CMOS Mixed-Signal Circuit Design
CMOS Mixed-Signal Circuit Design
CMOS Mixed-Signal Circuit Design
• IEEE
IEEE PRESS
I Visit www.wiley.com/ieee I
~WILEY
wiley.com
9780470 290262
I
54
Board
Chief
CMOS
a.Malik
S. Nahavandi
Mixed-Signal Circuit Design
M. S. Newman
W. Reeve
Second Edition
Information Services (SIS)
~ Editor
Editor
iety, Sponsor
rs
+IEEE
IEEE PRESS
~
WILEY
I A JOHN WILEY & SONS, INC., PUBLICATION
I
Brief Contents
Chapter 1 Signals, Filters, and Tools 1
Chapter 2 Sampling and Aliasing 27
Chapter 3 Analog Filters 73
Chapter 4 Digital Filters 119
Chapter 5 Data Converter SNR 163
Chapter 6 Data Converter Design Basics 203
Chapter 7 Noise-Shaping Data Converters 233
Chapter 8 Bandpass Data Converters 285
Chapter 9 A High-Speed Data Converter 301
vi
Contents
Preface xv
vii
viii Contents
322
325
Preface
Designs that combine analog circuits with digital signal processing, DSP, are called
mixed-signal designs, MSDs. Designs that use both digital and analog circuits but no
DSP, like a 555 timer or a pipeline ADC, are not, by this definition, MSDs. These
mixed-mode or mixed analog/digital circuits aren't as robust as circuits designed using
MSD techniques because they require precise components and often calibrations or
tuning. The use of DSP in a mixed-signal circuit relaxes the requirements placed on the
analog components (important) by overcoming the shortcomings, like low transistor gain
and poor matching, found in nanometer CMOS.
This book provides a tutorial introduction to MSD techniques. The content is
suitable for use in a senior/graduate electrical engineering course or as a reference for a
working engineer doing MSD. The assumed background of the reader is a course in
signals/systems, and courses in digital and analog integrated circuit design in CMOS
technology.
Chapter I covers basic signals, filters, and tools. The reader may be inclined to
skip this material; however, the author would recommend against it unless the reader can:
1) Give an example of an imaginary signal (answer using words and without math or
equations). 2) Explain why imaginary numbers are used. 3) Explain why I/Q signals are
used. 4) Describe the differences between delaying and non-delaying integrators. While
we could keep listing questions, even the most seasoned systems person will likely
benefit from reading and thinking about the material in Ch. 1. Don't skip it! Make an
attempt to understand what is going on in the discussions and not just to understand the
math. Once the reader feels they have mastered this material they should try to provide
physical metaphors to describe a concept or equation (for example, use a cup, water, and
a bucket to describe Eq. [1.56]). Remember that math is easy. Understanding what's going
on is tough.
XVI Preface
Preface
1
.g.
are key topics because sampling is
down-sampling) and interpolation
cessing and thus MSD. In addition,
3 and 4 cover analog and digital
flat can be used in MSD. Chapter 5
chapters in the book describe the
Ie associated trade-offs.
Signals, Filters, and Tools
and information with the goal of
ight. Hopefully, this will lead the
.sign problems. For example, the
ictive integrator. Why? Why can't a
converters using passive elements. Mixed-signal circuit design requires a fundamental knowledge of signals, signal
Ie topologies are discussed leading processing, and circuit design. In this chapter we provide an overview of signals, filtering,
st delta-sigma data converters. and the mathematical tools. The chapter may be a review for the reader; however, we use
ons of MSD techniques is speed. it to ensure a good foundation to build on in the coming chapters and to provide a quick
results are circuits that are precise, reference for the mathematical formulas we'll use throughout the book.
ample of a fast circuit that needs to
techniques, its design can be more 1.1 Sinusoidal Signals
( design, requiring special layout
Let's take a fundamental look at the sinewave. While there are many ways (equations and
. in the book, Ch. 9, presents a
formulas) of representing a sinewave, we must remember it is an empirically determined
flat uses MSD techniques that may
function. Naturally occurring signals, shapes, or constants are determined or described
= in nanometer CMOS technology
through empirical measurements or observations. For example, rt is determined by
so provide a good summary of the
dividing the circumference of a circle by its diameter
circumference
rr= (1.1)
diameter
, students, colleagues, and friends The goal of this section is to provide intuitive discussions that will help create a deeper
Ambrose, Hemanth Ande, Jeanne
understanding of what's going on in a circuit or system.
e Bermak, Bertan Bakkaloglu, Joe
-lhardt, Gilda Garreton, Shantanu 1.1.1 The Pendulum Analogy
Richard G. Lyons, Pui-In Mak, Consider the (ideal, that is, lossless) moving pendulum seen in Fig. 1.1a. In this figure the
Tewksbury, Donna Welch, Thad pendulum is moving back and forth between Points I and 3 repeatedly over time. As the
pendulum leaves Point I it starts out slow, gaining maximum speed as it passes Point 2,
and finally reaching Point 3. At Point 3 it stops and reverses direction. The time it takes
R. Jacob (Jake) Baker to make this complete journey back to the starting point, Point I in this discussion, is the
period, T. In Fig. LIb we plot the movement of the pendulum along the arched path. We
record the position to define a function, j(t), that indicates the pendulum's position at a
specific time
Position =j(t) =j(t + nT), where n is an integer ( 1.2)
This signal, we should all recognize, is a sinusoid or sinewave which repeats its position
with a frequency.j, of liT.
2 CMOS Mixed-Signal Circuit Design Chapter 1 Signals, Filters,
+1
Figure 1.1 Physical interpretation of a sinewave. II
4
Next, consider the circle seen in Fig. 1.2. One complete rotation around this circle
(360 degrees or 2n) is analogous to one complete movement (swing) of our pendulum. 'v
We started plotting the pendulum's position at Point 2 in Fig. 1.1b (Point 2, t = 0, in Fig. 31t
2
1.2). After TI4 we reach Point 3 in Fig. 1.1 b. This corresponds to a 90 degree, or n12,
I 27(
I movement in our circle. After another TI4 seconds we pass back through Point 2. In the y
I circle we've moved 180 degrees. This continues with each swing of the pendulum
I t = TI8
-I- corresponding to a complete revolution around the circle. Note that we do have some
I
I
.1
I
I
-. Point 3
Length ofO_7C
angle of 45 de;
(c) Point ii
y
-+:'
-I
-.
I
Point 2
I
I
/
-I- t = Tl2, T + Tl2, ... Length of 0 an
t=O, T, 2T, ...
I
I
I angle of 180 d,
I
I
t = 3T14, T + 3TI4, ...
»> Point I (e) Point iv
limitations when representing the movement of the pendulum with this circle. For
example, what is the amplitude of the sinewave (what is the relative position of the
pendulum along the arched path)? We'll address these concerns in a moment. For now
let's write, assuming we are using radian angular units,
t3 Position = sin (211' t) = sin(211/o • t) (1.3)
------- Pendulum This function, the sine function, tells us our relative position along the arc (the argument
of this function is the angle which relates to the position on the circle in Fig. 1.2). Point 2
ill
lmo!ion. corresponds to the function having a value of 0 (and times, t = 0, T/2, T, 3T/2, 2T, ... ),
oint 1 Point 3 to a value of + I, and Point 1 corresponds to -1. Finally, remember that the values
he period T. of the sine function in Fig. I.l.b, and Eq. (1.3), are determined empirically from
measured data (e.g., plotting the pendulum's position along the arched path against time).
Describing Amplitude in the x-y Plane
Examine the sinewave in Fig. 1.3a. For the moment we won't concern ourselves with the
actual distance the pendulum swings. In Fig. 1.3b we represent the sinewave, at Point i
(and Point vi), as a zero length vector along the x-axis (the amplitude of the sinewave is 0
t, time
at this point in time). As we move towards Point ii in Fig. 1.3a the length of the vector
---~--~x
,lete rotation around this circle time
ent (swing) of our pendulum.
ig. l.lb (Point 2, t = 0, in Fig.
~
Length of 0 and angle of 0
:l3. 0 or 211
ponds to a 90 degree, or rrJ2, 2
0
0° or 360° (b) Point i
270
s back through Point 2. In the y y
~ach swing of the pendulum
t= Tl8 t= TI4
. Note that we do have some
-----¥'---"~~x ----+----"~~x
Figure 1.3 A vector swinging around the x-y plane changing both length and angle
is used to represent a sinewave.
the pendulum in Fig. 1.1.
4 CMOS Mixed-Signal Circuit Design Chapter I Signals, Filters, a
---_ - _ . _ - - -
..
increases (here indicating an increase in both the x and y directions), Fig. 1.3c. At T!4 we
are at Point iii in Fig. 1.3a. As seen in Fig. 1.3d the length of the vector is 1 and the angle
The resulting signal, whe
is 90 degrees. Continuing on towards TI2 (Point iv) in Fig. 1.3a the vector is shrinking
once again finally reaching a length of 0 and an angle of 180 degrees, Fig. 1.3e. In Fig. (sinewave) with a length (p
1.3f, Point v, the length is I and the angle is 270 degrees. The key point here is that a in a circle with a frequency
sinewave is represented in the x-y plane by a vector that is changing length and rotating
around in a circle. Knowing this we can simply represent the sinewave by its peak value
and the associated angle between the x-axis and the vector as indicated in Fig. 1.3. For
example, as seen in Fig. 1.3d, the peak value of the sinewave occurs when the angle is 90 where we define the magnit
degrees (TI4) so we could write
sin2n/o ' t ~ lL90° ~x O,y= 1 (104)
knowing the vector representing the sinewave is actually rotating around in the x-y plane (a) Shifti
with time (and a frequency J;, ). Note that we could plot the sinewave more accurately by Fig. 1
adding a third axis, time (the z-axis), and showing the corresponding, corkscrew looking,
3-dimensional plot.
In-Phase and Quadrature Signals
Consider moving the sinusoid in Fig. 1.3a a quarter of a cycle (90 degrees, rrJ2, or a
delay, tdela )" of T14) earlier in time, Fig. lAa. We write, for this time-shifted signal,
. (2n
sm . (360
T· t +2n ') =sm T' ( 1+'4T))' =sm
T' l + 90) =sm. (2n . (2.njo'I+2
r n) (1.5)
As the reader probably already knows, this signal function is called cosinusoidal or
simply a cosine signal and is described using Figure 1.4 Shil
to represent the signal and not show the rotating vector with time. Note that we can say
the cosine signal is in-phase (l) since we are representing it above with a zero degree Figure 1.5 Show
phase shift. We could also say that the sine signal, since it's shifted in time by a quarter
cycle, Eq. (1.4), has a quadrature (Q) phase shift. Note what happens, looking at the Notice that the sigr
vector representations of the sinusoids in Figs. 1.3b-f and lAb, if we create a signal by multiplying them together ar
adding a sine signal to a cosine signal or
)S Mixed-Signal Circuit Design Chapter I Signals, Filters, and Tools 5
lirections), Fig. 1.3c. At TI4 we S IQ(t) cos 21".:/0 . t + sin 211:fo . t (1.9)
of the vector is I and the angle
The resulting signal, when plotted in the x-y plane (Fig. 1.5), results in a vector
1.3a the vector is shrinking
. 180 degrees, Fig. 1.3e. In Fig. (sinewave) with a length (peak amplitude) of Jl2;li or J2
that simply rotates around
s. The key point here is that a in a circle with a frequency of liT (= 10 ). In more general terms, 1.6,
is changing length and rotating
the sinewave by its peak value SIQ(t) = AI' cos 211:fo . t+ AQ' sin2yifo . t = JA; +A~ . cos (21if" . t+tan- 1 Al (1.10)
.or as indicated in Fig. 1.3. For
ave occurs when the angle is 90 where we define the magnitude and phase as
I 2 2
ISIQ(t) I LSIQ(t) = JA I +AQ Ltan- I AI (1.11)
),y= 1 (104)
rotating around in the x-y plane (a) Shifting the sinewave in
1e sinewave more accurately by Fig. 1.3a earlier by Ti4. y
+1
responding, corkscrew looking, t= TI2
'. unit circle
time
\"
_~_.......:;;ii.:a·Ir'~.~ x
, . t) (1.6) Before going too much further, we should ask why we would want to add I and Q
signals? The answer is that since the IIQ sinusoids are at the same frequency, shifted in
tle signal in Fig. 1.3a by a phase time by a quarter cycle, we can transmit them with, for example, changes in their
at the sine signal lags the cosine respective amplitudes and increase the information sent for a given bandwidth. Note that
1 we talk about phase shift it's we can't shift the sinewave seen in Fig. 1.3a by 180 degrees, or Tl2, and add it to an
11. It doesn't make sense to talk unshifted sinew ave since we would get no signal at all!
requencies. Finally note that the
y
(1.7)
)eriod.
e signal. The peak value of the
uld write
,y 0 (1.8) Angle is 45°
Angle is tan- 1 ~~
Further, with a little algebn
Figure 1.6 Again, showing how an I1Q signal can be represented in the x-y plane. The next thing we h
Plotting the real componen
to simplify the math!, To move towards this goal we'll develop the complex, or z-, plane we've always used. To plot
and the frequency-domain representation of signals. the factor ) in the compl,
shouldn't be included when
1.1.2 The Complex (z-) Plane
e, of the complex number, 1
Let's attempt (and fail using the x-y plane) to simplify our mathematical description of the
IQ signal given in Eq. (1.10). Recall the following Taylor series expansions
(1.12) z
6 z'
k
-+ - (U3)
6! 8!
7! + 9! - ... (1.14)
cosk+sink k 2 k 3 J(4
l+k----+-+
_ k6 e 8 9
- +k- +k- ( 1.15) Figure 1.7 Th(
2! 3! 4! 5! 61 7! 8! 91 - ...
Comparing Eq. (LIS) to Eq. (1.12) we see that we are close to writing the Taylor's series
for e k • Why is this important? Perhaps the simplest explanation is that if we can represent Example 1.1
sinewaves using exponentiation, then mUltiplying two sinewaves, or shifting a sinewave In the complex plane, pi,
in time, can be performed using simple addition (of exponents). As seen in Fig. 1.8
I
I The question now is how do we modify things to ensure that all terms are added
Le f2 11/Q'/ = 2rc/o • t. The ~
so that Eq. (1.15) matches Eq. (1.l2)? Let's look at the first discrepancy (-1)· ~. The
only way to change the polarity of this term is take the square root of -1 and move it
I
I
inside with k 2 • As the reader may know instead of writing R
for all of these terms we
simplify things and write
}=R (1.16)
Numbers using} (or i) are called imaginary or complex numbers (the reason for using the
name imaginary will be explained in Ex. 1.1). Tmaginary numbers are invaluable for
time-shifting and scaling sinusoidal signals. We now rewrite Eq. (1.12) using) as
e j2n!o-t = cos 2rc/a . t +} . sin
.ks k6 .e
3!+4!+J5T 6!-J"7T+'" (1.17)
Figure
VlOS Mixed-Signal Circuit Design Chapter 1 Signals, Filters, and Tools 7
I be represented in the x-y plane. The next thing we have to discuss is plotting complex numbers. Examine Fig. 1.7.
Plotting the real component, x, of a complex number, x + }y, follows the same methods
develop the complex, or z-, plane we've always used. To plot the imaginary component, we now use the y-axis. Note that
the factor } in the complex number simply indicates the imaginary component and
shouldn't be included when finding the magnitUde of the number, + y2 or the phase, jx2
8, of the complex number, tan-If.
)tir mathematical description of the
,or series expansions
im,y
+ ... (1.12) z - plane
z=x+}y y
(1.13)
(1.14) x
(1.15) Figure 1.7 The complex plane, plotting the imaginary number
---------......------------------------------
8 CMOS Mixed-Signal Circuit Design Chapter 1 Signals, Filters, a
plane, on the unit circle, without changing amplitude (both the real and the towards this method of ch,
imaginary components oscillate back and forth between + 1 and -1). One complete Eq. (1.25) as vin(f) or sim
rotation takes lifo, or T seconds. We need to pause a moment and ask "If the discussed in the next sectiOl
magnitude of e j2nfo-t is a constant value of 1 isn't it a DC signal?" (answer: no) Returning to our cor
When a DC signal is represented in the x-y plane it doesn't rotate or change
amplitude like a sinewave does. (Well, since the frequency of a DC signal,/, is 0,
you could say it does rotate around the x-y plane just like a sinewave but since
or, using Euler's formula,
T == 00 (= llf) it never leaves the x-axis). What this, eJ2 rr!o'/, imaginary signal can
be used for is to introduce or represent delay (phase shift).
To understand this last statement in more detail let's write
A . cos 2rr./o . t Re{ A . e l2 "J" , } --1> AL.O (1.21 ) The magnitude response of 1
Suppose we want to delay this signal by td' We can represent a delay at a particular
frequency, 1" using
I':.:1 I J(1 + cos 21
or using
ej2rj~'(-/d) --1> 1L.2rc/o . (-td) --1> 1L.2rr.· -:.;: (1.22)
To simplify the notation we can drop the real indication, Re {}. We could also describe the
delayed cosine signal in terms of angle notation as The phase response is given
Again noting that when we use angle notation, the x-y plane, or the complex plane, we Notice at / == 1!(2td) the ph,
need to know the frequency,l" of the input signal since this information isn't present in
(1.29) and
these representations.
lmplitude (both the real and the towards this method of characterization let's write the frequency domain description of
Jetween + 1 and -I). One complete Eq. (1.25) as vin(f) or simply v,n' Representing a signal in the frequency domain will be
pause a moment and ask "If the discussed in the next section.
m't it a DC signal?" (answer: no)
Returning to our comb filter in Fig. 1.9 we can write
plane it doesn't rotate or change
! frequency of a DC signal, f, is 0, (1.26)
me just like a sinewave but since
his, e j21t!o'/, imaginary signal can or, using Euler's formula,
.ase shift) .• Real Imaginary
r-.. ~ ~ ..--------.,
can represent a delay at a particular I~:I I J(l +cos2rrf· (-td»2 + (sin2rrf· (-td»)2 b(l +cos2nj- td) ( 1.28)
or using
(1.22)
I+cosx 2COS2~ ( 1.29)
simply
(1.23)
(1.30)
jon, ReO. We could also describe the
The phase response is given by
(1.24) tan-l [ sin2rrf· (-td) ] (1.31)
1 +cos2nj- (-ld)
x-y plane, or the complex plane, we
since this information isn't present in
Notice at 1/(2td) the phase is tan- 1(010), which evaluates to ± 90 degrees. Using Eq,
(1.29) and
sin x 2 sin E. cos E ( 1.32)
2 2
the phase response is
useful, filter called a comb filter, Fig.
this filter let's discuss notation. If our LVVO.UI
In
= n(-td) I for J< l/(2t d) (1.33)
Note that the phase response is linear indicating constant delay through the filter,
if· t ( 1.25) important for distortionless filtering. Figure 1.10 shows the magnitude and frequency
lry time, which we've done up to this response for a comb filter (at this point the reason the filter is called a comb filter should
it itself (at the same frequency as the be obvious).
: signal's frequency,!, and looking at
ude and phase shift vary). To move Example 1.2
Figure 1.11 shows one possible implementation of a comb filter. A 50 Q
characteristic impedance co-ax transmission line with an electrical length of 5 ns
(delay) is used for the delay element (note how the transmission line is terminated
with a 50 ohm resistor and it is assumed that 50 n « 5ill so that the 5ill
resistors don't load the output of the transmission line). Determine the comb
filter's characteristics (transfer function, vojvin ). Verify your answer with SPICE.
ilter.
10 CMOS Mixed-Signal Circuit Design Chapter I Signals, Filters,
L VOU (
Vi"
degrees
2
- 45- -
,.
- 90 - .
_1 l f(Hz) _1 3
1.0V--,--;-:--f-
1
-2td
O.9V
4td 2td td 2td 2t d O.6V
4td O.7V
O.6V
O.5V
Figure 1.10 Magnitude and phase response of the comb filter in Fig. 1.9. O.4V
O.3V .... __
ww -#-- .
O.2V -~~-- .. ---~-
O.lV .. ---- --~~-
In order to perform the addition operation seen in Fig. 1.9 we use the two 5k o.OV-'----+-
100
resistors seen in Fig. 1.11. The current through the top resistor must equal the
50 MHz
current through the bottom resistor or
Figure 1.12 Si
Vinej2n{(-ldl VOIlI VollI - Vin
(1.34)
5k 5k output will have an amI
and thus Vout is the average of the signals on the top and bottom of the resistors output is lagging the inl
Fig. 1.13. Note, in this
(1.35) signal to propagate thro.
filter's behavior follows
The result is our derivation of the comb filter's magnitude response in Eq. (1.30)
is scaled by 2 or
I how the filter eliminates from its output, input signals at frequencies that are o.ovt-·
1-
multiples of 1I2td. This filter can be very useful in communication systems where
it's used to isolate, and prevent crosstalk between, transmission channels.
In order to ensure that we can sketch time-domain signals from frequency
::::1+
I
1-
-n.BV ------;.-
domain plots consider the case when the input has I V peak amplitude and a
frequency of 50 MHz. We can see from these plots, and the equations, that the
-l.ov-l----i--
Ons 4ns
1.0V T"""'<:,...--r-.-..,.---~/~;;:_'_'c:..c..:"'--,---__:"....,..~--__,
0.9V
O.8V
0.7Y
O.sv
O.SY
the comb tilter in Fig. 1.9. O.4V
D.3Y
O.2V
D.W
in Fig. 1.9 we use the two 5k O.OV-'---i--4-----i--------¥----i------4
the top resistor must equal the
50 MHz
Figure 1.12 Simulating the operation of the comb tilter in Fig. 1.11.
Voul-Vin
(1.34)
5k output will have an amplitude of 0.707 V and a phase shift of -45 degrees (or the
I and bottom of the resistors output is lagging the input by 2.5 ns). Time domain simulation results are seen in
Fig. 1.13. Note, in this figure, that there is a start-up time, the time it takes the
(1.35) signal to propagate through the transmission line (note the kink at 5 ns), before the
filter's behavior follows the equations we derived .•
agnitude response in Eq. (1.30)
OS1l:200~Hzl (1.36)
Figure 1.13 Time domain input and output (50 MHz) for the comb filter seen in Fig. 1.11.
(plotting eIlrfT, in the z-plane simply plots the unit circle). Note that plotting a delay, Z-I
or e j2nf (-T.• ), also results in a magnitude of one but the phase shift is now -2rc/- Ts. As we
move through the book we will regularly use Eq. (1.37) to evaluate the frequency
response 0/ a discrete-time system, see Fig. 1.21 and the associated discussion, and to
relate the complex number, z to the /requency,.f
t
Js
14 5; t 14
, 'lsi
9lt 14
~t.
1m, Y
Plottinge
J'on£
- f, = cos 2nf + 2nf:
j . sin
1
z x-"-jy /~ I . /
i
1
Figure 1.17 ~fagn
-~-.-~="----+---t:---~ Re, x
Note that if we di\ic
1
/ == /sI2, 3Js12, 5/5 12, 0, j" 21s, 31s, ... can think of it as an averagi
by two). Let's pause and th
output, which is the sum c
3Js14, 7/5 /4, 11/5 14, ...
resolution of the filter's outp
1 all 9-bits and not perform
Figure 1.15 Plotting a delay in the complex plane.
filter" ... or comb filter). Thif
Figure 1.16 shows the digital implementation of the comb filter. We can write (which is simply lowpass til
size should increase!) Noti:
(1.38)
number of input words and (
and thus the transfer function is
Finally, to show ho\\
(1.39) go to zero, consider averag
spaced apart by Ts , Fig. 1.1
or, after reviewing Eqs. (1.27) to (1.33), we get signal frequency and plot the
MOS Mixed-Signal Circuit Design Chapter 1 Signals, Filters, and Tools 13
equency domain
~2}+-Vin(f) = . VOUI(/)
~epresentinga delay
n the complex plane.
_-I
/.. =e Vin~VO"t
:Ia:,
Figure 1.16 A digital comb filter using one delay. Note how the
ify the notation, since delays are so output is the average (sum) of consecutive inputs.
can use
Ivoull I Is/ I
VoUl
Vm
2 . cos 1t- art d L - , -rr . -/ "lor / Is
< (l.40)
e j2 !t{(-·T,) e (L37) Vm /s 2
main, is seen in Fig. 1.14b (some The magnitude and phase response for this digital comb filter is seen in Fig. 1.17. Note
o differentiate between the general that if we apply a constartt (DC, / 0) value of 1 to the input of our digital filter, the
(1.37) is plotted in the z-plane (the output goes to 2 verifying that what we get with Eq, (l.40), at least at DC, matches what
its phase shift is 2rcI T, or 2IT .1: we see in Fig. 1.17,
2V7~·~
---)0;; Re, x Figure 1.17 Magnitude and phase response of the comb filter in Fig. 1.16.
~ = 0, Is, 21s, 31s, ... Note that if we divide the output of the digital comb filter in Fig. 1.16 by two we
Cart think of it as an averaging filter (taking two inputs, summing them, and then dividing
by two). Let's pause and think about this for a moment. If the input is S-bits then the
output, which is the sum of two 8-bit words, is a 9-bit word. To avoid lowering the
, 3/,/4, 7fs!4, 11/,/4, ...
resolution of the filter's output (the number of bits in the filter's output word) we can keep
:omplex plane. all 9-bits and not perform the divide by two (but still call the filter an "averaging
filter" ...or comb filter). This is important because we will regularly pcrform averaging
:'the comb filter. We Cart write (which is simply lowpass filtering a signal) to increasc the signal's resolution (so the word
(L3S) size should increase!) Note that our definition of averaging herc is simply adding a
number of input words and outputting the result.
Finally, to show how it's possible for the output of the comb filter in Fig. 1.16 to
t-l go to zero, consider averaging two points of an input signal, at a frequcncy of 1s!2,
z (L39)
spaced apart by T, ,Fig. l.lS. It may be a good exercise, at this point, to vary the input
signal frequency and plot the corresponding outputs to vcrify the results scen in Fig. 1.17,
14 CMOS Mixed-Signal Circuit Design Chapter I Signals, Filters,
time »
Figure 1.18 Showing how the averaging filter in Fig. 1.16 can have zero output atJsl2.
or using
Unit
. 2X mag:
I -cosx == 2 sm 2" (1.45)
we get
(1.46)
Q
J2lr:\cs'
VOlll I v 1
I-V , 1/1
...£I!.!.-(l-
- Z -l)_~
time )
I •••
, ,
Ts , " )
.16 can have zero output at/s l2. 1s)4f,12 j, 3//2 /(Hz)
Figure 1.20 Magnitude and phase response of the digital differentiator (also a comb filter).
:alled an averaging filter since its
filter at DC was, as seen in Fig. The phase response is
~ital difJerentiator or difJerencer,
/ Vout TI: / cO/ f
L - = - - T I : - lOr < <J' (1.47)
;;ent input signals. Vin 2 Is S
Figure 1.20 shows the magnitude and phase responses for the digital differentiator in Fig.
1.19. Notice how, for this comb filter, the gain at DC is O.
1.2.3 An Intuitive Discussion of the z-Plane
It will be very helpful in our discussion of mixed-signal circuits and systems to gain an
intuitive feel for the frequency response of a discrete-time system, like the digital filters
we've discussed in this section (or the switched-capacitor filters in Ch. 3 that are
Vout discrete-time, continuous-amplitude, circuits), by looking at the z-plane representation of
the system. To move toward this goal, consider the transfer function of the simple digital
igital differentiator.
averager seen in 1.16 with a transfer function of
sampled at discrete times n~, we
] vin[(n-l)Tsl (1.41) It is very useful, for an intuitive understanding of the frequency response of a discrete
time system, to plot the poles and zeroes of the transfer function in the z-plane, Figure
(1.42)
Im,y
z x
unit circle indicates z = e f2n/ T
,
(1.43)
x-
o-
Indicates a pole
Indicates a zero
l~
6
X Indicates six poles at a location
2 -----ft~----~-------1-----Re.x
)s (1.44) 0 - Indicates two zeroes at a location
! -
1.21 also shows how Eq. (lA8) can be displayed on the z-plane. A pole is located at z = 0 angle from the pole alo1
(at the location the denominator goes to zero and the transfer function goes to infinity) 45 0 resulting in an overa
and a zero is located at z = -1 (at the location where the numerator goes to zero).
Note that (1) an
The z-plane is usually used to describe the frequency response of a discrete time periodic with period Is
system, H( I), by assuming the input to the system is a unit magnitude sinusoid with are only concerned with
'0 f
varying frequency,.f. This input, 1 . e'-rtJ" evaluates the output of the system or .(.12 [the Nyquist freqt
magnitude response of}
-
phase
magnitude ~ is a delay of one clock
H(J) == Vout(J) when Vin(J) 1 e j2"Lis ( lA9) greater than or equal to t
in hardware (the output (
We should now see that the unit circle, shown in Fig. 1.21, indicates the relationship
between z and I when specified by Eq. (1A9) (this is important). Therefore, to determine Example 1.3
H( I) from a plot of a transfer function's poles and zeroes in the z-plane, we simply Determine, using the
evaluate H(z) along the unit circle. To show how this transfer function evaluation is of the transfer functi,
performed, consider Eq. (l.48) and the corresponding plot of its pole and zero shown in frequency of 1s14.
Fig. 1.22 along with the magnitude of Eq. (1.48) or Eq. (lAO) plotted against frequency
in Fig. 1.17. At DC (f == 0 and z == 1 . eO = 1LO) point A in Fig. 1.22, the gain of the
circuit is two and is calculated using
distance to zero
IH(J)I (1.50)
distance to pole
The distance from the zero to point A is 2 while the distance between the pole to point A
is 1. Therefore, as shown in Fig. 1.17, the magnitude of H( I) is 2. The phase of the
transfer function is calculated along the positive x-axis using
LH(J) = L of zero - L of pole (1.51)
which, as seen in Fig. 1.17, results in a phase angle of zero. Next consider evaluating H(z)
at 1s14 J ==1s14 and z = 1 . el'i
(
n \
1L90) , point B in Fig. 1.22. The distance from the pole
to point B is 1 while the distance from the zero is fi resulting in a magnitude fi . The Magnitude rep
I
I
Figure
z-plane
-------.L-----
z==x+jy If we label the lengtr
Point B, fsl4,5fs14 ...
I-
I
H(z) == z: J Also at point B, z :::: e
jru2 magnitude of the tram
Figure 1.22 The z-plane pole and zero for Eq. (1.48).
S Mixed-Signal Circuit Design Chapter I Signals, Filters, and Tools 17
Jlane. A pole is located at z = 0 angle from the pole along the x-axis to point B is 90°, while the angle from the zero is
Jsfer function goes to infinity) 45° resulting in an overall phase response of -45° (verify with Fig. 1.17).
merator goes to zero).
Note that (1) any digital filter's or discrete-time system's frequency response is
ICY response of a discrete time periodic with period I, (one complete revolution around the unit circle), (2) we normally
unit magnitude sinusoid with are only concerned with evaluating H(z) over the top half of the unit circle (from DC to
put of the system or 1,/2 [the Nyquist frequency, In]), and (3) a pole at the origin has no effect on the
phase
magnitude response of H(z) but does affect the phase response (z-l , a pole at the origin,
Ie .--"----. is a delay of one clock cycle). Finally note that the number of poles in H(z) must be
L
. ei 21tf r ( 1.49) greater than or equal to the number of zeroes if the digital filterlsystem is to be realizable
in hardware (the output ofthe system cannot occur before the system's input).
1.21, indicates the relationship
lrtant). Therefore, to determine Example 1.3
roes in the z-plane, we simply Determine, using the graphical approach just discussed, the magnitude and phase
transfer function evaluation is of the transfer function indicated by the poles and zeroes plotted in Fig. 1.23 at a
t of its pole and zero shown in frequency of 1s14.
lAO) plotted against frequency
. in Fig. 1.22, the gain of the
z-plane z-plane
Figure 1.23 The z-plane pole and zero plot for Ex. 1.3.
Point B, fs/4, 51s14 ... If we label the length from a pole (zero) to the evaluation point P (z), then the
Also at point B, z e iJti2 magnitude of the transfer function is given by
""
j _Z_l_
~ ,..; IH(z) I PI 'P2
-1···'··.···.1.
" Labeling the angles for the poles and zeroes as indicated in the figure, we can
I
•
write the phase response as
\ Point A, DC or 0, Is, 21s ...
L.H(z) == 83 -8 2 81
~
1.2.4 Comb Filters with Multiple Delay Elements
i Examine the digital filter seen in Fig. 1.24. The transfer function for this filter is given by
l d'
18 CMOS Mixed-Signal Circuit Design
I,
I H(z) Vtn~Vout
in the frequency domain
Figure 1.24 A digital comb filter using multiple delay elements.
(1.54)
Figure 1.25 shows the magnitude responses and pole-zero plots for comb filters with
varying numbers of delay elements, K. For K 4, for example, the zeroes are located at: v,"[n~
I,j, -1, and -j. In other words the transfer function goes to zero at, DC,J/4,f/2, 3j/4,fs, vout[(n L
I )TdT~
clk I
t--Yn-=:-,,)
K 3
H(z) == I _Z-3
2 Figur
,
, K= 16 used to plotting, Fig. 1.7, for
,
, H(z) l_z- 16
, 2
1/\ 1{1 'AA ~,,' iA:" ,II'
, •••
'/'I
~ To get Eq. (1.57) into the f,
f,f4 f,f2 ~fs/4 f, jCHz)
1s116 using Eq. (1.58) to simplify t
15f,116
Figure 1.25 Frequency response and z-plane plots for various values ofK in a comb filter.
)S Mixed-Signal Circuit Design Chapter I Signals, Filters, and Tools 19
fs +fs14, etc.). For K = 3 the zeroes are located at: I, . " and T'
[f ' Because
- j. 2
of their simplicity, comb filters are very useful in mixed-signal design (and used to design
lowpass, bandpass, and highpass filters that will be discussed in Ch. 4).
1.2.5 The Digital Integrator
Examine the circuit seen in Fig. 1.26. The output of the circuit, in the time domain, is
given by
vin[nTsl + vout[(n
Vin~VOUI Vout[nTsl = l)Tsl ( 1.55)
or, in the frequency domain,
pie delay elements.
H(z) H(f) = Vaat (1.56)
Vm
This circuit adds the current input to the sum of the past inputs, performing integration.
(1.53) Note that there isn't a delay in series with the input signal so this circuit is often called a
non-delaying integrator or accumulator. We'll see in a moment that a delaying integrator
does delay the input signal so it has a transfer function of z-I/( 1 - Z-I).
K=3
j,
H(z) l-z-3
H{z) =
1 Vin ) 11 _lz_ll) Valli
Figure 1.27 shows the z-plane representation of the non-delaying integrator along
with its magnitude and phase responses. To determine equations for the magnitude and
K 4 phase responses we can write
H(z) l-z-4 Phase shift
H(z) =z (1.57)
1 ( 1.58)
a+jb
To get Eq. (1.57) into the form x+jy we can multiply it by its complex conjugate or,
using Eq. (1.5 8) to simplify the notation,
IH(f)1
z-plane
+---~----~--~.~
Is 3fs!2 f
LH(f)
degrees
A
90-
F
- 90 - -
Figure 1.27 The z-plane representation along with magnitude and phase response
for a non-delaying digital integrator.
The magnitude response it
Ia- jb a .-b
~~; ~.---'
~
=1 Real, x Imaginary,)
1
21sin nt I CMOSedu,com,
Evaluating the phase response directly from the z-plane plot, 1.3 Representing Si
From zero From pole
~ ,~ ~, .. In the past two sections we
fixed frequency and chan!
LH(f) 2nL (nL+ 11) 180 L - 90 (degrees) for 0 < I < Is (1.63)
Is Is 2 Is z-plane, Sees, 1.1.2 and 1,
with varying frequency, Fi~
At DC the phase contribution from the zero is 0°, while the phase contribution from the and non-periodic, that are
pole, at a frequency just above DC, is 90 0 • The result is an overall phase response of function can be represen
I
I -90 0 , At Is 14 the phase contribution from the zero is 90 0 , while the phase contribution frequencies (but all sinusoi
from the pole is 135°, resulting in an overall phase response of -45 0 • signal we are modeling wi
The Delaying Integrator exponential Fourier series r
transform which is used tc
Figure 1.28 shows the delaying integrator. We get this topology by moving the delay from assumed this material is a
the feedback path in Fig, 1.26 to the forward path. The output is related to the input using the material useful for refer
JS Mixed-Signal Circuit Design Chapter I Signals, Filters, and Tools 21
Out Out
Vout(f)
(Vjn + Vout)
/s
()2 /s 3/s 12
~
l
H(z) Vin ~l Vout
I-z-
( 1.64)
or
lagnimde and phase response H(z) H(f) = Vow = (1.65)
Viti 1 Z-l z-1
The magnitude response is exactly the same as the non-delaying integrator, Fig. 1.27 and
. -b ( 1.59) Eq. (1.62). The phase response is given by
) a 2 + b2
~
(1.61 ) When a digital signal is generated from an analog signal using an ADC (analog-to-digital
converter) signal aliasing can occur. Signal aliasing comes from the sampling process and
is discussed in detail in the next chapter. The importance here in this section is that if we
use the output of an ADC, clocked at Is , to provide the input signal to a digital filter we
have to restrict ourselves to input frequencies below the Nyquist frequency,f;2, in order
( 1.62) to avoid aliasing. Simulation examples using the filters in this section are available at
CMOSedu.com.
00
where, as just mentioned, n f represents the integer multiples of the original signal's
t+T
Cn == 1
T
Jg(t)· e- j21U1fi
dt (1.68) Dirac Delta Function (Unit
t
The key thing to note is that we use the Fourier Series to represent signals that are
discussed in the next chapter. Using the Fourier Transform to look at the spectrums of
periodic signals can be very challenging for all but the simplest waveforms.
The discrete version (defin{
Dirac delta function is the 1<
As an example of the use of the Fourier Series consider the squarewave seen in
by
Fig. 1.29. The amplitude and period of this waveform are A and Ts respectively. The time
the output pulse is high is Tp. When Tp = T/2 the squarewave has a 50% duty cycle. Using
When n = 0 we can use l'Hospital's rule to get As the amplitude tends tow
equal to 1. Also,
Co 4 (1.70)
2
Using the results from Sec. 1.2.2 we can write
(1.73)
1 where the magnitude is ] a
of a sinusoid is determined 1
D D
Ts
)
then, using Eq. (1.74)
nji
( 1.67)
G(f) = f g(t)·
-00
e-
j2nji
dt (1.74)
ies to represent signals that are The Dirac delta function, sometimes called the unit impulse response, is defined as
,Ie-and-hold and track-and-hold oCt to) = 00 when t= to and 0 ift-:/.: to (1.76)
form to look at the spectrums of
mplest waveforms. The discrete version (defined only at discrete time points, that is not continuously) of the
Dirac delta function is the Kronecker delta function, or simply delta function, and is given
consider the squarewave seen in by
'e A and Ts respectively. The time
~ave has a 50% duty cycle. Using oCt nTs) = I when t == nT, and 0 if t -:/.: nTs (1.77)
e get Some properties of the Dirac delta function are that
w
As the amplitude tends towards 00 the width of the function moves to 0 keeping the area
equal to 1. Also,
(1.70) OJ
A few other key properties are that the Fourier Transform of a constant is
(1.71)
00
f K· . dt = K . o(f) ( 1.80)
os 1tn) (1.72) The Fourier transform of a delay (noting a delay of to is written as e-j2njIO) is
f
00
---------......_ - - - - - - - - - - - - - - - - - - - - - - - - - _ ...
24 CMOS Mixed-Signal Circuit Design
[6] E. P. Cunningham, Digital Filtering: An Introduction, John Wiley and Sons, Fig
1995. ISBN 978-0471124757
Repeat question 1.1
QUESTIONS
forward path of the (
1.1 Suggest an alternate physical example, to the swinging pendulum in Fig. 1.1, of
Determine the expo
sinusoidal motion.
in Fig. 1.29 if it is c(
1.2 Add the z-axis (time) to the representations of the sinewave seen in Figs. 1.3b-f as
1.14 Detennine the expo
discussed following Eq. (1.4).
in Fig. 1.29 for the g
)S Mixed-Signal Circuit Design Chapter I Signals, Filters, and Tools 25
1.4 Figure 1.7 shows how the magnitude and phase are calculated for an imaginary
number that resides in the first quadrant of the plane (both real and imaginary
-/J)] ( 1.85) components are positive). Show how we calculate the magnitude and phase of an
imaginary number in the other quadrants.
1.5 If the output of a system occurs after the corresponding input to the system, is the
phase shift positive or negative? Why? What does linear phase indicate?
(1.86)
1.6 Using the SPICE files found at CMOSedu.com, verify, in the time-domain, the
~quencies only we could rewrite frequency response information seen in Fig. 1.10 for input frequencies of 1= 0
(DC), 1/4td , and 1I2td .
(1.87) 1.7 Repeat question 1.6 for the digital comb filter (averager) seen in Fig. 1.17 for
input frequencies of DC,j/4, and//2.
1.8 Plot the magnitude and phase frequency responses of a discrete-time system
"+II) + 8(f- 11)] (1.88) having the transfer function (I + z-I . Next, show the location of this system's
poles and zeros in the complex plane and verify, using the intuitive method
1S since they simply represent a discussed in Sec. 1.2.3, the gain and phase of the response match the frequency
response plots when the input signal frequency is O.
ft + n:) = sin (2n:lt n:) (1.89) 1.9 For the 3 delay element comb filter seen in Fig. 1.25, repeat question 1.6 for input
frequencies of 0,1,16, and/;;3.
fATLAB and Wavelets, Infinity 1.1 0 Show how to plot 1/(4+ 3j) in the complex plane. What is the magnitude and
phase shift of this complex number?
'on to Analog and Digital 1.11 Determine the z-domain representation for the circuit seen in Fig. 1.30. Also, plot
and Sons, 2006. ISBN 978 the frequency response, both magnitude and phase, and the location of poles and
zeroes for this system.
fuction, John Wiley and Sons, Figure 1.30 Circuit used in question 1.11.
1.12 Repeat question 1.11, and sketch the resulting circuit, if a delay is added to the
forward path of the circuit seen in Fig. 1.30.
inging pendulum in Fig. 1.1, of
1.13 Determine the exponential Fourier series representation for the squarewave seen
in Fig. 1.29 if it is centered around ground.
sinewave seen in Figs. 1.3b-f as
1.14 Determine the exponential Fourier series representation for the squarewave seen
"*
in Fig. 1.29 for the general case where Tp T,/2.
26 CMOS Mixed-Signal Circuit Design
LI5 What is the Fourier transform of the signal seen in Fig. 1.29?
1.16 What is the area under the Dirac delta function bordered by the x-axis? Why?
1.17 Show how to take the Fourier transform of sin (2rrJo +0) and cos(2rrJo+~). Plot
the magnitude and phase responses of the transforms.
-, Pvt
t Anti-aliasing filter
-I
~AA;1 Ism
I
• Sample an
in Fig. 1.29')
bordered by the x-axis? Why? Chapter
2
l (2nfo + 8) and cos (2n:/o + $). Plot
anTIs.
The block diagram for one example of a mixed-signal circuit design used to process
analog signals is seen in Fig. 2.1. The system's input and output are continuous-time
(analog) signals. The input signal is applied to an analog filter, called an anti-aliasing
filter (AAF) that limits the input spectral content to avoid aliasing (discussed in Section
2.1) when the signal is sampled and held using the sample-and-hold, SIH (also called a
zero-order hold, ZOH), circuit. The output of the SIH is connected to the analog-to-digital
converter (ADC) where it is converted to a digital word that changes with time. The
digital word is then passed through a digital signal processing (DSP) block where it is
manipulated (e.g. lowpass filtered using a digital filter like the one seen in Fig. 1.16). The
output of the DSP block is connected to a digital-to-analog converter (DAC) that changes t
the digital words back into an analog voltage. At this point the signal does have a
changing amplitude but it is still discrete in time (jagged as seen in Fig. 2.1). The output
of the DAC is fed to a reconstruction filter (RCF) to smooth it out. We'll see that this step
removes the spectral content above the Nyquist frequency,/" or/)2.
Note that analog circuits are an important component of this system. The AAF
and RCF must be completely analog in design. Further, the ADC and DAC may contain
precise analog circuits.
Figure 2.1 Signals resulting from AID and DfA conversion in a mixed-signal system.
28 CMOS Mixed-Signal Circuit Design . Chapter 2 Sampling and AI
2.1 Sampling
V'L
In this section we discuss how sampling a signal changes its spectrum. Impulse sampling, Volts f.
decimation, the samplc-and-hold (S/H), the track-and-hold (T/H), interpolation, and
K-path sampling methods are discussed.
2.1.1 Impulse Sampling
Consider the sampling block diagram shown in Fig. 2.2. Let's assume that we apply a fm f
sinewave input, x(t), to this sampling gate of the form, Vpsin (2rrJin . t) (for the moment, a (a) Input spectrum
single frequency input). The output of the sampling gate (a.k.a. sampler),){t), is the
product of the input and a sampling unit impulse signal (Dirac delta function), 8(t nTs)
or
00
Figure 2.3 One-side
yet) L Vpsin(2rrfin' t)· 8(t- nTs) (2.1)
sampler is connected to an
Noting that the frequency of the input is 1;n while the sampling frequency is Is(-= lITs), . fn [the Nyquist frequency]),
the spectrum of the input signal is seen in Fig. 2.3a. Note that the Fourier series so that only /In remains (thi!
representation of the series of sampling impulses above is
Example 2.1
f 8(t
n~-oo
nTs) == f
.Ts n=--co
ej2rcn'IIT, == f
Ts n=--<:o
eI21tnkl (2.2) A sampling gate is strol
MHz Us 100 MHz an,
If we take the Fourier transform of the input signal after sampling (we take the Fourier the resulting output freq
transform of Eq. [2.1)), that is, we look at the spectrum on the output of the sampler, we sampler. Also, sketch th,
get
The resulting frequency
output of the sampler thr
results in an output sine'
situation, that is, to avoic
or, knowing the Fourier transform of ej2rr./ol is 8(/- fo)
and reconstructing, we I
00
sampler are less than lsi
Y(f) = 2jTs . k,,!-}8(f /In - lif,) 8(f + Jin kIs)] (2.4) we see that this is the pUl
both the AAF and RCF (
The sampled spectrum is repeated, at intervals of Is, as seen in Fig. 2.3b (the one-sided
a cutoff frequency equa
spectrum is shown, which is what we will use throughout the book). Note that if an ideal
Figure 2.5 shows the tim
lowpass filter (LPF) is applied to the output spectrum of the sampler (the output of the
I
I
:').I~
v h x(t) == Vpsin (2rrJin . t)
1
rVt v~
I. , , ~ In >~ Sampler ou~u: Aliased signal 40 60
~ t I yet) = x(t) . 8(t- nTs)
50
f,• -_t.:.2
Sampling impulses:
Figure 2.4 Spe.
8(t- nTs)
It should be clear [TOT
Figure 2.2 Impulse sampling a signa\. in a reproduction of the sam
signal's spectrum should hav
v1ixed-Si~'11al Circuit Design Chapter 2 Sampling and Aliasing 29
IY(f)1 (2is-fm)
Volts
,pectrum. Impulse sampling, In is \2fs
d (T/H), interpolation, and
Ts
Figure 2.3 One-sided spectrum of a sinewave (a) before and (b) after sampling.
(2.1)
sampler is connected to an ideal LPF) with a bandwidth greater than /;n (and lower than
ling frequency is is( = liT,), In [the Nyquist frequency]), then the higher-order frequency components can be removed
fote that the Fourier series so that only /;n remains (this is our smoothing or reconstruction filter shown in Fig. 2.1).
00 Example 2.1
L eJ21tn/s1 (2.2) A sampling gate is strobed with an impulse train running at a frequency of 100
n=-CO
MHz (is = 100 MHz and the time in between the impulses, Ts, is 10 ns). Sketch
mpling (we take the Fourier the resulting output frequency spectrum if a 60 MHz sinewave is applied to the
he output of the sampler, we sampler. Also, sketch the time domain input and output of the sampler.
The resulting frequency spectrum is shown in Fig. 2.4. Notice how connecting the
:Uin-k/s)/] . e -J2rtft . dt (2.3) output of the sampler through an LPF, with an ideal abrupt cutoff frequency of In,
results in an output sinew ave with a frequency of 40 MHz. In order to avoid this
situation, that is, to avoid ending up with the wrong, or alias, signal after sampling
and reconstructing, we need to ensure that the signal frequencies applied to the
sampler are less than I sl2 (the Nyquist frequency, again, In). Reviewing Fig. 2.1,
J + /;n -kis)] (2.4)
we see that this is the purpose of the antialiasing filter (AAF). Notice how, ideally,
both the AAF and RCF (reconstruction filter) in Fig. 2.1 are both ideal LPFs with
1 in Fig. 2.3b (the one-sided
a cutoff frequency equal to half the sample frequency (the Nyquist frequency).
