solutions
Trimming or calibration
Measure DAC levels during test or at power-up
Apply correction values to each level using auxiliary DAC
Dynamic Element Matching Algorithms

Shuffle DAC unit elements to obtain fairly precise "average"
output levels
Two ways
Data independent shuffling
Data dependent shuffling
Data dependent shuffling algorithms allow to push most of

the DAC "noise" outside the signal band
See e.g. [Carley, JSSC 4/1989], [Galton, TCAS II 10/1997],
[Vleugels, JSSC 12/2001]
Single bit DAC
B. Murmann
EE315B - Chapter 11
18
Data Independent Shuffling (1)
Carley, L.R., "A noise-shaping coder topology for 15+ bit converters,"
IEEE JSSC, vol.24, no.2, pp.267-273, Apr. 1989.
B. Murmann
EE315B - Chapter 11
19
Data Independent Shuffling (2)
Carley, L.R., "A noise-shaping

coder topology for 15+ bit
converters," IEEE JSSC, vol.24,
no.2, pp.267-273, Apr. 1989.
B. Murmann
EE315B - Chapter 11
20
Data Dependent Shuffling (1)
Select elements such that

each one is used (on
average) the same number
of times
DAC errors quickly sum to
zero; errors are pushed to
high frequencies
R.T. Baird, and T.S. Fiez, "Improved

DAC linearity using data weighted
averaging ," IEEE ISCAS, pp.13-16,
May 1995.
B. Murmann
EE315B - Chapter 11
Code 3
Code 4
Code 2
21
Data Dependent Shuffling (2)
B. Murmann
EE315B - Chapter 11
22
Single-Bit DAC
A single bit DAC has only two output levels
Even if these two levels are imprecise, the errors will only affect
gain and offset of the DAC and modulator
Tolerable in many applications
EE315B - Chapter 11
B. Murmann
23
Modulator with Single-Bit Quantizer (1)
Model
1-bit
code
Expected SQNR (from slide 15 with B=1)

2
SQNR
Psig
Pqnoise
1

9
2 2
= 2 2
= 2 M3
2
1
3
12 3 M
= 3.4 + 30 log(M )
[dB ]
E.g. M=128 SQNR=60dB
B. Murmann
EE315B - Chapter 11
24
Modulator with Single-Bit Quantizer (2)
Implementation example
[Schreier, p. 31]
Not all that great in terms of achievable SQNR, but sufficient for
some applications
E.g. digital voltmeter
See [van de Plassche, pp. 469]
B. Murmann
EE315B - Chapter 11
25
Simulated Response
B. Murmann
EE315B - Chapter 11
26
Spectrum
[Schreier, p. 39]
f/fs
Looks like there is some noise shaping, but SQNR=55dB is

lower than the expected 60dB
B. Murmann
EE315B - Chapter 11
27
Amplitude and Frequency Dependence

[Schreier, p. 40]
[M=256]
Erratic dependence on amplitude and frequency

Simple linear model fails to predict this behavior
Issue: Quantization error sequence is not "sufficiently random",

as assumed in the beginning of this discussion
B. Murmann
EE315B - Chapter 11
28
Quantization Error in 1st Order Modulator
Input x(n)
0.5
Quantization error e(n)
-0.5
0
200
400
600
800
1000
200
400
600
Sample index n
800
1000
0.5
-0.5
0
A complicated, but deterministic function of the input
B. Murmann
EE315B - Chapter 11
29
Aside: Quantizer Gain
Another issue is that the gain of the single bit quantizer is illdefined, but we assumed it to be unity in our analysis
The actual quantizer gain can be found from simulations, and

then plugged back into the linear model for better agreement
(and stability analysis using root locus, etc.)
References
Schreier, Sections 3.2 and 4.2
S. Ardalan, and J. Paulos, "An analysis of nonlinear behavior
in delta - sigma modulators," IEEE TCAS, vol.34, no.6, pp.
593-603, June 1987.
T. Ritoniemi, T. Karema, and H. Tenhunen, "Design of stable
high order 1-bit sigma-delta modulators," Proc. IEEE ISCAS,
pp. 3267-3270, May 1990.
B. Murmann
EE315B - Chapter 11
30
Tones
Since the quantization error is correlated with the input, the

shaped quantization noise contains spurious tones, some of
which lie in the signal band
The linear model cannot predict these tones
It is generally difficult to predict tonal behavior for arbitrary

inputs, even with a nonlinear model
Analytical results exist for DC and sine inputs, see e.g.
R.M. Gray "Spectral analysis of quantization noise in a singleloop sigma-delta modulator with DC input," IEEE Trans.
Comm., pp. 588-599, June 1989.
R.M. Gray et al., Quantization noise in single-loop sigma-delta
modulation with sinusoidal inputs," IEEE Trans. Comm., pp.
956-968, Sept 1989.
Interesting and intuitive to look at DC input as a worst case
B. Murmann
EE315B - Chapter 11
31
DC Input (1)
E.g. x(n)=0
Modulator generates an alternating sequence of 1s and 0s
Single tone at fs/2; no low frequency component
E.g. x(n)=0.001/2
Compared to previous example, only one in 1000 outputs
will change
Output has period of 1000T, and hence contains a low
frequency, in-band component
B. Murmann
EE315B - Chapter 11
32
DC Input (2)
For a DC input, the modulator output consists of discrete tones

("idle tones") with power and frequency given by
sin ( fkT )
Pk =
fk = k DC + 0.5 fs

where k is an integer, and <r> represents the fractional part of r

(r modulo 1)
Strongest tones occur for small k, due to reciprocal dependence
The plot on the following slide shows the total mean square
error due to in-band idle tones as a function of DC input (M=16)
B. Murmann
EE315B - Chapter 11
33
MSE due to Idle Tones
X/
B. Murmann
EE315B - Chapter 11
34
Idle Tone Considerations
Idle tones are known to be a significant issue in audio applications

The human ear can detect tones ~20dB below the
thermal/quantization noise floor
If idle tones are an issue, there are several options for mitigating their
impact
Larger oversampling ratio
Multi-bit quantizer and DAC
Dither
Superimpose a pseudorandom signal at the quantizer input to "whiten"
quantization noise
See e.g. Chapter 3 of Delta-Sigma Data Converters by Norsworthy, Schreier & Temes.
Overdesign by making quantization noise much smaller than

electronic noise from integrators
Noisy integrator(s) help randomize quantization error sequence
Higher order modulators

Naturally produce "more random" quantization error sequences
B. Murmann
EE315B - Chapter 11
35
Higher Order Modulators
Motivation: better SQNR for a given oversampling ratio, plus

improved idle tone performance as a side benefit
Commonly used architectures

Single quantizer loop with higher order filtering
Essentially a logical extension to the first order noise shaping
concept discussed previously
Cascaded, multi-stage modulators

Contain a separate quantizer in each stage
B. Murmann
EE315B - Chapter 11
36
Higher Order Noise Shaping
Lth order noise transfer function
HE ( z ) = 1 z 1
EE315B - Chapter 11
B. Murmann
37
In-Band Quantization Noise

Pqnoise =
fB
fB
f
2 2
2 sin
12 fs
fs
2 2
12 fs
f
2
fs
2 2L 2fB
12 2L + 1 fs
2 2L 1

12 2L + 1 M
2L
df
2L
df
2L +1
2L +1
For an Lth order modulator, every doubling of M results in an

increase in SQNR of 6L+3dB (L+0.5bits)
B. Murmann
EE315B - Chapter 11
38
SQNR [dB]
SQNR with Single Bit Quantizer
EE315B - Chapter 11
B. Murmann
39
Building a Second-Order Modulator (1)
General idea: Start with a first order modulator and replace

quantizer by another first order loop
-1
-1
B. Murmann
-1
-1
EE315B - Chapter 11
40
Building a Second-Order Modulator (2)
More general structure

E(z)
X(z)
+
z
1-z
a1
a2
Y(z)
z
1-z
a a z 2
HX (z ) = 1 2
D (z)
D( z ) = 1 z 1
B. Murmann
(1 z )
(z ) =
1
HE
+ a2bz 1 1 z 1 + a1a2 z 2 = 1
D (z)
a = a2 = 1 and b = 2
e.g. for 1
a1 = 0.5,a2 = 2 and b = 1
EE315B - Chapter 11
41
Boser-Wooley Modulator (1)

(Single-bit)
[Boser & Wooley, JSSC 12/1988]
a1=0.5 and b=1, but a2=0.5 (instead of 2)
Since the integrator is followed by a single-bit quantizer, a2 can

be scaled to reduce swing requirements in second integrator
B. Murmann
EE315B - Chapter 11
42
Boser-Wooley Modulator (2)
EE315B - Chapter 11
B. Murmann
43
Performance of 2nd Order Modulator

[Schreier, p.70]
Compared to first
order modulator,
SQNR is in "better"
agreement with
simple linear model
Improved idle tone

performance
-20
O
-30
1st order
O
-40
O
O
-50
-60
X
-70
-80
2nd order
-90
-100
-110
-120
8
16
32
64
128
256
512
1024
Oversampling Ratio
B. Murmann
EE315B - Chapter 11
44
Single Quantizer Modulators with Order >2
Most general (mathematical) filter decomposition
He ( z ) =
L0 ( z ) =
B. Murmann
1
1 L1( z )
Hx ( z ) =
Hx ( z )
He ( z )
L1( z ) = 1
L0 ( z )
1 L1( z )
1
He ( z )
EE315B - Chapter 11
45
Stability
Having more than two integrators in a feedback loop means that

the loop can be unstable (criterion = BIBO)
From the diagram of the previous slide, it is clear that the

stability of the loop mostly depends on L1(z), and therefore the
characteristics of the NTF
How about the nonlinear transfer characteristic of the quantizer?

Unfortunately, there is no crisp mathematical result that
would address this question for all possible configurations
One important, and general aspect of having a nonlinearity
in the loop is that the stability becomes dependent on the
signal (and also L0)!
In practice, designers rely on a combination of stability analyses

using the linear model (!), established heuristics, and time
domain simulations of the nonlinear model
B. Murmann
EE315B - Chapter 11
46
Stability Heuristics
Single-bit
First order modulator is stable (bounded integrator output)
with arbitrary inputs of less than /2 in magnitude
Second order modulator is known to be stable with arbitrary
inputs of less than /20 in magnitude, and for "reasonable",
slow varying inputs of magnitude <0.8/2, integrator outputs
are "likely" to stay within bounds
Lee's criterion: modulator is "likely" to be stable if
max[He()]<1.5
Multi-bit
A modulator with Nth order differentiation using an N+1 bit
quantizer is stable for arbitrary inputs with amplitude less
than half the quantizer input range (Schreier, p. 104)