!book). Note that if an ideal
Figure 2.5 shows the time domain sketch of the sampler's output. •
e sampler (the output of the
It should be clear from the preceding discussion that: (1) sampling a signal results
al. in a reproduction of the sampled signal's spectrum at DC,/s, 2/s, 3/s, etc., (2) the input
signal's spectrum should have no significant spectral content above J" in order to avoid
30 CMOS Mixed-Signal Circuit Design Chapter 2 Sampling and Ali.
aliasing, (3) to avoid aliasing both filtering the input signal using an AAF and increasing Sampler input, 20 MHz
the sampling frequency should be used, and (4) to reproduce the sampled signal from the
output of the sampler (which is nonzero only during the sampling impulse times) a ,J
lowpass RCF should be used.
Note that our discussion illustrates the operation of a sampling gate driven with ,,
impulse signals. As shown in Fig. 2.1, a practical system would have other building
blocks. We would rarely, if ever, sample a signal and then reconstruct it without
processing it first.
A Note Concerning the AAF and the RCF
Before going any further, we should discuss the ideal characteristics of the AAF and the
RCF. The ideal characteristics of these filters are shown in Fig. 2.6. Note that both of 10 ns 30 r
these filters must be analog by design. The ideal cutoff frequency for the filters can be no
greater than In (assuming the sampling rate on the input of the system is the same as the Figure 2.7 1m]
sampling rate on the system's output) and the filters should ideally have linear phase. Let's
discuss these two ideal characteristics.
The transfer function
domain impulse response (w
~
n=fs!2 large amplitude, very short t
IH(jro) I Ills for RCF
A o ; ) output in the time domain. T
!
r----, ~ f
: slope ;-2TCl o
IrnPUlse-L
input to
RCF t
In =fs!2 f o
(a) (b)
Figure 2.6 (a) Ideal magnitude and (b) phase responses for the AAF and RCF.
Figure 2.81
[OS Mixed-Signal Circuit Design Chapter 2 Sampling and Aliasing 31
------~---------------------------------------------------
The ideal magnitude response, shown in Fig. 2.6a, passes all spectral content
Alias 40 IV1Hz sinewave below the Nyquist frequency while removing all signals above this frequency. The ideal
phase response, shown in Fig. 2.6b, provides a constant delay, to' to all signals below I..
\
In other words, the filters remove all unwanted signals while not distorting the wanted
signals.
Time Domain Description olReconstroction
In this section we show why the filter shown in Fig. 2.6, an ideal brick walliowpass filter
with linear phase response, is the ideal RCF on the output of our impulse sampler. Shown
in Fig. 2.7 is a 20 MHz sinewave sampled at 100 MHz. Suppose we want to reconstruct
the original input 20 MHz sinew ave from the sampler output (the weighted impulse
functions). After reconstruction, the output of the RCF should be a single-tone, 20 MHz
sinewave (it should be an exact replica of the sampler input). To determine what happens
75 ns lOOns when the output of our sampler is applied to the ideal RCF, we need to determine the
time-domain response of the RCF when its input signal is an impUlse.
put for Ex. 2. L
The transfer function of a system is the Fourier transform of the system's time
domain impulse response (what we are trying to find here). In other words, to determine
ic.o) the transfer function of the system, we apply an impulse to the input of the system (a very
In =/,/2 large amplitude, very short time duration pulse, Fig. 2.8). We then look at the system's
~
-~~--f output in the time domain. Taking the Fourier transform of this output gives the system's
~
: slope = -21tto
Impulse
input to
ReF -
1 ti~e_ --tIlls
tL
Ideal ReF ~
10
~+-.~.
1
1
.
tIme
o 1----- t=O
(b)
In f
transfer function. Therefore (in the reverse order), to determine the time-domain impulse
response of the ideal RCF, given the transfer function, we take the inverse Fourier
transform of the transfer function. The ideal RCF's transfer function (Fig. 2.6) can be
defined by
IH(j)1 Ills for If I <fn else IH(j)1 = 0 (2.5)
The time-domain impulse response is then given, remembering 2fn , by
In
h(t) f 1. ej21ljl . df = "'----~-
-:(.,is
sin2nfn . t
Sinc( nis . t) (2.6)
nis' t
where
Figure 2.10
smx == Sinc(x) (2.7)
x
An Important Note
The time-domain impulse response of our ideal RCF is shown in Fig. 2.9. Notice that our
impulse is applied to the system's input at t 0 and that the output actually anticipates, or It is important to note that 01
starts, before the application of the input! This indicates that the filter can't be built in a amplitude (unlike an analog
practical analog circuit. Before we discuss the implications of this severe limitation (an and amplitude). The amplitu
ideal reconstruction filter can't actually be built because its impulse response is infinite in amplitude input to the samI
time), examine Fig. 2.10. Figure 2.10 shows the individual impulse response outputs of used to describe systems usi]
an ideal RCF with the impulse train of 2.7 as the input. The output of the RCF is the words, whether we are disc
weighted sum of the individual responses. While this figure is "busy," the basic concept waveforms with amplitudes
of reconstruction can be seen. represent the discrete-time s
for continuous-time systems.
sin nis . t ,
= Sinc(nfs . t)
nifs . t 2.1.2 Decimation
IHU)I
,., In the last section we focu
1-
I concepts to digital signals. v
is fromis to a lower rate off)K
2,4, 8, 16, etc.). This reduct
and is illustrated with the b
x[nTsl
In
,
I Input word rate
, Figure 2.9 (a) Ideal RCF frequency response and (b) impulse response (time).
I Figun
1
What can we do to ease the requirements on the RCF? One answer is to increase is
so it's much larger than the maximum sampled frequency. This solution is the basis for
oversampled data converters or noise-shaping data converters (studied in detail later in I Quantize: to limit the possib
the book). This solution also eases the requirements placed on the AAF. Another idea for time, for example, means that tl
as the sampling impulse times f
easing the design of the RCF is to increase the sample rate of the digital data coming out
during certain discrete time inl
of the DSP block in Fig. 2.1. This technique is called intelpoiation.
discussed in the next section).
- a
~~;"
, ?:'k""
fen;
MOS Mixed-Signal Circuit Design ~+,~'"
", Chapter 2 Sampling and Aliasing 33
else IH(f)1 = 0
mbering 2/n = Is , by
(2.5)
j
e-j-2n!n"(
Tt I,' t
Sinc( Tt/• . t) (2.6) ~==~~~~~~~~~~~~~~~~~~~==~~~-->
time
I ~K
~I.n:.:..-_-++--_ _ Out
- I'> / );
(b) -I3.5dB
Input word rate, Is Output word rate,f,IK
nverters (studied in detail later in I Quantize: to limit the possible values of a quantity to a discrete set of values. Quantizing in
rate of the digital data coming out as the sampling impulse times for the ideal impulse sampler) or that the amplitude is unchanging
during certain discrete time intervals (such as seen in the output of the ideal sample-and-hold
terpolation.
J
discussed in the next section).
34 CMOS Mixed-Signal Circuit Design Chapter 2 Sampling and P
decimate) can be confusing since, among other uses, the dictionary definition "to select 2.1.3 The Sample-and.
by lot and kill one in every ten." The origin of the word comes from a method of
.Understanding the opera
punishing military troops by selecting one in every ten for execution. Our much more
understanding the concep'
kind-hearted definition will mean that we are passing the input word through a lowpass
most mixed-signal syster
digital filter and then down-sampling the result (discarding samples). This procedure is
, sampler so that the samI
effectively passing the digital data through an anti aliasing filter and then resampling the
times. Having tl
result at a lower rate, Fig. 2.12. Note that the sampling gate is simply a register so it is
1n't""."-t,,,,t for proper ADI
trivial to implement decimation.
see, is that it will il
Sampling gate :
Out y[Ki· Ts]
f..----'-I-_+_-
Output word rate,/sIK
••• Figur,
1
1
f
-1-
(c) ••• @j;./8
~~:~.;
VIOS Mixed-Signal Circuit Design 1iC' Chapter 2 Sampling and Aliasing 35
f~i?<
~"\"r!k
Vin
Output clock
~ clock
m after AAF : f
__---'-O~,----' . ~.~.-.-@~). Is
1ator output : f
time
CO •••
f
@fs/S
).
:imation is employed. Figure 2.15 Sample-and-hold output with return to zero format.
-------------~ .. -~ .. ~.--.<--
vm(nTs) I co
IX; r""·----~--··
== ( - 2: Vin(f- kfs)
yet) 2: [Vp sin(21tfin·t)·o(t-nT,)]®[u(t)-<u(t 1)] Ts k=-oo
(2.9)
or
Weighting from S/H, IH(f
Y(f) == (V in (f) ®P(f»' H(f) (2.10)
T· ISine(n· T ·f)1
where, see Eq. (2.4),
Vp 00
vm(f) ® P(f):= 2jTs . k~'" [o(f - fin - kfs) - o(f + fin - kfs)] (2.11)
We can determine the spectrum of the sampling pulse h(t) seen in Fig. 2.16a, IH(f)I, by
taking its Fourier transform
T,
H(f) f [u(t) - u(t - T)]e-: f 2n: f f • dt (2.12)
o
which is evaluated as
phase shift
T· Sine(n I T)
rre(luCJncy response by Sine
(2.13) the output of the ideal SIR
The magnitude of Eq. (2.13), !H(f)I, is plotted in Fig. 2.16b. The phase response input signal is attenuated b
corresponds to a shift in time of TI2. Since mentioned above, adds dist
S/H's frequency response usi
I Y(f) (Vin(f) ® P(f» . H(f) (2.14)
1
00 T
h(r)
'T .
2:; s
L [0(1 -
k;-oo
fin -/ifs) - 0(1 + lin - k/s)] . T· Sincere I T) . ej21thj
oc
f) (2.10)
Iru")1 = T·ISine(re. T 2Ts . k~OO [8(1 - lin lfs) 8(1 + lin lfs (2.16)
sse shift magnitude For most circuit designs, T= Ts so that, as Eq. (2.16) and Fig. 2.17 show, the
~ r·~--~------'
sample-and-hold operation weights the amplitude of the ideal impulse sampler's
-j~/T . T. Sine(re I T)
frequency response by Sine ; ( or Sine ~1. Note that at the sampling frequency (fs=l/Ts)
(2.13) the output of the ideal S/H goes to zero. Also note that at the Nyquist frequeney, J", the
g. 2.16b. The phase response input signal is attenuated by 0.64 or 3.9 dB. This "droop" in the S/H's response, as
mentioned above, adds distortion to the input signal. Let's illustrate the effects of the
S/H's frequency response using an example.
f) (2.14)
Sample-and-hold with T = Ts
Desired spectrum
(b)
l~ =/s12
~I~~
0.13
f
2 1
T T
10.211 =-13.5dB
; spectrum.
Figure 2.17 The frequency response ofa S/H.
38 CMOS Mixed-Signal Circuit Design Chapter 2 Sampling and A
Example 2.2
Using an ideal SPICE model for the S/H show, and discuss, the spectrum resulting Ultimately the (processed)
from sampling a 3 MHz sinewave at 100 Msamples/s. an RCF to recover the inp
The results of passing a 0.5 V (peak) sinewave centered at 0.5 V (-9 dB using 2.19. The response peaks
RMS voltages) through the ideal S/H are shown in the SPICE simulation results introduced by the S/H Sim
seen in Fig. 2.18. Note that SPICE uses a one-sided spectrum so, for example, we the requirements placed 0
must multiply Eq. (2.16) by nvo to ensure proper signal levels. The attenuation the sampling, of having the br
97 MHz image sees is the peaking type response
signal power). Note that in
7 droop introduced by th.
Attenuation Sinc( 1\'06 ) = 0.031 -30.2 dB
The amplitude of the 97 MHz image is -9 dB below the attenuation resulting from
using a S/H or -39.7 dB. At the Nyquist frequency of 50 MHz, an input signal is
attenuated by -3.9 dB as seen in Fig. 2.17.
Note that the S/H cannot be used as an AAF since any aliasing that occurred
using the impulse sampler still occurs using the S/H. For example, sampling a 60
MHz sinewave at 100 MHz still results in a 40 MHz alias signal in the base
spectrum (the spectrum from DC to In), as shown in Fig. 2.3. Now, however, the
signal is attenuated by the SiH (the attenuation is -2.4 dB at 40 MHz when
sampling at 100 Msamples/s.) In other words, the S/H can be thought of as an
ideal impulse sampler followed by a Sinc response filter.
Figure 2.19 I(
An important thing to note is that repetitively sampling and holding a signal
results in only one S/H attenuation hit (assuming the timing is such that a Circuit Concerns for Imp/e
sampling operation is not occurring when the previous SIH stage's output is
changing). This means that topologies that use several SiH operations on an input Figure 2_20 shows a sing]
signal, such as a pipeline ADC, only attenuate the signal by Sinc(ref ifs) once. This switches are closed while t
is important to understand .• the bottom plate (left side c
CH , while the right side 0
.,, . op-amp through the $1 swi1
13 11) the op-amp operat,
short that the capacitor do
1 capacitor) is always at gn
,1
197 MHz
-45dB
I:
1 -5J1dB
'1
-63dB
-72dB n-::=::::-::j:-:.,",",-~ --.
-81 dB fJL_ _-:-_ _~--'-"~-:-::.r'+:----;----;-.::::o.._~""I"""'-i
: : , : : : I
• I~
-90dB
-99dB
OMHz
----- - - - - -
30MHz
-r---------. -~ -----------r- ----------j- -------- ---( ---- -- ---t-----j
I ,
60MHz 90MHz
I
120MHz
I
150MHz
I
160t.lHz
Bottom plate
/ ~
Figure 2.18 Output of a S/H after sampling a 3 MHz sinewave at 100 MHz. Fi
10S Mixed-Signal Circuit Design Chapter 2 Sampling and Aliasing 39
MHz
to tI t2 t3
~l
~
~2
~3
vow
~3
resulting from the 4>1 switch turning off is independent of the input signal. When the 4>2
switch turns off its charge will, ideally, be injected into the low-impedance input, vin ,
since the impedance looking into the right of the 4>2 switch is large. This leaves the
voltage across the hold capacitor unaffected. The sequence of turning off the switch to the
right of CH (the top plate) followed by turning off the switch connecting v in to CH (the
bottom plate) is often called bottom plate sampling.
Bottom plate sampling is illustrated in its simplest form in Fig. 2.21. In this figure Note droop
the switch connected to the bottom plate of the capacitor, the 4>] switch, is turned off first.
When this happens the charge is injected into the eircuit independent of the input signal
(each side of the switch is at ground). When the <1>2 switch turns off, its charge can be
injected into the low-impedance node, the input v", , or into the series combination of Cll
and the off <1>1 switch. Again, the charge takes the lowest impedance path to ground and
thus most of the charge injection resulting from the 4>2 switch turning off flows through ReF res
v in ' leaving the voltage across the hold capacitor unaffeeted. The name "bottom plate
sampling" can be confusing. Reviewing Fig. 2.20, we see that the (physical) top plate of
the hold capacitor is connected to the <1>: switch while, in Fig. 2.21, the (schematic
representation) bottom plate of the hold capacitor is connected to the 4>1 switch.
Figure 2.22 Ex
,I. Turns off last
'l'2~
It:~lCH
:!' the lowest
impedance
sampling circui
employing both analog- a
The TlH is implemented u
When the gate of the MO
path to <1>J directly drive the capacit
ground.
7
I ~
Turns off first R..r: ''':',~rv
CH,
assuming that the produc
Rch . is much smaller 1
An Example
_I
Before leaving this section let's give an example of the speetrums associated with (ideal)
I sampling and reconstruction using a S/H, Figs. 2.1 and 2.22. In Fig. 2.22a we represent an
input signal as a continuous spectrum that is not bandlimited (that is, the speetrum is
completely occupied). In a real spectrum the spectral components don't have a constant
-I-
amplitude (e.g. noise at high frequencies may be considerably smaller than desired
I content at low frequencies) but here, to simplify things, we assume a constant amplitude.
The first step in our example is to pass the input signal through an ideal AAF to
limit the spectral content to the Nyquist frequency, //2, Fig. 2.22b. At this point the
spectrum hasn't been sampled so it doesn't repeat at multiples off;. The output of the AAF
is then passed through the S/H. In Fig. 2.22c we show the Sinc weighting from the
sample-and-hold process but we don't show the effects of sampling. Note the droop in the
response at frequencies approaching//2. In (d) the output of the S/H is seen. Note how
the cntire spectrum is occupied. Finally (e) shows the output after passing the S/H's
output through an ideal RCF, Fig. 2.19. Note that the spectrum is no longer periodic.
Mixed-Signal Circuit Design Chapter 2 Sampling and Aliasing 41
~
the series combination of CH
npedance path to ground and
lch turning off flows through
Spectrum
(d) -t-!_ _ _ _ _ _---+_ _
ReF response
__ after
t
_ideal
_S/H •••
_'--_ _
f
)
Track Track
~
:trums associated with (idcal) Sinewave in Vinet) !
. In Fig. 2.22a we represent an
I 0J:ut
yet)
~""Uit~~
o •
tited (that is, the spectrum is
:lonents don't have a constant
derably smaller than desired
Issume a constant amplitude.
time
gnal through an ideal AAF to :.e
2.22b. At this point the T
:S ofIs. The output of the AAF :( T! >
the Sinc weighting trom the Track
Impling. Note the droop in the
of the S/H is seen. Note how
utput afier passing the S/H's Figure 2.23 Track-and-hold output.
1m is no longer periodic.
42 CMOS Mixed-Signal Circuit Design
To determine how the TIH affects a sampled signal let's first notice that the hold total T/H output spec
portion of the output is exactly thc samc as the SIH with RZ format seen in Fig. 2.15, Eq. Eqs. (2.15) and (2.22) or
(2.15). Knowing this we can focus on the track portion of the output and then sum the
responses to get the overall TIH response. For the track portion of the ideal T/H we can Y rlH(f) = Ts . Sinc( rt I T
write, sce Eq. (2.8),
Ts T cr:
r h,ul -Ts' L
l
k=
YI(t)=n~oc Vp sin(2rt/in· t )· [u(t-nTs-T}-u(t-nTs Ts
1/2j, then we
(f1-'I Sine [
n~ro O(t- nTs)l L
=k~-ro
00
Knowing
00 00
we can write
(2.19)
(2.21 )
The track portion ofthe T/H's output spectrum for an input signal, vmC!), is conversion then'
yH.~U'."'Hwe don't want tht
fl(f) = Vin(f)@l TSTs T. k~m . Sine(rt· kf, . (T, - T). 8(f- kf,,) j the track portion,
(discussed in Sec. 5.2.1).
2.1.5 Interpolation
,'
One of the main assumpti!
For the single-tone input sinusoid used in Eq. (2.17) discussions, is that an RCF
YIU) = and a cutoff frequency c
',- Vp(Ts T) ~ ( .
challenging analog filter de
2T
'.),
. Smc(rt . kfs . (Ts
£...,
k=···",
T). [8(f-.fin kfs) - 8(f+fin - kfs)] . ) we can increase is while k
words, by increasing!, the (
, or (2.23) wanted frequency of interes
-, Ideal impulse sampler response
Weighting from ht(t)
Decreasing the cloc
cover increasing the cia
Ifl(f)1 L
k=-oo
en
-T)·Sinc(rt·kj,·(Ts T»· 2;, .[8(f-fin-kfs)-8(f+fin
V
interpolation can be used
introductory section, we foc
goes from is to a high!
(2.24) nec:es!;ari a power of 2. In
)S Mixed-Signal Circuit Design Chapter 2 Sampling and Aliasing 43
lal let's first notice that the hold The total T/R output spectrum, YTIH(/), for a general input signal VinCI), is the sum of
RZ format seen in Fig. 2.15, Eq. Eqs. (2.15) and (2.22) or
of the output and then sum the T 00
Ts-T. f (2.25)
Ts k=-oo
:1- nTs Ts
IfT= TsJ2 = 112/s then we can write this equation as
00 ({ 1 f 1\
L'" o(t-nTs)l (2.17)
YT!H(f) = L
k=-oo
I -. SincllI· -) . e
\. 2 2 Is
"=-oc J
We know from Fig. 2.17 that a sinewave at nearly J;, will see an attenuation of 0.64 (-3.9
00
dB) when using a SIR with T = Ts. Using the T /H with T = T,/2 the attenuation the
Is L 8(f-kfs) (2.18) sinewave sees at nearlyJ;, (f ~ /,/2) where k = 0 (the frequencies from DC to 1/2) is
k=-oo
IY TlJ{(f) I = ! ISinc(~) . + 1 . eO I (2.27)
or
(2.19)
(2.28)
We know we can't directly add the polar representation of these numbers so let's rewrite
reI(Ts T) (2.20)
this equation, noting 0.9 cos (-~) = 0.636 and 0.9 sin (-~) = -0.636, using the Cartesian
representation as
L~ _ _~
Input word rate,is Output word rate, K is
There are three basic interpolation schemes (ways of increasing the sampling, or
clocking, frequency): zero padding, using a hold register, and linear interpolation, Fig.
2.25. Figure 2.25a shows an example input to an interpolator. In (b)-(d) we show
interpolation with K 4.
Zero Padding Figure 2.26 A ze
The benefit of using zero padding, Figs. 2.25b and 2.26, is that the desired spectrum
remains unchanged. We'll discuss this more in a moment. The drawback of zero padding the word size and shift ea
is that the amplitude of the output signal drops by K. If our input is a constant value of 1 power of 2). For the pre~
and we insert 3 zeroes (K 4) then our output has an average value of 0.25 (average of 1,
two places reSUlting in t.
0, 0, 0). This is normally not a problem for a digital signal since we can simply increase original input). In the foil
x[nTs] .' the input or output wor
padding.
~ • (a) input signal
•
• •
I
• i Returning to Fig.
bas the assumed spectrUl
+-_-'-_-'-_--'-_.-L~_---,-
I
TI]
Y[/.. K
2
!
•
_
I
___,--'---,~--'----+) n ..L.
.~
"counter and AND gate aI
zero padded waveform Sf
Nyquist frequency has me
" (b) zero padding
I
I !...12 4
T,]
Y[I'. K
T•••,••• I...i.......r...t•••,•••I...
>1' Kt
' Ts
(a)
;. New Nygu
jS ?
1
t (c) hold register
(b)~
tiTr ~m""IIIIIIII ttTt,,,,1m~,
I
1
tllli
y[i'~]
.ll.llil •
. Kt
(c)
~C~~
~1"~'
1
; I
(d)
I
.1'T . ['::"''"
• ~ll." Ts
Jle rate)
y[i· In (a)
Out ---f/~)-----i~ I ...
/ )
Input sample rate,x[nT,] ..§ '
clk! (c) or (d)
o ;:s I------,f--_ y[ i .
Output word rate, K I, Is o e:>CJ
OU)
Sel .-----~KIs
terpolation block.
_-L,__IL-_~") n ...L
Returning to Fig. 2.26, we see that our input signal changes every Ts seconds and
has the assumed spectrum seen in Fig. 2.27a (marked (a) in Fig. 2.26). A log2K size
, Ts counter and AND gate are used to select the input signal one of K times so we get the
zero padded waveform seen in Fig. 2.25b (spectrum shown in Fig. 2.27b). Note that the
Nyquist frequency has moved from 1/2 to Kj;/2. The problem is that the images are still
(b) zero padding
• K=4
•••••• I.~i, (a) •••
I, New ~uist frequen~y f
(c) hold register
(b) •••
it""HU ") i, T
s
(c)
removal filter'
f
ation. Figure 2.27 Example spectrums when zero padding interpolation is employed.
46 CMOS Mixed-Signal Circuit Design
present in the wavefonn and thus need to be removed. This is the point of using the image . Ifwe multiply the top an,
removal filter seen in Fig. 2.26. Figure 2.27c shows the spectrum after increasing the
sample rate and passing the signal through an ideal image removal filter. Also seen in this
figure is the response of the RCF. It now starts rolling off at is 12 and can extend up to
Kf, - f,/2. Note that the RCF's response seen in Fig. 2.27a must cut off abruptly at is 12.
Finally, Fig. 2.27d shows the spectrums when a non-ideal digital image removal filter is
used. The figure shows the nonzero spectral content remaining in the spectrum after
filtering.
Hold Register
Figure 2.28 shows the block diagram of interpolation using an input hold register (which
may simply be the input of a digital filter). The benefit of this topology is simplicity. We
simply clock the input word K times faster into a register to increase the clocking
frequency as seen in Fig. 2.25c. Note how the wavefonn seen in Fig. 2.25c looks similar
to the output of the S/H discussed earlier in the chapter (so we should be expecting some
sort of Sinc response effect on our input signal as seen in Eq. [2.16]).In order to quantify
we label the interpolato
this last comment we can write the output of the hold register in tenns of its input using
T ] K·(n+ 1)-1 1 T ] r
r
Yu i·..2. = L -·x i·..2. =x[nTs]
_ K i~Kn K ~ K
(2.30)
noting, K· n :::; i ~ K . (n + 1) I , or
ile point of using the image Ifwe multiply the top and bottom of this equation by I we get
ctrum after increasing the
)val filter. Also seen in this H (z) = Yu(z) (2.34)
u X(z)
ls 12 and can extend up to
ust cut off abruptly atls 12. or
ital image removal filter is
ling in the spectrum after (2.35)
(2.32) This equation is plotted in Fig. 2.30 for varying K. Note how the maximum amount of
attenuation approaches 13.5 dB (the same as the Sinc response seen in Fig. 2.17).
(2.33) IHu(f) I
Desired sl"""~"""-_
It's also of intcrest to determine how much droop the filter will introduce into the
~ spectrums in 2.32. signal frequencies of interest at the Nyquist frequency. Figure 2,31 shows the droop
48 CMOS Mixed-Signal Circuit Design
YU</)~ Is
This equation is plotted in Fig. 2.31. Note that as K gets large the amount of droop
approaches 3.9 dB (= 0.64) or the same as the Sinc response seen in Fig. 2.17.
(b)~
RCf (;
dB
\- ~
..
(C) • . . . .
Main lobe I
4 13.5 dB
.
'(d)
~.. '
JIII\~
IFirst sidelobe
I:!.
-tlill-i-I 1 !~
f,/2
Kfsl:
346 8 10 K
Interpolation
Figure 2.32 shows some example spectrums when K = 4 and using an input hold
register for interpolation. In (a) we see a representative input spectrum that repeats every r
is. In (b) the input is upsampled to a clocking ratc of K -j,. The effects of the input • u vi
holding register's Sinc response are seen. At this point we have distorted our desired
K· n ~ i < K· (/1 +
information with the drooping response seen in Fig. 2.31. In (c) the signal's spectrum,
after passing through an ideal image removal filter to remove the undesired frequency Writing this equation in th
components, is seen (see also Fig. 2.28). The RCF can have a slow roll-off ultimately
passing negligible content at frequencies above K·1s - I s l2 where the image around the
-I-
I
new clocking frequency, K·j" exists. Finally, in (d) we show what the spectrums may
look like with a non-ideal image removal filter. The design of the RCF depends on the
allowed amount of unwanted spectral content that can be tolerated in the final output
_1- spectrum. which is the same as Eq. (:
interpolation to introduce
IHu(f) I Droop, dB
. T,]
:::t ·
[
I~Dmop
YI'K
t
-1-
I
~-.-----:- --.--.--.~
!s,new/2K
or
!
-3.8 -1
-4,0 t--t--t-+-+
3 4 6
+-+-1
8
-+--+I~
10 K
11,
Interpolator 11
f,/2
Figure 2.32 Example spectrums when interpolation using a hold register is employed.
Linear Interpolation
The final interpolation scheme we'll look at is linear interpolation, adding samples in
between the interpolator's inputs that linearly change with time, Figs, 2.25 and 2.33. We
1( = 4 and using an input hold can describe linear interpolation mathematically using
ut spectrum that repeats every '
Yu [C 1+
\).Ts]_ [ .. Ts] x[(n+l)Tsl x[n·Ts]
K - Yu 1 K + K (2.40)
.j;. The effects of the input
ife have distorted our desired
. In (c) the signal's spectrum, where K·n:5,i<K·(n+l). Note that when i K·n, Yu[i.1-]=x[nT,]=yu[nTs].
Gove the undesired frequency Writing this equation in the z-domain we get
ave a slow roll-off ultimately
Yu(z). zn+llK z". [ Yu(z) + X(z) . ~- X(z) ] (2.41 )
: where the image around the
how what the spectrums may
Yu(z) \
n of the ReF depends on the (2.42)
: tolerated in the final output
X(z) = K' 1
which is the same as Eq. (2.34). After reviewing Fig. 2.25 we might have expected linear
interpolation to introduce less distortion into the input signal than an interpolator made
••
i
i • • ••
6 8 10 Interpolator inputs
l-,==~~==~~====~~==~~~~~~l
+"
.K
' - Ts
K, n :5, i:5, K· (n + \) - \
dwidth. Figure 2.33 Using linear interpolation.
50 CMOS Mixed-Signal Circuit Design
using a hold register. However, the fact that the output of an interpolator made using a
hold register sees abrupt changes or steps over time (indicating higher spectral content)
Let's show how to apply K
results in the equivalence in both interpolator's spectral responses. Note that this type of
analog signal processing).
interpolator is more commonly found in analog signal processing using discrete-time
switch is closed when a no
analog circuits.
. high at the same time) sign
2.1.6 K-Path Sampling on the capacitor is
In the last section discussing interpolation (upsampling our signal or increasing the clock
rate) the input clock rate was Is and the output clock rate was Kis. If we define z in terms
and when the q,2 switch ope
of the input clock, z e j21t .{, then a delay on the output of the interpolator is written as
= e-
j2rtf
r: and a delay on the input of the interpolator is written as . If, on the Q2 vout[nTs]' (el
other hand, we define z in terms of the output clock, z then a delay on the output
is written while a delay on the input of the interpolator is written as . In simpler
terms, a delay on the input lasts T, seconds while a delay on the output lasts T)K seconds.
In analog signal processing we can get the same type of behavior, an increase in the
output clock frequency (or upsampling the input signal) by using more than one path.
In order to understand this last statement, consider the parallel paths of S/Hs seen
in Fig. 2.34. Each S/H is clocked on an opposite phase of an input clock. The resistors are If we input a DC voltage
used to sum the outputs of the S/Hs. If the input clock is Is, then the output signal will magnitude of Eq, (2.47) is
change every T/2 seconds or at a rate of 21s Hz. By using two-paths we effectively realize simple RC circuit (as seen i
an interpolation rate, K, of 2. In other words, we can think of the two-path S/H topology
as a single path topology clocked at 21s Hz. For K paths we can write, again,
/s,new K (2.43)
and, knowing the ten
cycle (which is negligible f.
Sample and
x(t)Jn---..-~
hold (S/H)
I
1
clock@/s phaseJ
1 Sample and
hold (StH)
'---___t...rL-=-='--_...J phase2
'1-
x(t) -"In!.L-_ _ _ _~ Sample and
I 1---»--y(t)@2J,
hold (S/H)
clock@2/s
(2.49)
~I)@f, or
Yl (t) + Y2(t)
2
1 ,
'0)0 time
-y(t)@2/s
i
rS~l
R,,- f,C, ~ CF
VO"t(t)
(n (n+ 1I2)Ts
We can also write the exact frequency response, Eq. (2.47), of this switched-capacitor
circuit
(2.52)
(~~ + I - cos 2rr.flf,) + j(-sin 2rr.f/fs)
or
(2.53)
Note that for f «fs the cosine terms is approximately one and the sine term is (b) Equivalent Cil
approximately 2rr.flf, so this equation simplifies to Eq. (2.50).
Next let's examine the two-path version of Fig. 2.35 shown in Fig. 2.36. The Figure
effect of using two paths is to double the output sampling rate, fs.new = 2fs. Using Eq.
(2.47) we can write
Overlapping Clock GI2
(2.54) Figure 2.38 shows a circl
signals. A shift register is I
At the new sampling rate we can write figure is used to detect ifm
'Z-l
flip-flops. The amount of I
'1- H2-path(Z) = C C C -2 (2.55) through the two inverters C(
/+ F F'Z
or, in generic terms of K, replace Z in the transfer function of the single-path topology
with zK to get the transfer function in the K-path topology,
.1
(2.56)
'1
elk
Shift register
Figure 2.37 shows K-paths and the equivalent single path topology, a time interleaved
(2.50) topology. Note that at this point there are several important topics we can discuss
including path matching and the effects of clock jitter (more later).
(2.51) ~~~.----------:~
~~--------~---
~3 : n _ _ _ _- - ' _ _
2.47), of this switched-capacitor L
~4 nL______T-__
••
(2.52) •
.....j( -sin 2:rf is)
Non-overlapping docks
(2.53)
I
z :~
ately one and the sine tenn is ej21tjl(Kj,)
(b) Equivalent circuit Ts/K
2.50).
g. 2.35 shown in Fig. 2.36. The Figure 2.37 A K-patb topology and its equivalent circuit. -I
)ling rate, /s,new = 2is. Using Eq.
/.
(2.54)
Non-Overlapping Clock Generation
Figure 2.38 shows a circuit useful for generating four-phase, non-overlapping, clock
signals. A shift register is preset so that only one bit is high. The logic block seen in the
I
figure is used to detect ifmore than one output is high. Note the clear and set inputs of the
flip-flops. The amount of non-overlap, or dead time, between pulses is set by the delay
(2.55) through the two inverters connected to the outputs of the NAND gates.
elk
Shift register
,
~2 In nanometer CMOS the I
Vin+~ giving a gain-bandwidth pn
,,
The SIB is an exarr
Vin- ··~--.L""-C-L- Vont···
classic feedback equation
to t1 ~2
$3
t2 tJ
,
~1
~2
n Since, during the hold open
Can write, for DC,
1
~3
I ?~-
Indicates plate closest to the substrate As the op-amp's open-loop i
that the closed-loop gain w
Figure 2.39 Fully-differential SlH differential topology.
;; Mixed-Signal Circuit Design Chapter 2 Sampling and Aliasing 55
Example 2.3
md discrete analog integrators Simulate the operation of the S/H building block seen in Fig. 2.39 assuming Cp
grams for the circuits that will 1 pF and j; 100 MHz. Show that the sampled signal isn't affected by the
mmed, in this section, that the matching of the two capacitors in the S/H or by an op-amp offset.
lue seen in 2.20 and the The simulation results are shown in Fig. 2.40. In part (a) the clock signals are
shown. Note the non-overlap time. Unlike the clock signals shown in Fig. 2.39
where the falling edge of ~2 is delayed from ~I' the simulation sets the signals so
they go low at the same time. This was to avoid the outputs of the op-amp
)gy seen in Fig. 2.20 is seen changing to very large values for the small amount of time the ideal op-amp
'hen the ~; and ~c switches are operates open-loop with an input signal applied.
hes are closed the ~I and ~2
;tance is "held" until the next In part (b) we show the op-amp outputs. Note how, when ~I goes high, both
outputs are set to the common-mode voltage by forcing the op-amp into a follower
configuration (which may lead us to use switches to short the terminals of the
Itof the SHand its output by op-amp to VeM when ~I is high if offset isn't important, more on this in a moment).
are closed (the ~3 switches are When ~3 goes high, the circuit behaves as an S/H. Part (c) of the figure shows the
outputs connected through ~3 switches, as seen in Fig. 2.39, driving 10 pF load
(2.57) capaci tances.
t Vos)
Ie op-amp is in the follower Introducing a 100% mismatch in the two capacitors by changing one of the
values of Cp from 1 pF to 2 pF doesn't affect the simulation results seen in Fig.
put voltages of the op-amp go
2.40. Also, an unrealistically large op-amp offset of 100 mV doesn't affect the
switches close we can wTite
S/H's operation.•
± Vos) (2.58)
Finite Op-Amp Gain-Bandwidth Product
led on the falling edge of ~I' The previous derivations assumed the op-amp had infinite gain so the one op-amp input
was driven to precisely VCM + Vos while the other op-amp input was held at V CM - Vos.
Op-amp open-loop gain is an important parameter when designing a Sill. Let's write the
open-loop gain of an op-amp, assuming a dominant pole, as
AOLDC V oul+ - V oul
AOl.(f) . f Vin+ - Vin
(2.59)
1 +J'
hay
In nanometer CMOS the DC gain, AOLDC' may be 500 while the j;dB may be 100 kHz
giving a gain-bandwidth product, or unity-gain frequency, of 50 MHz.
The S/H is an example of a feedback system that can be characterized using the
classic feedback equation
AOL
ACL = 1 + AA (2.60)
I-' OL
Since, during the hold operation, all of output signal is fed back to the input P 1 and we
can write, for DC,
AOLDC
ACL (2.61 )
I +AOLDc
the substrate
As the op-amp's open-loop gain becomes veri large the SIB's gain moves towards I. Note
that the closed-loop gain will always be less than the desired value. If, in a mixed-signal
:ial topology.
56 CMOS Mixed-Signal Circuit Chapter 2 Sampling an
circuit, VDD is I-V and a SIH with a resolution better than 1 mV is required, then the or AOLDC > 1,000. We
open-loop gain of the op-amp can be estimated using
VDD Resolution ..!...1_--",0-,-".0,-",0..!...1 < A OLDC
(2.62)
VDD 1 1 + AOLDC We can estirr
frequency /."') Aowc f"d
4.0V
(noting that this assUI
response is first-order)
3.5V
l.OV
~1
2.5V
2.0V
1.5V
1.0V
O.5V
~J
O.OV
(a)
This first-order system
~O,5V
Ons 4ns 8ns 1608 20ns 24ns Clock signals
n 1 m V is required, then the or AOLDC > 1, 000. We can write this equation in a more useful form as
or
!,Ij~,I
Ts
t settling < 2is 2
(2.68)
removes the offset voltage. When the ~I switches are closed the op-amp is in the
Inputs and outputs
~---r----l unity-follower configuration and its inputs move to VC,~f + Vos + Vino/se(t). When the ~J
switches close the voltage on the inputs of the op-amp is VCl..f + Vos + Vino/se(t J)' lJ sing
Eq. (2.57) we can write, assuming tl "" lJ,
fthe SIH in Fig. 2.39
(2.70)
58 CMOS Mixed-Signal Circuit Design Chapter 2 Sampling and J.
Iassuming
~
Vonoise
~I
:=
V mOlse
clock signals h.
~2
~3 a near 50% duty cycle.
Vou!
~3
I
Fig
Figure 2.41 SiH with input-referred offset and noise shown.
In the preceding di
At a time T/2 later, thc ~3 switches open and the ~I switches close again to sample the noise spectrum. However,
input signal and the noise. Writing the charge on CF between t3 and t] +T)2 and weighting the output 1
(2.71 ) Having said this, howev(
band limited to the Nyquis
Qualitatively, we can see that if the noise is moving slowly (e.g., Flicker noise) it is DC tof)2 range (the desin
removed from the output signaL However, fast moving noise isn't subtractcd out during
the autozero process. Ultimately the bandwidth of the circuit (say switch resistances and
capacitors) and op-amp finite bandwidth limit the frequency content of the noise.
(a)
To get a quantitative idea for how the autozero process affects noise in the S/H's
output signal we can write
vout(t) = Vin(t3) + Vinoise(t) - Vinoise(h) for t3 :s; t:S; t3 + Ts l2 (2.72) (b)
Focusing on the noise and taking the Fourier Transform of each side of this equation
gives
.,- Vonoise(f) := Vinoise(f) . (e /2rr!(t-lj) - 1) (2.73)
I
1
Note that when t is close to t3 the output has little noise. The worst case situation is right Desire
I before the ~3 switches open at a time t3 +TJ2 (the ~3 switches are on for T)2 seconds). If
.1_
we look at this worst-case situation only, then Figure 2.4:
~l
:l Vono;se
Vlnolse 2
assuming clock signals have I
~2
a near 50% duty cycle.
~~.~ Sample-and-hold with T = Ts
~3
I )
In the preceding discussion we didn't include the effects of sampling on the output
litches close again to sample the
noise spectrum. However, at this point, replicating the noise spectrum at multiples off,
leen 13 and t3 +TJ2 and weighting the output noise by the S/H response should be straightforward, Fig. 2.43.
Vas vinOlse(t» (2.71) Having said this, however, note that if our op-amp's input-referred noise PSD isn't
bandlimited to the Nyquist frequency, //2, as it is in Fig. 2.43 then noise will alias into
slowly (e.g., Flicker noise) it is DC to//2 range (the desired range).
noise isn't subtracted out during
ircuit (say switch resistances and
ncy content of the noise.
process affects noise in the S/H's (a) Bandlimited op-amp input-referred noise spectrum
f
3) for t3 :s; 1 :s; t3 -'- Tsl2 (2.72) . After autozero process
m of each side of this equation
So why would,
.~.~_ •.__
which is a signi
since the inputs of
driven to a known
VeM) that ensuring good
Note that in a two-stage
diff-amp together so th<
voltage of the (
as is our next topi
Noise and offset. 'ol",.,t,,,,,O' Capacitor Size
Figure 2.44 Showing how CDS reduces noise and offset in the SfH in Fig. 2.39.
Figure 2.45 shows a S/H that doesn't employ CDS. This topology is used to help
ensure the CMFB circuits and biasing in the op-amp are more tolerant to offsets. When
the 4>1 switches are on, the inputs of the op-amp are shorted to the common-mode voltage, Table 2.1 (
VeM' and the outputs are shorted together. Note that the op-amp settling time isn't a factor
in the design during this hold portion of the S/H process. At this point in time we can
write
(2.76)
When the 4>3 switches turn on, the op-amp moves into the follower configuration and the
op-amp inputs move to VCM± Vas. During this time we can write, see Eq. (2.58),
(2.77)
Since charge must be conserved
V oul = V in ± Vas (2.78)
I
1
Figure 2.45 Fully-differential S/H differential topology without using CDS. ~~.-------""--~.--.~----~-------..
So why would we use this configuration? Neither the offset or noise would be
reduced which is a significant disadvantage over the topology in Fig. 2.39. The answer is
that since the inputs of the op-amp are driven to a known voltage (VCM) and the outputs
are driven to a known voltage via the op-amp's common-mode feedback circuit (again
VCM) that ensuring good biasing and stable CMFB loops in the op-amp is easier to attain.
Note that in a two-stage op-amp design we would also short the outputs of the first-stage
diff-amp together so the inputs of output buffer are forced to a known value (the ideal
output voltage of the diff-amp). Design of fully-differential op-amps is discussed In
detail, as is our next topic, in the book CMOS Circuit Design, Layout, and Simulation.
MV~. Noise and offset. Selecting Capacitor Sizes
~~d
L8~s O.9ps 1.0ps
The selection of the capacitor sizes in the SiR is based on thermal noise considerations,
kTIC (kay tee over cee), and settling time. Small capacitors result in lower power circuits
and faster settling times but, at the same time, increase the thermal noise floor. For a
:et in the SIR in Fig. 2.39. detailed discussion of kTIC noise, as well as other circuit noise topics, see CMOS Circuit
Design Layout, and Simulation. Table 2.1 shows the relationship between various
. This topology is used to help
capacitor sizes and corresponding thermal noise for quick reference.
more tolerant to offsets. When
1 to the common-mode voltage, Table 2.1 Capacitor size and corresponding kTiC noise at 300 OK.
-amp settling time isn't a factor
!
'. At this point in time we can Capacitor I JkTIC, IlV JkTlC, mV
size, pF RMS peak-to-peak
i
i
6.4 0.038
(2.78)
2.2.2 The 5tH with Gain
Figure 2.46 shows a SiR with gain. Following the derivations from the last section we
can write
'l ~, fC L
T
and when the ~J
Q~~ =: CI' (Vin
switches tum on
VCM± Vas )+ . (Vin VCM± Vas) (2.79)
A Qr
Vout+
J
= c/, (VCM VCM Vas) (2.80)
Vout-
~CL and
~3
Q~: =CF'(VoUi Vcw =Vas) (2.81 )
Knowing charge must be conserved
600mVr
500mVl··· .. ·
4oomVT· .. · ..
30omv-'",'~
2oomvJ ..
100mVr-f· ..
v
OmV --.--
·IOOmV ...... ·
·200mV~ .. · .. -
I .
~300mV ..... -
I .
·400mV- .. ··":
.50omv) ....... :
-600mv-I----i
Ons 501
Figure 2.48
Figure 2.46 A S/H with gain.
A block diagram for the S/H in Fig. 2,46 is shown in Fig. 2,47. Note that this op-amp
topology employs CDS. Also note, though we didn't derive it in the last section (because
it ends up being negligible in most circuits), the residual offset after autozeroing is
VosIAoLDC. An op-amp DC gain of 1,000 and an op-amp offset of 50 mV results in a 50 Vout = Vouf+
j..lV residual offset when employing CDS.
1 +_C_,
CF
I
"1
"1
Example 2.4
Simulate the operation of the data converter S/H building block shown in Fig. (Partial
2,46. Assume C1 = CF = 1 pF and Is 100 MHz.
The simulation results are shown in Fig. 2,48. The gain, as we would expect, is 2.
It may be useful at this point to simulate this circuit with an offset or noise like we
did in Ex. 2.3 .• Figur
; Mixed-Signal Circuit Design Chapter 2 Sampling and Aliasing 63
$3
i
TCL
~
l Vout+
V OUl
I~CL
- I
Figure 2.48 Simulating the SIR seen in Fig. 2.46 with a gain of2.
~)
or, after reviewing Eqs. (2.82) - (2.63)
. (Vin+-Vin-) (2.84)
( CI)
Vou!+ ~ 1+ C • Vjn+ - C • VCI+ (2.86)
g. 2.47. Note that this op-amp F F
e it in the last section (because The differential output voltage is then given by
ual off'let after auto zeroing is
offset of SO mV results in a SO Voul = Vuu/+ Vout = (1 + (2.87)
If Fig. 2.46.
1, as we would expect, is 2.
th an offset or noise like we
Figure 2.49 implementing subtraction in the S/H.
64 CMOS Mixed-Signal Circuit Design
Rearranging the block diagram seen in Fig. 2.49 results in the topology seen in Fig. 2.50.
::!~~
.... 2.rl
D.W . . /
O.OY ...... ; ..
Vow
:::~~~lIt:::::::"
3V
-O.4V
-tl· I··· .: .
... , ..
-o.SY ....... ;..
-n.SY ....... ; ..
-O.7Y ....... , ..
-O.BV ......-: ..
-O.9Y ....... j ..
Figure 2.50 Block diagram of Fig. 2.49 with bottom plates ofC/ tied toVC/.
·1.0Y+----+--
Oos 50ns
Example 2.5
Simulate the operation of the SIH shown in Fig. 2.51 i(t; 100 MHz, CF = C1 = I
Single-Ended fo Differem
pF, VCl+ = 1.5VCM> and VC1 _ is O.5VCM (VCM= 500 mY). Comment on the resulting
output.
Many input signals are sing
is desirable in the first sta.
. change it into a fully-dif
differential converter will 1::
distortion (important). In 0
seen in Fig. 2.53. Agai
the $1 switches are on)
VCI- Vout-
Vin
1-
Figure 2.51 S/H used in Ex. 2.5.
,-1
The simulation results are shown in Fig. 2.52. We only show the situation when
we would want to subtract Vc ",/2 from the differential input signal. The inputs are
fully-differential, swinging around the common-mode voltage of 500 mV with an
l
amplitude of 100 mV (so they swing between 600 mV and 400 mY). The largest
$2
Vin+ .. ~- - - .
differential voltage is 600 mV ~ 400 mV or +200 mV while the smallest
, differential signal is 400 mV 600 m V or ~ 200 m V (so the differential signal
1-
swings around ground with an amplitude of 200 mY). Reviewing the block
diagram in Fig. 2.50 with the values in this problem shows that the circuit takes
this input signal, sample-and-holds it, subtract VCM!2 (250 mY) then multiplies it
by 2. We've scaled the output in Fig. 2.52 to show that this sequence of events is I
L
indeed what is happening. Note that if we were to switch VCl_ and VCI_ we would
add VCM!2 to the input signal. • Figure 2.5
S Mixed-Signal Circuit Design Chapter 2 Sampling and Aliasing 65
J: CL
J
Q1 =2C/.(v2n VCMtVOS)
noting that the charge on the C[ capacitors (half ofEq. [2.89]) redistributes through the
(2.89)
5.