EE315B - Chapter 11
B. Murmann
47
The Cost of Stability (Single-bit)
M=40, L=6
Max. Input [/2]
[Norsworthy, pp.156]
max[He()]
max[He()]
Higher out of band gain means higher attenuation in the signal

band and hence better SQNR
Unfortunately modulator becomes "less stable"
B. Murmann
EE315B - Chapter 11
48
Achievable SQNR
Diminishing return for order greater 5-6
EE315B - Chapter 11
B. Murmann
49
Topology Example: Single Feedback Loop
He ( z ) =
B. Murmann
1
1 + A( z )
Hx ( z ) =
A( z )
1 in band of interest
1 + A( z )
EE315B - Chapter 11
50
Topology Example: CIFB Architecture
[Schreier, p. 115]
Cascade of Integrators with Feedback
All zeros of NTF lie at DC for this structure, i.e. HE(z) = (1-z-1)N
Can create complex conjugate zeros by adding feedback paths

around pairs of integrators (cascade of resonators, CRFB
structure)
B. Murmann
EE315B - Chapter 11
51
Typical Design Procedure
"Cookbook design"
See e.g. Delta-Sigma Data Converters, by Norsworthy,
Schreier & Temes, Sections 4.4 and 5.6
Choose order based on desired SQNR and M
Design NTF using filter approximations (e.g. Chebyshev2)
Make sure to obey Lee's criterion
Determine loop-filter transfer function and evaluate

performance and stability using simulations
Determine implementation-specific coefficients
Scale coefficients to restrict integrator outputs to stay within
available range (dynamic range scaling")
Delta-Sigma Toolbox for MATLAB (by Richard Schreier)

http://www.mathworks.com/matlabcentral/fileexchange
Look under "Controls" and find "Delsig" toolbox
B. Murmann
EE315B - Chapter 11
52
"Cookbook" NTF Design Example (1)

% design parameters
L=4;
% order
M=64; % oversampling ratio

% stop-band attenuation; reduce if needed to make max(|He(w)|<1.5)
Rstop = 80;
[b,a] = cheby2(L, Rstop, 1/M, 'high');
% normalize to make He(z->inf)=1; needed for realizability
% makes first sample of impulse response of He equal to 1
% makes first sample of impulse response of A equal to 0
% (must have at least one delay around quantizer)
b = b/b(1);
% check Lee's rule; want max(|He(w)|<1.5 )
NTF = filt(b, a, 1)
[mag] = bode(NTF, pi)
EE315B - Chapter 11
B. Murmann
53

Transfer function:
1 - 3.998 z^-1 + 5.995 z^-2 - 3.998 z^-3 + z^-4
-----------------------------------------------------1 - 3.247 z^-1 + 4.013 z^-2 - 2.231 z^-3 + 0.4699 z^-4
mag = 1.459
He [dB]
-20
-40
-60
-80
-100
10
B. Murmann
-3
-2
10
Frequency [f/f s]
EE315B - Chapter 11
10
-1
54

% Loop filter transfer function
A = inv(NTF) - filt(1,1,1)
Transfer function:
0.7505 z^-1 - 1.982 z^-2 + 1.766 z^-3 - 0.5301 z^-4
--------------------------------------------------1 - 3.998 z^-1 + 5.995 z^-2 - 3.998 z^-3 + z^-4
% Check realizability
a = impulse(A);
a(1)
ans = 0
B. Murmann
EE315B - Chapter 11
55
Commercial Example
B. Murmann
EE315B - Chapter 11
56
Cascaded Modulators
Analog
In
DELAY
Digital
Out
DIGITAL
DIFFERENCE
Concept
Cascade of two or more stable (low order) modulators
Quantization error of each stage is quantized by the
succeeding stages and subtracted in digital domain
B. Murmann
EE315B - Chapter 11
57
Second Order (1-1) Cascade
Y1(z) = z1X(z) + (1 z1)E1(z)

Y2(z) = z1E1(z) + (1 z1)E2(z)
Y(z) = z1Y1(z) (1 z1)Y2(z)

= z2X(z) + z1(1 z1)E1(z) z1(1 z1)E1(z) (1 z1)2E2(z)
Y(z) = z2X(z) (1 z1)2E2(z)
Second order noise shaping using two first order loops!
B. Murmann
EE315B - Chapter 11
58
Properties
Order of overall noise shaping is equal to sum of modulator

orders
No stability issues
Improved idle tone performance

Input of second stage is "noise like"
Remaining quantization error from second stage is very
close to white noise
Cancellation of first stage quantization noise depends on

matching between analog and digital signal paths
Hard to suppress first stage quantization error by more than
40dB
Mismatch will affect idle tone performance
EE315B - Chapter 11
B. Murmann
59
1-1-1 Cascaded Modulator (MASH)
1
z-1
1
z-1
1
z-1
B. Murmann
EE315B - Chapter 11
60
Mismatch Sensitivity
B. Murmann
EE315B - Chapter 11
61
2-1 Cascade
B. Murmann
EE315B - Chapter 11
62
Mismatch Sensitivity
EE315B - Chapter 11
B. Murmann
63
Circuit Level Considerations
Finite OTA gain

Integrator leak
Dead zones
Nonlinearity
Electronic noise
OTA dynamic settling error, nonlinearity due to slewing
Capacitor voltage coefficients
Comparator hysteresis
Usually not a problem; simulations show that up to a few %
hysteresis can be tolerated
Unwanted mixing effects

E.g. if Vref contains fs/2, out of band noise will be mixed down
into signal band
B. Murmann
EE315B - Chapter 11
64
Integrator Analysis (1)
t/Ts
Qs
QI
n-1
CsVi(n-1)
CIVo(n-1)
n-1/2
CIVo(n-1/2) = CIVo(n-1) + CsVi(n-1)
CsVi(n)
CIVo(n) = CIVo(n-1) + CsVi(n-1)
n+1/2
EE315B - Chapter 11
B. Murmann
65
Integrator Analysis (1)
Assuming that Vo is sampled during 1, we have
CIVo ( n ) = CIVo ( n 1) + CsVi ( n 1)