(2.91 )
or
Vin+
CF'(VOUf+- os == C1+'2
VCM± V) (2.92)
Assuming C1+ = C1_ and taking the difference in the SIR outputs gives
(2.93)
Note that this topology doesn't employ correlated double sampling (CDS). Also note that
if we model the op-amp's offset with a single voltage source in series with the
n-l
non-inverting input of the op-amp then one of the inputs will go to VCM + Vos 12 while the n -1/2
other input will go to VCM - Vosl2 (we've just indicated the inputs of the op-amp are at a
potential of V CM ± Vos). Rence the factor of two in Eq. (2.93). In other words, if we
Figure 2.54 S
re-write all of the equations in this chapter by replacing Vas with Vas 12 then the factor of
two in Eq. (2.93) will go away. Simulations at CMOSedu.com are invaluable to
To begin, let's assu
understanding the operation of the circuits in this chapter. For example, see the
~] switch. When the ~I
simulation for Fig. 2.53.
shut off), the char
2.2.3 The Discrete Analog Integrator (DAI)
The final sampling circuit we'll discuss in this chapter is an analog building block that we
will find useful in implementing our data converters using feedback. The discrete analog and the output of the integI
integrator, DAI, is shown in Fig. 2.54. The two clocks signals, ~J and ~2' form ;,.,;;",?:c",;;stored on C 1 becomes
nonoverlapping clock signals. The common mode voltage, VCM ' falls halfway between
the mixed-signal system's high- and low-reference voltages (generally VDD and ground).
Note that the parasitic capacitance to ground associated with the bottom-plate of C1 is remembering that the Op-a!
charged back and forth between v] and v2 but doesn't affect the amount of charge in these charges, Q2 - QI ,
transferred to the feedback capacitor, Cr For this reason this DAr is often called a a~ou,tput voltage change. 1
,
-,
Table 2.2 Discrete analog integrator (DAI) input/output relationships.
The transfer function of the
Input
VI = input and V2 == VCM
Similarly, if we connect th,
shift in time by Ts!2) \
[OS Mixed-Signal Circuit Design Chapter 2 Sampling and Aliasing 67
(2.91)
Vout
VI
..-, Vin4
~'+'2
(2.92)
V2------------------~
Similarly, if we connect the output through the ~2 switches (the edges we label n in Fig.
2.54 shift in time by Ts12) we can write
QJ = C/(VCM vl[(n 1I2)Ts]) (2.99)
Q2 CICVCM - v 2[nTs]) (2.100)
and
68 CMOS Mixed-Signal Circuit Design
Example 2.6
Determine the transfer function of the DAI of Fig. 2.54 without the switches on noting the input signals ca
the output of the op-amp. discussed further in the ne)
Reviewing Fig. 2.54 we see that charge is transferred to the feedback capacitor DAI
only when the <\12 switches are closed. Therefore, the output only changes states
during the time interval when the <\12 switches are closed. The transfer function of
the DAI, when no switches are used on the output of the op-amp, is given by Eq.
(2.102). Using the <\11 switches simply adds a half clock cycle delay, Z·112 , to the
integrator's transfer function (instead of the output changing with the rising edge
of <\12' the output changes one-half cycle later on the rising edge of <\11)' •
A Note Concerning Block Diagrams
As we draw block diagrams describing our modulator topologies in this chapter and the
next we often show a circuit like the one shown in Fig. 2.55. The summation, gain, and
integrating blocks are implemented with a single switched-capacitor DAI having the
,, transfer function given by Eq. (2.102). The gain, G, of the DAI is set by the ratio of
,,
,
(2.101) capacitors as indicated in the figure. It's important to realize that this circuit is entirely
analog and is interfaced to, in general, both ADCs (vout[z] is connected to the input of an
) the CP2 switches is then ADC) and DACs (v2 [z] is eonneeted to the output of a DAC).
(2.102) It should be clear from both Fig. 2.S4 and Table 2.2 that many different
combinations of discrete analog building bloeks are possible. Figure 2.S6 shows two
other possibilities. In part (a) the capacitors used have the same parasitic capacitance on
each plate so there is no benefit to using a parasitic insensitive topology. The transfer
(2.1 03) function of this DAI is
7. Note that the factor C1ICF VOU1(Z) = C· -J . (V2(Z)- VI (z)) (2.1 04)
F l-z
then the magnitude response,
umerator simply modifies the noting each input signal sees the same delay, i.e., Z-I when the outputs are connected
)r -180 degrees) and has no through CPI controlled switches and delay when no switches or CP2 controlled switches
3 could discuss the frequency are used. If the integrator inputs must see the same delay and the capacitors available
Never, we would see that the have asymmetric parasitic capacitance, the topology of Fig. 2.S6b can be used. Its transfer
'ator would apply to the DAI function is
(a)
.logies in this chapter and the
5S. The summation, gain, and
!d-capacitor DAI having the
Ie DAr is set by the ratio of
READING
V2+ CF R. J. Baker, CMOS
Vl+--/
~1
:CI
I ;J Edition, Wiley-IEEE.
C. C. Enz and G. (
T Op-Amp Imperfectio
-t
/"
VCM ; Stabilization," Proc(
~ ::r Vou/ November 1996.
VI_--~
'CI 1. W. Couch, Mod,
~
Prentice-Hall, 1995. I
CF
V2
C 1 . v I (z) . z1I2 - V2(Z)
VOUI(Z)
l-z- 1
Qualitatively, using f
CF
an alias of the sample
Figure 2.57 Fully-differential discrete-analog integrator (OAI) implementation. Re-sketch Figs. 2.12
logic to implement th
DAI Noise Performance
Explain why retumin
Figure 2.58 shows the DAI with kTIC noise sources shown (see Table 2.1). The mean a signal. What is the \
squared input- referred noise is given by
Sketch the input and
(2.106) the DC component oj
4 MHz with a peak
in series with both VI and v1 . A total of 2kTICI is sampled onto C1 during each clock cycle. MHz.
If the input signal can swing from VDD to ground (peak value of the input is VDDI2
while the RMS value of this input is VDD!(2.fi ) ), then we can estimate the SNR using
f \ In Sample and
VDD/~2.fi ) hold (S/H)
SNR = 20 log -;::::===- (2.107)
This equation is useful for determining the capacitor values used in a DAI.
~1
ADDITIONAL READING
[l] R. 1. Baker, OvIOS: Circuit Design, Layout, and Simulation, Revised Second
Edition, Wiley-IEEE, 2008. ISBN 978-0470229415
[2] c. C. Enz and G. C. Temes, "Circuit Techniques for Reducing the Effects of
·---Vaul+ Op-Amp Imperfections: Autozeroing, Correlatcd Double Sampling, and Chopper
+ Stabilization," Proceedings of the IEEE, Vol. 84, No. 11, pp. 1584-1614,
/ " ' ' ' + - - - Volll- November 1996.
[3] L W. Couch, Modern Communication Systems: Principles and Applications,
Prentice-Hall, 1995. ISBN 978-0023252860
QUESTIONS
:z) 2.1 Qualitatively, using figures, show how impulse sampling a sinewave can result in
an alias of the sampled sinewave at a different frequency.
. (DA1) implementation. 2.2 Re-sketch 2.12 and 2.13 when decimating by 5. hint: use a counter and some
logic to implement the divide by 5 clock divider.
2.3 Explain why returning the output of the S/H reduces the distortion introduced into
m (see Table 2.1). The mean a signal. What is the cost for the reduced distortion in a practical circuit?
2.4 Sketch the input and output spectmm for the following block diagram. Assume
(2.106) the DC component of the input is 0.5 V while the AC component is a sinewave at
4 MHz with a peak amplitude of 100 m V. Assume the clock frequency is 100
lto C[ during each clock cycle. MHz.
value of the input is VDDI2
: can estimate the SNR using
In Sample and Sample and Sample and Out
hold (SIB) hold (SIB) hold (SIH)
107)
clock
used in a DAI.
when $2 switches close 2.8 Verify, with simulations, that the topologies seen in Fig. 2.34 are equivalent.
2.9 Determine the transfer function, and verify with simulations, the behavior of 4
.e DAl.
paths of the switched-capacitor topology seen in Fig. 2.36.
72 CMOS Mixed-Signal Circuit Design
2.1 0 In your own words discuss why the ~2 switches are shut off after the ~, switches in
the S/H seen in Fig. 2.39.
2.11 Sketch the op-amp's open loop response, both magnitude and phase, specified by
Eq. (2.59).
2.12 What is the voltage across CH in Fig. 2.41 in terms of the input-referred offset and
noise? Verify your answer with simulations commenting on the deviation of the
frequency behavior of the input-referred noise to the frequency response of the
voltage across the capacitor.
2.13 Provide a quantitative description of how capacitor mismatch will affect the
operation of the StH seen in Fig. 2.46. Verify your descriptions with simulations.
nalog F
2.14 Is it possible to design a SIH with a gain of 0.57 How can this be done or why
can't it be done? Use simulations to verify your answer.
2.15 For the first entry (v, input, v 2 = VCM) in Table 2.2 derive the frequency response,
magnitude and phase, of the DAr. Use simulations at a few frequencies to verify
your derivations. the last chapter we disc[
(RCFs) are an imr
2.16 Repeat Question 2.15 for the second entry. pertonm signal processing
2.17 Repeat Question 2.15 for the third entry. RCFs are still require
wider bandwidths)
2.18 Does the DAr use CDS? Why or why not? Use simulations to support your sWltclled-caLpa1c\t()r) counte
answers. cannot be fabricate.
,"~~}"'''l<'L' y true if passive n
±20%. By using activ.
elements, we can e
order filters while IT
I'
I I In this chapter Wt
I integrators (CArs or acti'
I
.......
capacitor (gm-C) integrator:
I practical analog filters f(
I'
I
fully-differential inputs ant
the active filter remains co
. signal using fully-differen
outputs. Single-ended tope
such as Sallen-Key, and
, inversion such as Tow-Th.
I
and Schaumann [2].
I,
where ro 2n -f andfis the frequency of the input (and thus the output). Next, consider
the block diagram in Fig. 3.2. This figure shows an integrator and a summing block. The
output of the block diagram can be determined by solving Figu
I
1
1 Vin+
Out
f---~-........-~-- voU/(J)
Vin-
, Integrator
1
IOf3dB
Vin+----.j
(3.2)
Reviewing Fig. 3.2, we see that the CAl of Fig. 3.3 alone will implement the
needed integration but not the summing (difference) block. By adding an additional
feedback path, the entire block diagram of Fig. 3.2 can be implemented. Figure 3.4 shows
(3.3) the integrator-based implementation of the circuits in Figs. 3.1 and 3.2 (noting the
op-amp must be able to drive a resistive load). This filter is called an active-RC filter
1 to Eq. (3.1), we see that if we because the RC is used with an active element (the op-amp). At this point there are
several practical and useful modifications that we can make to this filter. However, let's
work an example before moving on.
Q (3.4)
2rc
er filter (the filter has a single
Vout
?"'-_.-- VOUl+
Example 3.1
Simulate the operation of the filter in Fig. 3.4 from DC to 100 MHz if R 10k
and C = 10 pF. Show both the magnitude and phase responses of the filter.
It's important to notiee (fe
Assume the op-amp is ideal.
the voltages are first chanl
on the inputs of the op-am
From Fig. 3.1 we know the 3 dB frequency of the filter is 1.59 MHz. The
simulation results are shown in Fig. 3.5. The magnitude and phase response
follow, as expected, the responses for the simple RC filter shown in Fig. 3.1. •
the block diagram iJ
that at DC where s
diagram with the forward
from classic feedback thee
3.6, R/Rr Of course, all
capacitor is an open result:
» (
I
r
I
I
Figure 3.5 Magnitude and phase responses for the first-order filter in Fig. 35.4
ifR= 10k and C IOpF.
.- What would happen if we switched what we define as Vout+ and Vout- in the filter
described in Ex. 3.1 without changing any other connections? Perhaps it is trivial, but the
Vin+
answer is that the output will be inverted. We can modity the block diagram of Fig. 3.2 by
simply mUltiplying the output by -1, as seen in Fig. 3.6. The phase shift in Fig. 3.5 would
. shift up or down by 180 degrees. It would vary from 180 to 90 degrees, or from -180 to
-270 (because +180 degrees is the same as -180 degrees) instead of from 0 to -90
degrees. If we allow the resistors used in the filter to have different values, as seen in Fig.
3.6, we can add a feedback gain to our block diagram. Assuming the outputs are labeled
so that we don't have an inversion in the output of the filter (i.e., they are labeled as seen
in Fig. 3.4), we can write
d-Signal Circuit Design Chapter 3 Analog Filters 77
Note that at DC where s ~ 0, the block diagram in Fig. 3.6 becomes the classic feedback
diagram with the forward gain approaching infinity and a feedback factor of O2, Then
from classic feedback theory, the closed-loop gain becomes 1/02 or, for the filter in Fig.
3.6, R/R r Of course, analyzing this circuit (using loop equations) at DC when the
capacitor is an open results in the same gain.
Integrator
Example 3.2
or, with A OLe!) Wunfs
ModifY the filter in Ex. 3.1 so that the low-frequency gain is 20 dB.
Using Eq. (3.7) or Eq. (3.8), we leave C = 10 pF and RF = 10k. To get the gain of
10, we make R[= lk. The simulation results are shown in Fig. 3.7 .•
This equation is very re,
can be in a filter. For the
the i term in Eq. (3.14
VOUl(j) I
I v;n(j)
Example 3.3
Suppose a first-order
Figure 3.7 A first-order filter with gain, see Ex. 3.2. where Wun 10IRFC
how the magnitude aI
Effects ofFinite Op-Amp Gain Bandwidth ProductJ"" op-amp,!,,".
In the previous two examples we assumed a near-ideal op-amp. The open-loop gain of the The op-amp's unity g;
op-amp can be written, assuming a dominant-pole compensated op-amp, by the filter. This means
equal to the desired b,
Vaul
Ao[.(j) v+ -v. = (3.10) circuit is fun 11 0, wl
1I2rr.R F C. The magnit
I where v+ and V. are the voltages on the noninverting and inverting op-amp input
I ' terminals, respectively. Note that we are using hdB in both Figs. 3.6 and Eq. (3.10) to
1
indicate the 3 dB frequency of a frequency response. While the symbol is the same the
...
I
actual values vary from circuit to circuit. When a practical op-amp is operating at which shows the filter
frequencies above a few kHz, we can approximate the open-loop response (knowing the
imaginary part of the denominator is much larger than the real part) as The point of the
I minimize the effects of tb
1-' A (j)_AOWC·/3dB _fun 2rr.fun Wun (3.11) be used. In the remaini
I OL - jf - j S s
approximated as
I where!,," is the frequency where the op-amp's open loop gain is unity (0 dB). Rewriting
1
. Eq. (3.6) to include the op-amp's finite gain bandwidth product (that is,!,,") and assuming,
...
without the loss of generality, that the op-amp is operating with a single-ended output (v+
tied to VCM [AC ground]), results in The poles of this transfer·
.
Vin V-+Vaul v-+SC'(Vaul-V_) 0 (3.12)
RJ RF
After some algebraic manipulation with v. -V au/fA OLC!), we get
noting that if Wun ~ 00, th
VaUI = -----_-.::'-----;----:: (3.13) Fig. 3.1).
Vin
1+
Desired response
xed-Signal Circuit Design Chapter 3 Analog Filters 79
I I
vo.ut
Vm
(3.15)
Example 3.3
Suppose a first-order filter is designed based on the topology seen in Fig. 3.6,
~x. 3.2. where O)"n = 10IR F C and RFIR/ 10. Assuming 0)2« O)unICRF, comment on
how the magnitude and phase responses of the filter will be affected by the finite
op-amp,fun'
The open-loop gain of the The op-amp's unity gain frequency is only 10 times larger than the bandwidth of
op-amp, by the filter. This means the op-amp's closed-loop bandwidth (with a gain of 10) is
equal to the desired bandwidth of the filter. The bandwidth of a gain of lOop-amp
(3.10) circuit is / un I10, which here is equal to the ideal filter 3 dB frequency of
1I2rr.R F C. The magnitude of the filter's response can be approximated as
I inverting op-amp input RF
lop-amp is operating at
which shows the filter's 3 dB frequency is off by a factor of2.•
,op response (knowing the
,art) as The point of the preceding example is, in general lowpass filter design, to
fun _ O)un minimize the effects of the op-amp's finite fun a low value of closed-loop DC gain should
- S (3.11 ) be used. In the remaining discussion let's assume RdRJ = I, so Eq. (3.14) can be
approximated as
is unity (0 dB). Rewriting
(that is, j~n) and assuming, (3.16)
1 a single-ended output (v+
Spl,p2 (3.17)
set
noting that if O)"n --'t 00, then Spl 00, and Sp2 = -lICRF (the ideal position of the pole, see
(3.13) Fig. 3.1).
80 CMOS Mixed-Signal Circuit Design Chapter 3 Analog Filters
To get some idea of the required op-amp fun «(flun 2rr/ul1), let's assume that we
want the pole to vary no more than 1% from the ideal location due to finite op-amp
bandwidth
it
v_ _
,,]I -l
1.01 (3.19) l(
Repeat Ex. 3.1 if an op-amp is used with a DC gain of 10,000 and an /"n of 10 ·15dB ..• _+
-20dB .. _..,
MHz.
-25dB -... -~.
Because the op-amp's AOLDC is 10,000 and fun = 10 MHz, the op-amp's open loop -30dSJ-.·-:-·
hdB is 1 kHz (see Eq. [3.11]). We can use the circuit shown in Fig. 3.8 in our J
SPICE simulation to model an op-amp with finite fun' The RC in Fig. 3.8 is
selected to give an op-amp open loop hdB of 1 kHz.
:::::8
::::::
-----r
-45dB
-5UdS -.-._;-
. The 3-dB frequency of the filter described in Ex. 3.1 is, under ideal conditions, -55dB .
I 1.59 MHz. Because our op-amp's unity gain frequency is only 10 MHz, we would 10KHz
expect, from Eq. (3.16), the op-amp to affect the frequency response of the filter.
Fil:,'Ure 3.9 shows the simulation results using the op-amp model of Fig. 3.8. Figure 3.9 Magnit
Comparing Fig. 3~9 to Fig. 3.5, we see differences in both the magnitude and ifR=j(
phase responses of the filters. The magnitude response of Fig. 3.9 initially rolls off frequer
at 20 dB/decade below the ideal 1.59 MHz. Around 10 MHz the response
l-Signal Circuit Design
~ _ _ _ _ Vo+
81
-
. due to finite op-amp
(3.18) !
op-amp outputs
(3.20)
Figure 3.8 SPICE modeling of a differential input/output op-amp with finite bandwidth.
olving Eq. (3.20) for x,
lUst be 100 times larger transitions to 40 dB/decade. Clearly this faster roll-off is the result of the op-amp's HAl
Ie pole) to deviate less closed-loop pole coming into play. The limiting behavior of the op-amp, when
e in the filter's cutoff 'dt
looking only at the magnitude response, may be welcome (the filter's response
hdB" Clearly, from Eq. rolls off faster) in a lowpass filter. However, it is not welcome in other filters (a tar
econd-order when the highpass filter, for example). Figure 3.10 shows what happens if we decrease the
lapes of the magnitude filter's 3 dB frequency to 159 kHz by increasing the resistors used to lOOk. What
;een in Fig. 3.1. We can we are doing here is showing how making the op-amp's bandwidth much larger J-c
ble to fabricate precise than the filter's affects the frequency response of the circuit. The magnitude th'
:te bandwidth may still response starts to fall off at -40 dB/decade at the op-amp's unity gain frequency,
)r-based filters. Tuning 1: of 10 MHz. Also seen in Fig. 3.1 0 is the phase response of the filter. The
adjust the precise filter ~;-amp's (closed-loop) phase response, which starts rolling off one decade below
fuses or, if possible, fun' results in the final phase shift of the filter approaching -180 degrees.•
o times larger than the
) DC gain of unity, i.e.,
,veil-defined magnitude
-20 dB/decade
Consider the following
o· -40 dB/decade
00 and anJ,m of 10
I
82 CMOS Mixed-Signal Circuit Design Chapter 3 Analog FilteJ
Figure 3.11
-20 dB/decade
-40 dB/decade The size of the integrat
converters or modulate
Figure 3.10 Increasing the resistance to lOOk and replotting Fig. 35.9. stored in the capacitor u
c
+2:: "--"'1
Vout
/I.
VDD ___ _
+-20'
:r
itj
-40'
-60'
-80'
; -100'
V~_-----J+
R
VCM
1 .
I
VDD
VDDi2- --
; -120'
j -140'
i -160'
~ -180' Figure 3.11 Estimating maximum possible SNR of an active-RC filter.
: -200'
OMHz
The size of the integrating capacitor fundamentally sets the SNR in integrator-based data
converters or modulators. But consider the following: the maximum electrical energy
Ig Fig. 35.9. stored in the capacitor used in an integrator is
Vtul1e
MOSFET-C filter
Vin+
Vin+----'
Vin-
Rn= / (3.26)
KP \' lv,"", - V ~l
[
mN -
Ii";;
Vin+
= Vin+ . KP . L • lune -
W(U V THN - Vin+
) (3.27)
Figure 3.14 Firs
Some improvement in the linearity of the MOSFET resistors, say 10 dB (resulting in a
I DR of 50 dB), can be achieved by utilizing the fully-balanced signals available in the Why Use an Active Circu
I circuit. Consider replacing MI in Fig. 3.12 with the pair of MOSFETs, MIA and MlB,
I,
shown in Fig. 3.13. The resulting current is now Before going any furth
performance using a sim]
Vin+ Vin- KP W[ (V (3.28)
-R +-R = . -L Vin+' lune+ use these elements with
nlA niB
question comes when we
the output of the filter (Wl
»:ed-Signal Circuit Design Chapter 3 Analog Filters 85
MOSFET-C filter
- VouJ.·~
V OUl+
Vin------....l
MIB
G
Knowing Vin+ = -Vin and letting V tune V tune+ - V rune -, we can write the equivalent
current through MIas
~~\
The result is that the nonlinear behavior of the MOSFET's channel resistance due to the
changing drain-source voltage cancels to a first order. Figure 3.14 shows the
implementation of a first-order MOSFET -C filter using linearized MOSFETs.
\
I channel Parasitic
ce channel
capacitance
r.
(3.26)
:J
Vin+) (3.27)
Figure 3.14 First-order MOSFET -C filter using linearized MOSFET resistors.
say 10 dB (resulting in a
j signals available in the
Why Use an Active Circuit (an Op-Amp)?
)SFETs, MIA and MIB,
Before going any further, let's realize that we can get the exact same frequency
performance using a simple resistor/capacitor or MOSFET/capacitor as we get when we
(3.28) use these elements with an op-amp. So, "Why use the op-amp?" The answer to this
question comes when we realize that when a capacitive or resistive load is connected to
the output of the filter (without an active element), the frequency behavior changes. Using
86 CMOS Mixed-Signal Circuit Design Chapter 3 Analog Filten
the op-amp allows us to drive an arbitrary (within reason) capacitive or resistive load. Before discuss in,
Using an active element will also allow us to cascade first-order sections to implement of gm-C filters; namely,
higher order filters. precision application, tl
associated with the amp
3.1.4 gm-C (Transconductor-C) Integrators
be held at a precise valli!
The operational-transconductance amplifiers, aTA, is an amplifier that has only two is not too useful if use,
high-impedance nodes: the amplifier's input and its output. Figure 3.15 shows a schematic resolution is less than t
symbol, transfer curves, and a possible implementation for an OTA (based on a continuous-time signal p
fully-differential diff-amp). Transconductor-C, or g",-C, filters use a circuit, a
We can relate th(
transconductor, that provides a linear voltage-current transfer curve. Our aTA in Fig.
circuit in Fig. 3.15a usinl
3.15 does behave like a transconductor over a portion of the input voltage range but
becomes nonlinear for large input voltage differences, Vin+ - Vin-. By increasing the
lengths of the NMOS diff-pairs used in the aTA, we can increase the linear
common-mode range of the aTA, making it appear as though it were a transconductor. Comparing this result to
The fundamental problems with increasing the lengths of the diff-pair MOSFETs are the
increase in the OTA's input capacitance (affecting the location of the filter's poles and
zeroes) and, perhaps more fundamentally, the inherent reduction in the MOSFET's/r-
The big benefit of the gn
by adjusting the transcon
The circuit of Fig
Fig. 3.3. However, as set
Vin+
first-order filter. In order
Vin ( consider the transconduc
Vin+ - Vin transconductor is gm(Vin,
the second transconductc
(a) Schematic symbol of an OT A
3.l6b, the outputs of th
or transconductor. (b) Transfer curves for an OTA. trans conductor subtract.
we can write
VDD
T-=r~
I '
...;
I;
! '1 ~s;io.ut+ Voul+
I:
I:
I ,
Vin+ r------~ t-- Vin-
,-,
I!
L--+---'~-<I-~
I'
I'
I
I'
~I-'
: I (c) One possible implementation
: I of an OTA.
!I
.I
.citive or resistive load. Before discussing these issues in more detail, let's look at an important limitation
r sections to implement of gm-C filters; namely, the fact that the transconductor's input voltage must vary. In any
precision application, the input voltage must remain constant because of the roll-off
associated with the amplifier's CMRR (unless, of course, the common-mode voltage can
be held at a precise value). This limits the DR ojgm -C filters to around 50 dB. Again, this
ifier that has only two is not too useful if used as an antialiasing or reconstruction filter unless the system's
3.15 shows a schematic resolution is less than eight bits (48 dB SNR). The gm-C filter finds extensive use in
an OT A (based on a continuous-time signal processing.
ters use a circuit, a
We can relate the input voltage difference to the output voltage difference for the
urve. Our OTA in Fig.
circuit in Fig. 3.15a using
input voltage range but
v in-. By increasing the (3.30)
Vout+ VOIII == j(f)C
an increase the linear
: were a trans conductor. Comparing this result to Eq. (3.5), we can use the same design techniques if we require
f-pair MOSFETs are the
of the filter's poles and G _1_ == gm or j3dB =Q (3.31)
n the MOSFET'sjr RC C 211:
The big benefit of the gm-C filter over the active-RC filter is the ability to tune the filter
by adjusting the transconductor's gm'
+ iout-
The circuit of Fig. 3.15a implements an integrator, as does the active-RC circuit of
Fig. 3.3. However, as seen in Fig. 3.2, we also need to implement a summing block in a
first-order filter. In order to move toward the goal of implementing the summing block,
consider the transconductor circuit shown in Fig. 3.16a. The output current of a single
Vin+ -Vin transconductor is gm(Vin+-Vin-). We can sum this current with the output current from
the second transconductor to implement the summing block in Fig. 3.2. As seen in Fig.
3.16b, the outputs of the two OT As are combined, so the output currents from each
;urves for an OT A. transconductor subtract. Assuming each transconductor has the same transconductance,
we can write
(3.32)
Vin
(b)
ransfer curves.
Figure 3.16 Implementing a first-order filter using transconductors.
88 CMOS Mixed-Signal Circuit Design Chapter 3 Analog Filters
or
V out+ - V out-
(3.33)
Vin+ V;n 1 +J<roC . ...L
g",
If the transconductors have different gms, we can design a filter with a DC gain (see
Problem 3.10), which can be characterized using Eq. (3.8).
Example 3.5
Repeat Ex. 3.1 using a gm-C filter with agm of 100 IlAN.
To simulate a transconductor using SPICE, a voltage-controlled current source
can be used as seen in Fig. 3.17. In order to set the output common-mode voltage
to VCM in the simulation, we add the large resistors (whose values can be changed
to simulate the finite, nonideal, output resistance of the OTA) connected to VCM" If
we didn't use these resistors, after reviewing Fig. 3.16, it would result in an F
unknown common-mode voltage on the second transconductor input.
loops. Figure 3.19 shows
components to provide th
r---*------""-Vou/+
Vin+ vinp
V in- --:-.---\....''J
vmm
'WE ________ '--_---..",.---Vou/
SP~ , which is the same result af
G 1 voutm voutp vinp vinm lOOu
~ VCM
Notice how SPICE defines positive current flow as current
flowing from the + terminal to the- terminal
In order to have the same time constant, and thus pole location, as in Ex. 3.1,
,:
I '
let's set the capacitor value, in the schematic of Fig. 3.16 to 10 pF. The value of
the transconuctance, lIgm , is 10k. The simulation results are shown in Fig. 3.18. Figure 3.19 Sl
",-,, As we would expect, the shape follows, exactly, that of the active-RC filter in Fig. th
,,
'
'
3.5. Also, although not shown, the phase response matches as welL.
A High-Frequency Transc
Common-Mode Feedback Considerations
, A transconductor, Fig. 3.:
In most OTAs the load capacitance is used for compensation. These capacitances common-mode range of
compensate both the normal, forward, differential signal path as well as the CMFB path. increased as we discussec
Reviewing Fig. 3.15, we see that the capacitance in part (a) indeed does provide a load for location of the parasitic p<
differential signals. However, any signal that is common to both outputs (a circuit, that other than the
common-mode signal) doesn't cause a displacement current to flow through the capacitor. of nodes: the input and th{
Because both sides of the capacitor change at the same rate for common-mode signals, sum with the load capacita
the change in voltage across the capacitor is zero. This can result in unstable CMFB
~ed-Signal Circuit Design Chapter 3 Analog Filters 89
(3.33)
VeM
[lg a voltage-controlled
Vtune In
I Vine/) ---,)0-)
.
Vin+--{50
INVl [ ~
~l~ I
V aUf-
Vin-
Ifwe apply the filter's input to the noninverting input of the DAI and feed the output back The schematic of
to the subtracting input, see Fig. 3.21, we get clocking frequency
pF, the size of the i
(3.38) 10k, I pF. The 3 dI
Vin(Z) VOUI(Z) Because of the tim(
we can write, see Eqs. (1.62) and (1.66) analysis for the SPI'
filter at 1.59 MHz
I== C· I n DOl f,
2 Isin ny;
C1 1 dL
Iv In
Vout
-v oul F
1.1 an
Voul
. -v out ==
v In -nf,-s -"2 or < < s (3.39) peak-to-peak). The'
The sampling frequency (the frequency the discrete-time filter is clocked at) is is ' while
the filter's input frequency is labeled.! If we require I «Is (say at least sixteen times less
so sin nis "" nis),then we can rewrite Eq. (3.39) as
(3.40)
where
1
G z;--R and iJdB 2n (3.41 )
F sc
placing the gain of the integrator in the same form as Eqs. (3.4). The variable Rsc is a
switched capaeitor resistor
R Ts (3.42)
sc == C1 Is == C1
The equivalent block diagrams for first-order filters using CAIs and using the DAIs
(again assuming! «is) are compared in Fig. 3.2l. Figure 3.22 A switct
Ked-Signal Circuit Design Chapter 3 Analog Filters 91
r []
In Out (!)
Vine!) ) )to ) ) Vout
Integrator
i
(a) C] .J,
G CF S
V OU (- I
"
In Out
Vine!) Vou/(f)
Vout+
Integrator
(b)
,r.
I and feed the output back The schematic of the filter is shown in Fig. 3.22. Here we are assuming the
clocking frequency of the filter is 100 MHz. If the feedback capacitance, CF' is 10
pF, the size of the input capacitor, C/. is then, from Eq. (3042) and knowing Rsc is
(3.38) 10k, 1 pF. The 3 dB frequency of the filter is, once again,1I2n:RscCF 1.59 MHz.
Because of the time-domain (clock) component of the filter, we can't use an AC
analysis for the SPICE simulation. Let's apply a 1 V peak-to-peak sinewave to the
filter at 1.59 MHz and verifY the output of the filter is 3 dB down (0.707 V
: for 0 <I<.Is (3.39) peak-to-peak). The results are seen in Fig. 3.23.
n ~ Indicates plate
is clocked at) is J, , while 10V closest to the
I at least sixteen times less
,-----1 substrate
··--VoUI
(3040)
'----------i+
)
(3.42) 1
The impiementatic
:'::JJ ' .. ,
o.zv - ---+---- -;-- Note that the DAI 1
0.1'1 -- --+-.----: ---
0.0'1 ---i- --i:..;:,..L-i----i---r-'-L--i--i------i---'Lj---.
o.o~s O.2~s O.4~s O.6~s O.8~s 1.0~s l.~~s 1.6~s 1.8PS~~s
-
1.2ps
Let's comment on the exact transfer function of the DAI in Fig. 3.22. We see
in Fig. 3.22 that the output is indeed fed back to the input through a ~l controlled
switch. The result is that the output, through the feedback loop, sees one clock
cycle delay, Z·l. The output is assumed settled on the falling edge of ~2 during
each clock cycle. Because of this, the input, which is settled on the falling edge of
~l' sees only a half clock cycle delay, Z·I!2. This means that the DAI in Fig. 3.22
really has a transfer function of
(3.43)
If we think of the input signal arriving a half-clock cycle earlier, then the only
difference in the transfer function here, when compared to the continuous-time
equivalent and assuming!« is, is a small phase difference. We can delay the
input signal a full clock cycle by adding a q,l controlled switch on the output of the
Figure 3.2!
filter. However, because this switch may be part of the next filter section, we don't
discuss this option further.
The big benefit
Example 3.7
clocking frequency ca
(changing the filter's ch
the 90 dB range have b(
In
Vin(f) --~--{ + I-~----I .9J. Out
s I---:--"-~--- Vout(j) In the previous 1
with the potential alias
input of the filter. An ~
must be used to remo
noise-shaping modulate
the sampling frequency
relaxed AAF design. In!
Figure 3.24 General implementation ofa lowpass first-order filter. the switched input capa
switched-capacitor filtel
:d-Signal Circuit Design Chapter 3 Analog Filters 93
G1G2 (3.45)
/3dB=-
FZ] 21t
Note that the DAI used in this filter has a transfer function of
[J
Ips 2.0ps
. [Cll . vm(z) . Z1I2 ~~. Vout(z)] (3.46)
I Fig. 3.22. •
in Fig. 3.22. We see
'Ough a 4J 1 controlled ~l 4J2
oop, sees one clock
g edge of 4J2 during
.n the falling edge of
VCM---,-\ .
C
I \~-<!'~_l---l
. VCM
Vin----t--~ II T_ ''----t-_-+----j
--VOUl
(3.43)
The big benefit of switched-capacitor-based filters is the fact that the filters' poles
-time lowpass (first- and zeroes are determined by a ratio of capacitors and an external clock frequency (which
is often a precise frequency set by a crystal oscillator). No tuning is needed. Varying the
clocking frequency can precisely set the filter's characteristics for adaptive filtering
(changing the filter's characteristics on the fly). Switched-capacitor filters with SNDRs in
the 90 dB range have been attained at audio frequencies.
Out
~ vou,(J) In the previous two examples we used ideal op-amps and didn't concern ourselves
with the potential aliasing resulting from the analog sample-and-hold operation on the
input of the filter. An analog antialiasing filter, AAF, (that is, not a discrete-time filter)
Vout(f)
- - - == --"--::- must be used to remove the potential aliased signal prior to sampling. As with the
Vin(J) noise-shaping modulator-based data converters we'll study later in the book, the fact that
the sampling frequency, J" is much larger than the input frequencies of interest allows a
relaxed AAF design. Indeed, the resistance of the MOSFET switches used combined with
·order filter. the switched input capacitance (CI ) form a lowpass filter. This filtering may serve as the
switched-capacitor filter's AAF in many designs.
94 CMOS Mixed-Signal Circuit Design Chapter 3 Analog Filters
...
I,
Out
I:
I
Figure 3.26 Digital block diagram of an integrator-based discrete-time lowpass filter. This filter's transfer fum
functions. Using this to
,ed-Signal Circuit Design Chapter 3 Analog Filters 95
Figure 3.27 shows the z-plane plot and magnitude response for this first-order
filter. Remembering that all discrete-time filters have a periodic frequency response (a
overed in this chapter we
period off, or one complete revolution around the unit circle), we can compare this filter's
omponents. High-speed,
response to the filters in the previous examples. To do so, let's assume fs 100 MHz and
:ching or absolute values
write, from Eq. (3.41),
: our discussion on digital
:ad of filters using analog
g waveform without first
f3dB ~' = 1.59 MHz "" 0.1 . Ig~ MHz that is, A= 0,1 (3.49)
digital filters that use We can write the magnitude response of Eq. (3.48), with z = cos 2nfl/S + j sin 2njifs , as
rge (take up a significant
cal in a general purpose A
(3.50)
~d specifically for digital J(cos2njlj~ -1 +A)2 +sin2 2rrfifs
Figure 3.28 shows the responses of the first-order discrete-time filter. The maximum
,cussed later in the book, attenuation of the filter occurs at f,12 or 50 MHz here and is, from Fig. 3.27 with A 0.1,
Its in the circuit. Isn't it 0.0526 or - 25.6 dB.
J digital filtering for the
:"
pator
~ ~,
VOUI(f)
Consider the block diagram of the general first-order filter shown in Fig. 3.29. We can
relate the filter's output to its input using
Voutif) (3.51 )
r
: Out
,
s
VOuf(f)
I +1.iG;"
---= (3,52)
I+ G,sG
2
:ete-time lowpass filter. This filter's transfer function is termed "bilinear" because it is the ratio of two linear
functions. Using this topology we can implement lowpass, allpass (used for phase
96 CMOS Mixed-Signal Circuit Design Chapter 3 Analog Filters
Active-RC Implementatio
The active-RC impleme
Again, as mentioned earl
repeatability from one p
fsl2 on-ehip, with fuses (or s'
the summation is impler
3dB currents are summed a
OdB
common-mode voltage,
-3dB
switehed-eapacitor imp Ie
-6OB
-!ldB We won't discuss
dB VB log of -12dB function. It should be ob'
frequency -15dB linearized MOSFETs (set
-lUdB
-21dB
-2~dB
-27dB
-lOdB
10KHz 100KHz lMHz 10MHz 100MHz
fsl2 3fs12
shifting), and highpass filters (keeping in mind that the highpass filter will ultimately
Vjn - - - _ . _ - -......... 1
change into a bandpass filter response because of the op-amp's or transconduetor's high
frequency rolloff). The location of the filter's pole is given by
(3.53)
In Out
Vjn({) --+--+--1 i-----clo>--+--+--- Vout({)
Figure 3.30 1m
Transconductor-C Implef
- Figure 3.31 shows tht
transconductor stages. A
Figure 3.29 Implementation of a bilinear transfer function using an integrator. currents. Summing the cu
~d-Signal Circuit Design Chapter 3 Analog Filters 97
Active-RC Implementation
The active-RC implementation of the bilinear transfer function is seen in Fig. 3.30.
Again, as mentioned earlier, the resulting active-RC transfer function suffers from poor
repeatability from one process run to the next. The RC time constants must be tuned,
on-chip, with fuses (or switches) and adding/removing resistors or capacitors. Note how
rs the summation is implemented by changing the input/output voltages to currents. The
o· currents are summed at the inputs of the op-amp (which remain, ideally, at the
-10'
common-mode voltage, VCM)' This is important to note in both the active-RC and
-140'
switched-capacitor implementations.
-210'
-280' We won't discuss the implementation of the MOSFET -C-based bilinear transfer
-350' function. It should be obvious that replacing the resistors in Fig. 3.30 with MOSFETs or
-420' linearized MOSFETs (see Figs. 3_12-3.14) provides a MOSFET-C implementation.
-490'
-560'
RF
CI
3fs12
RI
G j =_1_
RICF
(3.53) I
CI G2 =:: RI
RF
G3 RICI
Transconductor-C Implementation
Figure 3.31 shows the implementation of the bilinear transfer function using
transconductor stages. Again, as with the active-RC filter, signals are summed using
using an integrator. currents. Summing the currents at the output nodes results in
98 CMOS Mixed-Signal Circuit Design Chapter 3 Analog Filter:
where we know Vout+ = -Vout- and Vin+ = -Vin-. It will be helpful to write
~ (3.57)
I/s2C) IIsC) I/sC)
Using this expression, we can write Eq. (3.56) as
(VO!lf+ Vout-)' (s(C) + C 2)+ gm2) = (Vin+ Vin-)' (gml +sCJ) (3.58)
or finally
Vaut+ - V out-
Vin+ - Vin
(3.59)
It's important to note that when looking at this equation, the location of the pole and zero
can be adjusted by changing each transconductor's gm independently. The ability to adjust
one variable in a filter's transfer function and only change the position of a single pole or
zero is called orthogonal tuning.
Figure 3.3
and so
and G3 (3.61)
C11 or
Note how, in this discrete-time filter, the passband gain is C I3 IC F when the filter is
designed for a highpass response (and the filter no longer behaves like a discrete-time
filter). The gain at DC in all situations is Cn/Cn.
ed-Signal Circuit Design Chapter 3 Analog Filters 99
V ou /+ =0 (3.56)
lis2C 2 ~l
o write
(3.57)
(3.58)
~:::-i'--..o-- VOUI
Mr '
/'""~---Voul+
(3.59) C
(3.63)
(3.61 )
ell or
'13/CF when the filter is
aves like a discrete-time rt/o,
PhP2 == - ±). 2rt/o .11 -
1(1\12 (3.64)
Q V ,2Q/
100 CMOS Mixed-Signal Circuit Design Chapter 3 Analog Filter:
In order to move toward the goal of implementing a biquad, consider the block Active-RC Irnplernentati
diagram in Fig. 3.33. This block diagram is essentially the cascade of two first-order
Figure 3.34 shows the
blocks (as seen in Fig. 3.29) except that instead of feeding the output of the second stage
associated design equati
back to its input, we feed it back to the input of the first stage. We can determine the
If, for example, a low
transfer function of this filter by writing
simplified.
Voul' (3.65)
or
(3.67)
V out
(
i Vin •
\
(1 +SG3)G 1
G G -
S+ I 2
Vout
S+
GIGS ) (1
G G
1 2
.
G )G4
+s 6 S (3.68) Vin- --r---vVV'-
I Rn
or , 11---
II GIG4G5(l+SG6)) _ '1/(I+SG3)G,(l+SG6)G4) Cll
VOUII +
"SZ+sG1Gz
- Vm
\
2
s +SGIG 2
(3.69)
(21t/d G, G 4 G S (3.75)
Figure 3.34 fm!
Vou!1
I .
--lou[-S
GIGS (3.66)
Cn Cl2
(3.67)
Voutl Rl2
Vout+
+ +
(3.68) ('------.----1 + ('------.----1 + Vout-
Rn V ou tl+
Cl2 Cn
(3.69)
~4G6)+GIG4
(3.70)
GI G4 GS
(3.71 )
C Cl2 1
(3.72) a 2 =n- - - ao=--~--
R n C Fl R l2 C n
CFICn
(3.73)
(00
-=--=---+
2n[0 1 Cl2
(3.74)
Q Q RFICFl CFlRnCn
(3.75)
Figure 3.34 Implementation ofthe active-RC biquadratic transfer function filter.
I--+-_V_o:..::ut(f) Figure 3.35 shows the frequency response, pole-zero locations in the s-plane, and
transfer function for a second-order lowpass circuit made using an inductor (L), capacitor
(C), and resistor (R). This LRC circuit has the same frequency response shape as a
lowpass biquad filter. However, the DC gain of the LRC circuit must be unity while the
DC gain of the biquad filter can be set to ao/(2n[0)2 . Note that if the pole quality factor,
Q, is greater than 1I.j2 the response will show peaking. Setting Q to 0.707 results in the
Butterworth or maximally flat response.
1 using two integrators.
102 CMOS Mixed-Signal Circuit Design Chapter 3 Analog Filters
From Fig. 3.35 we have two equations we need to solve Using the basic top'
Cn = Cn 0 and then
/0 = 1.59 }';{Hz -+ LC = lOx 10- 15 and Q
2rc!iC
We can set C 100 pF, then L = 100 IlH, and R 1.4l4k (definitely not practical
values if the circuit is going to be purely integrated). The response of the resulting which we shall use tc
LRC circuit is shown in Fig. 3.36. Note how the cutoff frequency is set by the filter is given by
IdS 40'
-21 dB
-20'
-40'
- 40 dB/decade here. The simulation n
t
it
-2adB
-35dB
O'
-80'
Notice that at DC,
integrator, vourl+ and VO"l]. ,
":, :
~
-42dS
-49dB
-100'
-120'
section. As the frequency
inductor and capacitor, while the Q of the circuit is set by all three elements
(variations in the resistance having the largest effect on the circuit's Q). Higher Q
10 indicates the poles are moving toward the imaginary axis (keeping in mind that a
2rr.JLC
system with right-half plane poles is unstable [oscillates]) and more peaking, Fig.
11[ 3.37. •
Q
R~C
Q M-+-'-++i-i"+---i--i--i-~ii--- Q 5
max
+--:---:--:+:+~~~~ ~~~=i=i~#T--Q = 1
)1 + 4~2
f (Hz) Q 0.707
Q = 0.5
I
x=/o' i1
V
Figure 3.37 The effect ofQ on the frequency response of a second-order lowpass filter.
r.
Example 3.9
Simulate the design of an active-RC filter that has frequency characteristics
ncy (10) of 1.59 MHz. similar to Fig. 3.36.
Using the basic topology of Fig. 3.34, we see that for a lowpass filter,
C lI = C n 0 and therefore G3 = G6 0. Further
L
I
10 = 1.59 MHz = 2rr. C R C' R
'Fl n Fl Fl
:lefinitely not practical
,ponse of the resulting which we shall use to set Rn = RFl = lOk and C Fl CFl lO pF. The Q of the
'equency is set by the filter is given by
dB downat/o
Knowing a 2 = a l = 0, the gain at DC is ao/(2rr/o)2 or RFl/RIl (lIG s), which is I
- 40 dB!decade here. The simulation results are shown in Fig. 3.38 .•
Notice that at DC, when used in the lowpass configuration, the outputs of the first
integrator, v outl + and vouil_ , must be equal. If not, the difference is integrated by the second
section. As the frequency increases, so does the difference in these voltages.
Figure 3.39 shows the second-order bandpass response. Again, as with the
second-order lowpass response, the center frequency (resonant frequency) is set by the
values of the inductor and capacitor. The Q of the filter indicates how narrow the
bandpass response is; higher Q indicates a narrower response. Note how the response
eventually rolls off at -20 dB/decade. At low frequencies the capacitor can be thought of
leseribed in Ex. 3.8. as an open (resulting in a first-order RL circuit response), while at high frequencies the
inductor can be thought of as an open (resulting in a first-order RC circuit response).
104 CMOS Mixed-Signal Circuit Design Chapter 3 Analog Filters
n
OdB 20' OdB
-7dB o· -fidB ---,
-ti3dB
-7UdB
-77dB
-160'
-180'
-200'
:::::14
-G6dB '
1KHz 10KHz 100KHz It.tHz 10t.tHz 100MHz 1KHz
Figure 3.38 Magnitude and phase responses for the active-RC filter of Ex. 3.9. Figure 3.40 Bal
Example 3.11
I Repeat Ex. 3.10 if the
Vow sliC
-=
Vin S2 +s...L + Figure 3.41 shows th,
RC
20k with the inductor
OdS-,
: -
fo/Q -9dB-+---<
~~~~u1>
Im,iro
One zero at infinity
and one at the origin
--'---'m----';loo- Re,cr
·12dB ....;.
.!£Q. -81dB ....:.
s plane
X2Q 20 dB/decade -90dS - ,~.
s = cr +jro
/0 ·99dB~
1KHz
Example 3.12
Example 3.10 Use an active-RC filt{
Repeat Ex. 3.8 for the bandpass LRC circuit. Let's begin by writing
Again, we can set C = 100 pF and L = 100 /IF. Solving for Q using the equation in Vow
Fig. 3.39 results in Vin
C filter of Ex. 3.9. Figure 3.40 Bandpass response of a second-order circuit with a Q of 0.707.
Example 3.11
Repeat Ex. 3.10 if the Q is increased to 20.
fo
2rcJLC Figure 3.41 shows the simulation results. To attain a Q of20, we use a resistor of
20k with the inductor and capacitor values remaining unchanged.•
Q=Rff
J
~
~---c»> f (Hz)
- 20 dB/decade
Example 3.12
Use an active-RC filter to implement a filter with the response shown in Fig. 3.41.
Let's begin by writing the filter's transfer function
, using the equation in Vout
Vin
Looking at this equation, Eq. (3.62), and Fig. 3.34 we see that a2 = ao = 0 and so
Cn. 0 and RlI 00, Further then
106 CMOS Mixed-Signal Circuit Design
C ll
al = - - ' -
RnCFlCn
'-- _1-J
Jio -
271: CFIRnCnRn
1 == 1.59 MHz
Technically, the filter is
filter.