CIVo ( z ) = z 1CIVo ( z ) + z 1CsVi ( z )
Vo ( z ) Cs z 1
=
Vi ( z ) CI 1 z 1
Unfortunately, this ideal expression holds only for infinite

amplifier gain
Let's look at impact of finite gain
B. Murmann
EE315B - Chapter 11
66
Finite Gain (1)
t/Ts
Qs
QI
n-1
CsVi(n-1)
CIVo(n-1)[1+1/A]
n-1/2
CsVo(n-1/2)/A
CIVo(n-1/2)[1+1/A] = CIVo(n-1)[1+1/A] +
CsVi(n-1) - CsVo(n-1/2)/A
CsVi(n)
CIVo(n)[1+1/A] = CIVo(n-1)[1+1/A] +
CsVi(n-1) - CsVo(n)/A
n+1/2
B. Murmann
EE315B - Chapter 11
67
Finite Gain (2)
Again, assuming that Vo is sampled during 1, we have

C
1
1
CIVo ( z ) 1 + = z 1CIVo ( z ) 1 + + z 1CsVi ( z ) s Vo ( z )
A
A
A
1 C
z 1 1 1 + s
A CI
Vo ( z ) Cs
g z 1
=
Vi ( z ) CI
1 Cs 1
1 [1 ] z 1
z
1 1
A CI
Vo ( z ) = [1 ] z 1Vo ( z ) + g z 1Vi ( z )
Finite gain results in "leaky integrator"

Some fraction of previous output is lost in new cycle
B. Murmann
EE315B - Chapter 11
68
Frequency Domain View
Limited gain at low frequencies (0, z 1)
H0 = H( z ) z =1 =
g
g
= A
1 [1 ]
But noise shaping relies on high integrator gain at low frequencies
B. Murmann
EE315B - Chapter 11
69
Required DC Gain
[Boser & Wooley, JSSC 12/1988]
Good practice to make OTA gain at least a few times larger than
oversampling ratio
B. Murmann
EE315B - Chapter 11
70
Integrator Noise
Noise splits into an input referred and output referred component

The two noise components are correlated
Luckily, only the input referred component is relevant
Well see that output noise is first-order shaped by the modulator
[Schreier, TCAS1, 2005]
EE315B - Chapter 11
B. Murmann
71
Integrators Input Referred Noise

2
=
v in,tot
kT kT 1
kT 1
+
+
Cs Cs 1 + 1
Cs 1 + x
{
x424444
1444
3
1
x = gm 2Ron
[Schreier, TCAS1, 2005]
2
Analysis of phi2 noise:
Calculate noise charge left behind
at Y after switch has turned off.
Charges at X and Y are equal and
opposite, so calculating charge on
Cs accomplishes the task.
B. Murmann
EE315B - Chapter 11
72
Second-Order Modulator
2
v in,tot1
PSD(f)
PSD(f)
2
v in,tot2
fs/2
fs/2
Y(z)
= 2 1 z 1 z 1
N2(z)
Y(z)
= z 2
N1(z)
EE315B - Chapter 11
B. Murmann
73
In-Band Noise
PSD(f)
PSD(f)
Digital filter
f
f
fb
fs/2
+
fb
PSD(f)
fs/2
fs/2
f
PSD(f)
OSR =
f
fs / 2
fb
fs/2
B. Murmann
f
fb
EE315B - Chapter 11
fs/2
74
In-Band Noise
PSD(f)
PSD(f)
fs/2
2
vin,tot1
fs/2
1
OSR
2
v in,tot2
2 1
3 OSR3
Total in-band thermal noise at modulator input:

2
v in,tot
B. Murmann
vin,tot1
1
2 1
2
+ vin,tot2
OSR
3 OSR3
EE315B - Chapter 11
75
Example of a Continuous Time Modulator

[Mitteregger, ISSCC 2006]
B. Murmann
4-stage amplifier with

feedforward compensation
Impractical for SC circuits
Great for CT delta-sigma
modulators
EE315B - Chapter 11
76
Multi-Mode Modulator
[Ouzounov, ISSCC 2007]
B. Murmann
Delta-sigma ADCs are

more amenable to BW
and DR reconfiguration
Very hard to
reconfigure pipelined
ADCs
Great for flexible, multistandard wireless

receivers
EE315B - Chapter 11
77
Decimation Filters
References
J. Candy, "Decimation for Sigma-Delta Modulation," IEEE
Trans. Communications, pp. 72-76, Jan. 1986.
Chapters 1 and 13 of Delta-Sigma Data Converters, by
Norsworthy, Schreier, Temes.
B.P. Brandt and B.A. Wooley, "A low-power, area-efficient
digital filter for decimation and interpolation," IEEE J. SolidState Circuits, pp. 679-687, June 1994.
E. Hogenauer, "An economical class of digital filters for
decimation and interpolation," IEEE Trans. Acoustics,
Speech and Signal Processing, pp. 155-162, Apr 1981.
Objectives
Remove out-of band quantization noise
Re-sample at lower frequency
Ideally at Nyquist rate
B. Murmann
EE315B - Chapter 11
78
Example
Decimation Filter (M = 256)
fS
Analog
Input
11.3
MHz
Modulator
1-Bit
(2nd-Order)
11.3
MHz
fN
44.1
kHz
Digital
Output
(16 Bits)
Quantization
Noise
Signal
Noise
Frequency
Frequency
Frequency
Filter must attenuate spectral components around NfN,

Otherwise they will alias onto signal after re-sampling
B. Murmann
EE315B - Chapter 11
79
Filter Requirements
Pass band 020kHz, transition band 2024.1kHz (f=4.1kHz),

stop band 24.1kHz5.65MHz
A digital FIR filter that meets these requirements would require

more than fs/f = 11.3MHz/4.1kHz 2800 coefficients
Impractical!
B. Murmann
EE315B - Chapter 11
80
Multi-Step Decimation
Key idea: Dont try to decimate down to fN in one step