If, again, we set CFl = Cn = 10 pF and Rn = Rn = 10k, we get anfo of 1.59 MHz.
Further then, with a Q of 20, we can set RFJ to 200k. Finally, setting the passband
gain to unity results in
2n{o Cn
al = - - = ~ Cll 0.5pF
Q R 12 C F1 C n
While these values do result in a biquad with the shape seen in Fig. 3.41, the
values are not practical. Redoing the calculations while trying to minimize the
component spread gives another possible solution: RI2 = lOOk, CFl 20p, RFI
lOOk, Cn = 5p, ClI = 5p, and Rn = lk. The simulation results are seen in
3.42.•
120'
100'
80'
60'
Figure 3."
40'
20' The biquad in F
0' understand the topology,
,20'
more than two bilinear
·40'
switched capacitance, C
·60'
feedback resistance, R n ,
-liD'
. filter implementation, ge
-100'
1KHz 10KHz 100KHz 1MHz 10MHz 100MHz
Figure 3.42 Outputs ofthe biquad of Ex. 35.12 using active-RC elements. We have a major concel
Fig. 3.34 or 3.43 with a
Switched-Capacitor Implementation within a factor of 4 of e:
The switched-capacitor implementation of the biquad circuit is shown in Fig. 3.43. Note magnitude different (100
G6 0,
how this circuit is a simple translation of the active-RC circuit of Fig. 3.34. Again, if the
filter designed using this section has a lowpass or bandpass response, it can be simplified.
For example, from Figs. 3.35 and 3.39 (the implementation of lowpass and bandpass and 271:10
filters), we see that a2 is zero. This indicates that G6 can be set to zero (removing Cn in
Figs. 3.34 or 3.43). The resulting second-order filter response can be written as This equation shows tha
Vout =:: als+aO SGIG3G4+GIG4 of RFI results in a smalle
(3.76) input of the first op-amp
Vin s2+e~o)s+(271:1o)
-Signal Circuit Design Chapter 3 Analog Filters 107
4>1
llated by replacing
--:I Cn !
v~ '~~~-4-
Vout-
Cn
20' ·is Gs Cf12 ·is
00'
80'
60'
Figure 3.43 Implementing a biquad filter using switched capacitors.
40'
23' The biquad in Fig. 3.43 can look confusing until we start to dissect it. If we
3' understand the topology of Fig. 3.32, we see that the switched-capacitor biquad is nothing
23'
more than two bilinear filters connected in cascade. The only difference is that the
43'
switched capacitance, Cm' is fed back to the input of the first op-amp to simulate the
liO'
feedback resistance, Rn , in Fig. 3.34. This circuit, for the general lowpass or bandpass
80'
filter implementation, gets much simpler when the unused components are removed.
00'
HighQ
RC elements. We have a major concern alluded to in Ex. 3.12 when using either of the topologies of
Fig. 3.34 or 3.43 with a large Q. As we saw in this example the capacitor values were
within a factor of 4 of each other (20p and 5p) but the resistors used were two orders of
)wn in Fig. 3.43. Note magnitude different (lOOk and 1k). This large difference can be traced to, again assuming
Fig. 3.34. Again, if the G6 0,
;e, it can be simplified.
lowpass and bandpass (3.77)
zero (removing Cn in
e written as This equation shows that RFI has the largest direct dependence on Q. Using a large value
of RFI results in a smaller feedback signal (a smaller amount of current is fed back to the
(3.76) input of the first op-amp). In other words G2 in Fig. 3.33 is small.
108 CMOS Mixed-Signal Circuit Design Chapter 3 Analog Filte
In order to minimize the amount of signal, V QutP fed back and summed with the
input signal, while at the same time forcing the components to have similar values,
consider the modified, from Fig. 3.33, biquad block diagram shown in Fig. 3.44. All we
and
have done here is added a separate signal path in parallel with the G2 path. Instead of
subtracting, though, we are now adding the signal to the input summing block. Equation
(3.70) can be rewritten, assuming G6 is zero (a bandpass or lowpass response), as
Voul s(G 1 G3G4) + Gl G 4 In an attempt to mi
(3.78)
Vin s2+SGl(G2-G2Q)+GIG4GS 4p, CF2 to 20p, CF1
simulation results a:
or, equating the coefficient of s in the denominator of this equation with the coefficient of
s in the denominator ofEq. (3.62), results in
(3.79)
The implementation of the "high-Q" biquad is seen in Fig. 3.45 (with G6 included). The
CII
additional gain from the figure is
Rn I
G 2Q=--Q-
RFi Q-1
(3.80) i -Rll
v ..
In-t-L-;\N'
Rewriting Eq. (3.79) results in
2n/o G1G2
(3.81 )
Q Q
Let's use this result in the following example.
In VourV> L._ _ _ __
Example 3.14
Repeat Ex. 3.13 usin
is clocked at 100 Ml
Figure 3.44 Implementation of a "high-Q" biquadratic transfer function.
To implement the fil
capacitors. Howevel
Example 3.13 ratios of capacitors. '
Repeat Ex. 3.12 using the high-Q circuit of Fig. 3.45. values in Fig. 3.46. I
The passband gain is I so we know that all resistors by 10. T
OJ
resistors can be calct
~ G1G2
Gl= Q = GjG3G4 =G 1(G2 G 2Q)==Q
(3.78) In an attempt to minimize component spread let's set RFI to 5k, Rfl. to 20k, CII to
Gs 4p, to 20p, CF ] = 20p, Rn = l.25k, and finally RFIQ = 5.25k (roughly). The
on with the coefficient of simulation results and schematic are shown in Fig. 3A6.•
(3.79) RFI
RF2
; (with G6 included). The
CIl
(3.80)
Vin+ j :£ ~c--_-++-
Vout+
V(}ut
(3.81 )
I
~=G2'
Voutif)
Q Q
RFI _
Q 1
Figure 3.45 Implementation of the "high-Q" active-RC biquadratic transfer function filter.
The bold lines indicate the added components.
Example 3.14
Repeat Ex. 3.13 using a switched-capacitor implementation. Assume that the filter
is clocked at 100 MHz.
msfer function.
To implement the filter, we need to replace the resistors in Fig. 3.46 with switched
capacitors. However, we notice in the gain equations that the resistors are all
ratios of capacitors. This means we can reduce the size of the filter by scaling the
values in Fig. 3.46. In order to do this let's divide all capacitors by 10 and multiply
all resistors by 10. Therefore, we can write C'l OAp, CFI 2p, and Cn = 2p. The
resistors can be calculated using
G 1G2
Q RFI = 50k-). C m =. O.2p
C/2l'1s
~ 1.25k
and Cn)'
5.25k
20p 4>1 4>
(O.16p) ,
O.2p O.I9p ,
~--+H-- Voul+
~/..L I /
Vin- --1 -_-i
r'
(O.8p): T T
vin+~O~P! ••.
5k
I.25k
VeM
-1
ffiI--::::
~
',1'
./""
I!,·,.
- Vin- O.4p'
9da-..,.....,..,.....,.,.."---,-,-.~~~~L--,-,-~.,--:-:-,..,..,.,,CT 120' (O.8p): !
1-
100'
CQT"
80'
SO·
O.2p O.~
40'
(O.16p)
20'
o'
'20' Figure 3.47 Switc
.+.l-j.:.,,:i1- -.40'
-no'
,..;.; ~.~~:~I- ·80' Again, as in Ex.
·100' transient analysis, we
we get the correct Ot
signal to the filter at
However, as seen i1
Figure 3.46 Bandpass filter discussed in Ex. 3.13. simulations it's easy t
outputs oscillate at fo (
as the Q of the bandpl
1 plane. If, for some re<
R12 -C-- == 200k -+ Cm == 0.05p (!)
!l21s become an oscillator (
the filter we approxin
Rn == C I" == 12.5k -+ Cm == O.Sp
122 'Js «t:
-s
cos 2,g
j,
'" I and sir
Looking at the value of Cm , we see it may be too smaIL Let's change its scale
factor from 10 to 4. This means
We have to scale Cn as well (so that G4 remains constant). Now Cn = 5 pF. frequency J,. As we di
in excess of the sampJ
the combination of th.
[xed-Signal Circuit Design Chapter 3 Analog Filters III
Figure 3.47 shows the implementation of the filter. Note how easy it was to
implement the high-Q circuitry. All we did was add two capacitors to the circuit.
Also note how the circuit is simplified after removing the unused components (Rn
and Cl2 ).
~l
2p
(O.16p)
Again, as in Ex. 3.6, because this circuit can only be simulated using a
transient analysis, we will input a sinewave at a known frequency and verify that
we get the correct output. Looking at Fig. 3.46, we see that if we apply a I V
signal to the filter at 1.59 MHz we should get a 1 V signal out at 1.59 MHz.
However, as seen in Fig. 3.48, the filter is unstable and oscillates. Using
3.13. simulations it's easy to show that even if we ground the inputs of the filter, the
outputs oscillate atfo(1.59 MHz). To understand why, remember in Fig. 3.39 that
as the Q of the bandpass filter is increased, the poles move closer to the right-half
plane. If, for some reason, the poles move into the right-half plane, the filter will
become an oscillator (unstable). It's important to remember that when we designed
the filter we approximated our discrete-time variable z as 1 + = It (when f
cos "" 1and sin ~; "" ~J). We could be more exact and write
Let's change its scale
2nf . 2nf
z "" cos-+ jsm- (3.82)
j, Is
which clearly will not follow 1 + ~:for frequencies f approaching the sampling
..row Cn = 5 pF. frequency J;. As we discussed in Ch. 2, sampled signals will have spectral content
in excess of the sampling frequency. Practically, the spectral content is limited by
the combination of the switches "on" resistance and the capacitors in the circuit.
112 CMOS Mixed-Signal Circuit Design Chapter 3 Analog Filtel
gain-bandwidth produc
ideal integrator gain of
Output
Input-_..:.:c:.:~
As f gets larger, the cosine term will decrease from one, causing the real portion to
get smaller. A decrease in the real component, as seen in the complex plane in This subtraction results
Fig. 3.39, causes the pole to move closer to the right-half plane (causing the Q to of the circuit; Fig. 3.50
increase). (3.85) to estimate the sl
Practically, the maximum Q we can design for (but not attain) is in the r
neighborhood of 5. If we redesign the biquad of Fig. 3.47 to have a Q of 5, we see
that all we need to change is CII (from O.4p to O.8p) and CI21Q (from O.19p to
O.16p). The simulation results are seen in Fig. 3.49. In this figure, we adjusted the
input frequency until we reached a 3 dB point (the filter's center frequency was
1.59 MHz, as expected). This occurred at 1.52 MHz and 1.66 MHz. The Q of the
circuit is not 5, but is, from Fig. 3.39, 1.59/(1.66 - 1.52) or 11.36.•
So, for the filter Q to re
0.8V
O.2V
Input
o.ov
Output
-,uv
·1.llV
Figure 3.49 Output ofthe filter in Fig. 3.47 after lowering the Q to maintain stability.
gain-bandwidth product U;m) of the op-amps. We can model these effects by replacing the
ideal integrator gain of lis with
l~ 1 (3.83)
s s ( 1+_'_1\
2tt/un I
Using this result, we can rewrite the pole locations of Eq. (3.62) (see Eqs. [3.63] and
[3.64]) as
(3.84)
« I (3.88)
fun
Let's use this result in the following example.
s 2.01ls 'l
-------~--~~*_---------~~
splane
Re
a maintain stability.
~ ~ X I ~ Ideal distance is ~
:>ice for high-Q filter Figure 3.50 Showing Q peaking resulting from the op-amp finite gain bandwidth product.
the effects of the finite
114 CMOS Mixed-Signal Circuit Design Chapter 3 Analog Filters
Example 3.15
Resimulate the filter in Ex. 3.13 using op-amps that have anfu" of 100 MHz.
The center frequency,lo, of this filter is 1.59 MHz and the Q is 20. Using Eqs.
(3.87) and (3.88), we can estimate the increase in Q due to op-amp finite
gain-bandwidth product as
Q2fo
0.636 ---? Qshift 55
fun
Practically, this is too high of a Q (the poles are too close to the right-half plane),
and the filter will be unstable (noise in the circuit, or simulation noise in the
simulation, will push the poles into the right-half plane). Figure 3.51 shows the
simulation results (see also Ex. 3.4). The inputs to the filter are grounded. The
unstable oscillation frequency is close to the ideal,lo, but is shifted by a small
amount. •
Figure 3.5
Figure 3.51 Showing how the filter of Ex. 3.13 becomes unstable due to
finite op-amp bandwidth.
Transconductor-C Implementation
Let's redraw the bilinear filter in Fig. 3.31 as seen in Fig. 3.52. We redraw it like this to
show how the feedback gain, G2, is implemented. In the block diagram of the biquad filter
ShO\\<l1 in Fig. 3.43, we used a similar scheme to implement the feedback gain, Gs- Figure
3.53 shows the implementation of a biquad filter using transconductors where we have Figure 3.53 •
drawn it so that the trans conductors appear to be connected in series. This topology can
be redrawn so that it looks similar to Fig. 3.52 (showing a direct correspondence between
it and Fig. 3.43). Note that we could have drawn the schematic without the crossing wires
if we switched the output polarity of two of the transconductors (that is, put the minus S. Franco, Design
output on the top of the output instead of the plus output). Third Edition, Mc(
R. Schaumann am
University Press, 2
P. R. Gray, D. A.
Circuits, Wiley-IEl
ed-Signal Circuit Design Chapter 3 Analog Filters 115
~n of 100 MHz.
Q is 20. Using Eqs.
ue to op-amp finite Vin+ 1-+-----_-----_--- VOIit+
G3 .f.L
gml
Ips
mstable due to
r4] P. R. Gray, B. A. Wooley, and R. W. Brodersen (eds.), Analog MOS Integrated Derive the transf
Circuits Il, Wiley-IEEE, 1989. ISBN 0-87942-246-7
[5] P. A. Lynn, An Introduction to the Analysis and Processing of Signals,
Hemisphere Publishing Corporation, 1989. ISBN 0-89116-981-4 Vinl+ ----"'\
, Analog lvIOS Integrated 3.9 Derive the transfer function for the following first-order transconductor filter.
Processing of Signals,
6-981-4
Continuous-Time Filters:
· ISBN 0-7803-0425-X f - - + - - - - V OUl+
c
mcies, Kluwer Academic I---+_---Voul-
3.10 Show the derivation details that result in Eqs. (3.44) and (3.46).
3.11 Show the details of how the gains, G, are derived in Fig. 3.30.
3.12 Is it possible to tune the gain, Q, and cutoff frequency of the lowpass biquad
independently? If so, how? Give examples using the simulation netlist used to
generate Fig. 3.38.
· and a resistor. What happens to the poles in the biquadratic equation, Eq. (3.62), if the Q is less
than 0.5? (Hint: The filter behaves like the cascade of two first-order filters.) Is
I outputs are referenced to thef.nax equation in Fig. 3.35 valid?
uld also be at VCM') Compare the size of the elements used in Exs. 3.8 and 3.9. Is there a benefit to
· using a CAl with a 3 dB using an active element for monolithic implementation?
e your design to verify it Show, using the simulations from Ex. 3.14, that increasing the switch resistance,
and thus the spectral content present in a switched capacitor circuit, can help to
the actual pole locations stabilize high-Q switched-capacitor bandpass filters.
~scribed in Ex. 3.4. Redesign and simulate the operation of the filter discussed in Ex. 3.14, with a Q
, op-amp model shown in of 5, while trying to minimize the difference between Cll and Cn . Suggest a
possible modification to the filter topology (similar to how we add G ZQ in Fig.
3.45) to reduce this component spread.
it data converter. Further
.C topology. If VDD = 1.0 Show how to derive the transfer function of the transconductor-C biquad filter
citor that should be used, seen in Fig. 3.53. Can this filter be orthogonally tuned? If so, how?
, thermal noise. Is it wise, Repeat Ex. 35.9 using the transconductor-based biquad.
at its SNR is equal to the
How would a "high-Q" biquad be implemented using trans conductors? Repeat Ex.
35.12 using the transconductor-based biquad.
> a linear output swing of
Show, using biquad sections, how the lowpass ladder filter seen in Fig. 3.56 can
be implemented.
U 6 ifthe trans conductors
Ithe one seen in Fig. 3.6,
118 CMOS Mixed-Signal Circuit Design
Figure 3.56 Implementing a ladder filter using biquads, see problem 3.20.
Digital F
Chapter
-_.----VOUI
4
lee problem 3.20.
Digital Filters
In this chapter we study the implementation of digital filters specifically useful for
mixed-signal circuit design. As in the last chapter, our filters will be based on the
integrator and, to a lesser extent, the differentiator. Prior to studying the material in this
chapter the reader may want to review Sec. 1.2.
If, for example, VREF+ 1.25 V and V REF- 0.25 V and N = 3, then our LSB, the vertical
distance between adjacent points in Fig. 4.1, is 0.125 V. Note that in our discussion of an
ideal DAC we are assuming that the output of the DAC ranges from VREF- up to
V REF+ - 1 LSB. We could just as easily have assumed that the output ranged from
V REF- + I LSB up to V REF+. The important thing to notice is that the DAC output range is
1 LSB smaller than the difference between the positive and negative reference voltages.
For the DAC developed in this chapter, we will assume VREF+ = VDD = 1.0 V and
VREF- 0 V. Selection of the power supply rails, which are noise free in a SPICE
120 CMOS Mixed-Signal Circuit Design
VREF+ VREF
We can implement this
VRE~ 8/8 1 source). For a 3-bit, idee
like
1 ~~ 7/8
VREF+ -
• >Nonlinear depend€
6/8 • Bout Vout 0 V=«v(v
4/8
Example 4.1
3/8
Write the nonlinear dt
2/8
• The statement follow,
VRE~ 0
118
.'
"
I I
Bout Vout 0 V=«v(v
+(v(B11l)*2048)+v(
+v(B7L)* 128+v(B6l
000 001 010 011 100 101 IJO III +v(B2L)*4+v(B1 Lt:!
VDD
Figure 4.1 An idea13-bit DAC
simulation, allow the maximum output range for the DAC (assuming the reference
voltages are indeed the maximum and minimum voltages in the system, i.e., no charge
pumps or external, larger, power supply voltages). Ifwe need more resolution when using
our ideal DAC, we will simply increase the number of bits, N, used and thus decrease the
value of the DAC's LSB.
SPICE Modeling the Ideal DAC
We can write the output of the ideal DAC in terms of the reference voltages and digital Figure 4.2 '
input codes b" (which are logic "0" or "I "), and assuming that an input code of all zeroes
results in an output voltage of VREF-, as 4.1.2 The Ideal ADC
bl"l-l bN-2 b bo)
VOUT (VREF+ - V REF-)· ( Y + 22 + ... + 21"1-1 + 2N + V REF-
l
(4.2) The characteristics of our
is shifted to the left. Ifwe
or intersection of the analog
DAC transfer curve of
ixed-Signal Circuit Design Chapter 4 Digital Filters 121
TJ"
rOUT = (VREF+ V ) 21
REF-' (b N-I 2 N- 1 + b N-2 2 N- 2 + ... + b l ' ~')1 + b)
0 + V REF- (4.3)
N '
Write the nonlinear dependent SPICE source statement for an ideal 12-bit DAC.
Digital +v(B7L)*12B+v(B6L)*64+v(85L)*32+v(B4L)*16+v(B3L)*B+
remembering that a "+" in the first column of a line indicates that the text on the
remainder of the line behaves as if it were typed at the end of the previous line. It
doesn't indicate addition .•
The next thing we need to concern ourselves with is the digital logic levels. We
want to use our ideal DAC with real circuits where the logic voltage levels may not be
!'OUT well defined. We need to determine and use a switehing-point voltage based on the
power-supply voltage VDD. We will assume the input logic code is a valid logic "1 II if its
amplitude is greater than VDDI2 and a logic "0" if its amplitude is less than VDDI2. The
switch implementation used to generate the logic signals (l V and 0 V) used in our
dependent source from real signals is shown in Fig. 4.2.
~ lOOMEG
(
lore resolution when using Close4. \V~en BX; < triQ
renee voltages and digital Figure 4.2 Generating logic levels using voltage-controlled switches.
input code of all zeroes
III
converting a (normalized) input voltage of 0.1 V will result in an output code of 000 If the input signal j
which is the same output code resulting from converting 0 V. Unlike the ideal DAC, the input signal is mult
ideal ADC quantizes its input with the practical result of adding noise to the input signal.
This output is usee
This noise is called quantization noise.
repeated. This cont
The reference vol1
Digital calculating the midpoint b
output code, b2 b 1 bo I LSB
~
the VeM is referenced to 0 '
111 VCM
110
We also want to level-shit
101 want to shift the transfer (
this we use the following
100
output voltage of the ideal
011 • Level shift by VREF
BPIP PIPIN 0 V=V(C
010
001
000 1/2
~O 1/8 2/8 3/8 4/8 5/8 6/8 7/8~8/8
Analog input voltage ""
V REF- We are level-shifting the:
VREF+ -1 LSB
model as flexible as possi'
0.5 V and VREF = 0.25
simplified.
4.1.3 Number Represe
Suppose we have an ideal
and a V REF- of ground.
minimum output of the [
LSB or 0.9961 V. An inpl
1----- bo the common-mode volta)
corresponding digital cod
offset format. The output
format. A more useful fc
1
Figure 4.3 An ideal 3-bit ADC. complement format, Fig. L
3. If the input signal is greater than the reference voltage, the output bit is set to a
Figure 4
high, and the reference signal is subtracted from the input. The difference is
multiplied by two and passed to the output of the stage.
ixed-Signal Circuit Design Chapter 4 Digital Filters 123
in an output code of 000 4. If the input signal is less than the reference voltage, the output bit is set low. The
Unlike the ideal DAC, the input signal is multiplied by two and passed to the output of the stage.
Ig noise to the input signal.
5. This output is used as the input to the next stage and steps 2, 3, and 4 above are
repeated. This continues for N stages (where N is the number of bits in the ADC).
The reference voltage, or common mode voltage VCM ' can be determined by
calculating the midpoint between VREF+ and V REF- followed by subtracting so that
the VCM is referenced to 0 V. This can be written as
VCM VREF+ + V REF V VREF+ + V REF V
2 -t CA10 = 2 - REF- = (4.4)
Example 4.2
0111 1111 (0.9961 V)
Show how to conv
(+127)
complement number
time
represent 0 as -I an<
1000 0000 (0 V) 7 VCM 128· VLSB
(-128)
-----I
;l Ideal
3-bit ADC
100 (4)
101 (5)
110 (6)
0000 0000 (O)
0000 0001 (1 )
000000 10 (2)
B (3-bits) 4-bits
~
felk I 111 (7) 0000 DOll (3) Extend the sign-bit
,
ADC Output
is binary offset with sign extension + 110(
1 100(
Figure 4.6 Showing how to change the output of an ADC to two's complement and Overtl(
how to extend the sign bit.
Figure 4.~
xed-Signal Circuit Design Chapter 4 Digital Filters 125
Example 4.2
Show how to convert a I-bit binary offset number into 2-bit and 4-bit two's
complement numbers.
J~m_
_ VCM"" 0.5 ..... time)
VC.H-VLSB
A I-bit binary offset number has values of 0 or I (common for the output of a
noise-shaping ADC). To change this into a 2-bit two's complement number we'll
represent 0 as -I and I as +I or
dU
1~01(+1)
.ent format.
For 4-bit representations we can write 0 ~ 1111 (-I) and I ~ 0001 1) . We'll
plement and binary offset use these results frequently when discussing noise-shaping (delta-sigma) data
converters.•
s represented using III I
~ MSB corresponds to the A comparator is an example of an ADC (or quantizer) that will generate the I-bit
e common-mode voltage, number used in this example. Reviewing Eq. (4.1) we see that this equation doesn't work
(a code or voltage below for determining the LSB of a I-bit ADC. We can modify Eq. (4.1) for this situation
1 LSB VREF+ - VREF "" VLSB for N 1 (4.6)
Adding Numbers and Overflow
Nord size to increase the
Figure 4.7 shows how two's complement numbers are added. Note that the first thing we
L6 shows how we would
do, when adding two digital signals, is to extend our sign bit to avoid harmful overflow
-bit ADC to 8-bits while
(see the allowable overflow examples in the figure). This increases the word size and
v1.SB of the ADC's output
allows the final output word size to always be large enough to accommodate the sum of
C's input signal were VeM'
the inputs. Note that this is important. In filters employing feedback, like the digital
1000. The common-mode
integrator seen in Fig. 1.26, we may increase the word size even more to avoid, or delay,
zero, level. Since we'll be
harmful overflow. For an integrator a DC input will always, after some time, result in
k the reader should spend
overflow (unless the integrator's input is always zeroes). Also note that harmful overflow
.5, and 4.6.
is easy to detect since it will only occur when the two inputs are the same polarity
(positive or negative). If the MSB of the two input words is a 0 (I) then the output word's
Qillp.u.t After extensjon MSB must be 0 (l) . If it's a I (0) then we know overflow occurred. In simpler terms, if
1(0) 11111100(-4) the two inputs are positive (negative) then the output must be positive (negative).
(1) 1111 1101 (-3)
1(2) 1111 1110 (-2) Adding two's complement numbers A Out (A+B)
(3) 11111111 (-1) 0011 (+3) 0011 0110 (+6)
A (3-bits) 1-----"...... ( 4-bits)
1(4) 00000000 (0) ---'---3I>---l 1100(-4) 1100 (-4) 1000 (-8) OF!
(5) 00000001 (1) 1111(-1) 0001 (+ 1) 0000 (0) OF!
B (3-bits) ,
1(6) 00000010 (2) -'--i--)-----' 0011 (+3) Illl (-1) 0110 (+2)
(7) 00000011 (3) 0001 (1) 1100 (-4) 1101 (-3)
Extend the sign-bit
A (3-bits) Figu
--'--:-~---I
B (3-bits)
-:...,~.-~~~
Extend the sign-bit
, Invert the word we are subtracting
-I
subtracting 1
~·(I+z-1
H(z) + ... (4.9)
l-z- 1
t (A-B) The transfer function for the counter is then
1 (+7)
K
)0 (0) OF! H(z) = I-z- (4.10)
0(-2) OF! 1 -z -I
)0 (+4)
We'll see this equation frequently so let's spend some time on it. Note that the counter
II (-1)
employs decimation, see Sec. 2.1.2, since the input word rate is K times faster than the
output word rate. More on this in a moment (since aliasing will be a concern). The
are subtracted.
magnitude of the frequency response ofEq. (4.10) is given by
. (K 'X)i\,
sm 1t
IH(j) I (4.11 )
when doing mixed-signal
n by discussing one of the
Sine( 1t¥)
. with a I-bit input. If we IH(j)I=K. (4.12)
ld out and reset every KT, Sine ( 1tf;)
e Fig. 6.3)) will result in a
Figure 4.10 shows the frequency response of the counter (see, also, Fig. 2.29). Note that
:onds. A constant input of for large K and small frequencies this equation can be approximated using
g 1s and Os (the input is a
.n an output of K12 (0 for (4.13)
i to its inputs using
Figures 2.30 and 2.31 in Ch. 2 relate the amount of attenuation and droop we can expect
(4.7) from a Sinc response filter for various values of K. Again, these equations can be used to
characterize all of the Sinc filters in this section.
ing them together, and the
ike any lowpass filter, as
Note, again, if we input a DC value of constant 1s then the counter's output is K (see Fig.
4.10). If our input is 101010 ... or 11001100... or 111000111000 etc. then the counter
-2+ ..• + Z 1-K (4.8) output is K/2. However, looking at the filter response seen in 4.10 and knowing an
input of 10 10 lOis a squarewave with a frequency of f)2, we would expect the counter's
128 CMOS Mixed-Signal Circuit Design Chapter 4 Digital Filter
Using a counter
/H(f)/
A Sinc(Kn . f,) the significant amount
/H(f)/ didn't discuss problem
K
sinin· L) counter's uses, howevel
" Is
See Fig. 2.30 as some sensing appli(
shrunk down to limit th
(review Fig. 4.10) .
•••
K~ K K K
The counter's input wa
2K Counter's output word rate can be used, Fig. 4.12,
K clock cycles (note
Figure 4.10 Frequency response of a Sine-shaped digital filter. accumulator, seen in Fi
hold register and the su
output to be zero (and it is if we think in terms of two's complement numbers). However, filtering behavior of th'
let's look at the operation in terms of binary offset numbers.
maximum input word's
Let's assume K is 8 and the counter is clocked at a frequency of 100 MHz. The summation of K, N-bit,
I-bit word rate coming into the counter is at 100 MHz while the counter is read out every wide (rounding up
12.5 MHz (decimation by 8). Figure 4.11a shows the frequency response of the counter accumulate-and-dump f
and the first-harmonic of the input of 1010101... or a squarewave at 50 MHz. We only
concern ourselves with the first harmonic of the squarewave to keep the figures from
getting too cluttered. A sinewave at 50 MHz, see Eqs. (1.85), having a peak-to-peak
amplitude of I, or a peak amplitude of 0.5, is represented at ± 50 MHz with dirac-delta
functions having amplitudes of 0.25. Figure 4.11 b shows the input signal spectrum after Input, N-bits
resampling at the output rate of 12.5 MHz (note the aliasing of the sampled signal at DC).
In (c) we see that after including the counter's Sine-shape response all of the tones
disappear except for the aliased tone at DC (and tones at 100 MHz, 200 MHz, etc.) The
output of the counter, which is read out as the counter is reset every KTs seconds, (that is Input clockJlJ
we don't look at the output of the counter while it's changing) is a constant value of 4 (that Is
never changes) even though the input is a 50 MHz squarewave.
Reset every K cl
Counter's frequency response
Using a counter as a digital filter, as we've just seen, has limited uses because of
the significant amount of aliasing that occurs with the inherent decimation. Note that we
:lU)1 didn't discuss problems that occur when the input signal contains noise. One of the
counter's uses, however, is in applications where the desired signal is constant (DC) such
as some sensing applications. By using really large values of K the bandwidth can be
shrunk down to limit the effects of noise while at the same time amplifying signals at DC
(review Fig. 4.l0).
f
, .. The Accumulate-and-Dump
The counter's input was a I-bit word. For N-bit input words the accumulate-and-dump
can be used, Fig. 4.12, in place of a counter. The input signal is summed in a register for
K clock cycles (note that this is the non-delaying integrator, also known as an
.igital filter. accumulator, seen in Fig. 1.26). At the end of this time the sum is clocked into an output
hold register and the summing register is reset (the integrator is reset). The equations and
ment numbers). However,
filtering behavior of the accumulate-and-dump are exactly the same as the counter. The
maximum input word's value is 2N - I (all N bits high). In order to accommodate the
quency of 100 MHz. The summation of K, N-bit, input words the register size used must be at least N + log2K bits
e counter is read out every wide (rounding up to the nearest integer). The frequency response of the
:y response of the counter accumulate-and-dump filter or the counter at DC is K.
rave at 50 MHz. We only
to keep the figures from
i), having a peak-to-peak : Accumulator or
. non-delaying
50 MHz with dirac-delta : integrator
I1put signal spectrum after Input, N-bits x[nTs]: + }--~--t-~--l
:he sampled signal at DC).
response all of the tones
\IlHz, 200 MHz, etc.) The
:very KTs seconds, (that is Input clockJ1Sl.fL· Dump sum into
this register.
a constant value of 4 (that Is .
r-L:...:..:..: lsiK
Res-et-e-v-e-ry-K-=-=c=lo=c=k=-c....:y=c=le=s=.~.>-------!.!-~4 . -.-~
~f,MHz B= (4.14)
2K
:ults in aliasing.
130 CMOS Mixed-Signal Circuit Design
Y(z) ~8-bit
ADC
~
8 .
Comb filter, Fig. 1.25
IH(f)1
I z-plane
i, Is K=2 H(z) = 1 900mV
'C?v:
I-z
810mV
••• nOmV
» fin (Hz) 6]OmV
~
5~OmV
4 • ,,
450m'l
, K=4 360mV
L087~~~ •••
270mV
) 180mV
OmV
C
Integrator
810mV
720mV
:::::1/::.
1-- H(z) = 1- ,-'
36omVr:::
270mV- .......... ..
180mvi- ............ :
90mV4... .
omv+ i I i I
Oos 1000' 200ns 300m; .400ns 500ns
1--- H(z) = 1
Figure 4.15 Input and output of the filter in Fig. 4.14 at a frequeney of6.25 MHz.
Example 4.4
Determine, and sketch, the time-domain impulse response of an averaging filter
withK= 8.
The transfer function of the filter is given by Eq. (4.1 0) or
g values of K.
H(z) = 1 I+Z~1+z-2+ + + z·5
l-z
The time domain relationship between the input and the output is then
the Sinc-shaped filter
y[nTsJ x[nTsJ+x[(n I)TsJ+x[(n-2)TsJ+ ... +x[(n-7)TsJ
Note that the output is simply a sum of the inputs over time, KTs' as seen in Eq.
(4.7). The time-domain impulse response of the first-order averaging filter is
the filter's frequency shown in Fig. 4.16. Note the rectangular shape .•
MHz.
le filter's output word x[ nTsl
1• Impulse input
(4.15) -++-+-1-11-+1""'IH-I-+-I-11-+1""'IH-I-+-I-II-~) time, niTs
o
8-bits the final output y[nTsJ
~.
x(l) + x(2) + x(3) + x(4) + x(5) + x(6) + x(7) + x(8) + x(9) + ... (4.16)
where the output clocking frequency is, as seen in Fig. 4.9, J.JK. For the lowpass
Sinc-response (averaging) filters discussed in this section
Finst output
Figure 4.17
x(l) + x(2) + x(3) + x(4) + x(5) +x(6) + x(7) + x(8) + x(9) + ...
Second output
I ,
Example 4.5
x(l)+ x(2) +x(3) +x(4) + x(5) + x(6)+x(7)+x(8)+ x(9) + ...
Repeat Ex. 4.4 if!
Third output
The transfer funct
x(l)+x(2)+ x(3)+x(4)+ x(5)+x(6) + x(7) +x(8) + x(9)+ ...
~---~---~
Fourth output
(4.18)
or, again, a filter response given by Eq. (4.10)
Cascading Sine Filters
The transfer function of a cascade of L of these Sine-response filters can be written as
1
H(z) = [ 1 =:-1-KJL (4.19)
IH(f) I
L ·3.9 dB
.2.1 performed averaging
:1 K = 4, it summed four
~.~~ -"
esult to the output:
B = 0.5(1s1K) L· 13 dB
Figure 4.17 General frequency response ofa lowpass Sine (averaging) filter.
+ x(9) + ...
Example 4.5
~)+ x(9) + ...
Repeat Ex. 4.4 ifanL = 2 averaging filter is used.
The transfer function of the filter is
s)+ x(9) + ...
H(z) [ I - z-8
l-z-I
J2 = 1 + 2z- 1 + + 4z-3 + 5z-4 + ... + 3Z-1 2 + 2Z- 13
8) + x(9) + ...
The time domain relationship is
(4.17)
y[nTsl x[nTs]+2x[(n 1)Tsl+3x[(n-2)Tsl+ ... +2x[(n-13)Ts]+x[(n 14)TsJ
rate as the inputs,],. The
arne as the counter or the The impulse response of the L 2 lowpass Sinc filter is shown in Fig. 4.18. Note
the triangular shape of the curve and how the impulse response of the L == 2 filter
lasts twice as long as the L = 1 filter's response in Fig. 4.16 .•
(4.18)
y[nTs l
8
The integrator (also sometimes called an accumulator) shown in Figs. 1.26, 1.28,
and 4.14 is an example of an infinite impulse response (IIR) digital filter. Applying a unit
amplitude impulse to the input of the digital integrator, with zeroes the remaining times,
In --I-
causes the output of the integrator to increase to one and remain at one indefinitely. In
other words, the output response of the integrator is of infinite duration. The integrator is
also called a recursive filter since its current output value is a function of previous output
values. Any digital filter with a denominator is a recursive filter. Figl
While the averaging filters seen in Figs. 4.13 and 4.14 use an integrator (IIR
section) and a comb filter (FIR section), overall, as seen in the previous two examples, it The time-domain repre
exhibits FIR behavior. Further, since the averaging filter's output is only a function of its
y[nTs] =
inputs, as seen in Eq. (4.17), it is a non-recursive filter.
4.2.3 Bandpass and Highpass Sine Filters . It's desirable to determ
So far we've focused on lowpass filters. There are many situations, especially in value that results in a 1
communication systems, where we may want to filter or perform data conversion on a bandpass filter with tri'
range of frequencies that doesn't extend from DC to B (= j,/2K). Bandpass ADCs and delays and additions s'
DACs, for example, are popular in communication systems. In this section, we introduce investigate is /s/4. At
bandpass and highpass sine-shaped filters. comb filter transfer fun
l-z- 8 ", Z8 -I
8 1 +z-l z7(z+ 1)
j,/4 J
•••
t
fsl2
Figure 4.21 A ban,
is 14 Is f
3j,/8 We can determ
procedure used to deten
Figure 4.19 A highpass filter implementation using a comb filter.
We can generate a bandpass filter by canceling the zeroes at j,/4 and 3/s 14, or
some other frequencies, using a digital resonator. The general topology of the bandpass
digital filter is shown in Fig. 4.20. Keeping in mind that the digital resonator is used to
cancel the zeroes ofthe comb filter, we can write At the center of the pas~
first side lobe, on eithe
response and is calculat.
(4.23)
(ed-Signal Circuit Design Chapter 4 Digital Filters 135
It's desirable to determine at which frequencies the cosine term is an integer, zero, or a
situations, especially in value that results in a trivial multiplication, that is, a shift so that we can implement the
Jrm data conversion on a bandpass filter with trivial multiplications. In other words, we want a filter that uses only
uq. Bandpass ADCs and delays and additions so that its implementation is simple. The first frequency we will
this section, we introduce investigate is /s/4. At this frequency the cosine term is zero and the digital-resonator/
comb filter transfer function (the bandpass transfer function in Fig. 4.20) reduces to
1 -K
H(z) := -=L (4.25)
1 +z-2
lceling the zero at DC in a
The magnitude and z-plane response of this filter, for K 8, is shown in Fig. 4.21.
by canceling a comb filter
8. The same equations,
Jf this filter where, in the
,hen looking at Fig. 4.19, Z8 - I
lic with period is. I + z-2 Z6(Z2 + 1)
'\
DC
f
Figure 4.21 A bandpass filter implementation using a comb filter and digital resonator.
We can determine the magnitude response of Eq. (4.25) following the same
procedure used to determine Eq. (2.38). The result, for the f,/4 resonator, is
l eomb filter.
13t
I Main lobe I T
First sidelobe 9-1
11 T
.,
Bandpass response when I:: /,/6 orl,./3
x[nTs]
+
7
st
1
Ii;;,dP~' .ttenn.tion whenf = /,/4
Example 4.7
Fig!
Assuming an 8-bit If
3 I I ~ I ) bandpass filter
3 4 6 8 \0 12 16 K
Figure 4.22 Lowpass and bandpass filter attenuation versus number of comb filer zeroes, K.
V.rhat is the output v
The cosine term in Eq. (4.24) can be set to ±l when 1= Is 16 or 1s13 resulting in a response is given by E
bandpass filter. It should be clear that with the appropriate choice of sampling frequency,
number of zeroes K used in the comb filter, and value of the cosine term, many different The sketch of the fil
combinations of simple bandpass filters can be implemented using these techniques. Again, the word size ~
There are, however, only a few resonators that can be implemented using binary numbers. is 8, the word grows:
The resonators' coefficients must be such that the poles perfectly lie on the unit circle and filter and 2-bits in tht
cancel the comb filter's zeroes on the unit circle. when the input is 25 .r-.
verify that Eq. (4.26)
The ratio of the main lobe to the first sidelobe for 1= 1s16 or /,/3 IS given, 12.5 MHz the filter's
assuming K = 12,24, ... , by frequency response). ,
is
Main lobe I K sin U;)
(~~)sin (~- i;)
sin
IFirst side lobe - ----si-n-=----:: 1.15Ksin (4.28)
IH(f)1
which is approximately l3.5 dB for K = 24,36,48 ... and 10.15 dB for K = 12, 4.22.
In order to increase the amount of attenuation between the main lobe and the first
where, since we norm
side lobe in a bandpass filter implementation, we can cascade filter sections (as we did in
equals the output), th
the lowpass filter implementations discussed earlier). For example, cascading five 1s14 we'll see that we can .
bandpass filters with K == 8 will result in an attenuation of 57 dB. Also, note that by
they don't contribute tl
changing the sampling, or filter clock frequency J.. , we can easily change the bandpass
filter's center frequency. A change in the clock frequency, and its selection, can easily be
implemented using a counter and some control logic.
Example 4.6
Sketch the block level circuit diagram for an 1s14 digital resonator.
From Eq. (4.24) the time domain representation of the 1,/4 resonator can be
written as
I
ed-Signal Circuit Design Chapter 4 Digital Filters 137
Example 4.7
Assuming an 8-bit input word, sketch the implementation of the Sinc-shaped
I )0 bandpass filter
K L£:
1 +z-2
er of comb filer zeroes, K.
What is the output word size? Using SPICE verify that the filter's frequency
fs/6 or 1s13 resulting in a response is given by Eq, (4.26) with K = 8. Assume j~ = 100 MHz.
:e of sampling frequency,
sine term, many different The sketch of the filter is seen in Fig. 4.24 (note the similarity to Fig. 4.14).
using these techniques. Again, the word size grows by log2K from the input to the output Here, where K
ted using binary numbers. is 8, the word grows 3-bits, to an output word size of II-bits, (I-bit in the comb
{ lie on the unit circle and filter and 2-bits in the resonator). Figure 4.25 shows the output of the resonator
when the input is 25 MHz (= J,J4). It can be useful to vary the input frequency and
verify that Eq. (4.26) is correct. For example, if we change the input frequency to
1=I s l6 orls13 is given, 12.5 MHz the filter's output is zero (this is the first adjacent zero point in the
frequency response). When the input frequency is 15 MHz, the output of the filter
is
I1t)sin(lL
!K \3 2K
(4.28) 3n) sin (Knf,) sin (8n&)
IH(f) I 0.5881
= -
0.588 = 1 -) 0.25 (scaled by 4)
IB for K 12, Fig. 4.22. cos (2nl1~o)
1
= cos (2n:f,)
he main lobe and the first
where, since we normalize the filter's output so that passband gain is I (the input
Iter sections (as we did in
equals the output), the output of the filter is 0.25 the input. In the next chapter
nple, cascading five 1s14
we'll see that we can throwaway some of the lower bits in the output word since
7 dB. Also, note that by
they don't contribute to an increase in the SNR.•
.sily change the bandpass
ts selection, can easily be Change into two's complement , , Change back to
and extend sign bit b . 1. Extend the SIgn-bit by 2 = \og2 K 1 binary offset format
~ ~ JC"
)nator.
Resonator
(.,J4 resonator can be
Figure 4.24 Digital filter sketch for Ex, 4,7.
138 CMOS Mixed-Signal Circuit Design
Digital resonators
Comb filter
In Out
Figure 4.2'
summation of the frequency domain Sinc responses to generate a bandpass filter with a
variable width. Note that every other digital resonator is subtracted rather than added to
the final result. This is to account for the phase reversal between adjacent resonator
outputs.
4.2.4 Interpolation using Sine Filters
We saw back in Sec. 2.1.5 that an interpolator is a circuit that increases the clocking rate
of the digital data. This was useful for, among other things, reducing the requirements
placed on the reconstruction filter. The simplest interpolator was the input hold register,
Fig. 2.28. Before we discuss interpolation using Sine filters let's answer the question
"Why can't we use an input hold register for interpolation here?"
1-10n5
Examine the block diagram in Fig. 4.27. The ADC is clocked at a rate ofls so the
anti-aliasing filter (AAF) limits the ADC's input spectrum to /,/2. On the input of the
1.7at/s 4.
ADC is a sample-and-hold (SIR) with a Sinc-shaped spectral response (-3.9 dB
attenuation at the Nyquist frequency of 1s12, see Fig. 2.17). We might expect, based on
Fig. 2.29 and the associated discussion, that the hold register seen in Fig. 4.27 would
IFig. 4.26. We are feeding provide an additional Sinc-shaped response in the signal path. Rowever, remember (see
ent center frequencies) and Ex. 2.2) repetitively sampling and holding a signal results in only one SIR attenuation.
esponses (the Sinc shapes) Since the input signal in Fig. 4.27 sees this response when it passes through the SIR it
ructing a waveform in the isn't affected by another Sine-shaped response as it passes through the hold register.
ept now we are using the
=/s,ne w
1-+-'10-- Out
Note that because of this assumption that the data has passed through a SIR and
experienced the associated Sinc-shaped filtering, we won't use zeroes padding in the
discussions in this section. It should be straightforward to extend the discussions to
designs that do use zeroes padding. Also note that if the digital data hasn't passed through
ler response a Sinc-response filter then passing it through a hold register, as discussed in Sec. 2.1.5,
does introduce a Sinc-response (which, as we'll fmd out now, may often be desirable
when performing interpolation).
The next question is "Why would we want the signal to see additional filtering?"
Sidelobes not shown
The answer to this comes from reviewing Fig. 2.26. The image removal filter is an
important component of an interpolator. We can implement this filter using
1 z_
H(z}= __ -K JL (4.29)
[ l-z-l
140 CMOS Mixed-Signal Circuit Design Chapter 4 Digital Filtet
The benefit, for an interpolator, comes from realizing that if we clock the comb filter with The attenuation thr'
a slower clock, fJK, then one clock delay is KTs so we can simplify the comb sections
using Sine!
IH(/) I ==
I-z-K -* 1 (4.30) Sine
Figure 4.28 shows an interpolator implemented using a cascade of Sinc Filters. The
attenuation through the interpolator is given, see Eq. (2.37), by or, for L = 1,
attenuation betweel
Sinc(K1tf'~"") L
2.30. For LIthe
IH(/) I == (4.31) dB.
Sinc(1tf,~J Figure 4.29 she
the interpolator'!
Larger attenuation can be achieved by cascading more stages, L. Again, the word size
points. For L = 2 v.
grows by I-bit through each comb filter and by log2K - 1 through each integrator. Also
a more "rounded" c
note that if we were to use zeroes padding instead of the inherent hold register present in It can be instructiv
the topology (see discussion on previous page) a selector, see Fig. 2.26, would be added the resulting outpu
in between the comb filters and integrators to introduce K - 1 zeroes. drops as the input
•
In I :ot"'or (it:'" T d
ret')I . Out S60mV,
880mV
800ml'
Input word rate, Is Output word rate,KIs == Is.new 120mI"
640ml'
560ml'·
L sections L sections 4BOml'
~
/
~
~mm
... ~
~~'~.~
" /=:
/ ,,~~~~
...
_ _ _ _ _ _
r=Ll.out
)~
"
~/
400ml'
320ml'
240mV,
160mV
Clocked atfs Clocked at Kfs = is",,,,,,, 80mI"
0.1
The Nyquist frequency is 6.25 MHz (the spectral content on the input of the ADC 320mV,
240mI"
should be < 6.25 MHz to avoid aliasing). The output clock rate is 100 MHz. For
HOmV
the case when L = 1 the word grows I-bit through the comb filter and 2-bits
80mV
through the integrators so the output word size is II-bits. When L = 2 the output 0,
word size is I4·bits. We'll see in the next chapter that the signal-to-noise ratio sets
the number of bits we keep on the output of the filter or in between adjacent Figure 4
stages.
'ted-Signal Circuit Design Chapter 4 Digital Filters 141
clock the comb filter with The attenuation through the filter is
mplify the comb sections L L
. (
Smc\Krc1;.new)
J I Sinc( 8rc;o~)
. (
sm \. 8rc]OO
II )
L
Out
~
tions L=l
~
.~ut
.~
--.~
Kis
0.2115 0.411S 0.611' 0.8115 1.01ls 1.211s 1.4110
~e removal. V!vout_no_lnterpotadonl
960mV,---=+=::.::.:::;..:::=::.c..,.--.---,--.--_--,
880mV
aOOmV
input frequency of 1.1 720mV
and
r f \, L
SinclKM· n;;:;;:) or we can implement deci
IH(f)1 -= (4.33) the associated discussion).