Perform a gradual reduction of sampling rate + some filtering
E.g. Two-step decimation
Example: M1=64, fs/M1=176.4kHz
EE315B - Chapter 11
B. Murmann
81
Sinc Filter (1)
A popular, low complexity choice for stage 1 is the so-called

sinc-filter
From a time domain perspective, this filter simply computes the

average of several samples
y( n ) =
Frequency domain
N
H( z ) =
1 1 z
N 1 z 1
1 N 1
x (n i )
N i =0
f
sin N
fs
1
H( ) =
f
N
fs
j f ( N 1)
e fs
Zeros at multiples of fs/N

Make N=M1 to attenuate alias components!
B. Murmann
EE315B - Chapter 11
82
Cascade of K Sinc Filters

0
-10
-20
-30
-40
-50
K=1
-60
K=2
-70
K=3
-80
-90
-100
0 f B fN
fs1
M1
2f s1
M1
3f s1
M1
Frequency
Higher order means better rejection

But also more in-band droop
Can show that for Lth order noise shaping, an (L+1)th order sinc
filter is the best choice
B. Murmann
EE315B - Chapter 11
83
Sinc Filter Performance

Noise penalty relative to "brick wall" filter
Droop
[Norsworthy, p.30]
[Norsworthy, p.31]
Only about 0.14 dB increase in baseband noise for decimation

to an intermediate oversampling ratio of 4
If droop is undesired, it can be corrected downstream, using a

separate post-emphasis filter
B. Murmann
EE315B - Chapter 11
84
Sinc Filter Performance (2)
In addition to suppressing quantization noise, the filter must

attenuate out-of-band signals present at the modulator input
Worst case freqeuncy is fs/M1-fB
E.g. 50dB for sinc3, and intermediate oversampling of 4x
Any additional desired rejection must come from analog filter
at modulator input
[Norsworthy, p.31]
EE315B - Chapter 11
B. Murmann
85
Sinc Filter Implementation

Numerator Section:
X
1 z -1
Delay
Denominator Section:
X
1
1 z -1
Y
Delay
B. Murmann
EE315B - Chapter 11
86
Complete Filter Implementation

THIRDORDER
SINC
FILTER
1
IN
11.3
MHz
20
176.4
kHz
FIRST
HALFBAND
FILTER
(R=18)
M 1 = 64
22
88.2
kHz
SECOND
DROOP
22
16
HALFBAND
CORRECTION
FILTER
FILTER
44.1
(R=8)
(R=110)
kHz
M2 = 2
OUT
44.1
kHz
M3 = 2
0
Passband Ripple
= 0.01 dB
-10
-20
[Brandt & Wooley, JSSC 6/1994]
-30
-40
Second
Halfband
Filter
-50
-60
First
Halfband
Filter
Third-Order
Sinc Filter
-70
-80
-90
-100
0
20
40
60
80
100
120
140
160
180
200
FREQUENCY (kHz)
EE315B - Chapter 11
B. Murmann
87
Droop Correction
0.6
0.5
0.4
0.3
Droop-Correction
Filter
0.2
First Halfband
Filter
0.1
0.0
-0.1
-0.2
Second
Halfband
Filter
Third-Order
Sinc Filter
-0.3
-0.4
-0.5
-0.6
0
10
12
14
16
18
20
22
FREQUENCY (kHz)
B. Murmann
EE315B - Chapter 11
88
Implementation
SINC FILTER
INTEGRATORS
20
20
20
20
20
43 multiplications and
84 additions per output
sample
Can use serial

arithmetic to minimize
hardware area
Since output rate is
usually fairly low
Modulator
Output
20
CLK (11.3 MHz)

Processor
Input
Virtual Addr.
Logic
7
7
Addr
8
7
Data RAM
Addr
(128 x 22)
PROCESSOR
Coefficient/
Control ROM
R/W
Dout
(256 x 22)
Din
22
13
22
Ctrl.
Filter
Output
22-Bit Arithmetic Unit (AU)
EE315B - Chapter 11
B. Murmann
89
D/A Conversion Revisited

TS
yH (t)
Time
N-Bit Digital lnput Word x(kT)
b1
b2
b3
y(t)
Time
bN
Digital-to-Analog Interface
B. Murmann
x*(t)
Zero-Order
Hold
EE315B - Chapter 11
yH (t) Lowpass y(t) Analog

Output
Filter
90
Frequency Spectra
Magnitude
SpectralImage
Images
Spectral
Bands (Dirac Reconstruction)
Baseband Signal
|X*(f)|
fS/2
fS
2fS
3fS
fS/2
fS
2fS
3fS
fS/2
fS
2fS
3fS
|P Z (f)|
|YH (f)|
Frequency
EE315B - Chapter 11
B. Murmann
91
Oversampling
Magnitude
Spectral
Image
Bands
Spectral
Images
Analog Filter
|X(f)|
fB
fN
2 fN
(a)
(M-2) fN (M-1) fN
M fN
Analog Filter
|XM (f)|
fB
(b)
fS = M f N
Frequency
Oversampling greatly reduces reconstruction filter requirements
How to create oversampled DAC input from a Nyquist rate signal?
B. Murmann
EE315B - Chapter 11
92
Interpolation (1)
Can increase the sampling rate of a discrete time signal by a

factor of M, by inserting M-1 zero-valued samples between the
actual Nyquist rate samples ("zero stuffing")
Causes an M-fold periodic repetition of the baseband
spectrum
Image Bands
Spectral Images
Nyquist-rate
Input
fB fS =fN
2 fN
After
Zero
Insertion
(M-2) fN (M-1) fN
M fN
fB
fN
2 fN
(M-2) fN (M-1) fN fS =M fN
EE315B - Chapter 11
B. Murmann
93
Interpolation (2)
Why is this a good idea?
Can remove images and get wide transition band to play with
Simple reconstruction filter
Possibility of noise shaping
Build a high resolution DAC using a low resolution D/A interface
B
Digital M
Lowpass
Filter
fB
fS =M fN
fB
fS =M fN
Interpolator
Output
B. Murmann
EE315B - Chapter 11
94
Example
DIGITA L
Digital 16
Input
fN
Digital
Digital
16
Interpolator M f
ANALOG
Analog
Reconstruction
Noise
N
Shaper M fN
Analog
Output
Filter
1-bit D/A Interface
Digital noise shaper is essentially a digital sigma-delta loop