Sinc( 1t f,:J that we are assuming thf
implementation of the deci
Figure 4.31 plots this equation. Note that the Nyquist frequency J/2) remains faetors of delay are implerr
unchanged since it's set by the sampling in the ADC. By using the factor M we can .an averager, 1 + , with a
achieve a narrower bandwidth but at the cost of more droop at the Nyquist frequency. The some authors, a CIC filtel
word size, in this filter, still grows by I-bit in every comb section and by log2K - 1bits in there aren't any integrators
the integrator sections. when using this decimatOl
increasing the order of the
iH(f)i
Sinc(Mn . f)
\H(f)1 =. j'
smc(n· K-jJ In
Desired spe:illJb!lI!._ Input word ralt
f h,ne'1' = K .Is
o M lvl
£
2M
Figure 4.31 Frequency response image removal tilter using a Sine interpolator, Fig. 4.30.
r
lC filters.
(4.29)
teristics using
H(z) ~z-n
= [K-l =[I+z-1+ + ... +z !-K]L
(4.32)
=[(1 + . (1 + z-2) ..... (1 + )Y
(1 + . (1 + z-2)L ..... (1 + )L (4.34)
or we can implement decimation using only non-recursive averagers (see Fig. 1.16 and
(4.33) the associated discussion). The word size increases by I-bit through each averager. Note
that we are assuming the decimation is a factor of 2 here. Figure 4.32 shows the
implementation of the decimator. The registers are used to re-sample the data. The higher
frequency 1/2) remains factors of delay are implemented by clocking the registers with the slower clock (clocking
using the factor M we can an averager, I +z-I, with a clock of IsIK implements 1 + z-K). This topology is called, by
Lt the Nyquist frequency. The some authors, a CIC filter (see comment at the bottom of previous page) even though
ction and by I bits in there aren't any integrators present in the topology. Figure 4.33 shows example spectrums
when using this decimator. Note that the sidelobes alias into the desired spectrum. By
increasing the order of the filter L, we can reduce the effects of aliasing.
f f"new K ·Is
.&
K
,en ted are often labeled CIC
mb filters. We'll continue to Figure 4.32 Decimation using Sinc anti-aliasing filters.
: next section.
144 CMOS Mixed-Signal Circuit Design Chapter 4 Digital Filters
Figure 4.33 Spectrums when decimating and a Sinc anti-aliasing filter used.
If, for example, we set al
(4.9) and (4.10),
Example 4.8
Using an 8-bit word, an input clock rate of 100 MHz, an input frequency of 1.1
MHz, and K == 8 determine the output word size, magnitude, and output clock rate
for Sinc-decimators with L 1 and 2. Verify your results with SPICE. or, again, a Sinc-shaped
earlier, Figure 1.24, this t
The transfer function of the anti-aliasing filter inherent in the decimator is given
by Eq. (4.34) with K 8 and L 1 or 2. The output word rate is 12.5 MHz. The
simulation results are seen in Fig. 4.34. Note that the decimated output with L 1
looks essentially the same as the output with L = 2 (except for more delay when L In~_-
2). This (the shape of the outputs being more or less the same) wasn't the case
for the interpolator outputs seen in Fig. 4.29. For the interpolator, where aliasing
isn't a concern, increasing L results in a reduction in the aliased signals present on
the output of the interpolator. Increasing L allows the interpolator output to move
closer to the original input (in the extreme, L ~ ct:), and we get the original analog
waveform). By increasing the order, L, in the decimation filter we reduce the
aliased signals present in our desired spectrum, Fig. 4.34 .•
V(vout_no_decimation)
As another examp
shifting the word left two
bit increases the word s
hardware). Figure 4.36 sl
first-order RC circuits stc
simply reveals the filter's
heuristic approach can be
The benefits of us
recursive so no feedback
drawback is that they, fc
recursive structures we'll
Structures are subject to il
Figure 4.34 Decimating by 8, an example. overflow. Let's discuss the
ixed-Signal Circuit Design Chapter 4 Digital Filters 145
x[nTs]
z-a
Ao =AI
A2 A3 =: 0.25
II.... !!!!!!!!
o
Step input
y[nTs]
/t.. Output
Figure 4.36 Step response of a 4-stage FIR filter "With all coefficients set to 0.25.
1
H(z) =: z-a (4.39)
Figure 4.38 shows the z-plane and magnitude plots for this equation. If a > 1 then H(z)
becomes unstable. So for a stable system we must require our poles to reside within the
unit circle. (There are no restrictions on the location of zeroes.) This sounds simple
enough; however, notice that we have, in most of the previously discussed digital filters,
placed poles right on the unit circle. If there is rounding in our digital numbers, we could
be faced with an unstable digital filter. This would be a very common occurrence in a
1000
digital filter implemented using software, if care was not taken to avoid rounding errors.
Since we use integer numbers in our hardware implementations, instability shouldn't be a (a) Integrator ove
problem unless we start to try to round numbers to decrease hardware complexity
(performing divisions or multiplications) without being careful. Figure 4.39
xed-Signal Circuit Design Chapter 4 Digital Filters 147
tput a .~ ..
:,nITs _ ____
I
_ ___ L____
•••
_ ____ _
,
l+a . )
Overflow
I)Ts] (4.37)
Recall that when we used integrators with comb filters earlier in the chapter we increased
the input/output word size of the integrators by log2K 1 while the comb filter outputs
:[(n-4)Ts ]+ ... , (4.38) increased by I-bit for an overall increase of log 2 K. We calculated these values by
determining the register size required to sum K words. In a recursive filter the output is
fed back and summed at various points in the filter, so determining the exact register size
required can be challenging. Figure 4.39a shows how an integrator overflows causing the
output to wrap around (go from the most positive value to the most negative value or
vice-versa). In Fig. 4.39b we show the more desirable situation where the output
z) = [aY(z)+X(z)]z-l saturates. This keeps the filter from becoming unstable; however, it will introduce
nonlinearities in the filter's response. Nonlinearity at some extreme is usually better than
'z) = Y(z) = --='--_ instability. The question is, "How do we determine overflow?" We know that we can't
, X(z) I-a·
look at the carry out bit because, as we saw in Figs. 4.7 and 4.8, the adder overflowing is
a normal occurrence when adding two's complement numbers. What we do know is that
r. if the input sign bits are both 0 then the output sign bit must be 0 (if both inputs are
positive then the output must be positive). The same can be said for adding negative
numbers. Figure 4.40 shows how we can modify an integrator to avoid overflow.
(4.39) Olill
quation. If a> I then H(z)
r poles to reside within the
roes.) This sounds simple time time
OOOOO-!L---------~
sly discussed digital filters,
r digital numbers, we could
y common occurrence in a
n to avoid rounding errors.
1000 - t--- lOOO-t-
lS, instability shouldn't be a (a) Integrator overflow (b) Clamping the output when
ease hardware complexity overflow occurs.
(4.40)
where s = jro = j2rr./. Re\'
and so
In 10k
-
Integmtor
- . - - .
Vin
I;
If
-
Out
. Vout(f)
Differentiater
Note that in order for our filter to be useful the frequencies of interest must be much
lower than the filter's clocking frequency, 1,. In other words, the frequencies where the
pole and zero are located must be much less than 1,. This means, assuming G2 1, that
GlD« 1 and G)D» I.
Let's attempt to simplify this tilter. If we look at the transfer function, Eq. (4.41),
we see that the feedback gain, G z, simply scales the amplitude of the transfer function and
can be used to further adjust the location of the filter's pole. Because we can scale the
amplitude of the signal either before or after the filter, and independent of the filter's
operation, and we can precisely set the pole of the filter using G iD , we ean, without loss of
functionality, set G2 to 1. We can then rearrange the summing, delaying, and multiplying
blocks, as seen in Fig. 4.43. Using these results we can write the z-domain representation
overflow. of the transfer function, Eq. (4.41), as
VOltt(z) G lD (1 +G 3D ) Z-G3D/(l +GJD)
(4.44)
Vin(Z) == z-(l GlD)' z
using an integrator and Before attempting to simplify the filter implementation seen in Fig. 4.43 further, let's
4.41 shows the digital show that, indeed, Eq. (4.44) is equivalent to Eq. (4.41) whenl«f,. It will be helpful to
}ortant, at this point, to see remember that
'ecUy implemented in Fig.
::::: 1 - §.. if 1 «Is or z "" 1 + §.. "" _1-s "" if "" 0 (4.45)
h h 1 h J?
(4.40)
where s == jm == j2rrf. Rewriting Eq. (4.44) in the s-domain results in
Vout(f)
Vjn(f)
= Gl~( 1+ G3D)
1 + f, 1 + GlD
·ll (1 -7J's
) G3D/(l + G3D)J (4.46)
(4.41)
1+
(4.47)
1+
10k
Integrator
VOllt(f) 1 1 +jro· lOOns
Out Vin(f) =2" l+jm·50ns
; Vout(f)
Out ~
LE
t:
Out
Out
H(z)
GlD
z-(1 G ID )
Comparing the transfer function in Fig. 4,42 to Eq. (4,47), we see that if
Out then we can use any of the filters in Fig. 4.43. The sketch of the digital filter is
z-ll---+-_---'l>-
seen in Fig. 4.44. The multiplication of the transfer function by 112 is nulled by
the mUltiplication by GlDG3D 2), To verify that the filter in Fig. 4.44 functions
as desired at DC, we see that the output of the first stage is 0.5 when the input is
0.1, and the output of the second stage is 0.05 (with a 0,5 on its input).•
~t
g-l
t::J 0,8
----"'--- . (1 . 10/11)
l-GID
Figure 4.44 Digital filter from Ex. 4.9.
V011f(Z)
--:
Vin(Z)
G lD G 3D II + G3D
G3D
.
1 Z-l
1
(1- GlD) 1
(4.48)
or
VOI"(Z)
--:
G ID (1 + G·3D ) . _z_G-,3:o:D....;/C,-1_+_G_3: .:; D, -) C4.49)
VinCZ) z-(l-GlD)
which is clearly the same response as derived in Fig. 4.43 or Eq. (4.44) except that the
le bilinear filter, output is one clock cycle, z, earlier. This reduced delay has no effect on the magnitude
response of the filter and little effect, assumingf«f" on the phase response of the filter.
152 CMOS Mixed-Signal Circuit Design Chapter 4 Digital FilteJ
Out Vout(Z)
f-"+---
Example 4.10
Using the canonic f
Comparing Eq. (4.~
2rrj3d8.p(
1
ADc=
Figure 4.45 Canonic form of a first-order digital filter. 2
The general form ofthis first-order canonic (or standard form) filter is seen in Fig.
4.46. The filter is termed canonic because the minimum number of delays are used. One and thus Bl == -0.9.
delay is used for each pole. Remember that in order for a digital filter to be realizable in implement the mult
hardware there must be fewer or an equal number of zeroes than poles in a filter's transfer
function. Let's do a quicl
apply 0.1 to the inp
4.42, the output of
Change to addition and negate B I input to the filter is
~~ AJ==I-G lD
According to Fig.
integrator is 1I[ 1
then be 0.5 - 0.45 0
Bo == GlD(l + G3D)
Bl =-GlDG3D
-I z+!!J.
H(z) = Bo+!i~=Bo' ~
I AlZ-I Z-Al
We can, again, derive the transfer function for the first-order bilinear digital filter.
This time, however, let's use the variables in Fig. 4.46. Again, assuming!« Is, and using
Eq. (4.45) results in
Figure 4.4
1+
H(f) = Bo +BI , _ _--"_ (4.50)
I-Al Example 4.11
Sketch the digital fi
where last chapter that has
the filter is clocked c
/,(l-A d (4,51)
/3d8,pole =: 2rr The filter's continu01
and
ADC (4.53)
Example 4.10
Using the canonic form of the first-order digital filter, repeat Ex. 4.9.
Comparing Eq. (4.50) with the transfer function in Fig. 4.42, we can write
I
2rcj3dB,pole == 50 ns
ADC=l=~..:.-;;;;;....!.. Bo+Bl
1-0.8 ~Bo+Bl 0.1
al filter. 2 I-AI
~ + !!J.
.. B
Bo'--O
Z-Al
igital filter.
Figure 4.47 Canonic form ()fthe first-order digital filter in Ex. 4.10.
(4.50)
Example 4.11
Sketch the digital filter implementation of the lowpass filter in Ex. 3.1 from the
last chapter that has a DC gain of 1 and a 3 dB frequency of 1.59 YIHz. Assume
the filter is clocked at 100 MHz.
(4.51)
The filter's continuous-time frequency response is given by
H(j) = ----"--
I +j.
(4.52)
Using Eqs. (4.50) to (4.53), we begin by calculating AI
154 CMOS Mixed-Signal Circuit Design Chapter 4 Digital Filt
and then
Bo+BI Bo+B)
ADC = - - = 1 = - - -? Bo +B) 0.1
. I-AI I 0,9
Let's put the zero at infinity so it doesn't affect the transfer function
f 3dB,zero = 00 fs l( 1 Bo
2n +!!J..) ~ BoO and B I = 0,1
Ifwe want to write the frequency domain transfer function, we write, again assumingf«
Is and using Eq, (4.45),
(4.55)
or In
~
(4,56)
While we can design higher-order digital filters using the topology of Fig. 4.49, we will
restrict our analysis to first- and second-order filters where hand calculations are
relatively easy to perform. Note that we can increase the attenuation of a filter by using Fi
several of these sections in cascade,
(ed-Signal Circuit Design Chapter 4 Digital Filters 155
= 0.1
mction ••
B1 =0.1
•
.t
•••
LII.
Number of poles n?: number of zeroes.
,nn of an n'h -order canonic Figure 4.49 General canonic fonn of a digital filter.
e transfer function). The
(4.55)
(4.56)
t) n-i
In order to translate this transfer function into the frequency domain, we use Eq. (4.45) lSdB
and assume our frequencies of interest are much less than the sampling frequency 10dB
odB
OdB
-!idB
·lOdB
(4.58) -15dB
·20dB
-25dB
-30dB
After some algebraic manipulation we can put this equation in the form seen in Eq. (3.62)
~=
Bo . S2 + ';;(2Bo + BI)' S +f.l(Bo + BI + B2)
j. ~m
~:::j1
-SOdB
s2+!s(2-AI)·s+f?(l-A I A2) -55dB
-60dB-+
where 10K
(4.60) Fi
01 .fs(2Bo +Bd (4.61)
00 =f}(Bo+Bl +B 2 ) (4.62) Comparing Biquads to ,
. Consider the frequency
21rlo = fs(2 - A J) (4.63)
Q . This filter uses a clockiI
and finally,
0.707 ~ Al
6
Q 2n· 1.59 x 10 = 1.859
100x 106(2-Ad
The block diagram of th
and
2rr ,\2
1o OdB"
l-AI-A2= ( Ts) ~A2=-0.869
I
J\
·20dB-i .
and finally, because the gain at DC is 1,
I
B2 l-A I -A2=O.OI
Note that if a scaling in the amplitude is allowable, we can remove this
mUltiplication or approximate it with shifts in the digital word.
The simulation results are seen in Fig. 4.51. In order to implement this
simulation in SPICE, we used transmission lines for the delay elements and
:::1
voltage- controlled voltage sources for both the multiplications and the adders. 'loodBl
This method allows us to simulate the filter's frequency response. Note that the
frequency response is periodic with the filter's clocking frequency.•
I?igure 4.52 I
xed-Signal Circuit Design Chapter 4 Digital Filters 157
-Bz
(4.58)
(4.59)
(4.60)
Figure 4.51 Simulating the digital filter in Ex. 4.12
(4.61 )
(4.62) Comparing Biquads to Sine-Shaped Filters
Consider the frequency response of the Sine-shaped lowpass filter shown in 4.52.
(4.63)
This filter uses a clocking frequency of 100 MHz and a K of 16 or
3
IH(z) I = 1 -1 1 (4.65)
l-z
(4.64)
Note the significant droop in the filter's response. It's desirable to design a filter that
doesn't have this droop or, even more desirable, contains a small amount of peaking to
compensate for a Sinc-shaped attenuation from a SIH, decimation filter, etc. Using
(4.59) - (4.64) we can set, for a digital biquad equivalent of this filter, B2 0.03 Al
I MHz and Q 0.707. 1.75, A2 = -0.78125 (there are several other solutions, depending on the desired
and a I are zero. This complexity or performance of the filter). The simulation results comparing the Sinc and
biquad filters are seen in Fig. 4.53. The transfer function of the biquad filter is
-~Od!!
20dB~----'-"-==~-,-----~- IOdS·
OdB OdB'
·20dB ..
·IOdB·
·40dB
-2DdB·
·60dB
·.1'··•.••••••••• ·8UdB
·SOdS
-24OdB·l i i i 1---- -100dB
20MH. 40MH. 60MHz 80MHz IOOMHz 5001
Figure 4.53 Comparing a third-order Jowpass Sinc filter to a third-order biquad. Figure 4
In Out
In
\AdjUst as needed
Example 4.13
Redesign the filter in Fig. 4.54 so that the response has a small amount of peaking ADDITIONAL REAl
at 3.125 MHz. Compare, with simulations, the biquad's response to the Sinc [I] M. Weeks, Di!
filter's response seen in Fig. 4.52. Science Press, :2
Reviewing Eq. (4.63), we see that to increase the Q we need to increase A j • [2] T. B. Welch,
Keeping in mind that we want to have simple multiplications relying heavily on Processing ji'O/
shifts, let's try increasing Aj to 1.78125 and A z to 0.8125. The simulation results 978-084937382
are seen in Fig. 4.55. In this figure we compare the modified filter response to the
response of the Sinc filter. Note that we have a couple of dB peaking in the [3] S. Haykin ar
cascaded biquad filter's output. _ Communication
0471432227
A Comment Concerning Multiplications
[4] R. G. Lyons,
While discussing the implementation of digital multipliers is outside the scope of this Prentice-Hall,2 1
book a comment is in order about simple mUltiplier implementations. While, in some
filtering applications, simple shifts can prove very useful, we can implement more useful [5] P. A. Lynn an
multipliers using adders. Figure 4.56 shows one possible implementation using a single Edition, John W
adder along with the associated multiplication factors. We could implement the [6] L. W. Couch,
coefficients in Ex. 4.13 using a similar scheme. For example, Aj = 1.78125 2 - 0.25 + Prentice-Hall, I'
0.03125 and A2 = 0.8125 I - 0.25 + 0.0625 (both requiring two adders). Other creative
ways can be used to implement multipliers. For example, a multiplication of 0.5625 by [7] E. P. Cunningh
cascading two simple mUltipliers with multiplication factors 0.875. 1995. ISBN 978
xed-Signal Circuit Design Chapter 4 Digital Filters 159
~ 100MHz
a third-order biquad.
Figure 4.55 Designing a biquad filter with peaking, see Ex. 4.13.
~out B= Multiply by
I o
0.5 0.5
"AdjUst as needed
0.25 0.75
to nonnalize
output word size. 0.125 0.875
0.0625 0.9375
0.03125 0.96875
. Eq. (4.66).
Figure 4.56 A simple multiplier using a single adder.
4.8 Repeat Ex. 4.3 for a filter with a transfer function of Show that the f
multiplier.
Also, plot the location of the filter's poles and zeroes in the z-plane.
4.9 Repeat question 4.3 for a filter with a transfer function of
4.14 Qualitatively explain why the desired spectrum of an input signal can't be
[lg SPICE. (Design a 3-bit increased by passing data through an interpolator. Using the simulations given in
s in decimal form. Ex. 4.8, verify that this is indeed the case.
for a I-bit DAC. Note that 4.15 In Fig. 4.32, which blocks serve as the AM and which serve as the S/H?
~ analog voltage out of the 4.16 For the FIR filter seen in Fig. 4.35 with all coefficients set to 0.25, sketch the
: is the voltage value of I filter's frequency response.
we get from the sketch? Is
)AC will be a ubiquitous 4.17 For the filter seen in Fig. 4.57 determine the range of values for a and b where the
book (see Fig. 7.15). filter will be stable. What is the filter's transfer function? Sketch the location of
the filter's poles and zeroes.
12 LSB to the left? How do
~. 4.9 as + I or -I (two's 4.18 Repeat Ex. 4.9 for a filter with a transfer function of
tal filter when the input is
orrect? Why? V 0111
I +j.
== _ _--'-"':.:.=_
in alternating sequence of
t is the output of the digital 4.19 Repeat Question 4.18 using the canonic form of the first-order digital filter.
:he frequency of the input
~ct? 4.20 Repeat Ex. 4.12 if the Q is increased to 1.
4.21 Show that the filter shown in Fig. 4.58 can be implemented using a single
multiplier.
the z-piane.
of
In ----...-~40--~~)_____ Out
Figure 4.59 A simple multiplier where A and B simply shift the data.
Data Con'
5.1 Quantization No
Examine the clocked comp:
VDD
T
Figure 5.1
ixed-Signal Circuit Design
Out Chapter
After studying Chs. 1-4 we should understand the sampling process, including analog
sampling, decimation, and interpolation, the operation of the ideal ADC and DAC, and
the basics of filtering. In this chapter we turn our attention towards quantization noise.
Quantization noise is the effective noise added to a signal after it passes through an ADC
(aka quantizer), a comparator (a I-bit ADC or quantizer)), or, for digital signals, a circuit
that reduces the word size (removes the LSBs of the word). One of the key things we'll
focus on is the shape of the quantization noise spectrum and how it's added to the
spectrum of an input analog or digital signal.
In,analog ~ 600mV
800mV
IV
IV
+400mV
+200 mV
i Out, digital
500mv~
200mV OV -200 mV
I~ clock 450mV OV -450mV
950mV IV +50mV
Is
Analog
vin_ _ _ _-I
Digital
i
I
Anal~ital
'm T
V Qe(f) (added noise)
360mV ...... .
210mV ...... .
Figure 5.2 Modeling ADC quantization noise.
180mVT· ....
90mvL
input signal has a spectrum of VQe( f). Determining the shape and range of this noise OmV.
Ons S{
spectrum is one of the goals of this chapter. Finally, while a good portion of our studies
will focus on the quantization noise introduced during the analog-to-digital process we
can also apply the same results when truncating a digital word's size, Fig. 5.3. Figure 5.5 Se
Udal
Figure 5.3 Quantizing a digital word and modeling the added noise.
-, OdS-!- i'''"'
'20d!!~'i""
5.1.1 Viewing the Quantization Noise Spectrum Using Simulations ' 30dB t ....
Consider the simple connection of an ideal 8-bit ADC to an ideal 8-bit DAC seen in Fig. -40d!!~' ... .
SA. If we apply a 7 MHz sinewave to the ADC with an amplitude of 004 V and an offset ~Ud!!~'
of 0.5 V (so the sinewave swings from 100 mV to 900 mY]) and clock the ADC at 100 -60dSJ -'"
MHz we get the signals seen in Fig. 5.5. Note that the output of the DAC looks very -10dS
similar to the output of an ideal SIR (see Fig. 2.14). Now, however, the amplitude of the -OOd!!..,
DAC output signal is quantized, that is, within 1 LSB (= 1.0/256 or 3.906 mY, see Eq.
-SOd!!
[4.1], for the present simulation) of the ADC input. This quantization is not obvious after lMHz
looking at Fig. 5.5 (the time domain response). However, looking at the spectrums of the
Figure 5.6 ADC inpu1
VDD = 1.0 VDD l.0
Analog I Digital --~ Analog
Bennett's Criteria
8
Vjn
Is I lrt'
:8-b:OC I
I
I
lldelli
8-bItDAC
Vout
In order to characterize tf
following assumptions (B
.~
l. The input (to t
Figure 5.4 Passing a signal through an ADC and then through a DAC.
VREF- so that no satura'
operating range of the A
spikes to the output spect
.fixed-Signal Circuit Design Chapter 5 Data Converter SNR 165
-{r~r\
V Qe(f) (added noise)
oise.
ADC input and the DAC output reveals the difference in the noise floor between the two,
VQeCz) (added noise) Fig. 5.6. The inherent noise floor in the simulation that is associated with the input signal
is approximately -80 dB (0.1 mY, RMS.) The noise floor associated with the DAC's
output (the signal + quantization noise) is approximately -70 dB (0.316 mY, RMS). It is
Y(z) Out
desirable to determine what sets this value and its spectral content. Again, note that the
N-Fbits ADC quantizes the signal which results in the quantization noise (an ideal S/H and DAC
::luantizer don't introduce quantization noise).
n9 Simulations
ideal 8-bit DAC seen in Fig; fu.\1S
)litude of 0.4 V and an offset'
]) and clock the ADC at 100
ltput of the DAC looks very
.owever, the amplitude of the
.01256 or 3.906 mY, see Eq.
mtization is not obvious after -90d~M+H-z--4-1-TI.!-Ht - - 81 MHz 121 MHz 161MHz
oking at the spectrums of the Is
Figure 5.6 ADC input and output spectrums for Fig. 5.4 with signals seen in Fig. 5.5.
1.0
Bennett's Criteria
In order to characterize the spectral characteristics of the quantization noise let's make the
following assumptions (Bennett's criteria) concerning the signal we are converting:
1. The input (to the ADC) signal's amplitude variation falls between VREF+ and
en through a DAC. VREF- so that no saturation of the digital output code occurs. Exceeding the normal
operating range of the ADC affects the quantization noise spectrum by adding spurs or
spikes to the output spectrum.
166 CMOS Mixed-Signal Circuit Design Chapter 5 Data Converter S:
2. The ADC's LSB is much smaller than the input signal amplitude. When this
80mV
isn't the case, the output of the ADC can appear squarewave-like (when converted back 60mV
40mV
into an analog waveform) and result in a spectrum, once again, that contains spikes or 20mV
spurs. We'll see later in the book that adding or subtracting a fed-back signal (from the OmV
-20mV
output based on the expected or past quantization noise) to the input modifies this ·40mV
.fiOmV
requirement.
3. The input signal is busy (not DC or a low frequency input). We define busy, for
the moment, as meaning that no two consecutive outputs of the ADC have the same
digital code. For the ideal ADC in Fig. 5.4 1 LSB 3.906 mV and Ts 10 ns so that the
input must change at least 3.906 mV every 10 ns. We'll see that adding a high-frequency
dither or pseudorandom noise signal to the input, which can be filtered out later (either
using a digital filter or when we pass the output through the reconstruction filter), can
:::::6
o.o~. 0.11
make the requirement on the input of being busy practical in an actual circuit. We use
these assumptions (Bennett'S criteria) in the following discussion unless otherwise
indicated.
:::::r··
--..--
-40dB
·48dB
..... .
.... --
An Important Note
-56dB .... ..
It's important to note that simply sampling an input waveform, using a SIH, does not
result in quantization noise. The amplitude into the ideal SIH, at the sampling instant, is
exactly the same as the amplitude out of the ideal S/H. In order to understand why this is :::::[1:
·BBdBL. ·.
important, consider the test setup shown in Fig. 5.7. If we apply a 3 MHz sinewave -96dB ------
centered around the common-mode voltage with a 400 mV amplitude we get the outputs -104dB --- .. -
seen in Fig. 5.8. Clearly there is a difference between the SiR's input and its output. -112dB .. - ....
However, this difference has nothing to do with noise, an unwanted signal, since passing -120dB
IMHz
the output of the SIH, VanIsh' through the ideal reconstruction filter of Fig. 2.19 results in an
exact replica of the SIH input Vi" .
Figure 5.8 Time-domain d
Difference spectrum
A section of the input and output, using the test setup of Fig. 5.9, is shown in Fig.
5.lOa. It's important to understand the input/output relationship between the ideal ADC
and DAC shown in this figure. (Note that clocking the ADC too slow or putting in a ramp
that rises too quickly will distort this waveform.) As an example, when the ADC input is
slightly above 482 mY, in this figure, the ADC output code (input to the DAC) changes.
The ADC output code can be calculated as 482 mVIl LSB ( 1 LSB 11256 = 3.906 mV
for the present simulation) or changing from 123 to 124. Looking at the transfer curves in
nput and output.
this figure it appears as though the output changes when the ADC code is 123 or 480.4
VISB
1I2LSB
Probability density function, p
E
," .. ,....... ~l LSB =
!.:'::::i'::':: 3.906 mY
f
-II2LSB
p' dQe= 1
Figure 5.11 Probability density function for the quantization error in an ADC
assuming Bennett's criteria hold.
!l2LSB V2
md DAC output.
PQe= f
-!l2LSB
p·(Qe)2·dQe= LSB
12
(5.2)
Fig. 5. lOa, is shown in Fig. Again, if our LSB voltage is 3.096 mY, then, once again, VQe,RMS = 1.13 mV or -59 dB
re that I) its average value is (RMS). If we look at Fig. 5.6, we see that the RMS noise voltage varies essentially over
we expect a wideband output the entire spectrum (white noise) and has a value ranging from around -70 dB down to
'orm), and 3) its peak-to-peak: less than -80 dB. Note that although the entire spectrum contains quantization noise it is
~ero average value, we can not because of the sampling process used in the ADC (and so quantization noise doesn't
its root-me an-square (RMS) experience aliasing). Quantization noise is added to the signal after the sampling process
during the analog-to-digital conversion process. In order to qualitatively understand why
the quantization error spectrum is white, in Fig. 5.6, we remember that there are abrupt
~dt = I LSB = VLSB (5.1) transitions in the DAC output, and if the quantization error is truly random, the times
JT2 JT2 between the changes have varying periods. We might speculate that by simulating a
longer time or using a multiple frequency input so as to "exercise" the ADC, the resulting
ific data converter. Note that quantization errors are further randomized and the resulting error spectrum will be flatter
esn't appear in the evaluated than what is seen in Fig. 5.6.
:Dcy, is ' isn't present in this
o.V, VQe,RMS = 1.13 mV or -59 5.1.2 Quantization Noise Voltage Spectral Density
If the quantization noise voltage spectrum is truly flat (Bennett's criteria hold) we can
determine the noise power spectral density of VQe.RMS, V8e(f) with units of V 21Hz, or the
noise voltage spectral density, VQe(f) with units of VI/Hz by solving
:age can be thought of as a
in Fig. 5.11. The probability V2 /,12
-,;'-.
. ,:'
with units of VI jii;. Note that the quantization noise spectral density is inversely sampling frequencies. F
proportional to the sampling frequency. Figure 5.2 shows how we can model the ADC as we assume Bennett's cril
a summation of the input signal and the quantization noise. In order to see the
After looking at Eq. 5.5 we might think that by simply increasing the sampling remove tones (input sigJ
frequency we can reduce the amount of quantization noise an ADC introduces into an in Fig. 5,13 the way
analog input signal. While increasing the sampling frequency spreads the quantization (remembering simply s
noise spectral density out over a wider range of frequencies (see Fig. 5.12) with a DAC's output. Using i<
corresponding reduction in amplitude, the sampling frequency doesn't affect the total zero. In most systems \\
RMS quantization noise voltage. However, bandlimiting the spectrum using a filter mixed-signal system. Tl
reduces the amount of quantization noise introduced into an input signal (this is system is linear (consl
important and the reason mixed-signal design is so powerful). In the simplest case a spectrum of the outpu
lowpass filter is used on the output of the ADC to reduce the amount of quantization wanted signal and its ali
noise introduced into the signal. We can write the amount of noise introduced into an
input signal over a range of frequencies using
IH
J
V~e.RMS = 2 V~e(f) . dl where IL <IH'S/s12 (5.6)
z,omV~
1.6mV
IL 1.2mV
Again, the factor of 2 is used to account for the contributions to VQ€,I1MS in the negative
frequency spectrum. Let's show that the sampling frequency doesn't affect the :::::
. L
quantization noise, assuming Bennett's criteria are valid. O.OmV
-O.4mV~-"-'
-o.8mv+ ...
All quantization noise is assumed to
-1,2mv+--- ...
fall in the frequency spectrum of
VLSB interest (the worst case situation).
-1, 6mV t-.. _- ~.
~2,OmV+---·~··T
J12/s +----::~/----,I Dos SOn
J" =/s12 I
Vin II
Figure 5.12 Quantization noise spectral density. i j, ] . j
Example 5.1 I
Using the SPICE simulation that was used to generate Fig. 5.IOb with sampling
frequencies of 100 and 200 MHz calculate and simulate the amount of
Doubling the sampling frequency has little effect on the output quantization noise.
Our calculated value was 1.13 mV using Eqs. (5.1) and (5.3) while the simulated
Figure 5.13 Dt
values are 1.127 mV for both 100 and 200 MHz sampling frequencies .•
Jectral density is inversely sampling frequencies. Figure 5.13 shows the simulation results. Again, note that
" we can model the ADC as we assume Bennett's criteria are valid (input busy, large compared to an LSB, etc.)
In order to see the quantization noise introduced by the ADC we have to
ply increasing the sampling remove tones (input signal and its aliases) from the DAC's output signal. As seen
an ADC introduces into an in 5.13 the way we do this is by subtracting the S/H input signal
cy spreads the quantization (remembering simply sampling a waveform doesn't introduce noise) from the
:ies (see Fig. 5.12) with a DAC's output. Using ideal components the delay through the system is nearly
ncy doesn't affect the total zero. In most systems we have to adjust the delay to match the delay through the
he spectrum using a filter mixed-signal system. This only works if the phase shift through the mixed-signal
J an input signal (this is system is linear (constant delay). A more useful technique is to look at the
full. In the simplest case a spectrum of the output signal and remove, using for example Matlab™, the
the amount of quantization wanted signal and its aliases .•
of noise introduced into an
2.0mV,~~_~_ _ _ _ _V(vq=,=,el_~""""""T"""""''''~_ _~
'f 5.f,J2 (5.6)
l.SI11V
)
;12 f
~VQe(t)
Ideal Ideal
lsity.
ADC DAC
L - _ _I I VQe(f)
I
where M is the number of points used in the FFT andhes is the resolution of the FFT (the 5.2 Signal~to-Noise I
distance between adjacent points). Note that here we are assuming that VFFr(f) is an RMS
value. If it's a peak value then the right-side of Eq. (5.7) should be divided by .fi . In the last section we devel
as a noisy circuit block wht
Figure 5.14 shows a portion of the quantization noise spectrum for the signal seen the input signal. Logically,
in Fig. 5.l3. We can estimate the RMS noise from this spectrum assuming, as seen in Fig. the mixed-signal system in
5.12, that the noise is bandlimited. While we can look at the specifics of the FFT in the
simulation to exact numbers, let's simply count the number of points seen in the plot If we apply a sine
(Fig. 5.14) and see how close we get to the exact answer of 1.13 mY. First note that -74 Vp l!2) to an ADC input,
dB is 200 !lV RMS and, as seen in Fig. 5.14, we are assuming a constant spectrum (white ADC input signal is VLSBI'i
noise). Looking at the plot there are roughly 32 points from DC to 50 MHz so
-74dB~
. -SOdB
(see Fig. 5.12)
Effective Number ofBits
9.dB
Equation (5.13) relates tho;
the input signal is a sine
·100dB+-----i----i------' ---il---~·
OMHz 20MHz 40M!lz 60MHz SNR, in most cases, will ~
When the SNR is measure
Figure 5.14 Spectrum of the quantization noise seen in Fig. 5.13.
and we assume that the largest possible amplitude sinewave is the ADC input (to
maximize the SNR), that is,
(5.12)
then (5.10) can be rewritten as
of (5.9) The effective number of bits, NelJ , is, from Eq. (5.14), 7.18 bits. Note that to
calculate VQe,RMS we either take the output of the DAC and feed it into a spectmm
tend beyond /,/2, unlike the analyzer or take the FFT of the digital output data of an ADC so that we get a plot,
in either case, similar to what is seen in Fig. 5.14 .•
174 CMOS Mixed-Signal Circuit Design Chapter 5 Data Converter S1'
Coherent Sampling
Example 5.4
Using thc ideal 8-bit ADC and DAC shown in Fig. 5.4 with a sampling frequency It's important to understand
of 100 MHz show, using SPICE, that applying a full-scale sinewave at 24 MHz to SNR that is different from tl
this configuration will cause the resulting SNR to approach the ideal value given input frequency is 10 MHz
by Eq. (5.10). coherent. The sampled poin
quantization error can then t
Let's begin by calculating SNR,deaJ' From Eq. (5.11), SNR ideaJ is roughly 50 dB, as
the data converters have 8-bit resolution. Coherent sampling (
namely spectral leakage. II
The time-domain input and output of the circuit and the corresponding DAC
consider the sinewave wit
output spectrum, are shown in Fig. 5.15. The input to the ADC in Fig. 5.4 is a 24
performed on a time-domai
MHz sinewave with a peak amplitude of 0.5 V centered on a DC voltage of 0.5 V
simplest window is the reet
(the peak-to-peak voltage of the input wavefonn is VDD or 1 V). The V Qe.RMS
is finite and set by the simu
measured with SPICE, in a separate simulation using a SIH to remove the input
the infinite duration sinewf
signal and aliased signals from the DAC's output spectrum, is 1.4 mV. The
Fig. 5.16b to obtain the wa
simulated SNR is then (0.5/.j2)/l.4 mV or 253 (48 dB), which is close to the means the resulting wave
value, 50 dB, calculated at the beginning of the example.• response (an impulse) and
wavefonn) in the frequenc
1.0V response being an impulse j
0.9\1 Fig. 5.16e. Note that th~ F
0.8V "leaks" into the frequencies
of the peak value of the Si
24 MHz input minimize these sidelobes,
ADC input
0.4V
0.]\1
0.2\1
O.W
DAC output
o.ov
(a)
_IV\; ,+,
(e)
(e) Frequency Sf
Figure 5.15 Signals from Ex. 5.4, simulating Fig. 5.4 with a 24 MHz sinewave input.
Figure 5.16 Showing h
riixed-Signal Circuit Design Chapter 5 Data Converter SNR 175
Coherent Sampling
th a sampling frequency It's important to understand that poor selection of the input frequency can result in an
sinewave at 24 MHz to SNR that is different from the ideal value calculated using Eq. (5.13). For example, if our
ch the ideal value given input frequency is 10 MHz while the clock frequency is 100 MHz then the sampling is
coherent. The sampled points repeat with every cycle of the input signal. The amount of
deal is roughly 50 dB, as quantization error can then be, repeatedly, near ± 112 LSB or much larger than VLss I!l2 .
Coherent sampling can be useful to minimize the undesired effects of the FFT,
the corresponding DAC namely spectral leakage. In order to understand what is meant by "spectral leakage,"
ADC in Fig. 5.4 is a 24 consider the sinewave with infinite duration shown in Fig. 5.16a. When an FFT is
n a DC voltage of 0.5 V performed on a time-domain waveform, the first step is to "window" the waveform. The
simplest window is the rectangular window. In a simulation the duration of the sinewave
o or 1 V). The V Qe.RMS
I/H to remove the input is finite and set by the simulation time or transient stop time, T"op' We can think of taking
;trum, is 1.4 mV. The the infinite duration sinewave of (a) and multiplying it by the rectangular waveform of
Fig. 5.16b to obtain the waveform used in the simulation, Fig. 5.l6c. This multiplication
), which is close to the
I
means the resulting waveform is the convolution of the original sinewave spectral
response (an impulse) and the frequency domain transform of the squarewave (a Sinc
waveform) in the frequency domain. The result is that instead of the sinewave spectral
response being an impulse function, as seen in Fig. 5.16d, it is a weighted Sinc waveform,
Fig. 5.16e. Note that the FFT spectral response of the sinewave in (e) is spread out or
"leaks" into the frequencies around the actual or continuous time response. The large ratio
of the peak value of the Sinc pulse to its first sidelobe is usually undesirable. Rather, to
!='I=--+-- 24 MHz input minimize these sidelobes, other windowing functions are used. One commonly used
ADC input
DAC output
Ts/op
900s 1DOns
(a) (b)
V Ts/op f
(c) (d) Frequency spectrum of (a)
Log amplitude
11 Ts/op
A
/ ~f f
Linear amplitude
f;n 11 Ts/op
(e) Frequency spectrum of (c) (t), Von Hann (Hanning) window
.24 MHz sinewave input.
Figure 5.16 Showing how spectral leakage in an FFT affects the spectrum of a waveform.
176 CMOS Mixed-Signal Circuit Design Chapter 5 Data Converter S
window is the von Hann (a.k.a. Hanning or Cosine) window represented, without the
sidelobes, in Fig. 5.16f. The response is shown on both linear and log amplitude scales
and the width of the window is 21T,ropat its base.
Selecting an input sinewave frequency, I.. ' such that Is If. is a whole number
creates a condition where an integral number of input sinewave cycles fit perfectly into
the simulation or measurement time (so the windowing function isn't important). This
results in an output spectrum that contains isolated tones (no spectral leakage). Coherent
sampling may be used to reduce the effects of spectral leakage when determining the
SNR by ensuring1.. «fs (to randomize the quantization noise).
Signal-to-Noise Plus Distortion Ratio
In a practical data converter the output spectrum contains not only quantization noise but
distortion resulting from nonlinearities and mismatch in the data converter circuitry.
Figure 5.17
When we calculate the RMS quantization noise voltage using Eq. (5.7) and nonideal
components, we are actually calculating the noise plus the distortion in the spectrum.
Measuring SNDR 1
Until this point we have only used ideal components, so that distortion in the output
DAC, or loading digital da
spectrums was absent. We can rewrite Eq. (5.7) to indicate that when it is used with a
output of an ADC. Trying
measured spectrum, both noise and distortion are included in the result as
oscilloscope, is usually a \
IM- 1 comparable to the dynami
VQe-l-D,RMS = IL V;"FT(k ./RES) (5.\5) utilize narrow band filteri]
~ 1.=0
circuit and can have dyna
The signal-lo-noise plus distortion ratio is then given by spectrum. Also note that tl
The effective number of bits is 3.7. In other words, a 5% gain error in the ADC
amplifiers results in an effective resolution of less than half of the ideal value of If a 1,000 to 1 dynamic n
8-bits.• bits is needed.
OS Mixed-Signal Circuit Design Chapter 5 Data Converter SNR 177
--
178 CMOS Mixed-Signal Circuit Design Chapter 5 Data Converter
Another way to specifY DR is as the ratio of the RMS full-scale input sinusoid resolution will be less th:
amplitude, Vpl fi , to the input sinusoid amplitude (RMS) that results in an SNDR of 0 section we relate the sar
dB. (The RMS amplitude of the input signal is equal to the RMS quantization noise plus effective number of bits.
distortion, VQe+N,RMS, when the SNDR is 0 dB.) This is nothing more than saying that the around the ideal value.
SNDR can be used to specify DR, Figure 5.19 show!
sinewave frequency is rut
Example 5.6
point (when the sinewave
Determine the DR for the ideal ADC in Ex. 5.4 using Eq. (5.19). Compare the
input signal. We assume t
result to the SNDR calculated in Ex. 5.5.
ifthe sampling clock freg
Using Eq, (5.19), the DR is 48.16 dB (the ideal value). The SNDR calculated in is 50 ps I1T s ), then tt
Ex. 5.5 was 24 dB. Clearly, the SNDR is a better indication of DR than is the (where parts per million [1
value obtained using Eq. (5.19) .•
Specifying SNR and SNDR
The SNR and the SNDR are usually specified as a function of input sinewave amplitude
at a fixed frequency, Fig. 5.18. The x-axis in Fig. 5.18 is normalized so that an input
sinewave with a peak-to-peak amplitude of VREF+ VREF- corresponds to 0 dB. We might
be wondering how we differentiate between SNR and SNDR as both, up to this point,
have been calculated in the same way (Eqs. [5.71 and [5.15]). We continue to calculate
SNDR using a data converter output spectrum, remembering to zero out the desired tones
and images, and Eq. (5.15) as was done in Ex. 5.5. When we calculate the SNR, we
follow the same procedure except that now we also zero out any spikes or ~purs (spurious
responses) in the spectrum that are "sticking up" above the noise floor in the spectrum.
These spikes come from imperfections in the data converter and result in distortion in the
output waveform. Note in Fig. 5.18 how the SNR and the SNDR coincide until the input
signal amplitude gets reasonably large (so the distortion tones increase in amplitude
above the quantization noise).
ill
"0
.€'" Figure ~
B 60
if;
'6
'" 40
S full-scale input sinusoid resolution will be less than ideal because of a problem known as aperture jitter. In this
~t results in an SNDR of 0 section we relate the sampling clock's jitter to the data converters SNR and thus the
v1.S quantization noise plus effective number of bits. Clock jitter is the variation in the period of the doek signal
'I more than saying that the around the ideal value.
Figure 5.19 shows the basic problem. In this figure we have assumed the input
sinewave frequency is running at the Nyquist frequency/" (= Is 12 ) so that the sampling
point (when the sinewave crosses zero in this figure) is seeing the fastest transition in the
1- (5.19). Compare the
input signal. We assume the peak amount of jitter in the clock signal is 6.Ts • For example,
if the sampling dock frequency is 100 MHz (T, IOns) and the peak-to-peak clock jitter
le SNDR calculated in is 50 ps 6.Ts ), then the specification of the sampling clock stability is 5,000 ppm
tion of DR than is the (where parts per million [ppm] = 10-6 and 6.Ts (stability, ppm)· (l(fs)).
~
Aperture
uncertainty/jitter
SNDR Figure 5.19 Data converter input signal and clock jitter.
The slew rate of the signal in Fig. 5.19, at the sampling point (when the clock
signal transitions high), is given by
~I
(5.20)
We can relate the uncertainty in the sampling instant, 6.Ts ' to the uncertainty in the
sampled voltage, 6.V" using
l. converter.
6.Vs .,
-T = TCj,Vp, or 6.Vs (5.21)
6. s
Ifwe require the uncertainty in the sampled voltage, 6.Vs' to be at most 0.5 LSB == (
1 system with "real world" -VREF_)!2N+land we remember Vp = (VREF+ VREF-)!2, then our maximum allowable
er of bits) limited by the peak-to-peak clock jitter can be determined for a particular data converter using
· signal used isn't ideal, the
180 CMOS Mixed-Signal Circuit Design Chapter 5 Data Conver
Example 5.S
For an ideal 8-bit
converter with 100
assuming the ADC'>
We can write the ,
peak-to-peak jitter a
or
Example 5.7 N
Suppose a phase-locked loop (PLL) is used to generate a clock signal for a data The effective numbe
converter. If the resolution of the data converter is 10 bits and the frequency of the
sampling clock coming from the PLL is 900 MHz, then specify the maximum Using Oversampling to J
jitter allowed in the output of the PLL. Assume that the maximum sampling error Suppose we limit the m.
allowed is 0.5 LSB and that the data converter is sampling a sinewave with a sampling frequency is re
frequency of 100 MHz.
Because the sinewave being sampled has a frequency below the Nyquist value,
Eq. (5.22) cannot be used directly. Instead, after reviewing the derivation of this
ill other words, we are f
equation, we can rewrite it in tenus of any input signal frequencY''!;n' as
(we are oversampling 1
1 1 (5.24) frequency ([). where K
I1Ts S N
2
'-
11:
. f,II!
2
signa\. Notice that Eq. (
noting that when'!;n 12, Eq. (5.24) reduces to Eq. (5.22), Using Eq. (5.24) frequency and data cony,
with the numbers in this problem results in a peak-to-peak jitter of 1.56 ps! The or the sampling frequenc
reader familiar with PLL design will recognize that this is a very challenging For a given maxi
requirement when designing a PLL (that is, to design a PLL with an output stability of the oscillator'
frequency of 900 MHz and an output jitter less than 1.56 ps) .•
St,
We're now in a position to answer how, given the peak-to-peak clock jitter, the
SNR ofa data converter is degraded from the ideal value (given in Eq. [5.13]) when the Ifwe were sampling at 1
input sampling clock isn't ideal. Rewriting Eq. (5.24) and assuming sampling clock would be
MHz, with lOps jitter w
1
I1Ts ;;:. N . 2 I' (5.25) increase the sampling cle
2 11: 'Jin
atleast 10,000 ppm (thq
xed-Signal Circuit Design Chapter 5 Data Converter SNR 181
Example 5.8
For an ideal 8-bit ADC clocked at 100 MHz, determine the SNR of the data
113.4
converter with 100 ps of peak-to-peak jitter in the input sampling clock, t1Ts ,
assuming the ADC's input is a full-scale sinewave at 25 MHz.
28.3
7.1
We can write the number of bits lost by solving Eq. (5.26) as a
function of
peak-to-peak jitter as
1.77
0.44 NLoss = N + 3.32· log (2n ·Jin . t1Ts) assuming t1Ts ;?: .-LN . 2 +. (5.28)
2 n 'jin
0.11
or
NLoss = 8 + 3.32 ·10g(2rc· 25MEG·IOOps) = 2 bits
dock signal for a data The effective number of bits, NeD' is then 5.5 and the SNR is 34.87 dB. •
ld the frequency of the
specify the maximum Using Oversampling to Reduce Sampling Clock Jitter Stability Requirements
.ximum sampling error Suppose we limit the maximum input frequency coming into an ADC to In , such that the
ng a sinewave with a sampling frequency is related to the maximum ADC input frequency by
..
L PLL with an output
~n in Eq. [5.13]) when the. If we were sampling at I MHz and the stability required was 10 ppm, then the jitter in the
ling sampling clock would be at most 10 ps, peak-to-peak. Increasing the sampling rate to 100
MHz, with lOps jitter would require an oscillator stability of 1,000 ppm. If we were to
(5.25) increase the sampling clock frequency to I GHz, then the stability of the clock would be
at least 10,000 ppm (the period is 1 ns and the jitter is lOps or 1% [10,000 ppmD.
182 CMOS Mixed-Signal Circuit Design Chapter 5 Data Converter SN
To-'oo 0 --Tol2
We need to point out that the effects of clock jitter are possible even if the clock is
The spectrum of the averag
perfectly stable because of the clock's finite transition times (rise and fall times). If the
transform of the autocorrelat
rise time of the clock signal in Fig. 5.19 is finite, say 50 ps, then the same derivations and
function (PSD) and is given 1
discussions concerning jitter in this section can be applied to determine how the SNR of
the data converter is affected. We would assume the aperture window is a function of the
transition times of the sampling clock signaL The slower the transition times, the larger
the sampling uncertainty. In any practical data converter the SNR, and thus the effective
number of bits, will be reduced because of the clock jitter and finite transition times as The power spectral density fi
the input signal frequency increases.
5.2.2 A Tool: The Spectral Density
The observant reader may have noticed, in the last section, that we only discussed the This is simply two impulses
of VZl4 (V2/Hz). The total a
peak-to-peak jitter, l1T" and how it effects the data converter's performance. It is very
useful, in many situations, to also have an idea how the spectrum or spectral
Mixed-Signal Circuit Design Chapter 5 Data Converter SNR 183
characteristics of the data converter's output change as a function of the random sampling
d at 44.1 kHz could have jitter or a random variable such as noise. In this section we discuss tools useful in
; uncertainty to 0.5 LSB. describing the spectrum of a random signal.
these circumstances was The Spectral Density ojDeterministic Signals: An Overview
lity requirements of the
$ed the sampling clock Consider the simple sinewave signal of the form
Vin(t) -= Vp sin2rgtnt (units, Y) (5.33)
Ie sampling frequency, is This signal is termed "deterministic" because the signal has a well-defined shape whether
of the input signal). We it is continuous or sampled. We can find the average power of this signal, as a function of
s at or below 22.05 kHz time Riit), using the autocorrelation junction (ACF) for continuous signals given by
define the oversampling
R,n(t) == lim
To-w"
i Tol2
or knowing
; given by
sinA . sin B t[ cos(A - B) cos(A + B)] (5.36)
(5.31 )
we can write
lsing oversampling, then we V2
Vi .sin 2rgtnt· sin 2nJinCt + t) -= f[cos2nJint cos 2rgtn{t + 2t)] (5.37)
(5.32) When we integrate this result, the term cos [2rif,n(t + 2t)] represents a sinusoid with a
frequency of 4rgtn (remembering our integration variable is t) and a phase shift of 2nJint.
K is the oversampling ratio. Over a long period of time this term averages to zero. Therefore, we can write the average
oversampling reduces the value ofEq. (5.33) as a function oftime (the autocorrelation function)
Rin(t) lim
To->m
i Toll
0 -To/2
V2
J 2' cos 2rgtnt . cit -= f .cos 2rgtnt (units, y2) (5.38)
that we only discussed the This is simply two impUlses in the frequency spectrum located at ± hn with an amplitude
er's performance. It is very of V;/4 (y2/Hz). The total average power of this signal is given by
the spectrum or spectral
184 CMOS Mixed-Signal Circuit Design Chapter 5 Data Converter ~
"" cc
PAVG = JPin(f)' dJ J
2· Pin(f)' dJ (units, V /Q or watts)
o
2
(5.41)
assuming a l-Q (normalized) load, which, for Eq. (5.33), is Vp2 (V2). The PSD is the Fourier tran
The voltage spectral density, with units of VI jHz , is simply the square root of
Eq. (5.39) (that is, the square root of the PSD Jpin(f) D. The root mean square
(R.\1S) voltage of a signal is given by The RMS value of the sam1
through an ideal reconstru
VRMS=JPAVG = 21P in (f).dJ= !21(voltagespectraldensity)2.dJ (5.42) again, Vpl Ii .
The PSD of
o , 0 Vi14 at frequencies of ± Is.·
The RMS value of Eq. (5.33) is simply, as one would expect for a sinewave, Vpl Ii. The Spectral Density ojRa1
Note the similarity between Eq. (5.42) and Eq. (5.7). Let's use our jitter discussi<
of a random signal. We'll d
Example 5.10
random variable that falls b
Determine the ACF, PSD, average power, and RMS value of a signal Vet) made
in the region (just as was a
up of three sine waves with peak amplitudes of V1 ' V2 , and V3 with frequencies of
when calculating the RMS
ft, J;, and J;. assume the jitter has a G
Using Eqs. (5.34) and (5.38), the ACF is practical and realistic situati
R(t) =
V2
T cos 2rrJI t +
V2
-t V2
cos 2rr/2t+ ; cos 2nJ3t (units, V2)
Consider the repres
figure shows the ideal po
represented on the probabi
The PSD (positive frequencies) is determined using Eqs. (5.39) and (5.40)
rising edge of the clock, tr;
P(f)
vf vi
4,o(f-JI)+4·o(j-J2)+4·U( -j3)(umts, V 1Hz)
vi s: J . 2
PDF as shown. We are aSSl
the rising edge of the clock
probability of being in the
The average power, using Eq. (5.41), is
boundary (as shown in traCt
PAVG
vl + vi + vi (units, watts)
2 Trace
Finally, the RMS value of the signal is given by Ideal cI
(units, V) 2~t{
Note that if we added phase shifts to OUf signals the results would be the same; the 3~t(
phase shift doesn't change the signal's average value, so we get the same results
whether sines or cosines are used in our original spectrum .•
4 _ _ _ _-!~al
Next, suppose that the sinewave specified by Eq. (5.33) is sampled at a rate ofls
Vin(nT.,) = Vp sin(2nf;n' nTs) (5.43) 5~ Edge cl
2
(5.41) Rin(nTs) = fV cos 2n/in . nTs (units, y2) (5.45)
is simply the square root of Pin(f) = 4;s k"'foo [8(f-.fin + kfs) + 8(f+/in + kfs)] (5.46)
) D. The root mean square
The RMS value of the sampled sinewave, Eq. (5.43), assuming we have passed the signal
through an ideal reconstruction filter (RCF) with a bandwidth of fs/2, is simply, once
~ctral density)2 . df (5.42) again, Vpl fi . The PSD of the signal, after passing through the RCF, has an amplitude of
Vp4 at frequencies of±/;".
lect for a sinewave, Vpl fi . The Spectral Density ofRandom Signals: An Overview
Let's use our jitter discussion of the last section to illustrate how to look at the spectrum
of a random signal. We'll do this in two parts: (1) we'll begin by assuming the jitter is a
random variable that falls between two limits and has equal probability of lying anywhere
ue of a signal Vet) made
in the region (just as was assumed for the quantization error probability density function
I1d V3 with frequencies of
when calculating the RMS quantization noise voltage in the last chapter), and (2) then
assume the jitter has a Gaussian distribution around some average value (the more
practical and realistic situation) and determine how the output of the ADC is affected.
Consider the representations of clock jitter shown in Fig. 5.20. Trace I in this
figure shows the ideal position of the rising edge of a clock signal. This point is
represented on the probability density function (PDF), p(t) , at time zero. On the next
(5.39) and (5.40)
rising edge of the clock, trace 2, the edge is a little too early and is represented on the
PDF as shown. We are assuming, probably incorrectly for most practical situations, that
f-/J)(units, y2/Hz)
the rising edge of the clock is falling within the peak-to-peak boundaries with the equal
probability of being in the correct position (as shown in trace 1) or at the edge of a
boundary (as shown in trace 4). We also know that the area under the PDF curve in Fig.
(units, watts)
Trace
: Ideal clock edge position p(t)
Probability density function, PDF
(units, Y) 2~tooearlY 5 1 3 4
Figure 5.20 Clock jitter assuming the edge falls with the same probability
anywhere within the peak-to-peak limits.
186 CMOS Mixed-Signal Circuit Design Chapter 5 Data Converte
5.20 must equal unity, and the average value (also known as the mean or the expected
value and denoted by < y > or y) of a PDF is given by
..
00 RMS JItter = (J '" 6/:iTs
Example 5.11
Detennine the average value of the jitter with the PDF shown in Fig. 5.20.
We can use Eq. (5.47) to detennine the average value of any PDF. Applying this
equation to the PDF shown in Fig. 5.20 results in
flT.,!2
This somewhat obvious result means that the average position of the clock rising Figure 5.21
edge is the ideal position indicated by trace 1 in Fig. 5.20. Any PDF that is
symmetrical about some center point will have an average equal to the center sampling clock affects a
point. • (5.20), (5.21), and (5.22).
given time, as
The variance of the PDF is defined as the average of the square of the signal's
departure fTOm its average value. For a random signal this can be written as
(x"
f (y_y)2 ·p(y)·dy (5.48)
where 8Ts(t) is a randon
time. (The variable 8Ts(t
and the expected transiti(
where (J is the standard deviation of the PDF (the square root of Eq. [5.48]). For our value of 8Ts(t) is /:iTs, v
purposes, in this book, we can think of variance as the average power of a random probability distribution fi.
(voltage) signal and the standard deviation as the RMS value of the signal (see Eqs.
[5.41] and [5.42]). Example random signals include the time difference between the Rewriting Eq. (5.
actual edge of a clock and the ideal edge location (jitter), the voltage difference between written as
the input of an ADC and the ADC's reconstructed output (quantization noise), and the
random fluctuations of electrons due to thennal motion in a resistor (thennal noise).
Nn in Fig. 5.20.
:my PDF. Applying this
o cr 2cr Time
. dto:=O
ition of the clock rising Figure 5.21 Sarnplingjitter with a Gaussian probability distribution.
5.20. Any PDF that is
1ge equal to the center sampling clock affects an ADC output spectrum with a single-tone input. Using Eqs.
(5.20), (5.21), and (5.22), we can write the sampling error voltage (review Fig. 5.19), at a
given time, as
f the square of the signal's
be written as !J. Vs(t) 0:= oTs(t)· Vp ·2rr.Jin . cos 2rr.Jint (5.49)
where oTs(t) is a random variable indicating the jitter in the sampling clock at a given
dy (5.48)
time. (The variable oTs(t) is the time difference between the actual clock transition time
and the expected transition times that are spaced by Ts [see Fig. 5.20].) The peak-to-peak
oot of Eq. [5.48]). For our value of oTs(t) is !J.Ts, while its average value is zero. Again, we assume that the jitter
,rerage power of a random probability distribution function is Gaussian, as seen in Fig. 5.21.
lue of the signal (see Eqs.
me difference between the Rewriting Eq. (5.49) using a discrete time step nT" the sampling error can be
voltage difference between written as
luantization noise), and the Sampling error amplitude Carrier term
sistor (thermal noise).
, .
!J. Vs(nTs) = oTs{nTs)' Vp ·2rr.Jin· COS 2rr.JinnTs (5.50)
We're interested in the spectrum of this error signal as it will add to our RMS
as a probability density quantization noise plus distortion voltage, effectively lowering the data converter's
SNDR. Notice that the spectrum of Eq. (5.50) will have aliased components (and so will
the sampled signal) so we need to filter out these components above Is 12 (with the
reconstruction filter.) Also note that multiplying the sampling error by the cosine term in
Eq. (5.50) simply shifts the error spectrum to a frequency 1;•. The cosine terms act like a
.) = __
(!J.T )2 ,
s _ (seconds")
carrier in an amplitude-modulated signal. This is illustrated in Fig. 5.22.
12 12
Example 5.13
Repeat Ex. 5.8 assuming the clock jitter has a Gaussian PDF.
seconds)
In this example the peak amplitude of the input signal, Vp ' is 0.5 V, the input
frequency, 1;., is 25 MHz, and the peak-to-peak jitter is 100 ps. The average
;k rising edge. Note the power in the sampling error amplitude spectrum is
2
= cr 2 • .....
-,--,,---:-=-....,;.... = (\. !J.Ts
6
\) 2 • (Vp . 2rr.Jin)
2 (5.51)
if we assume the jitter has a . PAVGJiller
Sampling error amplitude spectrum Data converter output spectral content resulting from jitter
~4) l.--~ o f o f
)
o
Figure 5.22 Modulating sampling error with an input sinewave frequency. Figure 5.23 Sampling an
Cf
2 -'..-(V-,,-P_'2-::-n..::.. if,:. . .,;11):-
•- 2 2
o
JPjitter(f) . d/ (5.52)
If the spectrum of the phase noise due to jitter is narrow, as seen in Fig. 5.22, then the
spectral density of the sampling error, Pjit/.,.(/), is concentrated around the frequency of
the input sinusoid. However, if we assume the phase noise spectrum is white and evenly
distributed throughout the base spectrum (so that we integrate Eq. [5.52] from DC to//2), o
we can write
2
Cf2 (Vp' 2rrjin)
Pjiller(f) = /s . 2 (5.53)
Fi~
The power spectral density of the sampling error voltage, assuming even distribution of
the noise throughout the base spectrum, is shown in Fig. 5.23.
ixed-Signal Circuit Design Chapter 5 Data Converter S:t-<R 189
)
f o
lewave frequency. Figure 5.23 Sampling amplitude error PSD assuming sampling error spectrum is white.
858 x 10-6 y2
It's important to note that we have been discussing clock signals that are square waves
(that is, have odd order harmonics) and so discussing jitter (a time-domain term) is,
generally, more appropriate than discussing phase noise (a frequency domain term).
mY. The quantization
However, because the terms are both widely used to indicate the same, basic, effect (a
variation in the period of a periodic waveform), we will briefly discuss phase noise
Y specification from measured oscillator data.
Consider the representation of a measured oscillator spectrum (power spectral
III effects is then given density) shown in Fig. 5.24. In general, oscillator noise is specified in terms of the carrier
voltage (or power) with units of dBc (decibels with respect to the carrier). The ratio of the
power of the fundamental (called the carrier or sampling clock) at is is taken to the noise
power in a bandwidth at some offset from the fundamental
IO.log(V Z)
r , 10· log (V2IHz)
B
Phase noise, dBc/Hz 10 . log r11 Posc(f) . df ] - 10· log P(ls ) (5.54)
) 7.53. Note that this is LrLJ
1/
(5.53) fm
Examine the two-path combination of ADCs and DACs shown in Fig. 5.25. The top ADC
and DAC are the path we had back in Fig. 5.4 clocked at 100 MHz. The bottom path is a
mirror image of the top except that its clock signal is inverted. The two resistors are used Further note that if we w(;
to sum (or more correctly average) the outputs of the DACs into a single output. This filter out part of the quanti
configuration effectively samples the input at 200 MHz (200 Msamples/s [2 .];]). this section to improve SN
Example 5.14
Repeat Ex. 5.1 but use
An,log ~~ 8 Ideal Analog seen in Fig. 5.25.
v in_~ Ideal I
:l
1
~~ Id~l
Voal
Digital
results in adding, using
@2f,
8 Ideal I I
Figure 5.26 shows the quantization noise PSD spectrums when the single path z.omV-r-
ADC/DAC is clocked atfs and then at 2fs. Notice how the quantization noise is spread out 1.5mVj······
over a wider frequency range when a higher sampling frequency is used. As seen in Ex.
5.1 the sampling frequency doesn't affect the RMS value of quantization noise. The area
of the spectrums in Fig. 5.26 is equal to VZsB/12 (negative frequencies are not shown) for
both sampling frequencies.
1. vIsB
@2f, 2 12];
ViSB
12]; +---------------------, ~ J
f Figure 5.27 Sh
Figure 5.26 Quantization noise power spectral density for two sampling frequencies. An Important Note
For averaging to effective]:;
Now if we were to compare the quantization noise added to an input signal using be linear to the final outpu1
a single path clocked at 21s to a system using two paths clocked at fs we would find that Fig. 5.28. In the ideal sitm
the two path system added less noise. In order to understand why, notice that when we that falls exactly in betwee
take two random uncorrelated variables, the quantization noise added to the input signal converter has a nonlinearit:;
from each path in Fig. 5.25, and average (sum the power from each) them we get is much different from t1:
o 2 contains a missing code (aT
VQe,RMSl + V Qe,RMS2 times of 1-LSB results in t
vQe,RMS.2-path
2
2
(5.55)
converter is nonmonotonic
Signal Circuit Design Chapter 5 Data Converter SNR 191
Example 5.14
Repeat Ex. 5.1 but use, to increase the sampling frequency, the two path topology
Analog seen in Fig. 5.25.
As discussed in Ex. 5.1, the RMS value of the quantization noise added to the
VOlil = VoutQ +2 Volilb input signal is 1.13 mY. Using two paths to increase the sampling frequency
results in adding, using Eq. (5.56),
@2/s
VQe,RMS= ~. 3.9},mV =0.797mV
v2 12
to the input signal. Figure 5.27 shows the simulation results (the quantization
noise added to the input signal when two paths are used). The simulated VQe,RMS is
. nOise. 0.787 mV .•
-1.0mV
2
LSB
-1.5mV ,,
, , ,, ,
.., , ,
,,
, ,
............ '. --- _._ ..... -. - -- -- _. - ...... ---- - -- -- --'- -... " .'- .. --- -
2/s
-2.0mV ~.-
,
,
,, .
,
,
,
,
,,
'-.
,
,
,,
-.~
•
,
. , ,
,
_.'-
I
I
,,
,
,
,
,
,
,
-2.5mV ······-r······-f-----··r-····
I , ,
-r-----·~-····--:-------f-------~---·· -~-"-----
, •
-3.0mV+--i-'--'i------i-'--'-i-. --+--+---j-'--i--'--1-----I
o.o~s O.2~s O.4~s O.6~s O.8~s 1.0~s 1.2~s 1.4~s 1.6~s 1.8~s 2.0~s
/ Figure 5.27 Simulated quantization noise using two-paths, see Ex. 5.14.
-,." -
192 CMOS Mixed-Signal Circuit Design Chapter 5 Data Converte
~AVeraged
Example 5.15
For the topology se
~points From this spectrum
Averaged
- Points input signal. Compa!
To begin let's write (
• Indicates output of the data converter
increase in its output) then the averaged value is meaningless. Finally, note that an input
DC value (a digital code that isn't changing for the DAC, or an analog voltage that isn't This PSD is plotted
changing for the ADC) or a value that isn't "busy" (not changing by at least I LSB in
between sampling instances) will not benefit from averaging.
5.3.1 Using Averaging to Improve SNR
2
VLS,
The averaging topology shown in Fig. 5.25 is not practical in most situations (one notable
exception is the parallel combination of noise-shaping converters discussed later in the 3.1
book). The silicon area required to implement the extra ADC and DAC generally costs
more than is gained by the reduction in quantization noise. Figure 5.29 shows how we can
add a digital averaging filter (in this figure averaging two ADC's outputs, or K = 2) to the
output of the ADC to reduce quantization noise. By lowpass filtering the quantization
Figure 5.30 QuaT
noise PSD seen in Fig. 5.26, we reduce the amount of noise added to our signal.
Unfortunately, this lowpass filtering also limits the range of allowable (wanted) signal To determine th,
frequencies. Notice that both the ADC and digital averaging filter are clocked at a rate of
Is. The averaging filter seen in Fig. 5.29 is described using Figs. 1.16 and 1.17.
We might, at this point, assume that we can use a low-resolution ADC, say 6-bits,
with a significant amount of averaging (K » 2, see the lowpass Sine-shaped filters seen or
in Ch. 4) to attain large resolutions (again the ADC must be linear). Assuming the input
to the ADC is busy and we place restrictions on the bandwidth of the signals coming into
the ADC then we can increase the resolution by averaging. We have to place restrictions
V;n-,
~ Ideal
ADC
'------'
~-+/-~~r-+/-
Lowpass filter
Digital
after we review the
the gain of the filte
Knowing this we ca
on the bandwidth of the signal coming into the ADC because, unlike Fig. 5.25, we haven't
nput signal increased the sampling rate of the signals. Therefore, the amplitude of the power spectral
density seen in Fig. 5.26 remains unchanged. For an averaging of two, we would have to
Ideal code level limit our desired input signal bandwidth to is 14, Fig. 1.17. If this wasn't the case, then an
input sinewave at is 12 would average to zero.
msfer characteristics
I + Z-l = 2 ·1 cos ni 1
It would appear that our RMS quantization noise has actually increased! However,
Digital after we review the transfer characteristics of our filter seen in Fig. 1.17 we see
the gain of the filter at low frequencies (where our desired signal resides) is two.
--+ Knowing this we can write the SNR as
~r
2V lJi V lJi
SNR = 20 log P j6 = 20 log P [i4
VLsBI 6 VLsBI 24
luantization noise.
. ..- .........
E -0~
t
Averag
ADC output 2
SNRideal = 6.02N + 1.76 + 10 10gK (5.58) Averaging two points (id
where N is the number of bits (the resolution) of the data converter whose output is being (a)
averaged. Using no averaging, that is K = I, results in this equation simplifying to Eq.
(5.13). Averaging two samples causes the SNR,deal to increase by 3 dB or the effective
resolution of the data converter to increase by 0.5 bits. The increase in resolution due to
averaging can be written as
1010gK
Increased resolution,}lhnc (5.59) black dot) is shifted dow]
6.02
offset from its ideal posi
Figure 5.31 shows how averaging the output of a data converter changes the effective we could use any numbf
resolution of the data converter. Again, note that the increase in resolution is based on the better than the desired fir
following assumptions: a busy input signal, the input signal is bandlimited, and the data
converter is linear to the final resolution (data converter resolution, N, + improvement in The number of
resolution, NIne> coming out of the averaging circuit. improvement in resolutio
(the number of bits comi)
Improvement in resolution,NIne
(bits added)
(5.57)
ADC output I
.}
•
i:.~Averaged point
~ Ideal level It 0.5 LSB
E
/DJ>.L
1 LSB ~.~ ~ ~ 1 LSB ~tf· ~
1 LSBC
~ Averaged point 0
~can "'Tite the ideal SNR
ed) lowpass filter as
(5.58)
t
ADC output 2
(c)
rter whose output is being (a) Averaging two points (nonideal)
uation simplifying to Eq. offset of 0.5
by 3 dB or the effective (b)
Tease in resolution due to
Figure 5.32 Linearity requirements when averaging.
K
(5.59) blaek dot} is shifted downwards by 0.5 LSBs and so the averaged point shows a 0.25 LSB
offset from its ideal position. While we used a single LSB difference to show averaging,
rter changes the effective we could use any number of LSBs to show that the ADC accuracy must be equal to or
resolution is based on the better than the desired final digital filter output accuracy.
bandlimited, and the data
ion, N, + improvement in The number of bits in the ADC (its resolution) N, and the number of bits
improvement in resolution after filtering, N/ne ' are used with the final, total number of bits
(the number of bits eoming out ofthe digital filter) to give
N Final = N + N fnc (5.60)
= 6.02(N+Nfnc ) + 1.76 The ADC output should ideally ehange in increments of the exact LSB voltage. In reality,
the changes will be different from the ideal output levels (as just diseussed). In order to
aehieve an increase in the number of final bits, the output of the ADC must be aceurate
(its aetuallevels must be spaced from the ideal levels) to within
+ V Rli"F+ VREF- 1
± (0.5 LSB) . - N (5.61 )
~ 2NFi"ul+ 1 2 J~
where no averaging (NIne 0 and K = 1) means the ADC is at least 0.5 LSBs accurate.
~ This is a significant limitation when using averaging to increase the resolution of an
mber of points averaged
ADC. This is especially true when a resolution greater than 10 bits is desired with INL
and DNL less than ±0.5 LSBs. In the next seetion, and in the next chapter, we will look at
er resolution. feedbaek topologies that may relax the accuracy requirements plaeed on the ADC and
allow averaging to more effeetively remove quantization noise.
5.3.3 Adding a Noise Dither
.2. In part (a) we show the
ltputs spaced by one LSB Our assumption, when diseussing the benefits of averaging or calculating the spectral
.) are located on the ideal density of the quantization noise, falls apart for DC or slow-moving signals (the ADC
lese levels (and hence our input is not "busy"). In order to help with this problem consider adding a noise signal to
1 where the ADC outputs the ADC input that has a frequency content that falls within the range
Following this offset, the
ltput of the ADC (the top is < f
2K~ 2
(5.62)
196 CMOS Mixed-Signal Circuit Design Chapter 5 Data Converte
so that it can be filtered out with the averaging filter (see Fig, 4.10). This noise is often amplitude be greater th:
called dither (a state of indecision or agitation) because it helps to randomize the spectral allowable range of input
content of the quantization noise, making it white. from dithering since the,
Fignre 5.33 shows the basic idea. In part (a) a DC signal is applied to the ADC Before we discus
that falls halfWay between two ADC transition codes spaced apart by 1 LSB. The output (a Gaussian PDF) for thl
code of the ADC remains unchanged with time. In part (b) a noise signal is added to the and DC input shown in I
DC input which has two benefits: (I) the quantization noise (the difference between the long time, we get the av
input signal and the reconstructed ADC output code) changes with time, and (2) the would also mean that we
output of the ADC has some variation which makes it "possible to determine the DC average enough ADC ou
voltage after averaging, generally a good idea to
Finally, it's important th
dither signal must averag
ADC Output code N offset (the known DC ofj
in the data convcrter's (ae
co
til DC Input signal
....J
Figure 5.33 (a) DC input signal and (b) DC input signal with dither added.
We can add the noise signal to our desired input signal with a circuit similar to
that shown in Fig. 5.34. Simple resistors add and reduce the noise signal applied to the
ADC input. The noise signal source is, most easily, derived from some sort of
asynchronous logic circuit and has a peak amplitude (before reduction) of VDD I V in
this book). In this figure note that we have indicated that the dither signal amplitude
Figure 5.35 Input to
should be approximately 0.5 LSB RMS (remembering the signal is, ideally, random and
band limited as specified by Eq. [5.62]). This number, 0.5 LSB RMS, is subjective, and no
exact rules as to its selection can be given othcr than the desirc that the peak-to-peak An example of ar
Thc outputs of the rows
and fight against each ot
Input with dither Input with dither
~
Input
,.~ Y I:"l ADC
Input 50
-~-r~----L ADC
I I
~~
r Dlth;l
It,ooo
Approximately 0.5 LSB RMS dither
--~
Dither
I nois;' I 0 toiY generating
I s.'ource i Id!git~l
-- l:lrcUit
Block diagram
Circuit implementation
~. 4.10). This noise is often amplitude be greater than I LSB. One disadvantage of adding the dither is that the
)s to randomize the spectral allowable range of input signals shrinks. A DC signal at VDD - 1 LSB will not benefit
from dithering since the ADC will be at its full-scale output.
gnal is applied to the ADC Before we discuss the implementation of a dither source, consider one possibility
apart by I LSB. The output (a Gaussian PDF) for the desired probability density function (PDF) of the dither signal
noise signal is added to the and DC input shown in Fig. 5.35 (the input to the ADC). [fwe average this signal over a
(the difference between the long time, we get the average or DC input signal since the dither averages to zero. This
ges with time, and (2) the would also mean that we can have some dither spectral content below f/2K as long as we
sible to determine the DC average enough ADC output samples to make its contribution to the SNDR small. It is
generally a good idea to use Eq. (5.62) as a gnide for allowable dither spectral content.
Finally, it's important that any dither signal we generate has a symmetrical PDF (the
dither signal must average to VDDI2 before amplitude reduction). If not, an unknown DC
offset (the known DC offset is the VDD/2 attenuated by the resistive divider in Fig. 5.34)
I~
in the data converter's (actually the filter's) output will result.
p(t)
Probability density function, PDF
time
)0 RMS dither = (J = 0.5 LSB
pCt) ~ .exp [
a,,2n:
5,000
Dither
Ot IV generating
digital
circuit
Circuit implementation
;igna1.
}'igure 5.36 One possible implementation of a dither circuit.
198 CMOS Mixed-Signal Circuit Design Chapter 5 Data Converter S
other than the nonnal logic levels of VDD and ground for significant amounts of time.
IH(f) I
The dither signal can be made more random by adding more rows of inverters. The A
challenge to this design is setting the number of inverters used in each row so that the
spectral content falls within the desired range (which may require a large number of
inverters) and keeping the output of the dither circuit uncorrelated with the sampling
clock. Other techniques for generating random noise, such as using linear serial feedback
registers, can be found in most books covering communication systems.
5.3.4 Jitter
/n =/s
We can apply the averaging discussion just developed directly to the jitter discussion
(a)
presented earlier in the chapter and answer the question, "How does averaging effect the
sampling amplitude error power (resulting fromjitter) in a data conversion system?" Ifwe
Figure 5.37 (a) AAF requiI
assume that the jitter has a Gaussian PDF, then the average power in the sampling error
r
amplitude, from Ex. 5.13, is
significance of this will be
In =fsl2 f In =fsl2 f
;tly to the jitter discussion
N does averaging effect the
(a) (b)
L conversion system?" Ifwe
,ower in the sampling error Figure 5.37 (a) AAF requirements without averaging, and (b) AAF requirements with averaging.
significance of this will be easier to see as the number of points averaged increases and
our averaging filter's response gets sharper with more attenuation (see Fig. 4.10). Of
(5.63) course, the penalty for the relaxed requirements of the AAF is reduced signal bandwidth
for a fixed sampling frequency.
It may be helpful to rewrite
5.4 Using Feedback to Improve SNR
(5.64) By averaging the outputs of an ADC, or interpolating between inputs of a DAC, the
effective data converter resolution can be increased. As specified by Eq. (5.58), every
doubling in (octave increase in) K (where K is the number of points averaged) results in a
0.5-bit increase in effective resolution. An effective ADC resolution increase of 6-bits
requires averaging 4,096 samples. If a 1 MHz signal bandwidth is of interest, our
(5.65)
sampling clock frequencY,h, will have to be 8.192 GHz!
------------
.J
Vout(f) Out
(to digital
filtering.)
In Vin(f)
feed back, and output, the average value of the input signal. This signal can be filtered
(averaged) to reduce the accuracy required of the analog circuit components. Fi~
Noise-shaping modulators effectively output the average of the input signal over time. In
a noise-shaping data converter the averaging and decimating filter, as discussed earlier, is
connected to the output of the modulator. Because of the averaging used in noise-shaping In a noise-shapir
modulators, the analog components, in the forward path of Fig. 5.38, require less signal bandwidth so that
accuracy. However, the DAC's output, in the feedback path (which is subtracted from the on the other hand, will f
input), doesn't experience the averaging so, once again, the DAC must be linear to the signal spectrum passe:
final desired resolution of the data converter. DAC linearity concerns have led to the use quantization noise spect
of a single-bit DAC (an inverter), in many noise-shaping data converter applications. The filter or analog compon
one-bit DAC is inherently linear. (Two output points determine a line!) Because of the DAC in the feedback pa
relaxed requirements placed on the analog circuit components, we will concentrate the an integrator, the quant
next chapters, in detail, on noise-shaping topologies for both ADCs and DACs. Notice removed with the avera
that both predictive and noise-shaping modulators utilize oversampling. modulator does not redt
pushes the noise to frequ
In order to understand these statements in more detail, let's use the additive
quantization noise model for the ADC developed in this chapter, Fig. 5.2. Figure 5.39 ADDITIONAL READ]
shows Fig. 5.38 redrawn using this model where the quantization noise is represented in [1] R. J. van de Plas
the frequency domain by VQeCf). We can relate the inputs (the wanted input signal and Converters, SpriI
the unwanted quantization noise) to the output of the feedback modulator by
Signal transfer function, STF(f)
[2] Engineering Sta
Noise transfer function, NTF(f)
~ Devices), Newne
A(f)
(5.66) [3] M. Gustavsson,
I + A (f) . B(f) 1 +A(f)· B(f)
Communications
In a predictive modulator the feedback filter, B( f), has a large gain so that, ideally, the
[4] S. R. Norswortl
fed back signal equals the input signal. If A(f) = 1 (a wire), then both the STF (signal
Converters: The
transfer function) and the NTF (noise transfer function) have a value of, approximately,
978-0780310452
IIB( f). Recovering the input signal requires passing the output of the predictive
modulator through an analog filter with a transfer function of precisely B(f) (noting that [5] J. C. Candy and t
B[f] is a digital filter in the modulator of Fig. 5.39). The required precision of the analog Wiley-IEEE Pres
filter (the matching between the filter in the modulator and the filter in the demodulator)
[6] S. K. TewksbUl
limits the attainable resolution when using predictive modulators. Notice that both the
Noise-Shaping (
input signal and the quantization noise experience the same spectral shaping (spectral
CAS-25, pp. 436
discrimination is absent in a predictive modulator). Also note that the name "predictive"
comes from the modulator attempting to predict the input signal in order to drive the [7] W. R. Bennett, ,
output of the summer to zero. If the prediction is perfect, the signal that is fed back Vol. 27, pp. 446
exactly matches the input signal.
: Mixed-Signal Circuit Design Chapter 5 Data Converter SNR 201
Voul(f) Out
(to digital
filtering.)
Vout(f) Out
Analog
modulator.
""j
202 CMOS Mixed-Signal Circuit Design
QUESTIONS
5.1 Develop an expression for the effective number of bits in terms of the measured
signal-to-noise ratio if the input sinewave has a peak amplitude of SO% of (VREF+
VREFJ·
5.2 When using Eq. (S.14) what is the assumed ADC input signal? Put your answer in
terms of the ADC reference voltages.
5.3 Describe, in your own words, the difference between specifying SNR and SNDR.
5.4 Using SPICE simulations with an ideal ADC and DAC, show how coherent
sampling can result in an RMS value of quantization noise larger than what is Data Con
specified by Eq. (S.3). Comment on the shape of the quantization noise's
spectrum.
5.5 Suppose a perfectly stable clock is available (AT., is zero in Eq. [S.2l D. Would we
still have a finite aperture window if the clock has a finite rise time? Describe why
or why not?
Mixed-signal design is pc
5.6 How do the number of bits lost because of aperture jitter change with the with analog circuit design.
frequency of an ADC input sinewave? If the ADC input is a DC signal, is aperture precision of the analog eire
jitter a concern? Why? is "Analog Design without
5.7 Why must Bennett's criteria be valid for the averaging filter in Fig. S.29 to reduce In the past five ch,
the quantization noise in the digital output signal? Give an example input signal mixed-signal circuits. For
where averaging will not reduce quantization noise. converter (ADC) using mi:
5.8 Assuming Eq. (S.S7) is valid, rederive Eq. (S.13) including the effects of through an anti-aliasing 1
averaging K ADC output samples. Is Eq. (S.13) or the equation derived here valid shaping modulator, as dis
name for an ADC with 1 tc
for a slow or DC input signal? Comment on why or why not.
feedback loop to get a ru
5.9 If Bennett's criteria are valid, does averaging ADC outputs (or DAC inputs) put output is extracted using
any restrictions on the bandwidth of the input signal? Why? Give an example. decimated. In simple wor
representation of the analo
5.10 How accurate does an 8-bit ADC have to be in order to use a digital filter to
various DC input signals.
average 16 output samples for a final output resolution of lO-bits (see Eq. (S.59])?
Assume the ideal LSB of the 8-bit converter is 10 mY. Your answer should be Note that this type
given in both m V and % of the full-scale. sampling frequency, f" mt
also see that these types 0
5.11 Show the detailed derivation ofEq. (S.66).
on the topology used, dell
5.12 Summarize, and compare, the advantages and disadvantages of predictive and greater detail later).
noise-shaping data converters.
Clock input
Figure
fixed-Signal Circuit Design
I-bit
Analog input I~
-----J AM
L-..J
H ~NS d~l ~
NS modulator r
.
Digital
Filter
l Digital Output
f.!........_. _'r=~
Clock input "Decimation filter
A VREF+ VI
!
1.0
Average = I~O 0.25
0
1.0
[ 0
nJ1JlJL Average = = 0.5 VREF- 0
::i 1.0 A one-bit DAC
0 (31)·(10)
.... (Repeats every 32) Average
2
ro
oi 32
;; Figure 6.3 Thh
~
~ ~
~I
. . ..
..
In this chapter we'll use, exclusively, a one-bit ADC and DAC. To understand why, let's
review Fig. 5.28 and 5.32. In all cases, unless the spacing of the codes is perfect in these
data converters, averaging the codes will result in non-linearity. However, by using an so,
ADC or DAC with only two output codes in our NS modulator we are guaranteed
linearity (two points determine a line!). If one of the reference voltages in the ADC or
DAC is offset from its ideal value then we'll get a gain error but no nonlinearity error. and therefore
In Chs. 4 and 5 we treated the clocked comparator as a I-bit quantizer (ADC) with
an LSB given by
I LSB VREF+ VREF- = VLSE (6.1) This is nearly the tran!
this response as integr
as an added noise source, Fig. 5.2. The PSD of the added noise, see Eq. (5.5), is moment. Using Eq. (5.
2
VQe(f) == 12/
vISE (6.2)
s
In this chapter we'll set VREF+ = VDD I V and VREF- == 0 V so that VDDI2 or 500
mV. We can think of the output of the comparator (the output of the NS modulator) as a Notice that the AAF i
binary offset number with a value of 1 or O. After a simple conversion to two's
function appears to be
complement numbers, the output can be changed to +1 (01) or -I (II). Since our input average the modulator
signals are referenced, or swing around, VCM it's useful to think in terms of two's zone issues for a passi
complement numbers. For example, in Fig. 6.2, if we apply 500 mV to the input of our to get a better represen
NS modulator we get 1010101... (second sketch) or an average of 500 mV (= VCM). Since
all signals are referenced to VCM it may be more useful to say that we aren't applying a
signal (meaning just the common-mode voltage is present at the input) and thus we get a
Next, let's write
sequence of + 1, -J , + I, -1, etc. that averages to zero (Vcu)' Figure 6.3 shows how we
think about this in more detail. If we pass a signal through an inverter we are multiplying
it by -1 when we think in terms of two's complement numbers.
ixed-SignaJ Circuit Design Chapter 6 Data Converter Design Basics 205
······-······OOorO
~
Two's complement
,.-'-..
II or-I
Average '= (31)·(1.0)
A one-bit DAC
~ 32
Figure 6.3 Thinking about the inverter in terms of the common-mode voltage.
Averaoe = (7)·(1.0)
e 8
6.1 Passive Noise-Shaping
Averaoe =.!Jl.
'" 8 Figure 6.4 shows a schematic and block diagram of a passive NS modulator. Reviewing
Average =0 V 5.38, we see that this is a modulator topology with 8U) = I. To calculate AU) let's
I > write the output of the summing block (the resistors), or the input to the A(f) block, as
40
Vi" - Vint -Voul - Vinl '
R + R =I~ (6.3)
: DC averages.
noting that we are assuming all voltages are referenced to VCM' The output of the A(f)
block is
(6.2) STF(f)
,
I 2
VoutU) = 3 . RC 'Vin(f) + . 3 +Jm
. RC .VQe(f) (6.7)
) that VDDI2 or 500 +Jm
of the NS modulator) as a
Notice that the AAF is built-in to this topology via the STF. At DC the noise transfer
rlple conversion to two's
function appears to be 2/3. However, we know that at DC, as seen in Fig. 6.2, we can
r -1 (11). Since our input
average the modulator's outputs and recover, exactly, the analog input signal (see dead
think in terms of two's
zone issues for a passive modulator on page 215). This means that VQe(O) ~ O. Let's try
)0 m V to the input of our
to get a better representation for the output of the modulator. In order to begin notice that
of 500 m V (== VCM)' Since
that we aren't applying a (6.8)
e input) and thus we get a
Next, let's write
:;'igure 6.3 shows how we
Iverter we are mUltiplying (Vin-Vint + -VOUl-Villl). + V (f) (6.9)
\ R R jmC Qe Vout
206 CMOS Mixed-Signal Circuit Design Chapter 6 Data Converter
Vin - Vim
Digital
Analog >-----.r--- Vout
DAC Is
-Vaul
F
(a) Circuit implementation of a passive NS modulator
Example 6.1
VQeCf) Simulate the operatior
10 pF, and the c10cl
simulation results and
Vin l---,--VOUI
Figure 6.6 shows the
circuit. Notice that as
stays low more often.
or a logic 1, more oft
~_-_V;;.;.OU,,-I--1- 1 code into an N-bit we
used to get our origina
In =fsl2 I
Figure 6.5 Modulation noise spectral density.
fS modulator
Example 6.1
f) Simulate the operation of the NS modulator seen in Fig. 6.4 when R is 10k, C is
10 pF, and the clocking frequency is 100 MHz. Comment on the resulting
simulation results and the operation/limitations ofthe circuit.
Figure 6.6 shows the input (a sinewave at 500 kHz) and (digital) output of the
circuit. Notice that as the input signal moves towards ground the output signal
stays low more often. As the input signal moves towards VDD the output is high,
or a logic I, more often. While we can use a digital filter to change this one-bit
code into an N-bit word (and we will!) let's show that a simple RC filter can be
used to get our original signal back from the digital data, Fig. 6.7.
::>dulator. The STF starts to roll-off at 1I2nRC or 1.59 MHz. While the circuit will
continue to operate at higher frequencies the SNR of the data converter will suffer.
We'll see that for proper operation the comparator's gain becomes very important.
L1lator.
Here we are using an ideal comparator. Finally, it's interesting to look at how Vint
changes with time for various inputs. Remember, as seen in Eq. (6.12), that this
term adds distortion and noise to our input signal. At high-frequencies this term is
small because of the pole at 1/2nRC and at low frequencies the contributions are
iroRC,vout (6.1 0) small since the feedback loop responds faster than the input signal changes.
What's of concern is the moderate frequencies and how the SNR is degraded .•
1+jroRC) (6.11)
-2· Vim
!e + 1 +jroRC
(6.12)
Extra noise/distortion
We spent a great ,
decimation. Let's use these
modulator. Let's set the bar
Figure 6.7 Using an RC circuit to filter the digital output in Fig. 6.6.
The value of RC must be rr
clock, T" to keep v int from
6.1.1 Signal-to-Noise Ratio
Eqs. (5.10) to (5.13) we get
Let's calculate the SNR for the first-order noise-shaping modulator seen in Fig. 6.4 and
characterized by Eq. (6.12). In the following we'll ignore the extra noise/distortion SNRidea/ 20·log
resulting from variations in Vim' Further note that the pole associated with both the STF v
and NTF won't affect the SNR since it's common to both. We know the output of the
modulator is passed through a lowpass filter with a bandwidth B to remove the
modulation noise, Fig. 6.8. The smaller B the lower the noise in the final digital output If we set RC 4.4(1; = 4.47
word and the larger the SNR. Again, the trade-off with using smaller lowpass filter
bandwidth is that the frequency range of the allowable input signals shrinks. Let's SNRideal = 21
calculate the RMS noise in the filter's output using
B 2 B For K 4 this equation i~
V;Oise,RMS = 2 f INTF(f)l21 V Qe(f)12 . df = 2· ~~~ f
. (2nj- RC)2 . df (6.13) beyond 4 we get an increa:
o 1s 0 dB
or
2 ViSB 2 B3
Vnoise.RMS = 2· 12fs . (2nRC) . 3 (6.14)
again noting that using a smaller digitallowpass filter bandwidth, B, reduces the noise. Figure 6.9 compares the fil
Eq. (5.58). It's important
oversampling feedback mo·
be busy).
6.1.2 Decimating and Fi
It's important to note that I
was passed through a perf
output through a Sine aver<
the higher frequency noise·
want to answer two questi(
should be used in the digit:
Figure 6.8 Filtering out modulation noise to calculate SNR. we use only this filter (no:
NS modulator be affected?
xed-Signal Circuit Design Chapter 6 Data Converter Design Basics
We spent a great deal of time in Ch. 4 using Sinc-shaped lowpass filters for
decimation. Let's use these filters here for filtering and decimating the output of the NS
modulator. Let's set the bandwidth of the filter, see Fig. 4.17, to
B j, (6.15)
2K
noting that we assumed, when deriving Eq. (6.14), a brickwall-shaped lowpass filter
response. We can reduce the noise further, after decimation, by passing the digital data
through additional lowpass filtering (perhaps using the biquad filters discussed in Ch. 4
with peaking to account for the droop introduced with the Sinc filter). Making the
substitution gives
1.4ps
V;oise,RMS = 12 (6.16)
Jut in Fig. 6.6.
The value of RC must be much longer (at least five times) than the period ofthe sampling
clock, Ts, to keep Vim from varying too much. Following the procedures used to derive
Eqs. (5.10) to (5.13) we get
llator seen in Fig. 6.4 and
the extra noise/distortion SNRideal = 20 . log 6.02N + 1.76 20 --::::::~ + 20 logK3/2
Dciated with both the STF Vnoise.RMS
Te know the output of the
dwidth B to remove the (6.17)
in the final digital output Ifwe set RC 4.4if, 4.4T, then we can write
ing smaller lowpass filter
put signals shrinks. Let's Vp / ,fi
SNRideal = 20 . log . 6.02N + 1.76 - 18.06 + 30 10gK (6.18)
Vnoise,RMS
2
For K = 4 this equation is the same as Eq. (5.13). However, for every doubling of K
Y·RC) ·df (6.13) beyond 4 we get an increase in resolution, Nine' of 1.5 bits or an increase in SNRideai of 9
dB
18.06
(6.19)
(6.14)
SNRideal 6.02(N + Nine) + 1.76 (6.20)
:h, B, reduces the noise. Fignre 6.9 compares the first-order NS modulator to simple oversampling, Fig. 5.31 and
Eq. (5.58). It's important to note that Bennett's criteria need not be valid using the
oversampling feedback modulator discussed in this section (e.g. the input doesn't have to
be busy).
Filter removes 6.1.2 Decimating and Filtering the Modulator's Output
_ _ this noise.
It's important to note that Eq. (6.17) was derived assuming the output of the modulator
was passed through a perfect lowpass filter with a bandwidth of B, fl2K. Passing the
) output through a Sinc averaging filter, see Fig. 4.14, will result in a poorer SNR because
'2 f the higher frequency noise components will not be entirely filtered out. In this section we
want to answer two questions: (1) what order, L (sec Eq. [4.19]), of Sine lowpass filter
should be used in the digital filter on the output of the NS modulator, and (2) assuming
late SNR.
we use only this filter (no additional filtering), how will the ideal SNR of the first-order
NS modulator be affected?
210 CMOS Mixed-Signal Circuit Design Chapter 6 Data Convel
the word size increased by 10gzK bits in each Sinc filter stage, For a cascade of L filters
we can require :::::j
120mv1
30 10gK - 18.06 640mvl
L . Iog2K~ 6.02 (6.21 )
56omV1
480mV-i
For K:S: 256 we only need one lowpass Sinc filter or L = 1.
400rnv~
Example 6.2
Sketch the implementation of a K 16 decimating filter for the NS modulator :::::-"1
16UmV
discussed in Ex. 6.1. Simulate the operation of the resulting ADC (modulator and BOrnV
filter). Estimate the SNR, number of final bits, and Nyquist frequency. 0.01
the SNR (= 25.78 dB), is 4-bits. Our filter causes the word size to increase by I-bit
;e shaping using a passive through each of the 4 stages so that our final output word size is 5-bits. We can
ltor and neglecting the throw the LSB of this word out; however, notice that the largest word we get out
listortion term in Eq. (6.12). of the filter is when the filter's input remains a 1 at all times. In this case the
filter's output (the output of the ADC made using the NS modulator and filter) is
16 or 1 0000. When the filter's input is a 0 at all times our filter's output is 0 0000.
The output can swing from I 0000 to 0 0000 around the common-mode code of 0
rsampling
1000 (8). By removing the MSB we can make the output swing from 1111 (15) to
0000 (0) around 1000 (8). However if the NS modulator's input moves close to
VREF+ (here VDD) then the output of the filter can overflow.
-'»
~umber of points averaged Figure 6.11 shows the simulation results. The input frequency, as in Ex. 6.1, is
500 kHz. Looking at the frequency response of our filter, Eq. (4.20) and Fig. 4.17,
,rsarnpling (Fig. 5.29). we should only see a minor amount of attenuation through the filter. However,
looking at Fig. 6.11 we see more than a minor amount of output signal reduction.
le increase in the number of We've, up to this point, ignored the extra distortion/noise term seen in Fig. 6.12.
uses a I-bit ADC, then the In the next section we'll discuss this term in more detail and how matching and
mverter is Nine + 1 bits. (An offsets effect the performance of the modulator.•
S modulator] would ideally
1 Ex. 4.3 and 4.14 that 960mV..,--,..-"=---....,--'-'?--,---,--~=:!....,.---,
(6.21 ) MOmV
560mV
480mV
400mV
320mV
240mV
r for the NS modulator
160mV
Ig ADC (modulator and
t frequency.