Shapes "truncation noise" that results from truncating 16-bit
word to a 1-bit output
EE315B - Chapter 11
B. Murmann
95
Spectra
Spectral Images
Digital
Input
fN
Interpolator
Output
Noise
Shaper
Output
2 fN
(M-2) fN (M-1) fN
M fN
Baseband Signal
Truncation Noise
Quantization
Noise
M fN
M fN
Analog
Output
Frequency
B. Murmann
EE315B - Chapter 11
96
Example
E(z)
16 +
X(z)
18
19
z1
z1
Y(z)
To 1-bit D/A Interface
+
18
Clipper
+
[Su &Wooley, JSSC 12/94]
Y(z) = z 2 X(z) + (1 z 1 ) 2 E(z)
Clipper prevents second integrator from overflowing

Digital "wrap around" would cause large errors
EE315B - Chapter 11
B. Murmann
97
Semi-Digital Reconstruction (1)

Digital
Input 1
DIN
n -Bit Shift Register
z 1
z 1
...
z 1
DIGITA L
ANALOG
a1
a2
...
an
[Su &Wooley, JSSC 12/94]

Analog Output, A OUT
HFIR (z) = a1 z1 + a2 z2 + ... + an zn
Attractive alternative to fully analog reconstruction filter

Build FIR filter with weighted analog outputs
B. Murmann
EE315B - Chapter 11
98
Semi-Digital Reconstruction (2)

Digital
Input
128-Bit Shift Register
DIGITA L
ANALOG
Weighted
Current
Sources
a
a
1
...
2
a 128
Current-to-Voltage
Conversion
I
out
I out
A OUT ( z ) =
a1 z
+ a2 z
+ a3 z
Analog
Output
+ + an z
DIN ( z )
H (z)
Linear if H(z) is independent of DIN(z)

EE315B - Chapter 11
B. Murmann
99
Measurement Results
Reconstruction Filter
Output
0
Noise Shaper Output
Analog Output
Power Spectrum (dB)
20
40
60
80
100
120
140
160
1k
10k
100k
1M
Frequency (Hz)
B. Murmann
EE315B - Chapter 11
100
Energy Limits in A/D Converters
Boris Murmann
Stanford University
EE315B - Chapter 12
B. Murmann
ADC Landscape in 2005

-6
10
-8
P/f s [J]
10
-10
10
-12
10

20
30
40
50
60
70
SNDR [dB]
80
90
100
110

B. Murmann
EE315B - Chapter 12

-6
10
-8
P/f s [J]
10
-10
10
-12
10

20
30
40
50
60
70
SNDR [dB]
80
90
100
110

B. Murmann
EE315B - Chapter 12
Questions
How much more improvement can we hope for? What is the

fundamental energy limit?
Will certain architectures continue to improve faster than others?

Are there any architecture-specific limits?
Will CMOS process technology scaling play a significant role in

future improvements?
How does absolute conversion speed affect energy

consumption and the projected limits?
B. Murmann
EE315B - Chapter 12
Fundamental Limit
1 VFS
2 2
SNR =
kT
C
Pmin = CVFS fs VDD
Emin =
VDD = VFS
Pmin
= 8kT SNR
fs
[Hosticka, Proc. IEEE 1985; Vittoz, ISCAS 1990]

EE315B - Chapter 12
B. Murmann

-6
10
-8
Energy [J]
10
-10
10
4x/6dB
-12
10

Emin
-14
10
20
B. Murmann
30
40
50
60
70
SNDR [dB]
EE315B - Chapter 12
80
90
100
110

-6
10
-8
Energy [J]
10
-10
10
4x/6dB
-12
10

Emin
-14
10
20
30
40
50
60
70
SNDR [dB]
80
90
100
EE315B - Chapter 12
B. Murmann
110
Normalized Plot
8
10

6
EADC/Emin
10
10
10
~10,000
100x in 8 years
~100
~2x in 8 years
10
20
B. Murmann
30
40
50
60
70
SNDR [dB]
EE315B - Chapter 12
80
90
100
110
Aside: Figure of Merit Considerations
There are (at least) two widely used ADC figures of merit (FOM)
used in literature
Walden FOM
Energy increases 2x per bit (ENOB)
Empirical
FOM =
Power
ENOB
fsnyq
Schreier FOM
Energy increases 4x per bit (DR)
Thermal
BW
Ignores distortion
FOM = DR(dB) + 10log
EE315B - Chapter 12
B. Murmann
FOM Lines
10
Energy [J]
10
10
10
10
-6
-8
-10
-12

Emin
-14
Walden FOM = 10fJ/conv-step

Schreier FOM = 170dB
20
30
40
50
60
70
SNDR [dB]
80
90
100
110
Best to use thermal FOM for designs with SNDR > 60dB
B. Murmann
EE315B - Chapter 12
10
Energy by Architecture
10
P/f s [J]
10
10
10
-6
-8
-10
Flash
Pipeline
SAR
Other
-12
20
B. Murmann
30
40
50
60
SNDR [dB]
EE315B - Chapter 12
70
80
90
100
11
Flash ADC
Ecomp
B. Murmann
Eenc
EE315B - Chapter 12
12
Encoder
Assume a Wallace encoder (ones counter)
Uses ~2BB full adders, equivalent to ~ 5(2BB) gates
Eenc 5 2B B Egate
EE315B - Chapter 12
B. Murmann
13
Matching-Limited Comparator
Cc
Cc
C
A 2VT
= A 2VT c
WL
Cox
Offset
Cc =
A 2VT Cox
+ Cc min
2VOS
Required
capacitance
3 VOS =
Simple Dynamic Latch