80mV i I I-~--t--r----i--i---i---+----l
0.0115 0.7115 1.4115 2.1115 2.0ps 3.5115 4.2115 4.9ps 5.611' 6.3115
Figure 6.11 Simulating the ADC (modulator and filter) discussed in Ex. 6.2.
(6.22)
SNR Calculation using a Sine Filter
lock rate is 100 MHz/I 6 Let's now consider how filtering with a Sinc filter instead of the ideal lowpass filter
.125 MHz. The increase effects the SNR of the modulator/filter implementation of a data converter. Remember the
tput word size, based on SNRideal was calculated in Eq. (6.18) assuming the modulation noise was strictly
bandlimited to B. Figure 6.12 shows the PSD of the NTF2(f) and IVQe (f)12 (the
modulation noise) in a first-order passive NS modulator. Also shown in this figure is the
shape of the averaging filter's magnitude response squared that is normalized to unity by
dividing by K, see Eq. (4.20). Here we are showing the shape of a filter with K = 16 and a
range ofJJ2 (see Fig. 5.12).
We can calculate the RMS quantization noise resulting from a cascade of a first
order modulator and an averaging filter, L I, using
/,/2
V Ri
'~
\ V2
IVQe(f) I2 • INTF(f) I2 = l~f. (2nj- RC)2 .
Figure 6.;
+=T-~---~--~--~:··--~==~-=~--~~f
~~ ~n
Resistor Mismatch
B = {~ , Ideal maximum input frequency
In order to detennine the ef
Figure 6.12 Showing modulation noise and filter response. neglecting the effects of the
or
Rr (Vin vinr)-Ri'
(6.24)
RrVin Vint·(Ri+J
fl rd"'M~
2
VQe,RMS = u'
viSB (
2nR
C)2
. 2n2K3 (6.27)
Comparing this result to what we got in Eq. (6.16) we see that using the Sinc lowpass
filter is nearly ideal for removing modulation noise. The more important concern, when
using the Sine lowpass filter, is the droop that the desired signal undergoes. Ideal shape Sh:
(ideal area is all(
6.1.3 Offset, Matching, and Linearity shaded)
Examine the passive NS-modulator with offsets and mismatched resistors seen in Fig. (a)
6.l3. In this section we want to discuss how non-ideal behavior affects the performance
of the modulator. To begin, notice that a comparator offset has the effect of causing the
average value of vint to be VCM ± Vas and thus the offset passes directly to the output of
the modulator. This means that if we apply the common-mode voltage to the input the Figure 6.14 C
average of the output voltage is VCM ± Vas.
(ixed-Signal Circuit Design Chapter 6 Data Converter Design Basics 213
Ri
0---~------~------------~+
et jroCRj -Vim'
Vout = 1 +jroCR . Vjn + 1 +jroCRf ' V Qe + 1 +jroCRj (6.31 )
J
(6.25)
The effect of mismatched resistors is a gain error (but no non-linearity, and thus,
distortion). Note that the absolute value of the capacitor isn't critical or important for
precision operation,
The Feedback DAC
The feedback DAC is the most critical component in the modulator for precision
(6.26) operation, Consider the waveforms seen in Fig. 6.14. In this figure we show the ideal
shapes of the DAC's output (remember our DAC here is a simple inverter so we get
'3
(6.27) ,
perfect linearity, Fig. 6.15). As seen in the figure, the shape of the pulse affects the
amount of charge, or current, we send back to the capacitor. In part (a) we see the ideal
pulse shape and the ideal area under the pulse (the shaded area). In part (b) we see how
1.0Vf
O.9V •...•
n.BV •••.•
the finite rise time and fall time can affect the actual area under the curve and thus the
output of the integrator. In order to minimize these unwanted effects we can use wider n.N]'.. '"
D.6V ,"'"
pulses as shown in parts (c) and (d), which means we run the modulator at a slower
D.SV
clocking frequency. Increasing the width of the pulses minimizes the percentage of the
O.4V
area affected by the transition times. Note that the feedback signal directly subtracts from
O.JV
the input signal so that any noise or unwanted variation in the fed back signal, such as an
D.N
amplitude variation, can be considered as adding noise to the input (and thus degrading
D.1V
the modulator's SNR). This is important. In the coming chapters we will use switched
D.W
capacitor circuits that simply have to fully-discharge in order to make the shape of the D~.
feedback pulse less important.
Figure 6.16
Ideal
/ MSB of the filter's outp
:> 1.0 through the selector to tb
,;:f
£I-
::I
\ Nonideal the selector to clamp the
0 Dead Zones
0 Besides the nonlinearity
0 0.5 1.0 Comparator input, V
the output doesn't move,
is seen when the input si
Figure 6.15 Ideal and nonideal transfer curves for a l-bit DAC. zones are the result of
DACOjJ,et 0000100001. etc.}. The
relatively wide range of i
Notice, in Fig. 6.15, that if our DAC's reference voltages are offset from their ideal values this question examine th!
of VREF+ and V REF- that, like mismatched resistors, all we get is a gain error (no the output of the modulal
non-linearity). Possibly the most important concern when using a single-bit inverter suppose the input voltag
topology DAC is the on-resistance of the transistors used to drive the feedback resistors. time, it's possible for the
The drive strength of the MOSFETs should be so high that little voltage is dropped across the fed back current [-\
them (so the output of the DAC swings between V REF+ and V REF-). This (limited DAC using an active integrator
drive) is, again, another reason we'll soon start to focus on switched-capacitor
implementations of NS modulators. 6.2 Improving SNR
Linearity of the First-Order Modulator
In this section we discw
We've ignored the effects of the extra noise/distortion term seen in Eq. (6.12) because it's We'll start out by discussi
a nonlinear error highly dependent on the input signal's characteristics (and thus hard to modulator. This topolog
quantify). Let's use simulations to look at the linearity of the first-order NS modulator and discussed in the last sectit
digital filter seen in Fig. 6.1 0 (with a DAC connected to the output of the digital filter to avoid dead zones), and Ie
show the digital data as an analog wavefonn). Figure 6.16 shows the input/output of Fig. we'll discuss using switcb
6.10 with a slow ramp applied to the modulator's input. Note that when the input ramp benefit of switched-capac
voltage is close to the power supply rails the digital output becomes nonlinear (too high remove the effects of
when the input is close to ground and to low when the input is close to VDD). Also seen non-overlapping clock g(
in this figure, as mentioned in Ex. 6.2, is the digital filter overflowing. In order to avoid discuss putting NS modu
this situation we can use a selector, like the one seen in Fig. 4.40, with the Set input lowering the final output
connected to the MSB of the filter (remember we do a mUltiply by 2 by removing the section by showing how a
filter's MSB to scale the filter's output to full scale as discussed in Ex. 6.2). When the and noise by removing th.
1ixed-Signal Circuit Design Chapter 6 Data Converter Design Basics 215
MSB of the filter's output is a 0 we simply pass the multiplied-by-2 output of the filter
through the selector to the final output. When the MSB of the filter's output is a I, we use
the selector to clamp the final output to 1111 and thus avoid overflow.
Dead Zones
)
Besides the nonlinearity seen in Fig. 6.16, notice that for some ranges of input voltages
Comparator input, V
the output doesn't move or, rather, the output is dead. A simple example of a dead output
is seen when the input signal is close to the common-mode voltage, 500 mY. These dead
a I-bit DAC.
zones are the result of a repeating modulator output code (e.g., 010 10 I, 110110,
0000100001, etc.). The question is why does the modulator output code repeat for a
relatively wide range of input voltages (e.g. 460 m V < Vi" < 540 mV)? In order to answer
ffset from their ideal values this question examine the modulator seen in 6.13. With an input voltage of 500 mV
re get is a gain error (no the output of the modulator is 101010... so that the average of the output is 500 mV. Now
using a single-bit inverter suppose the input voltage is increased to 510 mY. Because the voltage Vim varies with
[rive the feedback resistors. time, it's possible for the average current supplied from the input, [v m -V;nt(t)]IR; to equal
Ie voltage is dropped across the fed back current [-vout-Vint(t)]IR/ when the modulator's output is 1010101... By
VREF-). This (limited DAC using an active integrator, Sec. 6.2.4, we can hold Vin/ constant and eliminate dead zones.
:us on switched-capacitor
6.2 Improving SNR and Linearity
In this section we discuss techniques for improving signal-to-noise ratio and linearity.
,!llin Eq. (6.12) because it's We'll start out by discussing the second-order passive (uses two capacitors) noise-shaping
lcteristics (and thus hard to modulator. This topology is considerably more useful than the first-order topology
rst-order NS modulator and discussed in the last section because it has better linearity, randomizes the output code (to
ltitput of the digital filter to avoid dead zones), and lower modulation noise in the signal bandwidth of interest. Then
lWS the input/output of Fig. we'll discuss using switched-capacitors instead of resistors in the feedback paths. The big
~ that when the input ramp benefit of switched-capacitor circuits over continuous-time circuits is that, Fig. 6.14, we
~comes nonlinear (too higb remove the effects of non-ideal pulse shapes. The drawback is the need for a
s close to VDD). Also seen non-overlapping clock generator (which consumes power and layout area). Next we'll
:rflowing. In order to avoid discuss putting NS modulators in parallel. This is useful for increasing SNR while not
g. 4.40, with the Sel input lowering the final output clocking frequency (at all or as much). Finally, we'll end the
iply by 2 by removing the . section by showing how an active circuit, e.g. an op-amp, can be used to improve linearity
ised in Ex. 6.2). When the and noise by removing the extra term seen, for example, in (6.12).
216 CMOS Mixed-Signal Cireuit Design Chapter 6 Data Converte
SNRideal =
( ---.<-
Vl-V2
R ' R
I
Vat/I-V2) ·--=v,
jmC·
(6.33)
:::lu.. ;.
8V U
If the quantization noise added to VI and v2 is VQel and VQe2 then + V Qel = Vo"ll and
V2 + VQe2 = V out2. We can then write
VJ
n.1V
O. 1·· ..
"'.,u~.'
(
n.GV 'U"'+'
Vin - 3Vl +V2 +jmRC· V Qe1 == voutl(jmRC-I) (6.34) n.sv ....... j..
0.4'1 ....... ~U
VI 2V2 +jmRC· VQe2 = v ou l2(jmRC -1) (6.35)
O'JV~""'i
Adding these two equations gives O.2V .........
O.lV·.. ,"
Vin - 2vI - V2 +jmRC. (V Qel + V Qe2) = (vaul] + vat/a)' (jmRC - I) (6.36) o.nv .
o~s s~s
As a quick check to ensure that this equation is correct, notice that if we pass the output
signal through a 100vpass filter with a bandwidth approaching 0, 0) - ) 0 and we use a large Figure 6.18 Repeating thi
RC then for a DC input, Vin "" Vj "" V2 and Vin - 2Vl - V2 -2V out (so Vau! "" Vi")' Next,
adding the random signal powers together (see, for example, Eq. [5.55] and the associated
Example 6.3
discussion) gives Repeat the simulation
2 modulator seen in Fig. I
2 (2
(mRC) . "VQel + VQe2
2) =(mRC )2 . 2VQe (6.37)
The simulation results
resembles the input am]
'(ed-Signal Circuit Design Chapter 6 Data Converter Design Basics 217
and
ight wonder if we can get
:moving the comparator's
this circuit we have two Finally, we can write
k to the inverting input of
because the output is fed 2 (roRc)2
V . '=
1 + (roRC)2 2 (6.39)
I drive these nodes to the oul,nOlse
2 1 ·v·
2
(6.40)
VOut,signal
1+ (roRC)2 In
noting the quantization noise power is cut in half. Equation (6.18) can be rewritten for the
second-order passive noise-shaping topology as
Vlj2
P
SNRideal 20'10g =6.02N+1.76 15.13+3010gK (6.41)
VnOlse,RMS
The modest increase in SNR, however, is not the big benefit of this topology. Rather the
better linearity and reduction in dead zones are the major benefits of the topology. The
extra noise/distortion term can be written as
2vI + V2
(6.42)
or. 1 + (roRC)2
Note that this topology attempts to drive VI and V 2 to the same value, that is Vi.' The
rite drawback of this is that the comparator must operate with an input range extending from
1 VDD to ground. Figure 6.18 shows the simulation results where we repeated the
VI (6.32) simulation seen in Fig. 6.16 but used the second-order modulator. Clearly both the
DC
linearity and extent of dead-zones are much better using the second-order modulator.
(6.33) .
:hen VI + VQel
2)' UroRC-l)
5~s 1 O~s 15"s 2n~s 25~s 3n~s 35"s 40"s 45~s 50\ls
that if we pass the output
ro ~ 0 and we use a large Figure 6.18 Repeating the simulation seen in 6.16 but with a second-order modulation.
Voul (so V oul "" Vin). Next,
VI
1.0V
O.9Y
O.8Y .......
O.7V •.
,, 'i
O.6V
n.sv
O.4V -
O.3V
O.ZY
O· 1V T
o.ov
o.o~s
Figure 6.19 Repeating the simulation seen in Fig. 6.11 but with a second-order modulation. Figure 6.21 Regene
n-l n
n 112
Figure 6.22 Sv,;itche,
Figure 6.20 First-order passive modulator using switched-capacitors. gam err(
~ed-Signal Circuit Design Chapter 6 Data Converter Design Basics 219
second-order modulation. :Figure 6.21 Regenerating Fig. 6.7 using a passive switched-capacitor modulator.
rs The operation of this topology can be described by simply substituting Eq. (6.43) into Eq.
(6.31). The signal gain can then be written as
lemented using switched
placed with SC resistors). Rj == C;
(6.44)
5.7 but using this new R,
,ver the continuous-time
If C, isn't exactly equal to Cj then the input signal undergoes a gain error. In order to
ning out of the feedback
eliminate this gain error, and to simplify the implementation of the modulator, consider
lcitors fully charge and
sharing the feedback and input capacitors, Fig. 6.22. Simulations showing the operation
of this circuit are found at CMOSedu.com. Reviewing the derivations in Sec. 2.2.3, we
again see Fig. 2.35, see that the input signal, in Fig. 6.22, is subtracted from the fed back signal and summed
together using Clm • The inputs of the comparator are swapped, from Fig. 6.20, to ensure
the polarity of the output signal matches the input signal's polarity. Notice that we
connected the input and fed back signal's to the bottom plate of C/. Remember that this is
the plate with the largest parasitic capacitance. When the ~I switches are closed, the
bottom plate of C and its parasitic, are charged to VI•• Noise coupled into this bottom
"
plate, from the substrate, sees the low-impedance input so that ideally none of the noise
Digital
passes through to Ctnt" When the <1>2 switches close the bottom plate of C/ charges to vow!"
>----t>--- Vout Since the bottom-plate parasitic charges back and forth between Vi. and VO" ' , none of the
charge from this parasitic is transferred to C,.,.
~l
>----tr--- Vout
VCM
2 (J ViSB (6 48)
VQe ) = 12 . K .r
path J.Y
.
For the noise-shaping topologies we would replace K (for example, see Eq. [6.18]) with
Kavg' Kpalh while the bandwidth of our desired signal spectrum ranges from DC to B
when decimation is used where
B= (6.49)
2Kavg
The next question that we need to answer is "Is it practical to implement parallel
paths of noise-shaping modulators or Nyquist-rate data converters?" Clearly there will be
an increase in layout area using this approach. Is the increase in SNR worth the increase
in chip size, power draw, or complexity? Well, let's not spend any time on Nyquist-rate
converters (e.g., flash or pipeline) in a K-path configuration (called a time-interleaved
converter) since the input signal would have to meet Bennett's criteria (pages 165-166,
Figure 6.23 Two-patr
consider problems with a DC input). Reviewing Fig. 2.37 we see that using a SC
modulator like the one seen in Fig. 6.22 may be a possibility. However, notice that we Note that each pa
need to multiply up the clock frequency from!, to KJ; then pass the resulting signal at a different voltage, del
through the non-overlapping clock generator seen in Fig. 2.38. Further, the SC
modulators would have to settle (fully charge/discharge their capacitors) within TjK
seconds (an even bigger concern if the modulator uses an active element like an op-amp). so combining the input n
If this is possible why not simply use a single path clocked at the higher rate, KJ; (or feedback paths into K-r
better yet two paths clocked at KJI2 like the topology seen in Fig. 2.36)? We get the same mismatch, requires som,
performance as the K-paths but with a much simpler implementation. Since we have to we've added an amplif
direct, and process, the input signal to the various paths, again Fig. 2.37, at a high-rate comparator's offset is ref
limits the use of K-paths in SC circuits (more on this in a moment and in Ch. 9). G = 100, for example, th
(ed-Signal Circuit Design Chapter 6 Data Converter Design Basics 221
What about using the modulator seen in Fig. 6.4 in a K-path implementation?
:m noise out over a wider Figure 6.23 shows a 2-path, or time-interleaved, implementation. There are several
veraging filter bandwidth benefits to this topology including: 1) the input can be continuously connected to the
be reduced. As discussed K-paths of modulators, 2) no active device (e.g. an op-amp) with settling time issues, and
pIing frequency by using 3) the comparator's are triggered on the rising edge of a clock signal (so the feedback
he same variable, K, for operation can occur over the entire clock period, Ts for each path). The drawback of this
[tput of a modulator, Sec. topology is the mismatched gains through each path. Notice that we've combined the
ction, use variables with digital outputs of the modulator into a two-bit word (so the possible outputs are 00, 01,
llator implemented with and 10). This addition, we should recognize, is a filter with a transfer function
I-
H(z) = I +Z-l Z-Kpath
"'--=---
I _z-l
where /s,new = Kpafh ',Is, =e _j2n,_f_
/s,treW
Itor or an average of two filter. Ifwe were to use 8-paths we would get a 4-bit output ranging
(6.46)
from 0000 to 1000 (which may result in the same overflow problems we saw in Figs. 6.16
m be described using and 6.18 when the input of the topology is near or VDD, here). A simple solution to
the overflow problem, other than using a selector to keep overflow from occurring as
(6.47) discussed earlier, is to use only 7-paths (instead of 8) so the final output code ranges from
000 to Ill.
llh -Is so that the PSD of
quency range
Output
VOUll (2-bits)
(6.48) >---1"'=::"';
@2/s
pie, see Eq. [6.18]) with Carry
n ranges from DC to B
(6.49)
v:l~.
Path filter
T';C -.~---~
O.BV •• i'"
4R~ r~
1-z- 1 0.7'1 ..... .
0.4V .... ..
+ 0.3V ..... .
0.2V .... ..
0.1V .... ..
o.ov+-
o.ops
Analog i K-path
In l modulator
o.o~s
modulator. Figure 6.26 shows adding a register to re-time the output of the modulator
(decimate) back down to /,. Also seen in this figure is the digital filter seen in Fig. 6.1 that Figure 6.27 Regen
is used to remove the modulation noise. Regenerating the simulation results in Fig. 6.25
~ed-Signal Circuit Design Chapter 6 Data Converter Design Basics 223
V(out) V!dou~
Path filter 1.0V-.--""",nrri
. .
, , , , ,
4 ) O.9V <-.----------,----------,----------,---------- . . - -
, , , , ,,
H(z) = I-z-I O.BV
,,
,
,,
,,
,
,,
,
,
,
,,
,
I
,
--.----------,----------,--------.-,----------..,-
,,
,
,,
. -
I-z ,o· , , ,
, _______ ,, __________ •, __________ ,, _________ _
O.7V
O.6V
O.5V
O.4V ..........•.........
O.3V ..........•.........
O.2V
O.1V
o.oV'+---i---j-.LLLLLlU.tw'-WlllJ.lilljLlULJb...-T---'''''-"..1UfiWllllJlJW.Df'-".J.IJ
o.o~s O.2~s O.4~s O.6~s O.B~s 1.0~s 1.2~s 1.4~s
~
, 3-bits out Figure 6.25 Simulating the operation of the 4-path modulator in Fig. 6.24.
,
@41s
bits:
· , I _ Z-Kpa1h log2 K path + 1
K- b ItS, I _ Z-Ka"g ]L
1 - Z-I I---b-cif-ts---I [ 1-z- 1 j,'
1
H(z)=~
I-z
-4
· . 4.32 !
Figure K a",
)
I-*---'----"-~-----l-c-Kpath I Final output clock rate:
Note that decimation
is not used here or
: +
,
Kpath
1_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
Decimator:l ____________________
Kavg Decimator:
I
+ ,
:s shown. using decimation of 4 we get the results seen in Fig. 6.27. The rate of the digital word
output by the path filter is fs (just as it is when using a Nyquist-rate converter). Further,
led since the integrator
Delta-I-Sigma topology). the path filter in Fig. 6.26 has the same frequency response as the SIR used in a
Nyquist-rate ADC (a droop of -3.9 dB atfsl2, Fig. 2.17). Thus, with some thought, we
ly around each feedback can use this topology to implement high-speed Nyquist-rate ADCs.
or feedback paths, the
mplifier, like a diff-amp, l.OV-,--""-n-nrrVcm!v,;;,in,,.).",,-----,r-,_V-,-!o_ut.,..)- - c ._ _-,,~!c:cdO=-=ut,,-)_ : - - '
:quency response, of this O.9V .. . •.......• ; ..........•.•••...•.. ;.......... ; ..•......• ~ .... .
liminate the offset effects , , , , ,
-,----------,----------,-.--------,----------.,----
O.BV .. , , , , ,
, , , ,
Irab1e to Tsl Kpath). Using O.7V ..........•..•.
• ,,
______ 4 __________ •
,, ,, ,,
_ _ _ _ _ _ _ _ _ _ • _ _ _ _ _ _ _ _ _ _ .. _ _ _ _
,,, ,,, ,
,,
,
,,
: maximum speed of this O.6V ..........•....
,
--------_.- ... -.--.--------_
,
.. --------------
, ,
.
~. 6.7 and 6.21. The extra._ · ,
O.2V
: the modulators together,
O.W
ise in the signal. Because .
O.OV+-----r---r-
tly on the output of the O.O~s O.2~s O.4~s O.6~s O.8~s 1.0~s 1.2~s 1.4~s
using clock signals like those seen in Fig. 2.37 for high-speed conversion.
input to the comparator,
One may wonder why the STF is Z-K or such a long delay? Looking at the timing
of the clock signals in Fig. 2.37 we see that each path is only activated every To seconds.
If z == ej2rcfljs.new or z = e j2rcfl(K/.) then to represent a single path delay of T, seconds we have
to use Z-K. In this scenario, Fig. 2.37 where H(z) represents the transfer function of a In (b) we've replaced thl
modulator, each of the modulators is independent, that is, not sharing the integrating can write
capacitor (integrator) as we did in the topology seen in Fig. 6.24. With some thought we
can implement a practical K-path ADC using SC circuits sharing a common integrator,
Fig. 6.24, with clock signals. also seen in Fig. 6.24. having a width of Ts and leading
If the amplifier's gain, G.
edge spacing ofT/K seconds so that the effective samplingfrequency is K Is (Ch. 9).
Effects ofthe Added Amplifier on Linearity
We added the amplifier with a gain, G, in Fig. 6.24 in order to reduce the effects of
comparator offset and make combining K-paths of quantizers sharing a common
integrating capacitor a practical analog-to-digital converter topology. We haven't analyzed
how this added amplifier can affect the linearity and performance of the topology. Let's
do that here. Rewriting Eqs. (6.8) to (6.12) with this gain included we get
VQe(f) + G· Vint(f) = Vout(f) (6.54) (a)
(6.55)
Figure 6.28 ReI'
lixed-Signal Circuit Design Chapter 6 Data Converter Design Basics 225
Extra noise/distortion
(6.52)
It would appear that by using a very large gain, G, we can eliminate the lowpass filtering
ycle delay, ;:--1 while the effect of the input signal (and thus pass very wideband signals to the modulator'S output,
the noise-transfer-function useful for Nyquist-rate analog-to-digital conversion) and remove the modulation noise
) the differentiation, jwRC, from the output of the modulator. The extra noise/distortion term, however, remains in
he continuous-time passive the modulator's output. Again, the only way to get rid of the noise/distortion term is to
acitor topologies described hold V inl fixed so that it doesn't vary. This can be accomplished using relatively large 1. or
;orresponding output of the an active element discussed next.
Before getting too excited about removing the modulation noise we'll find out, in
the next chapter, that we can think of the comparator as also having a gain, Gc ' in series
(6.53) with the added amplifier that varies keeping the forward gain of the modulator (a
feedback system) equal to unity. In other words, if we increase the amplifier gain G we
at a rate of K·1s = fs,new, so
:ated in Eq. (6.48), will be get an effective decrease in the comparator's gain Gc'
F has the shape of a comb 6.2.4 Improving Linearity Using an Active Circuit
'lodulators in K-path ADCs
~onversion.
Figure 6.28a shows the integrating topology we've used in our passive modulator. The
input to the comparator, Vinl , in this circuit can be written as
~lay? Looking at the timing
. 1
activated every T, seconds. lin' j(J)C (6.59)
Iclay of Ts seconds we have
; the transfer function of a In (b) we've replaced the passive integrator (the capacitor) with an active integrator. We
Jot sharing the integrating can write
.24. With some thought we · Vin( Vc
- G . Vim an d lin = Jij(J)C (6.60)
'ring a common integrator,
a width of T, and leading
quency is K·1s (Ch. 9). If the amplifier's gain, G, is big we can write
. 1
Vc "" -lin' j(f)C (6.61)
~
tizers sharing a common i in Comparator c .c omparator
Jlogy. We haven't analyzed
ance of the topology. Let's
led we get
~f :t? .
~
Vim"
I _~
Vc·
~
>-
(f) (6.54) (a) (b)
(6.55)
.Figure 6.28 Replacing the passive integrator (a) with an active integrator (b).
226 CMOS Mixed-Signal Circuit Design Chapter 6 Data Converte
where we've switched the inputs of the comparator to account for the inversion.
Remember our goal is to keep v;nl constant so that we reduce the noise/distortion term in,
for example, Eq. (6.58). We can then write
Vi"'''''.fro C . G (6.62)
V,'. R
~"----------~~--~----~i'~------ Comparator
Analog C VCM-f + Digital
!\I'~__~v_·in_I _1_ Vouf
Figure 6.31
Example 6.4
Repeat Ex. 6.1 using the modulator seen in Fig. 6.29.
Treating the comparator ;
The simulation results are seen in Fig. 6.30. In the simulation we can show that Vim
doesn't vary (the point of adding the op-amp to the modulator). In order to show
the better performance in linearity/distortion let's pass the I-bit digital output we can write
through an average of 16 filter, see Ex. 6.2, and regenerate the results seen in Fig.
6.16 (see Fig. 6.31). Clearly, the active topology in Fig. 6.29 shows better linearity
than the passive topology seen in Fig. 6.4 .•
O.3V+....... :
lIJ---.;-1----.;-1----1-1----,1·
0.011" 0.2po 0.411. 0.6116 0.8110
-----i-
LOllS
Figure (
Figure 6.30 Repeating Ex. 6.1 for the first-order active NS modulator.
fixed-Signal Circuit Design Chapter 6 Data Converter Design Basics 227
(6.62)
:mentation of a first-order
grator. The same equations
lfiguration.
Comparator
Digital
Vaul Figure 6.31 Linearity of the active NS modulator, compare to Fig. 6.16.
Second-Order Noise-Shaping
Is Figure 6.32 shows a second-order active noise-shaping modulator based on the passive
topology seen in Fig. 6.17. Knowing that all voltages are referenced to VC.<,f (so that we
don't include VCM in the following derivations to keep the equations simpler) we can write
VI _(Vin-Voal) ._1_ (6.63)
ator. R jruC
_(VI + VOUI) . _1_ (6.64)
R jruC
Treating the comparator as adding noise to v2' Fig. 6.4,
ion we can show that VIRt
(6.65)
llator). In order to show
the I-bit digital output we can write
e the results seen in Fig.
29 shows better linearity (6.66)
Vout
R R
I
Is
-Vout Vaal
and finally
R R C
STF
NTF
v~~
h~
2
UroRC)
V out = -----=-----
2
'V rn + 2 • V Qe (6.67)
UroRC) + jroRC + 1 UroRC) + jroRC + I
The STF has a second-order response, Sec. 3.2.2. Poor selection of the resistor and
capacitor will cause instability. Note that for low frequencies the NTF increases as the R~V~
square of frequency so we expect better SNR, for a given bandwidth B (see Eq. [6.15]),
using this topology. Let's try to make this topology more robust by using only a single
feedback path, Fig. 6.33. The transfer relationship for this modified topology is then
NTF
STF ~
,.....-----A--- 2 Figure 6.
Voul ::;:; 1 'Vi" + UroRC) ,VQe (6.68)
UroRC)2 + 1 (jroRC)2 + 1 this equation reduces to, f,
Reviewing Fig. 3.35, the poles fall on the y-axis and so the topology is unstable. ofl/2,
R R
1'--+----11 VI
T NIl'
Unstable!! ! Signal-to-Noise Ratio
C ;
Following the procedure fi
R~
B
Vc
+
>-___-V"-2--; +
Vout
V;oise,RMS J
2 IN1
o
Figure 6.33 Second-order noise-shaping modulator with a single feedback path (bad).
V l.fi
Consider the single-feedback path modulator seen in Fig. 6.34. We've added a SNRideal 20 .log---!.P~-
Vnoise,Rk
resistor in series with the feedback capacitor in the first integrator. We can now write
Setting RC = 2.69/fs (notil
Vi =_(Vin~vour) 'C~C+R) (6.69) integrators, discussed in ill
STP NTF
~-~ , ~--~
2
jroRC + 1 . (jroRC) V
Voul = 'V m + 2 . Qe (6.71)
(jroRC)2 +jroRC + I (jroRC) + jroRC + I For every doubling in K
The benefits of this topology, other than the better SNR and only a single feedback loop, increase in SNRideal' To es
are that we can select RC so that the STF is closer to one over a wider bandwidth and (6.21 ),
there is smaller delay between the input and output of the modulator (which is of critical
importance in a K-path topology). Setting, for example,
RC 2Ts 2!fs (6.72) For 16 5, K '£ 1024 we can
~ed-Signal Circuit Design Chapter 6 Data Converter Design Basics 229
(6.67)
V I fi (21tRC -/,)2
iig. 6.34. We've added a SNRideal 20 . log V p == 6.02N + 1.76 20 log !80 + 20 logK5/2 (6.77)
,r. We can now write noise,RMS 80
Setting RC 2.69/fs (noting that we might have to increase the RC to avoid saturating the
(6.69) integrators, discussed in more detail in the next chapter)
SNRideai 6.02N+ 1.76-30.10+5010gK (6.78)
(6.70)
30.10
(6.79)
ADDITIONAL READIN(
Example 6.5
Simulate the operation of the second-order NS modulator in Fig. 6.34 clocked at [1] J. C. Candy and G.
100 MHz, with RC = 20 ns, and decimated with a filter having a transfer function IEEE Press, 1992. I:
ADDITIONAL READING
n Fig. 6.34 clocked at [1] J. C. Candy and G. C. Ternes (eds.), Oversampling Delta-Sigma Data Converters,
ng a transfer function IEEE Press, 1992. ISBN 0-87942-285-8
[2] S. K. Tewksbury and R. W. Hallock, Oversampled, Linear Predictive and
NOise-Shaping Coders of Order N> I, IEEE Trans. Circuits and Sys., Vol.
CAS-25, pp. 436-447, July 1978.
: in the number of bits,
[3] H. Inose, Y. Yasuda, A Unity Bit Coding Method by Negative Feedback, Proc.
IEEE, Vol. 51, pp. 1524-1535, November 1963
put clock frequency is
;.125 MHz. Using Eq. QUESTIONS
. Using Eq. (6.80) the
6.1 Suggest a topology for a passive-integrator NS modulator where the input and fed
Ig out of the filter is 9
w the lower 5-bits out back signals are currents. Derive a transfer function for your design. Does your
er 4-bits out (just not topology have the extra noise/distortion term seen in Eq. (6.12)? Why or why not?
.ply noise) .• Simulate the operation of your design .
6.2 Simulate the operation of the NS modulator seen in Fig. 6.4a but using a 4-bit
quantizer (ADC). Use a 100 MHz clock frequency and an input sinewave at 500
kHz (as used to generate Fig. 6.7).
6.3 Using Eqs. (6.14) and (6.17) compare the noise performance of passive NS
modulators using a I-bit quantizer to those using a 4-bit quantizer.
6.4 Repeat Ex. 6.2 if C is changed to I pF and a 1 GHz clock frequency is used.
Estimate the frequency where the output of the digital filter is -3 dB (0.707) from
the input signal. Verify your answer with simulations.
6.5 Suppose the comparator used in the NS modulator and filter used in Ex. 6.2 has a
50 mV input-referred offset voltage. How will this offset voltage affect the
6.3JlS
conversion from analog to digital? Verify your answer with SPICE simulations.
6.6 Figure 6.36 shows the implementation of an op-amp using mixed-signal design
5.
techniques. Assuming the comparator is powered with a 1 V supply, simulate the
circuit in the inverting op-amp confignration with the non-inverting input held at
0.5 V, a 10k resistor connected from the inverting input to the input source, and a
:lalog-to-digital converters feedback resistor of lOOk from the op-amp's output back to the inverting input (for
)r every doubling in K is a closed loop gain of -10). Set the input source to have a DC offset of 500 mV,
~ation, Fig. 6.24, replacing and a peak-to-peak amplitude of 20 mV at 500 kHz. Explain how the circuit
:du.com). The challenge is operates. Note that using an active integrator, instead of the passive integrator
:nodulator remains stable). results in more ideal behavior (less variation on the op-amp's inputs).
with moderate gains, e.g.,
.ng K-path topologies. The
es whose rising edges are
inverting input
erating these clocks and is
100 pF I Op-amp's
1 are that the matching of
output
: pulse rising/falling edges
n each feedback path, and non-inverting input
1 GHz
:he paths together. Clearly
j; however, we leave this
Figure 6.36 An op-amp implemented using mixed-signal techniques .
.amentals.
232 CMOS Mixed-Signal Circuit Design
6.7 In your own words explain why dead zones in the second-order passive modulator
seen in Fig. 6.17 are less of a problem than the first-order modulator seen in Fig.
6.4a.
6.8 Verify, using simulations, that the modulator seen in Fig. 6.20 suffers from
capacitor mismatch while the one in Fig. 6.22 does not.
6.9 What is a time-interleaved data converter? Why is a time-interleaved converter
different from the converter seen in Fig. 6.24?
6.10 Show the details of how to derive the transfer function of the path filter seen in
Fig. 6.24. Noise-S
6.11 Repeat question 6.7 if an active integrator, Fig. 6.28, is used in place of the
passive integrator.
6.12 Repeat Ex. 6.5 if K is changed from 16 to 8.
In this chapter we co
We start out discus:
topologies including SI
7.1 First-Order ~
Delt,
In VinCZ)
t +
Analog --"'"
L
Figurl
Consider what
2.2.3) as shown in the i
1.ixed-Signal Circuit Design
. time-interleaved converter
7
on of the path filter seen in
Noise-Shaping Data Converters
where, as in the last chapter, STF(z) is the signal's transfer function and NTF(z) is the
noise's transfer function.
Delta Sigma
In Vin(Z) Vaut(z) Out
! +
Integrator
Digital
Analog~
DAC
This equation is important! It shows that the input signal simply passes through the After careful review we sho
modulator with a delay while the quantization noise is differentiated (see Fig. 1.20 for the to the NS modulator represe
magnitude response of a digital differentiator with a transfer function 1 - Z-I). We can
We can use the SPII
think of the noise differentiation as pushing the quantization noise to higher frequencies.
the NS modulator of Fig. 7.
We'll come back to how NS affects the quantization noise spectral density, VQe( f), in a
a 500 kHz sinewave centen
moment. But first let's attempt to understand what's happening here.
input and output of the mod
In Fig. 7.1 the summer takes the difference (Delta) between the input signal and
the fed back signaL The integrator accumulates or sums (Sigma) this difference and feeds
the result back, via the ADC and DAC, to the summer. This forces the output of the
modulator to track the average of the input. Sometimes the fed back signal will have a
value greater than the input signal, while at other times the fed back signal will be less
than the input signal. The average signal fed back, however, should ideally be the same as
the input signal. Note that this type of NS modulator is often called a Delta-Sigma or
Sigma-Delta modulator. Also, at this point, we should see the need for the averaging
filters discussed earlier.
A circuit implementation of a first-order NS modulator is shown in Fig. 7.2. For
the moment we use a single-bit ADC and DAC (both implemented using the clocked
comparator) for gain linearity reasons (discussed in more detail later). The analog voltage
coming out of the integrator is compared to the common-mode voltage (this is our I-bit
ADC) using the comparator. For the I-bit DAC a logic-O has an analog voltage of 0 V, Figure 7.3 Sir
while a logic-I has an analog voltage of VDD I V here) so that the comparator's output
can be used directly (fed back to the DAI).
A Digital First-Order NS D(
So far our noise-shaping d
<PI
<PI ,<P2 I i~ Figure 7.4 shows a first-c
conversion. While the disc
VCM--~ I ~ ~ I reconstruction filters is giv
vd
-Voul
first-order NS demodulator.
PI""
1
~ v,.
""
: -l +
>----*------1'
Clod"d
+
00""'''''0'
:1
Digital inp~; I ~
j,
Figure 7.2 Circuit implementation of a first-order NS modulator. Clock input
The comparator is clocked on the rising edge of <PI resulting in a T, delay (Z-l) in Figure 7.
series with the fed back signal and a delay of Tj2 (z-112) in series with the input signaL
To understand this statement, remember that the nonoverlapping clock dead time (the Figure 7.5a shows a
time both <P 1 and <P2 are low) is short and, practically, the falling edge of <P2 occurs at the DAC. Figure 7.5b shows thl
same instance as the rising edge of <PI (and so we could also use (!)2 to clock the circuit and the circuit of Fi~
comparator). This results in a transfer function, see Table 2.2, to the input of the (see Fig. 1.28), and the qw
comparator (which can also be thought of as the ADC output since the fed back signal quantization by selecting (m
and modulator output are the same signal) of 011111 in two's complemeJ
-I current discussion, are VDD
Desired ADC input/output -zI _I (Vin - voua (7.3) implement a DAC using n
-z
later, is the preferred methm
ixed-Signal Circuit Design Chapter 7 Noise-Shaping Data Converters 235
imply passes through the After eareful review we should see that the eireuit implementation of Fig. 7.2 corresponds
iated (see Fig. 1.20 for the to the NS modulator represented by the block diagram shown in Fig. 7.1.
function l-z-l). We can
We can use the SPICE models developed earlier to demonstrate the operation of
oise to higher frequencies.
the NS modulator of Fig. 7.2. Assuming our sampling frequeney is 100 MHz, the input is
:ctral density, VQe( f), in a
a 500 kHz sinewave centered around Vel,! 0.5 V) with a peak amplitude of 0.5 V. The
Jere.
input and output of the modulator are shown in Fig. 7.3.
tween the input signal and
:) this difference and feeds
s forces the output of the
~d back signal will have a
~d back signal will be less
ould ideally be the same as
n ealled a Delta-Sigma or
he need for the averaging
:locked comparator
Digital inpu!
---'if--I
'L Digital
Filter
. r It
'-lS modulator. Clock input \ Interpolation filter
liting in a Ts delay (Z-l ) Figure 7.4 DAC using a NS modulator and digital filter.
~ries with the input signal.
ping clock dead time (the Figure 7.5a shows a block diagram of a first-order NS demodulator for use in a
19 edge of <1>2 occurs at the DAC. Figure 7.5b shows the practical implementation. The only differences between this
also use i!i2 to clock the circuit and the circuit of Fig. 7.1 is that the DAI is replaced with an all-digital integrator
2.2, to the input of the (see Fig. 1.28), and the quantizer (comparator) is replaced with a circuit that performs
. since the fed back signal quantization by selecting (using the MSB of the accumulator output word) digital VREF+ (=
011111 in two's complement) or VREF- 100000 in two's complement) which, for our
current discussion, are VDD and ground. Note that this topology isn't the preferred way to
(7.3) implement a DAC using noise-shaping. Using the error-feedback topology, discussed
later, is the preferred method.
236 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping.
which shows the output is a function of the first difference (order) of the quantization
· ·.·
2.1nV ....... .
1.8nV ....... .
noise vQe[nTs] - vQ.[(n - I)TJ Intuitively note that the smaller we make Ts (the faster we
1.5nV ....... .
sample since Ts = 1/f,), the closer our digital output vouJnTs] approaches the analog input
l.znv
E
vin[nTJ 0.9nV ...... ..
0.60V ...... ..
N ext, using Eq. (7.1), let's write the product of the noise transfer function and
0.3nV ..... .
vQe[z] (the modulation noise) of the first-order NS modulator in the frequency domain as O.OnV
OM.
NTF(z)VQe{Z):= (l-z-l)VQe(Z) --+ NTF(f)VQe(f) = (I_e--' f,).
i2rt
~
oJ 12js
Figure 7.1
(7.5)
where we have used, see Fig. 5.12, 7.1.2 RMS Quantizatioi
VQe(f)= ~ (units, V/JHz I forOSjSfs/2 (7.6) The RMS quantization no
..J 121s J using
B
and, once again (see Eq. [4.1] or, for the I-bit ADC [comparator], Eq. [6.1]),
V~e.RMS = 2 JINTF.
(7.7) o
Remembering that the rna;
frequency,;; , and the over
ixed-Signal Circuit Design Chapter 7 Noise-Shaping Data Converters 237
where N is the number of bits used in the low-resolution ADC/DAC in the modulator.
Using a single-bit ADCIDAC in a NS modulator, N = I, results in this book in
Quantizer
VLSB = I V. This again shows that we are not reducing the quantization noise, but are
Out pushing it to higher frequencies so that it can be filtered out. Using Eq. (l.46), we can
Analog write the PSD of the NTF as
INTF(f)12 . I VQe(f) I2 . 4 sm i
. 2 7t
(7.8)
121s is
Figure 7.6 shows the PSD of the first-order NS modulation noise for VLSB 1 V and
Is 100 MHz. Note that we are discussing modulation noise instead of quantization
noise. The modulation noise is the quantization noise after being differentiated by the NS
Quantizer
modulator. The modulation noise is the unwanted signal added to the input signal. After
MSB Out reviewing Fig. 7.6 we see that the magnitude of the modulation noise is significant.
However, after passing this signal through a lowpass filter, we can remove the higher
frequency noise resulting in a lower value of data converter RMS quantization noise,
VQe.RMS' By restricting the bandwidth of the modulation noise we can drive the RMS
quantization noise in our signal to zero. Of course, by lowering the bandwidth of the
digital filter on the output of the modulator we also limit the possible bandwidth, B, of the
input signal. Notice that we have violated Bennett's criteria, Sec. 5.1.1, by utilizing a
(b) a more detailed quantizer with an LSB that is comparable to the input signaL Now, however, we are using
feedback that adds or subtracts a signal from the input and ultimately affects the quantizer
input.
Remembering that the maximum bandwidth of our input signal is related to the sampling
frequency,!, , and the oversampling ratio, K, by
238 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping Dal
(7.14) we can solve for the SNRidea' as (knowing that the NS modulator of Fig. 7.1
(7.10)
uses a I-bit quantizer) 38.73 dB. This corresponds to an equivalent data converter
(ADC) resolution, using Eq. (5.13), of 6.14 bits (number of bits added is 5.14).
Note that the ADC is made with the components, modulator, and digital filter,
(7.11 ) shown in Fig. 7.4 .•
7.1.3 Decimating and Filtering the Output of a NS Modulator
(7.12) It's important to note that Eq. (7.14) was derived assuming the output of the modulator
was passed through a perfect lowpass filter with a bandwidth of B. Passing the output
n describe the ideal data through a Sinc averaging filter, see Figs. 4.10 and 4.13, will result in a poorer SNR
because the higher frequency noise components will not be entirely filtered out. In this
section we want to answer two questions: (1) what order, L (see Eq. [4.19]), of Sinc
averaging filter should be used in the digital filter on the output of the NS modulator, and
K 312 (in dB)
(2) assuming we use only this filter (no additional filtering), how will the ideal SNR of
the first-order NS modulator be affected.
We begin to answer to the first question by writing the increase in the number of
(in dB) (7.14) bits, Ninc' as
law every doubling in th~ N =3010gK-5.17
resolution (called simple inC 6.02
(7.15)
mpling ratio results in 1.5
If our NS modulator uses a I-bit ADC, then the final, after the digital filter, resolution of
\jRldeal)' A first-order NS
the resulting data converter is ~nc + 1 bits. (An NS modulator using a 5-bit ADC [often
; in Fig. 7.7 (see also Sec.
called a multibit NS modulator] would ideally have an output resolution of Nine + 5 bits.)
Further, we saw in Sec. 4.2.5 that the word size increased by log2K bits in each stage so
that we can require
3010gK-5.17
.I K> 3010gK 5.17
6.02 L og2 - 6.02 (7.16)
'deal
Solving this equation results in L being greater than or equal to 2. In general, we ean write
NS L= l+M (7.17)
where M is the order of the modulator. For a first-order modulator we use two stages in
the averaging filter, or,
Ie oversampling 1I
H(z) = [ _-=-.L. -KJ2 (7.18)
K l-z- 1
In the next section we discuss second-order NS modulators (M= 2). For these modulators
ersampling ratio we use a Sinc averaging filter with L = 3.
:r noise-shaping.
Example 7.2
Comment on the implementation of the digital decimation filter for the modulator
described in Ex. 7.1. Assume the final output clocking frequency is 100 MHzll6
mm signal bandwidth or 6.25 MHz.
6 of its output samples
The transfer function of the digital filter is
The block diagram of the filter is seen in Fig. 4.32 using 4 stages and L = 2. The and finally,
increase in resolution through each (I +z-I)2stage is 2 bits (see Fig. 6.10 for the
filter when L = I). The resolution calculated in Ex. 7.1 was 6.14 bits, which we
round up to 7-bits. Because the output of the digital filter is 9-bits (8-bits through
the four stages plus the I-bit coming out of the modulator), we drop the lower two This equation should be cc
bits (divide by 4) to get our final 7-bit resolution .• filter was ideal with a band
Next let's examine how filtering with a Sinc filter affects the SNR of the data NS modulator and a secone
converter. Remember the SNRideal was calculated in Eq. (7.14) assuming the modulation
noise was strictly bandlimited to B. Figure 7.8 shows the PSD of the NTF2 (f)·1 VQe(f)12
(the modulation noise) of the first order NS modulator. Also shown in this figure is the Comparing this to SNRjdeal I
shape of the averaging filter's magnitude response squared, Eq. (4.20). Here we are results in only a 2.16 dE
showing the shape of a filter with L 2 (set by Eq. [7.17] for a first-order modulator) and remember that using a Sin
K = 16. We limit our range to /, 12. SNR will be lower than wi
order modulators using Si:
4
valid the deviation from Sf.
sin (
Krr.L
/
)
7.1.4 Pattern Noise frol
111(f)12 = l. f,
f sinsin2((KnL)
4 550mV .....
V~,RMS == 12/ 8
f,l2
/' (7.20)
\
,
, 400mV'';-
2
o sin S Figure 7.9 Showi
ixed-Signal Circuit Design Chapter 7 Noise-Shaping Data Converters 241
rJ(f)l2· df 600mV,----,--,----,.--,----.:..c:;:.::L---,--~___:_-__,____,
4oomvl...~+_-t--+-+--i___;_-_T_-,___;_-....J
Ons 10nsJO~sZOns 40n5 SOns 60ns 70ns 80ns 90ns
KS) de
's Figure 7.9 Showing how filtering the output of a modulator results in ripple.
242 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping Dat
520mV
. . fit!IIII!!!I!!/ •. ·mmlj·. I ··:
..:..·..:
swing of the op-amp is beyOl
op-amp were to replace th(
~m~mjJ~l:it ....
500mV voltages less than VDD (= 1
op-amp saturation is not ne(
~80mV
~SOmV
..... . i desirable to understand how
~~OmV modulator. Note also, in Fi!