Assuming Ccmin = 5fF
for wires, clocking, etc.
Ecomp
2VOS
1 Vinpp
4 2B
Confidence
interval
3dB penalty
accounts for
DNL noise
SNR[dB] + 3
6
VDD
2B
2
2
2B 1
144 2 Cox A VT 2 + Cc min VDD
1
4
24
3
Vinpp
Matching
Energy
B. Murmann
EE315B - Chapter 12
14
Typical Process Parameters

Process
[nm]
AVT
[mV-m]
Cox
[fF/m2]
AVT2Cox /kT
Egate [fJ]
250
139
80
130
14
54
10
65
17
37
32
1.5
43
23
1.5
EE315B - Chapter 12
B. Murmann
15
Comparison to State-of-the-Art
-6
10
-8
10
Flash ISSCC & VLSI 1997-2012

Eflash65nm
Ecomp65nm
-10
Emin
[6]
P/f snyq [J]
10
[2]
[1]
-12
10
[5]
[4]
[3]
-14
10
-16
10
15
20
25
[1] Van der Plas, ISSCC 2006

[2] El-Chammas, VLSI 2010
[3] Verbruggen, VLSI 2008
B. Murmann
30
SNDR [dB]
EE315B - Chapter 12
35
40
[4] Daly, ISSCC 2008

[5] Chen, VLSI 2008
[6] Geelen, ISSCC 2001 (!)
16
Impact of Scaling
-6
10
Flash ISSCC & VLSI 1997-2012

Eflash250nm
Eflash130nm
-8
P/f snyq [J]
10
Eflash65nm
Eflash32nm
Emin
-10
10
-12
10
-14
10
15
20
25
30
35
40
SNDR [dB]
EE315B - Chapter 12
B. Murmann
17
SC Delta-Sigma Modulator
Assumptions
Model of first
integrator
Closed-loop gain =
C
C /2
Infinite transistor fT (Cgs=0)

Thermal noise factor ()
equals 1
Bias device has same

noise as amplifier device
Linear settling only
(no slewing)
Ci
Ci +
B. Murmann
Ci
2
2
3
Ceff = Ci (1 ) + CL =
EE315B - Chapter 12
1
2
Ci + CL Ci
3
3
18
SC Integrator Constraints
2
1 Vinpp
2 2
SNR
kT 1
4
Ceff OSR
Thermal noise
sets Ceff
Ceff
T /2
Ts / 2
= s
gm
1 ln SNR
ln
d
P = VDD
Settling time
sets gm
gm
gm sets power
gm
ID
EE315B - Chapter 12
B. Murmann
19
Pulling It All Together

Settling
Number
6474of
8
Eint DT = 64
{ kT SNR ln
Excess noise
Finite
VDD
penalty
2 }
VDD
1
1
SNR
V V
gm
DD
1
4inpp
24
3
ID
{
Supply
utilization Transconductor
efficiency
For settling with ~10 time constants and VDD=1V,

gm/ID=1/(1.5kT/q), Vinpp=0.5V, we have
E 200kT SNR
B. Murmann
EE315B - Chapter 12
20
Comparison to State-of-the-Art
-6
10
[8]
-8
[7]
P/f snyq [J]
10
-10
10
[1]
[2]
[4] [5]
[3]
[6]
DT ISSCC & VLSI 1997-2012
CT ISSCC & VLSI 1997-2012
First Integrator
EintCT
-12
EintDT
10
Emin
60
65
70
75
[1] Kauffman, ISSCC 2011

[2] Witte, ISSCC 2012
[3] Shettigar, ISSCC 2012
[4] Gealow, VLSI 2011
80
85
90
SNDR [dB]
95
100
105
[5] Chae, ISSCC 2008

[6] Pena Perez, ISSCC 2011
[7] Park, VLSI 2008
[8] Brewer, ISSCC 2005
EE315B - Chapter 12
B. Murmann
110
21
Overall Picture
10
P/f s [J]
10
10
-6
-8
Flash
Pipeline
SAR
-10
Other
E
flash32nm,cal
10
-12
E
E
E
10
-14
20
B. Murmann
30
40
50
60
70
SNDR [dB]
EE315B - Chapter 12
80
90
pipe
sar
CT
min
100
110
22
Discussion
Todays designs are very close to the predicted limits
The limit lines at high resolution are set by fundamental thermal

noise and architectural circuit overhead
Technology scaling wont help much
Sometimes it is even argued that technology scaling will

deteriorate efficiency, due to low VDD
Lets have a closer look at this
EE315B - Chapter 12
B. Murmann
23
A Closer Look at the Impact of

Technology Scaling
As shown previously:
2
1 VDD
1
E
VDD Vinpp gm

ID
Low VDD hurts, but this is not the only factor
Designers have worked hard to maintain (if not improve)

Vinpp/VDD in low-voltage designs
How about gm/ID?
B. Murmann
EE315B - Chapter 12
24
gm/ID Considerations (1)

30
180nm
90nm
gm/ID [S/A]
25
20
15
10
5
0
-0.2
-0.1
0.1
0.2
0.3
VGS-Vt [V]
0.4
0.5
0.6
Largest value occurs in weak inversion ~(1.5kT/q)-1

Range of gm/ID does not scale (much) with technology
EE315B - Chapter 12
B. Murmann
25

120
fT [GHz]
100
180nm
90nm
80
60
40
20
0
-0.2
-0.1
0.1
0.2
0.3
VGS-Vt [V]
0.4
0.5
0.6
fT is small in weak inversion region

Must look at gm/ID for given fT requirement to compare technologies
B. Murmann
EE315B - Chapter 12
26

30
180nm
90nm
45nm
gm/ID [S/A]
25
20
Example
fT = 30GHz
90nm: gm/ID = 18S/A
180nm: gm/ID = 9S/A
For a given fT, 90nm device

takes less current to produce
same gm
Helps mitigate, if not
eliminate penalty due to
lower VDD (!)
15
10
5
0
20
40
60
fT [GHz]
80
100
EE315B - Chapter 12
B. Murmann
27
ADC Energy for 90nm and Below

10
P/f snyq [J]
10
10
10
-6
-8
-10
-12
20
B. Murmann

90nm and below
30
40
50
60
70
SNDR [dB]
EE315B - Chapter 12
80
90
100
110
28
Efficiency versus Absolute Speed

=
10

2
FOM W [fJ/conv-step]
2006-2013
1998-2005
10
10
10 4
10
10
10
10
10
10
10
10
f s [Hz]
EE315B - Chapter 12
B. Murmann
29
Analysis
1

=
3
2

=
B. Murmann
Energy efficiency reduces with larger

gate overdrive
But, large gate overdrive is needed for

high transistor fT
fT is usually chosen as a multiple of the

sampling rate
EE315B - Chapter 12
30
Conclusions (1)
No matter how you look at it, todays ADCs are extremely well
optimized
The main trend is that the thermal knee shifts very rapidly toward
lower resolutions
Thanks to process scaling and creative design
At high resolution, we seem to be stuck at E/Emin~100

The factor 100 is due to architectural complexity and
inefficiency: excess noise, signal < supply, non-noise limited
circuitry, class-A biasing,
This will be very hard to change

Scaling wont help (much)
Some of the recent data points already use class-B-like
amplification
Can we somehow recycle charge?
B. Murmann
EE315B - Chapter 12
31
Conclusions (2)
At low resolutions, scaling will continue to help lower ADC

energy
Scaling will especially help improve the efficiency of GS/s A/D

converters
This is badly needed for high-speed data links
(electrical and optical)
For non-incremental improvements, we must explore new ideas

in signal processing that tackle ADC inefficiency at the system
level
Compressed sensing
Finite innovation rate sampling
Other ideas?
B. Murmann
EE315B - Chapter 12
32
Example: Finite Innovation Rate

Sampling of Ultrasound Signals
Low BW after filtering
High BW
Analog
Filter
Reconstruction
[Inspired by theoretical work of Prof. Yonina Eldar]
~100x reduction in sampling rate

B. Murmann
EE315B - Chapter 12
33

Ee315b Reader 2013 Stanford Dataconverters

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Ee315b Reader 2013 Stanford Dataconverters

Încărcat de

Drepturi de autor:

Formate disponibile

EE315B

VLSI Data Conversion Circuits

(This page is intentionally left blank)

Benefits of digital signal processing

Data converters often present a performance bottleneck

A/D Converter ca. 1954

Data Converter Applications (1)

Data Converter Applications (2)

Computing and Control

[Poulton, ISSCC 2003]

High performance digital

Schvan, ISSCC 2008

Time interleaved architecture using 160 (!) SAR ADCs for

A typical cell phone contains:

A total of 19 data converters!

Low-cost, single chip solutions require embedded data

Example: 802.11g Wireless LAN chip

Trend Toward Lower ADC Energy

ISSCC & VLSI 2006-2013

B. Murmann, "ADC Performance Survey 1997-2013," [Online]. Available: http://www.stanford.edu/~murmann/adcsurvey.html

Acquire a thorough understanding of the basic principles and

Staff and Website

Lecture videos are provided on the web, but please come to

Coursework web page

Please visit the discussion forum regularly

Prior exposure to Spice, Matlab

Analog Circuit Sequence

Midterm Project: (40%)

Final Exam (40%)

Please remember you are bound by the honor code

But if you are found cheating it is very serious

Save yourself and me a huge hassle and be honest

For more info

Tools and Technology

M. Pelgrom, Analog-to-Digital Conversion, Springer, 2010

Gustavsson, Wikner, Tan, CMOS Data Converters for

A. Rodrguez-Vzquez, F. Medeiro, and E. Janssens, CMOS Telecom

W. Kester, The Data Conversion Handbook, Newnes, 2005

B. Razavi, Data Conversion System Design, IEEE Press, 1995

R. Schreier, G. Temes, Understanding Delta-Sigma Data Converters,

J. G. Proakis, and D. G. Manolakis, Digital Signal Processing,

Ideal sampling, reconstruction and quantization

Nyquist-rate ADCs and DACs

Oversampled ADCs and DACs

Data converter performance trends and limits

Data converter testing, simulation techniques

Sampling, Reconstruction, Quantization

The Data Conversion Problem

Real world signals

How to "undescretize" in time and amplitude

We'll fist look at these building blocks from a functional, "black

Uniform Sampling and Quantization

Most common way of performing A/D

Aliasing Example (1)

v sig ( t ) = cos ( 2 fin t )

Aliasing Example (2)

Aliasing Example (3)

The frequencies fsig and Nfs fsig (N integer), are

In order to prevent aliasing, we need

The sampling rate fs=2fsig,max is called the Nyquist rate

Brick Wall Anti-Alias Filter

Practical Anti-Alias Filter

Need to sample faster than Nyquist rate to get good attenuation

How much Oversampling?