On. 200ns 400n6 600ns 800ns
change when it passes thrOUj
fed back signal, the comparat
Figure 7.10 Showing a lower frequency ripple in the modulator's output
frequency that is roughly 1/500 ns or 2 MHz, which is well within our base spectrum. The
resulting tone will lower the SFDR and the SNR of the data converter. The question now
:::l..·..·..
l.ZV ....... ".
becomes, "How do we minimize the possibility of unwanted tones appearing in the data l.OV
converter's output spectrum?" Looking at the digital output ofthe modulator we see that it 0.8V
would be better to spread or flatten the repeating data out so that we don't get repeating O.SV
tones (as discussed in the dead zone discussion in Sec. 6.1.3). Although we may still have OAV
a tone, or repeating sequence, at a frequency in the base spectrum, the amplitude of the 0,2V .
tone will be well below the VQe,RMS of the data converter (and so it won't affect the SFDR O.OV'+'---+
·0.2V ......... ..
of the data converter). In order to accomplish this spread or randomization we can add a
-o,~V
noise dither source (see Sec. 5.3.3) to our basic NS modulator, as seen in Fig. 7.11. By
-O,SV,-I---t
applying the dither to the input of the comparator (quantizer) the dither will be o.o~. 0.2~.
noise-shaped like the quantization noise (the spectral content of the dither, Eq. [5.62], is
less important).
Figure 7.12 01
VCM
>------1+ gain of the comparator is uni
+ modulator's forward gain as
Figure 7.11 Adding a dither source to a first-order NS modulator. We can rewrite Eq. (7.2) usi:
Z
Vaut(z) = - -
Finally, note that unwanted tones are usually not a problem if the input signal is l+z
busy and random (not DC as discussed in this section). Later in the chapter, we discuss
If GF approaches zero (the
second-order modulators that utilize two integrators. The second integration helps to
then the output of the mode
spread the repeating sequences out over a longer period of time so that, hopefully,
noise. (This is bad.) Since
negligible unwanted tone energy is present in the base spectrum.
difficult to filter the modula
Chapter 7 Noise-Shaping Data Converters 243
VCM Consider the linearized model of our first-order NS modulator shown in Fig. 7.13.
The gain of the integrator, see Eq. (2.103) or Fig. 2.55, is given by
~l
GI= (7.24)
CF
We have also drawn the comparator with a gain. Up until this point we have assumed the
+ gain of the comparator is unity. We'll comment on this more in a moment. Let's define the
modulator's forward gain as
(7.25)
\IS modulator. We can rewrite Eq. (7.2) using this gain as
greater than two, then, as seen in Fig. 4.38 and the associated discussion, the poles of the
YUUmV
transfer function reside outside the unit circle and the modulator will be unstable. We can
810mV
restrict the values of the forward gain to
720mV
(7.27) fj)nmV
5~OmV
Ideally, however, the gain is one. il50mV
360tuV
Example 7.3 270mV
Show, using SPICE simulations and the modulator of Fig. 7.2, that an integrator lS0mV
gain of 0.4 will result in an op-amp output range well within the power supply 90mV .......
range. OmV
O.Ops 0"
Figure 7.14a shows a schematic of the modulator with G1 0.4. Figure 7.14b
shows the output of the integrator (the output of the op-amp) in the modulator of
part (a) with the input sinew ave shown in Fig. 7.3. The output swing is limited to Figure 7.14 (a) First
roughly 80% of the supply range. For general design it is desirable to set our the outp
integrator gain to 0.4. This ensures our integrator doesn't saturate unless the input
to the modulator goes outside the supply voltage range. (precise integrator gain i~
followed by an ADC.
It's interesting to note that in both modulators, Fig. 7.2 and Fig. 7.14a, the
forward gain is unity. This is a result of the effective gain of the comparator Before leaving this
changing forcing the forward gain, controlled by the fed back signal, to unity. shaping modulator that use:
What this means is that our modulator functions as expected with a signal gain of is the input signal to the AI
one (Eq. [7.2] is valid) whether G1 is 1 or 0.4. Next we discuss how this change in range of ADC output code~
comparator gain occurs.• allowable range of modulal
Figure 7.15 shows the transfer curves for the comparator. The x-axis, the
comparator input, is the output of the integrator in our modulator. Shrinking the
integrator's output swing while holding the output swing of the comparator at the supply 1.0
rails (1 V) results in an increase in effective comparator gain. This gain variation, with the >
integrator output swing, helps to set the forward gain of the modulator to precisely I. We ~
:;
can write this using equations as 0
Integrator gain, G I Comparator gain, G c 0
'-'''~' ,
0
Integrator output Comparator(modulator) output Modulator output
(7.28)
Modulator input Integrator output Modulator input
If the modulator is functioning properly, then the average value of the modulator output Figure 7.15
will be equal to the modulator input and thus GF = I. It's interesting to note that this result
xed-Signal Circuit Design Chapter 7 Noise-Shaping Data Converters 245
:;>iz) 4>1
ADC Ip
VeM
I
Y(z) Out
Vaut
>--+----,------1 +
Clocked comparator
(a)
: forward gains,
o +=-~----'--+-----1- :>
ulator output o 0.5 1.0 Comparator input, V
Input signal
lulator input
Ie of the modulator output Figure 7.15 Comparator gain as a function of input voltage.
;ting to note that this result
246 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping I
In order to deterr
Digital requirements of the compa
output code comparator changes states.
111 Change in integrat<
110
For the modulator of Fig.
101 hysteresis is much less tha
100 Dashed lines indicate ADC gain enough so that the compar
then the modulator will 1
011 designs can be used while
010 7.1.7 Op-Amp Gain (In1
001 Now that we've discusse,
OOO~>~--------------------~
open-loop gain of the op·
o lI8 2/8 3/8 4/8 5/8 6/8 7/8 8/8 gain, some of the charge st
Analog input voltage (integrator output) the feedback capacitor, C
leakage. The charge on the
Figure 7.16 A 3-bit ADC. the feedback capacitance.
We can write the (
sizing of capacitors in the DAI). Next, notice in Fig. 7.16 that the variation in the gain of
The output voltage of tl
the ADC, with input signal, is more limited than the gains attainable with the simple
common-mode voltage Vc
comparator seen in Fig. 7.15. Limiting the range of ADC gains can result in modulator
v~ is the op-amp's inver
forward gains that are not exactly unity. This is especially true at high input frequencies
(2.102), we can rewrite Eq
where the gain of the integrator is low. However, if the integrator gain is high, the
effective gain of the ADC is not important. The point here is that using a multibit ADC
will increase the open-loop gain requirements of the op-amp used in the integrator.
7.1.6 Comparator Gain, Offset, Noise, and HystereSiS or, rewrite Eq. (2.102) to il
It's of interest to determine how the performance of the comparator influences the
operation of the modulator. Both the comparator's offset and input-referred noise, Fig.
7.17a, can be referred back to the modulator's input, Fig. 7.17b. By doing so we can
determine how they effectively change the input signal seen by the modulator. As seen in Using this result in Eq. (7
Fig. 7.17, the high gain of the integrator, A( f), reduces the effect of the comparator's gain of the modulator, GF ,
noise and offset on the input signal. For example, if the gain of the integrator at DC is
1,000 and the offset voltage of the comparator is 50 m V, then the input-referred offset is
only 50 !-lV.
In~~~fi\ /ut
~mparator
Integrator Integrator
Comparator
is ideally zero so that Eq.
gain, ClICn reduces the g
Note also that the denom
(a) (b)
term results in a data COl
Figure 7.17 (a) Referring the comparator offset and noise to (b) the input ofthe modulator. voltage that is a function
error] and frequency), but
fixed-Signal Circuit Design Chapter 7 Noise-Shaping Data Converters 247
For the modulator of Fig. 7.14 this equation can be evaluated as 0.2 V. As long as the
hysteresis is much less than this value and the gain of the comparator (110.2 or 5) is large
ed lines indicate ADC gain enough so that the comparator can make a full output transition with this input difference,
then the modulator will function properly. Very simple, low-performance comparator
designs can be used while not affecting the modulator's performance.
7.1.7 Op-Amp Gain (Integrator Leakage)
Now that we've discussed the gain of the comparator, let's determine how high the
~
open-loop gain of the op-amp must be for proper integrator action. With low op-amp
:/8 gain, some of the charge stored on the integrator's input capacitor, C1 , is not transferred to
It) the feedback capacitor, C F • This loss of charge is sometimes referred to as integrator
leakage. The charge on the input capacitance effectively leaks off when it is transferred to
the feedback capacitance.
We can write the open-loop, frequency-dependent gain of the op-amp as Aoi f).
t the variation in the gain of
The output voltage of the op-amp is then Vout=AOL(f)(v+-v_), where v+ is our
attainable with the simple
common-mode voltage VCM (the noninverting terminal of the op-amp), see Fig. 2.54, and
ains can result in modulator
v_ is the op-amp's inverting input terminal. Following the procedure to derive Eq.
ue at high input frequencies
(2.1 02), we can rewrite Eq. (2.100) with finite op-amp gain as
integrator gain is high, the
'i that using a multibit ADC
lsed in the integrator.
Q2 C/ (vCM- vout[nTs1 [T])
AOL(f) -V2 n s
(7.30)
or, rewrite Eq. (2.1 02) to include the effects of finite op-amp gain to get
comparator influences th~ Cl . VI (z) . V2(Z)
(7.31 )
ld input-referred noise, Fig. CF (I C, I
+Cr.AOLU)
)
-z
-I
7.17b. By doing so we can
,y the modulator. As seen iIi Using this result in Eq. (7.1) and, as discussed in the last section, assuming the forward
e effect of the comparator's gain of the modulator, GF , is one gives
n of the integrator at DC IS
n the input-referred offset is (7.32)
increase in the modulator's output noise (the change in the shape of the modulation noise) where, for the DAI (see Fig. 7,
we need to look at the noise transfer function including the effects of the gain error
(7.34)
or, in the frequency domain, The feedback factor is 0.714
must settle in a time, t « T)~
(7.35) this percentage using Eqs. (7.3
~
Following the same procedure used to arrive at Eq. (7.12) and assuming constant op-amp Vaul/inal
gain, AoL(f), from DC to B, results in
The output will only reach 67(
2 ViSB [ ,,2 1 (Is 13 + 2 f, ] the op-amp used has a unity g,
VQe,RMS 2· 121s . 4(1 +egain)f; . 3" \2KJ Egain ' 2K (7.36)
The largest degradation in the SNR, resulting from integrator leakage, can be estimated as In deriving Eq. (7.42)
0.5 dB if K:?: 8 (egain '" 1/8 neglecting GJ ). The minimum gain· bandwidth product of the (7.39) to determine the settli
op-amp is estimated as incomplete settling will resull
Op-amp unity gain frequencY'!un = K· B =Is 12 (7.39) output changes it will changt
the transfer function of our D,
assuming the op-amp is rolling off at 20 dB/decade at B (a dominant pole compensated
op-amp). Otherwise, the minimum gain of the op-amp can be estimated simply as the
overs amp ling ratio, K.
In order to illustrate typical op-amp requirements, let's consider the modulator of
Full, or complete, settling rec
Fig. 7.14 with K = 16 and B = 3.125 MHz (see Ex. 7.1). The fun of the op-amp is
larger than the sampling freq
estimated, using Eq. (7.39), as 50 MHz. If the open-loop response of the op-amp starts to
the required bandwidth of th
roll off at 10kHz, then the DC gain of the op-amp must be at least 5,000. However, we
error, resulting from incomp
could also use an op-amp with a DC gain of 100 (remembering low integrator gain
because the integrator is dire(
increases the undesirable effects [noise and offset] of the comparator on the performance
of the modulator topologies,
ofthe modulator) that rolls off at 500 kHz.
settling time becomes more
7.1.8 Op-Amp Settling Time integrator output must settle
linearized block diagram of
Equation (7.39) can be used, for the moment, to provide an estimate for the settling time
modulator is then evaluated t<
requirements of the op-amp in a first-order modulator. Assuming the settling time is
linear, and not slew-rate limited, we can write the change in the op-amp's output It's important to real
assuming a dominant pole compensated op-amp as slew-rate limitations. If slewi
_1_ not be a constant and will
Vaul == Vaulfinal(] (7.40) (whether a comparator follow
2rrfu . ~
Chapter 7 Noise-Shaping Data Converters 249
e of the modulation noise) where, for the DAI (see Fig. 7.18), the feedback factor is
:ts of the gain error
~= CF (7.41 )
(7.34) CF+CI
The feedback factor is 0.714 in the modulator of Fig. 7.14. The output of the DAI, vaUl'
must settle in a time, t « Ts 12), to some percentage of an ideal value, VaUl/inG'. Solving for
this percentage using Eqs. (7.39), (7.40), and (7.41) and assuming Ts l2 = t results in
(7.42)
The output will only reach 67% of its ideal final value in the modulator of Fig. 7.14 when
2 is ]
Egain · 2K
the op-amp used has a unity gain frequency of/, 12.
Ltio)
V out
:akage, can be estimated a In deriving Eq. (7.42), we used an op-amp unity gain frequency specified by Eq.
1· bandwidth product of th '
(7.39) to determine the settling response of the integrator. If the settling is linear then
incomplete settling will result in a constant DAI gain error (0.67 above). Every time the
=f,12 output changes it will change by some constant percentage of its ideal value. Rewriting
the transfer function of our DAI to include this constant gain error results in
ominant pole cOlnpem;atc:d Settling gain error, G s
)e estimated simply as CI '1 _~A.(f, jI")' VI (z) . z-1/2 - V2(z)
Z = -C . ( - e
Vout () P uni') . 1-1 (7.43)
F -z
consider the modulator
Full, or complete, settling requires that the op-amp's unity gain frequency, fun , be much
The f.," of the op-amp
larger than the sampling frequency, /, (in other words we can't use Eq. [7.39] to specify
mse of the op-amp starts
the required bandwidth of the op-amp if settling time is important). The constant gain
t least 5,000. However,
error, resulting from incomplete settling, can be tolerated in the first-order modulator
oering low integrator
because the integrator is directly followed by a comparator, as discussed earlier. In some
parator on the pel~tolm2lllcle
of the modulator topologies, though, the integrator is not followed by a comparator so
settling time becomes more important. In order to determine to what percentage the
integrator output must settle in these topologies, a gain term, say Gs ' is added to the
linearized block diagram of the modulator (integrator). The transfer function of the
:timate for the settling time
modulator is then evaluated to determine the allowable values of Gs for the application.
uming the settling time is
;e in the op-amp's output It's important to realize that we are assuming the op-amp doesn't experience
slew-rate limitations. If slewing is present, then the added gain term, in Eq. (7.43), will
1
not be a constant and will introduce distortion into the modulator's output spectrum
2nj;, . ~
(whether a comparator follows the integrator or not).
250 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping 1
In
\ .
~
first-order, the voltage cha
In addition fully-differenti:
substrate-coupled noise rej
Comparator
Integrator Figure 7.20 shows
The inputs are now differ,
Figure 7.19 The modulator's input-referred noise contributions from both the output of the integrator,
comparator and the integrator. transfer function as the sinl
Because the modulator's input-referred noise adds directly to the input signal, we
can use the derivations developed earlier in the chapter. As specified in Eq. (7.2), the
modulator's input, and thus its input-referred noise, pass through the modulator with a
delay of Z-l, If we assume the modulator's input-referred noise is white and bandlimited to
Is 12 such that
Vn,ckt(f) =: Ji for! <fs!2 (7.44)
then passing the output ofthe modulator through an ideallowpass filter with a bandwidth
Vl--
of B ( = fs/[2K] ) results in
(7.45)
Noting that not passing the output of the modulator through a lowpass filter results in an Figure 7.20 Fully-d
RMS output noise of Vn , we see that the averaging filter (the lowpass filter) reduces the
It's important to un
noise by the root of K. We could also think of the filtering as reducing the PSD of the
assume VCM 0.5 V and
modulator's input-referred noise by K. Remembering the jitter discussion from Ch. 5, we
Assuming the input is balm
see a direct parallel in the derivations of how averaging affects the RMS value of a
random signal (noise or jitter). maximum input voltage is
hand, is Vlm;n 0 I -I
range of the single-ended D
ixed-Signal Circuit Design Chapter 7 Noise-Shaping Data Converters 251
Finally, as used in Ex. 5.13, we can estimate the finite SNR of a data converter
from quantization noise, jitter, and circuit noise using
m be shown that this offset
e voltage, VCM' and thus vn.RMS I 2
" V Qe,RMS +
2
Vjiller,RMS +
2
Vckt,RMS (7.46)
resulting modulator output
:0 circumvent this problem, and
)ns affect the SNR of the where Vp is, again, the peak amplitude of an input sinewave and
.
~SD V;DA[(f). Figure 7.19.
, in series with the input Vjitter,RMS JPAVGJitter (see Eq. [5.51]) (7.48)
lth the integrator's and the' 7.1.11 Practical Implementation ofthe First-Order NS Modulator
e noise contributions from
ofthe integrator. Switched-capacitor circuits suffer from the problems of capacitive feedthrough and
charge injection. To reduce these effects, fully-differential circuit topologies are used. It
could be stated that if reasonable size capacitors and dynamic range are required,
fully-differential topologies are a necessity simply because they subtract out, to a
first-order, the voltage changes on the switched-capacitors resulting from these problems.
~ut In addition fully-differential topologies are used because they improve power supply and
substrate-coupled noise rejection and improve distortion (even-order harmonics cancel).
rator
Figure 7.20 shows the fully-differential implementation of the DAI of Fig. 2.54.
The inputs are now differential, that is, now VI VI+ - VI_ and V2 V2+ V2-, as is the
ons from both the output of the integrator, Voul = Vout~ - Voul-. The fully-differential DAI has the same
transfer function as the single-ended DAI assuming the input signals are differential.
~+----Vout+
/2
~+---- V OU /
lowpass filter results in Figure 7.20 Fully-differential discrete-analog integrator (DAl) implementation.
lowpass filter) reduces
IS reducing the PSD of
It's important to understand the signal levels in the fully-differential DAI. Let's
discussion from Ch. 5, assume VCM 0.5 V and the input voltages can range in amplitude from 0 to 1 V.
fects the RMS value of Assuming the input is balanced correctly ifvl+ = 0.85 V, then VI_ must equal 0.15 V. The
maximum input voltage is v tmax= I 0 = 1 V. The minimum input signal, on the other
hand, is V tmin 0 - I = -1 V. The range of inputs, or outputs, is then 2 V or twice the
range of the single-ended DAI.
252 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping Data
Figure 7.21 shows the implementation of a first-order NS modulation utilizing a 7.2 Second-Order Nois
fully-differential DAI. Let's simulate the operation of this modulator with the input
If we review Eq. (7.2), we mi
signals and capacitor sizes used in Fig. 7.14 (1. 100 MHz, C1 = 0.4 pF, and CF = I pF,.t:.
VQe(Z) , can result in an impTO'
= 500 kHz, and a 0.5 V peak input sinewave [the input sinewave, Vin+ Vin-, has a peak bandwidth B. The second-orde
amplitude of I VD.
quantization noise
v
Voul
The modulation noise may then
Ii
Figure 7.22 shows the simulation results for the outputs of the modulator of Fig.
7.21 after being passed through two RC filters with time constants of 100 ns. Passing a
single modulator output to the decimating filter would result in an output that is half the
input signal amplitude, which can be compensated for at the output of the filter by a
shift-left operation (multiply by two). Note that the input common-mode voltage of the
op-amp remains at 0.5 V. This is important as the design of the op-amp becomes more Figure 7.23 Comparing f
challenging if the common-mode voltage is not constant. The finite op-amp
common-mode rejection ratio (CMRR) can introduce distortion into the output of the Following the procedure usee
modulator. bandwidth B results in
1.OV,---:----,-:..>-=--:---.,---,--,----,~.::.:!.;-__:-__,
O.9V
with an increase in the SNR of
D.8V
D.N
s
O.6V Every doubling in the oversam
O.5V bits increase in resolution'
0.411 oversampling, first-order NS, ;
O.3V discussed earlier, the oversamp
O.2V
7.2.1 Second...()rder Modul.
D.1V
O.OV --i-i--+--+--i---r--tj---r----j
Consider the block diagram of
0.0115 0.2115 0.4115 0.61/5 1.6115 transfer function of this modula
Figure 7.23 shows a comparison between the modulation noise of first- and second-order
NS modulators. Notice that the modulation noise is "flatter" in the bandwidth of interest.
1.0n"',.--,----,="--,--,---,---"'="-~__,
order NS modulator.
I.UinV
fJ.<CnV'i·"··"'·"···;·""";···
0.1nll'+···,,;··..,,·,·····,
uts of the modulator of
0.2nll'+·" +."...;
(lstants of 100 ns. Passing D.1MV ,..... ;....+.. ~""
in an output that is half the ....v+--F::=....;,....-;;O;=c.:..;--f--+--+-;..---+--i
he output of the filter by
.....
tmmon-mode voltage of First-order modulation noise
: the op-amp becomes more Figure 7.23 Comparing first- and second-order NS modulator's modulation noise.
stant. The finite
Following the procedure used earlier to calculate the RMS quantization noise m a
bandwidth B results in
(7.51 )
*
""""~J""
7.2.1 Second-Order Modulator Topology
. - ..... - ~~
----t-'
Consider the block diagram of a NS modulator shown in Fig. 7.25 (see Fig. 5.39). The
liS 1.0\1s 2.0110 transfer function of this modulator may be written as
!ulator after RC filtering. A(z) 1
1 + A(z)B(z) . Vin(Z) + 1 + A (z)B(z) . VQe(z) (7.53)
254 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping I
13.3
10.0
6.66
3.33 o
o -+·-f<iL--+--+1--+1--+1..- --i")o
10 100 lk 10k K ,Oversampling ratio
+ \--4----1 A(z)
B(z)
Comparing this equation to Eq. (7.49), we can solve for the forward and fed-back circuit
VQe(Z)
Comparator
lise-shaping, M 2 A(z)
~----~ r-----~
I vou/(z) Out
ing,M
B(z) :
(a)
[= 0
~ Vout(z) Out
versampling ratio 1 -z -1
or resolution.
Vout(Z) Out
In v;n(Z)
odulator.
(c)
forward and fed-back
Figure 7.26 Block diagrams of second-order modulators.
modulator output. We can redraw the block diagram in (a), as shown in Fig. 7.26b,
resulting in the implementation of a second-order NS modulator shown in Fig. 7.26c.
The second-order (de) modulator topology of Fig. 7.26c can be used directly to
implement a NS DAC (see Figs. 7.4 and 7.5). However, this topology doesn't lend itself
directly to implementation using the DAL The major concern, as discussed in the last
section, is the op-amp's output going to the power-supply rails (integrator saturation).
This is more of a concern in the second-order modulator since the output of the first
integrator isn't connected directly to a comparator.
Figure 7.27a shows how we can add an integrator gain to the block diagram of
Fig. 7.51c without changing the system's transfer function. Figure 7.27b shows pushing
the gain, 1/G1 , through the second summer so that it is directly preceding the second
integrator. Notice that in Fig. 7.27b this (the second integrator's gain) is in series with the
lpology shown in Fig. 7.26a comparator's gain (not shown; see Fig. 7.15 and the associated discussion). This means
the differentiated, (1 - Z-l) we can arbitrarily change the second integrator's gain because the comparator gain
256 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping D
VQAz)
Comparator VCM
I
l~,
~2 ~l
I voU/(z)Out . C '
1 -Z-I
~ I ~.'V"qCM+
f~~
(a)
Figure 7.28 shows the DAr implementation of the second-order modulator of Fig. o.ova·
"j.:
.JJ.W , ....
7.27c. Note how the output of the modulator is fed back and immediately passes through .JJ.2V
C,Ops O,2~
the first integrator and is applied to the second integrator (no delay as seen in Fig. 7.27c).
This is a result of switching the phases of the clock signals in the first integrator. We
should also see how the input signal sees an added half-clock cycle delay. Note that at Figure 7.29 Sho'
this point it should be straightforward to sketch the circuit implementation of the the i
fully-differential, second-order modulator (see Fig. 7.21).
ixed-Signal Circuit Design Chapter 7 Noise-Shaping Data Converters 257
VCM
l~:
~2 4>. I Vout
-I
..,..'
~ VCM
f
':~: +
'\; Vin
cause
vn). V Qe(Z) Figure 7.28 Implementation of the second-order modulator of Fig. 7.27c.
immediately passes
delay as seen in Fig.
; in the first integrator.
;k cycle delay. Note that Figure 7.29 Showing integrator outputs in a second-order modulator with
:uit implementation of the integrator gains both set to 0.4.
258 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping Oat:
For a more quantitative view of how the gains in a second-order NS modulator are added to determine the
affect performance, let's consider a couple of different topologies. Figure 7.30 shows the schematically. Assuming we I
block diagram of the second-order NS modulator topology of Fig. 7.25, with an integrator look at the stability and for
gain coefficient, Gp and a comparator gain, Gc ' added. Deriving the transfer function of voltages.
this linearized model with G F = G I • G c results in
(7.58)
The poles of this transfer function are located at
Vvut(Z) Out 1 +z
We should make some observations at this point. Reviewing Eq. (7.27), we see
that the allowable range of forward gain, in the first-order modulator, is larger than the
allowable range in the second-order modulator. However, as long as the integrators don't
saturate (Gj doesn't approach zero), stability for either modulator is easy to attain. An Figure 7.32 Ger
analysis of the stability of higher-order modulators show that the range of allowable
forward gains decreases with the order of the modulator. For example, a third-order Notice that if G 1 = G~ =
modulator can have a forward gain of at most 1.15. Finally, notice that the input signal to Eq. (7.49). The poles of tt
range is more restricted for the second-order modulator, in order to avoid integrator
saturation, as seen in Fig. 7.29. We'll discuss methods to attain wider input signal range 2 GIG 2 G c G2G3
ZpJ,p2
and more robust stability criteria by adjusting the feedback gains later.
Notice that we are treating our modulator as a linear system even though it isn't When the modulator is func
linear; the comparator gain is a nonlinear variable. The linear approximation is useful in input, viiz) in (7.61), to
order to give an idea of the stability of the modulator under certain operating conditions.
Generally, a DC input is applied to the modulator in the simulation, while lowpass filters
ixed-Signal Circuit Design Chapter 7 Noise-Shaping Data Converters 259
~cond-order NS modulator are added to determine the average eomparator gain, Gc' Figure 7.31 shows this
~ies. Figure 7.30 shows the schematically. Assuming we know GJ (the gain coefficient ofthe integrators), we ean then
lig. 7.25, with an integrator look at the stability and forward gain of the modulator for varying DC input signal
ing the transfer function of voltages.
voAz)
}-....-+-....:
Out
Ilodulator with gains.
system even though it isn't, When the modulator is funetioning properly we require the (linearized) coefficient of the
approximation is useful in input, vm{z) in Eq. (7.61), to be unity
ertain operating conditions.
lation, while lowpass filters
I(z
G
Zpl )(z Zp2)
1-1 (7.63)
260 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping
Again, if we set Gl ::: G2 ::: G3 ::: 1 (and Gc = 1), then the poles are located at DC, that is,
Zpl,p2 0 (7.64) Vl(Z) J:ff
Equation (7.62) is useful to estimate the modulator's stability when scaling amplitudes by ~
adjusting the integrator gain coefficients, G G2 , and G3 •
j
,
G2 = C n (7.66)
Cn DA
we define our feedback gain, G 3 , as
G3 ::: Cn (7.67)
Cn Gl Cn
Figure 7.36 shows
Cn compared with Fig. 7
power supply range.
VCM: ,
VCM
1
1
SI vow
exceeds the power su
performance in the al
followed by a compar2
Let's attempt to ge
as seen in Fig. 7.31, t
"" 13 l ~ 7.37 shows how we
Example 7.4
Sketch the circuit implementation of a second-order NS modulator based on the
topology of Fig. 7.32, where G1 = G2 = G3 0.4. Comment on the stability of the
resulting configuration. Simulate the design and show the integrator output swing.
The block diagram of the modulator is shown in Fig. 7.35. We could dissect Eq.
(7.62) at this point to determine the transient properties of the modulator.
However, before discussing the transient characteristics of the modulator, let's
look at the integrator output swing. Figure 7.35 Impl
tlixed-Signal Circuit Design Chapter 7 Noise-Shaping Data Converters 261
1.._1_ I vour(z)
(7.65)
'2 1
want to implement a block
G2 = Cn
! Cn
ready defined
G3 = Cl3
Cn
(b)
Figure 7.36 shows the output swing of the integrators. This figure should be
compared with Fig. 7.29. The output of the first integrator now falls within the
power supply range. The output of the second integrator is reduced but still
exceeds the power supply range. This, as discussed earlier, has less impact on
performance in the actual transistor-based modulator because the integrator is
followed by a comparator.
Let's attempt to get an idea for the stability of the modulator by adding LPFs,
as seen in Fig. 7.31, to the simulation (with a DC input) to measure Gc ' Figure
7.37 shows how we will implement the LPFs. The voltage-controlled voltage
source is used to keep from loading the modulator with the RC circuit when it is
added into the general simulation. In our ideal modulator shown in Fig. 7.35 both
the comparator output and integrator outputs are ideal voltage sources, so we don't
) ourDAL
420mV
360mV
JOOmV
240mV
180mV
120mV
OmV
·GOmV
-120mV
·IOOmV
Gu.s lu!
Out
~I ~
~ lOOpF RC» l/fs I :;
360mV
300mI'
2AOmV
180mI'
120mV
SOmV
OmV
-SOmV
-120mI'
-180mV'+---i--i--i---i---j--i---+--r---i---i
Ous lila 2uI Jus 4ns 5us 6us 1us Bus Sus 10us
Figure 7.38 Average comparator input and output when using the modulator
of Fig. 7.35 with an input signal of 0.1 V.
hown in Fig. 7.35.
1.lY.--,.------,"-='-".--__ _ _ _ .---'-'-".--~_..,
LOY
G] 0.2
D.SY G2 0.4
This equation shows that
G3 = 1 increases by M + 0.5 bits.
D.1V -
D.SY Mih_Order Modulator TOF
n.sv
O.4V Reviewing the general N:
O.3Y forward transfer function,
0.2V
NS modulator. The transfe
D.W
O.W
-o.lV+--t---i---r--i----i---i--i---+--j---j
D.Ops O.2ps O.4ps O.6ps 0.8115 1.Ops 1.2115 1.~1i. 1.61'S 1.811S 2.0ps
Using this equation toge'
Figure 7.40 Integrator outputs for a modulator with first integrator gain of 0.2.
function of
This equation shows that for every doubling in the oversampling ratio, K, the resolution
increases by M + 0.5 bits.
M1h-Order Modulator Topology
Reviewing the general NS modulator topology of Fig. 7.25, we want to determine the
forward transfer function, A(z), and the feedback transfer function, B(z), for an Mth-order
NS modulator. The transfer function of a general Mth-order modulator is
(7.76)
Using this equation together with Eq. (7.53) results in a forward modulator transfer
ltegrator gain of 0.2. function of
A(z) (7.77)
;, (2) NS modulators using The block diagram of an Mth-order NS modulator is shown in Fig. 7 AI. Note, as we'll see
) ) cascaded modulators shortly, it's impossible (in an analog-to-digital converter) to implement this topology
cond-order modulators). using only one delaying integrator.
266 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping
Increase in
2 JMt-lI2 M
:grator VQe,RMS = {1";:;' [ -K . IT 2 (7.84)
..; 12 m=! m
oI.i-_---.
The change in SNR, when using the Sine averaging filter decimator instead of the ideal
filter with bandwidth, B, is given by looking at the ratio ofEq. (7.72) to Eq. (7.84)
For first-, second-, and third-order modulators, the difference in the SNRs is 2.16, 6.35,
NS modulator. and 10.39 dB, respectively. This shows that using a Sinc averager, theoretically, increases
the SNR if we neglect the decrease in the desired signal amplitude because of the droop
tor in the Sine-filters response, Fig. 4.17. To avoid the droop, as discussed earlier, the desired
signal content is often limited to frequencies well below fJ2K B). When the droop
$ that the number of Sine (reduction in the desired signal amplitude) is taken under consideration, the SNR, when
e modulation noise, is one using the Sine averaging filter, is worse than the ideal filter with bandwidth B.
q. (7.19)
7.3.3 Implementing Higher-Order, Single-Stage Modulators
H(f)12 ·df The single-stage, higher-order modulator of Fig. 7.41 can be difficult to implement
directly. It is impossible to implement a higher-order modulator, when using DAIs, where
all but the last integrator are nondelaying. However, as we saw with the second-order
modulator using two delaying integrators in Fig. 7.32 and Eqs. (7.61) and (7.62), the
stability criteria of a modulator using only delaying integrators is poorer than the criteria
of the topology shown in Fig. 7.41 (where only the last integrator is delaying). While we
(7.80)
can help the situation by staggering delaying and nondelaying integrators in a modulator,
the point is that implementing a higher-order modulator without modifying our basic NS
!lg topology will result in an unstable circuit. Intuitively, we can understand this by noting
that if the modulator's forward gain is too high and the delay through the forward path is
L) ]2(M+Il too long (because of the large number of integrators), the signal fed back may add to the
input signal instead of subtracting from it.
~) ·df In order to help with the stability of a higher-order modulator a topology that
feeds the input signal forward into additional points in the modulator (thereby reducing
the forward gain and delay) and feeds the output signal back as discussed earlier
(allowing scaling of amplitudes) is needed. In order to move towards this goal, consider
l(K1f.L) the modified NS topology for higher- order modulators shown in Fig. 7.42. The forward
~..::-Is-.df and feedback transfer functions can be written as
(7.86)
M
1t·IT2;:1 (7.87)
m=1
and
-A(z)B(z)
rder modulator followed by
(7.88)
268 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping Dat:
or
so that the feed forward coefficients, ax ' can be used to help make the STF(f) constant
over the region of interest (the STF can be made to have an overall lowpass response).
The NTF is given by
I
NTF(z) = 1 +A(z)B(z) (7.92)
or
NTF(z) = M
HPF(z) (7.93)
1-(z-1)-M. L: hi' (z_I)1-1
i=l
NS modulator.
The coefficients, bx ' are selected for a highpass response. Note also that our coefficients
are positive since the feedback paths, as seen in Fig. 7.42, are subtracting. The design of
(7.89) the modulator is performed by determining the feed-forward and feedback coefficients
using basic digital-signal processing filter design (and, to keep the algebra simple, a
19 to do with the modified, computer program of some sort), then to simulate the design to see if it exceeds
le NTF(z), for a general specifications. One challenge, among others, is to meet a given SNR without causing
7.43. At high frequencies harmful integrator saturation.
the magnitude of the noise
7.3.4 Multi-Bit Modulators
Ie modified NS modulator
Throughout this chapter we have assumed N I; that is, we have used a comparator for
s by changing the shape of our quantizer in the forward path of our NS modulator. The main advantage of single-bit
modulators, as discussed earlier, is the inherent linearity of the I-bit feedback DAC.
J Feedback DAC linearity is important because the output of the DAC is directly subtracted
=HPF(z) from the input signal. Any distortion or nonlinearity (or noise) in the output of the DAC
will directly affect the modulator's performance and, ultimately, limit the modulator's
mented with the feedback SNR. The benefits of using a multibit (N > 1) quantizer in a NS modulator are increased
nodulation noise (keep the SNR (see Eq. 7.73), better stability (the modulator behaves closer to the linearized theory
) at high frequencies and developed in this chapter), fewer spectral tones, and simpler digital-decimation filter. The
mique is that the signal no drawbacks of using multibit topologies, are the increase in ADC complexity (the ADC
the lowpass response. The must be a flash converter) and the need for the DAC to be accurate to the final accuracy
of the modulator. The ADC errors, like gain errors in the integrators, are less important
since they are in the forward, high-gain path of the modulator.
Simulating a Multibit NS Modulator Using SPICE
Figure 7.44 shows a circuit-level implementation of a first-order, multibit, NS modulator
using a 4-bit ADC and DAC. Figure 7.45 shows the SPICE simulation outputs of this
An example shape of a modulator. Most of the design effort, when developing multi-bit modulators, goes into the
higher-order NTF.
j design of the feedback DAC. Because it is nearly impossible to design highly accurate
-
LPF(z)· (1 Z-I)M
(say l2-bits resolution or better) DACs without trimming, or some sort of error
correction, methods have been developed that attempt to randomize DAC errors. If the
gain at high frequnencies) errors appear as a random variable, they may appear as white noise in the output spectrum
and not affect the SNR of the data converter.
Figure 7.46 shows one possible implementation of a DAC. This DAC utilizes
resistive unit elements, resistors laid out in a square that connect, on one side, between
VDD and ground while the other side connects to a rotating switch connected to the
'-order modulator. analog output. In other words, in one case we connect VDD to one comer of the resistor
270 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping D
~l ~2 VREF+ == VDD
Jp ~lClk VDD
_\ ~ ~
VallI
I
f\
VCM
T
O.4p
~
In
VREF =: 0
In
Is == 100 MHz '-------' 3·bit DAC inputs
bo
Figure 7.44 Circuit implementation ofa first-order multi-bit NS modulator. bl
b2
cube and ground to the opposite corner (assuming VREF+ = VDD and VREF 0). There
exist two voltage dividers along each of the sides of the resistor square. The output of the
DAC can change from zero, to (l/8)VDD, to (2/8)VDD, ... up to (8/8)VDD. Depending on
~o
Clock
the output of the decoder, one tap from each side is fed to the analog output. Because 1,9
U
there are two sides, the outputs from each side are combined and effectively averaged.
990mV'~-,-------,:..!.:-"""-,.---,---:--........,,----,--=""T---,.---,
900mV
810mV Connected to VDD wh(
720mV .. and connected to groun
tilOmV
~ 4-bits
VREF 0
)it In
\C r-----,f----'
Analog
i-bit NS modulator. Output
VDDbus
Ground bus
Connected to VDD when counter output is 0
and connected to ground when counter output is 8
Quantizer
Quantizer
Figure 7.47 Block diagram of an error feedback modulator. Figure 7.49 Show
(7.95)
A second-order modulator with a NTF(f) (I_Z-I)2 would use a feedback filter,
noticing from Eq. (7.94) that NTF(f) = 1 - F(z) , of
F(z) = 1-(1 (7.96)
Implementation of a second-order NS modulator is shown in Fig. 7.48. Note that when
trying to implement higher-order modulators using error feedback we run into the same Figure 7.5(
problem we encountered when using an interpolative modulator, namely. instability
resulting from a NTF that is too large at higher-frequencies. As with interpolative The number of bits u
modulators, we can design the NTF to have a high pass response. the maximum input signal
complement numbers, the we
word's MSB is used to incr
We'll comment more on this
In Vin(Z) VOIlI(Z) Out Figure 7.51 shows th(
if a I-bit output is used, the
filter (RCF). The I-bit DAC
Quantizer
multibit modulator and DAI
modulator order, as well as e
discussed earlier, is that the
since it is in series with the 0
Figure 7.48 Block diagram ofa second-order error feedback modulator. Interpolation fil
Digital input Digital
We've introduced the error feedback topology with the idea that it can be used in a filter
modulator (demodulator) that performs digital-to-analog conversion. We first introduced
a modulator for use in a DAC back in Fig. 7.5. At this point we need to answer the
question, "Why is the NS topology of Fig. 7.47 a better choice for DAC implementation,
in general, then the topologies of Figs. 7.5 and 7.41?" The answer to this comes from the Figure 7.
realization that the quantizer and difference block in Fig. 7.47 can be implemented by
simply removing lower bits from the digital input words. This is illustrated in Fig. 7.49. Let's look at how to
The resulting error feedback modulator will be simpler to implement than the modulators error feedback quantization
based on interpolative topologies. Figure 7.50 shows Fig. 7.47 redrawn to show the we know that
simpler implementation.
ixed-Signal Circuit Design Chapter 7 Noise-Shaping Data Converters 273
Vout(Z) Out
Nbits Out voutCz) Out
N bits N F bits vou/(z)
N-Fbits
~uantizer
Quantizer F
__ b_it_S---t:~'-)-<E----It ) /
r (
(7.95)
·uld use a feedback filter, Nbits Vout(Z) Out
N-Fbits
(7.96) Fbits
1 Fig. 7.48. Note that when
iback we run into the same Figure 7.50 Block diagram of an error feedback modulator.
dulator, namely, instability
:ies. As with interpolative The number of bits used in the modulator, N, is selected to avoid overflow when
se. the maximum input signal and fed-back signal are subtracted. When using two's
complement numbers, the words input to the adder must be the same length. The smaller
word's MSB is used to increase the smaller word's size until the word lengths match.
We'll comment more on this important concern in a moment.
Vout(Z) Out Figure 7.51 shows the block diagram of an NS-based DAC. As we saw in Fig. 7.4,
if a I-bit output is used, the modulator can be connected directly to the reconstruction
tilter (RCF). The I-bit DAC is perfectly linear so distortion concerns are reduced. Using a
Quantizer
multibit modulator and DAC gives a better SNR, for a given oversampling ratio and
modulator order, as well as easing the requirements placed on the RCF. The drawback, as
discussed earlier, is that the DAC must be accurate to the final desired output resolution
since it is in series with the output signal path.
I LSB
Nbits ,.---------A-- 7.3.6 Cascaded Modula
r~
VREF+ - VREF
0111111... == VREF+ (7.97) The NS modulators discus!
2N
loop topologies with the g
and topologies discussed in Sec
this section we discuss cas'
lOOOOOO... "" VREF (7.98)
discussed here are sometiI
For the DAC to function properly we must change the numbers back to binary offset While our focus in this sec
(complement the left-most or most-significant bit) unless the DAC input uses two's to modulators used in DAC
complement format. This is easy to see if the output of the modulator is a single bit
We indicated, in th,
(N-F 1) since an MSB of I == VREF+ and a O::!'-VREF- where VISB == VREF+ VREF-. By
to the input isn't practical il
dropping F bits the voltage weighting of an LSB in the modulator output can be written
analog subtraction, Vo.(z),
as
VQe(z) back to the input, (
(7.99) modulator. The second mO(
VLSB
own unwanted modulation
modulator, we can effectiv(
This result is used in Eq. (7.6) to estimate the quantization noise spectrum in a NS
modulator. The major benefit c
and second-order loops c
Implementation Concerns
addition, as we'll briefly di
We know from our discussions in the last chapter that most digital additions and I-bit ADC and DAC follov
subtractions utilize two's complement numbers because of the simplicity (see Sec. 4.1.3 with low oversampling rati,
and the associated discussion) in implementing the hardware. However, consider the
Second-Order (1-1) Modul,
two's complement N-bit input in Fig. 7.49. If we drop the lower F bits, the resulting
number fed back to F(z) (the quantization noise) is not in two's complement format. A second-order NS modu
modulators (called a I-I
In order to circumvent these types of problems, the topology shown in Fig. 7.52
modulator is given by
can be used. The input to the quantizer/subtractor is changed from two's complement
format into binary offset format. (See Figs. 4.4 and 4.5 for a comparison of the formats.)
Quantization is then performed; the lower bits are dropped from the output and fed back.
while the output of the sec(
The fed-back word (the quantization error) is then changed from a binary offset number
back into a two's complement number. The size of the word fed back is adjusted to match
the size of the modulator's input (knowing that the words used in two's complement
arithmetic must be the same size so that the sign bit is in the same location in each word,
see also Fig. 4.7).
/
Binary offset
Adjust the output word size so that it matches the modulator's input word size.
In vm(z)
N - Fbits Vout(z) .
Binary offset
Jits
ary offset
ut word size.
erence blocks.
Figure 7.53 Second-order (1-1) cascaded modulator.
276 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping Data Cor
(7.106) z
1
While we can set the second modulator's integrator gain coefficient, Gn> to 0.4 to avoid
integrator saturation, as discussed earlier, we must set GIl as close to unity as possible.
Using a unity gain coefficient results in a reduction in the modulator's overall dynamic
range (see Fig. 7.12 and the associated discussion). Note that the input signal appears in Figure 7.54 .
the unwanted term in Eq. (7.106). It should be obvious at this point that we can add
scaling parameters at various points in the modulator to attempt to maximize the
modulator's dynamic range. Also note that the number of bits in the 1-1 modulator's Third-Order (2-1) Modulators
output will be more than one bit (two bits if comparators are used in each first-order A third-order modulator formed
modulator). order modulator is shown in Fig.
vfixed-Signal Circuit Design Chapter 7 Noise-Shaping Data Converters 277
'Il)' Z-1
-1)Z-1 . VQel(Z) (7.104) .
'Qelout(Z) + (1
Unwanted tenn
(7.106)
:fficient, GI2 • to 0.4 to avoid
(7.1 09)
. _.. __ .. _-------------------------
278 CMOS Mixed-Signal Circuit Design Chapter 7 Noise-Shaping Data (
When this equation is compared to Eq. (7.106), we see that the undesired term is second
order differentiated. Also, we have more control over the integrator gains. Third-order
modulators using the 2-\ topology are much more robust than the I-I-I-based topology
and can provide output signals free of unwanted tones. Again, if integrator saturation (and
thus dynamic range) isn't a concern, then we can set G I G2 = I .
One of the interesting uses of the 2-1 modulator is the configuration where the
first (second-order) modulator utilizes a I-bit ADC and DAC, while the second
(first-order) modulator utilizes a multibit ADC and DAC. The overall linearity of this
topology is dominated by the second-order modulator, while the multibit modulator
provides an enhancement in dynamic range for a given oversampling ratio. These very
interesting data converters are discussed in greater detail in [7].
Implementing the Additional Summing Input
Before leaving our introduction to cascaded converters, let's discuss the implementation
dulator.
of the extra summing block used to generate the quantization noise, VQ.(z). Figure 7.56
shows the topology ofthe two summing blocks and how they can be combined.
(7.110) One way to implement the extra subtracting input and the integrator is shown in
(7.111)
put by first determining the VI (z)
ator). We'll use the topology
,tput, o/z), is (assuming that I--~-- Out to integrator
'---<E---- V2(Z)
~I] 'Z~I VQel(Z)
(7.112)
'hG c )
Figure 7.56 Showing implementation of the dual summing block as a single block.
[. [7.61])
Fig. 7.57. This DAI is a modification of the DAI shown in Fig. 7.33. The output of this
integrator is related to the inputs by
.-1
VQel (z) Cn Z~112 Cn 1 Cpo
(7.1l3) VOlt'(z) OI(Z)' - . - - V2(Z)' - . - - Vl(Z)' -=-. (7.116)
c) Cn l-z~1 Cn l-z~1 Cn 1
, we write the output of the If we set Cn = Cm '" Cn and we realize that the comparator in the second modulator,
assuming it is clocked with the rising edge of <P I (or the falling of <P2), adds a
half-clock cycle delay in series with the v1(z) input and a full clock cycle delay in series
:z) + (I
with 01(Z) and v2(z), then we can write
(7.114)
Vour(Z) [0 I (z) VI(Z)-V2(Z)]' (7.117)
1
Figure 7.58 shows the implementation of a 2-1 modulator.
We could also use the topology shown in Fig. 7.59 to implement the summing
block of 7.56. This topology has the benefit of using a single capacitor for a simpler
circuit and no matching differences between C n and . Unfortunately, the topology is
Z-I] . z-I VQel (Z)]
(7.115) no longer insensitive to the parasitic capacitance on the top plate of the switched
'c) capacitor. In the parasitic insensitive topologies, Fig. 7.57 for example, the top plate of
280
<1>1
VeM
VI(Z) -~.- .............
01(Z)---
VI (z) _ _ _ _ _ _ _ _ _ _ _ _----1
Figure 7.59 1m
res
Figure 7.57 Implementing the dual summing block for a cascaded modulator. ADDITIONAL READl
[I] R. Schreier and
the capacitor is always held at the common-mode voltage, VCAr In the topology of Fig. Wiley-IEEE Pres
7.59 the top plate is charged to yJ(t) when the <1>1 switches are closed and discharged to
VeM when the <1>2 switches are closed. The difference between these voltages combined [2] S. K. Dunlap an
with the value of the unwanted parasitic capacitance to ground on the top plate causes Delta-Sigma Mo
unwanted charge to transfer to the feedback capacitor and a gain error. This by itself isn't Papers, Vol. 51,
too bad. However, the unwanted capacitance can have a large depletion capacitance [3] A. Eshraghi ane
component, resulting in a voltage-dependent capacitance and thus nonlinear gain. ADC Architectu
Nevertheless, in some applications this topology may still prove useful. Theory and Appi
[4] S. R. Norswort
Converters: ThE
978-078031045:
VI (z) R. T. Baird and
[5]
Converters usinl
Systems II: Anal
December 1995
Vt(Z) - - -
Ot (z)
+
Figure 7.59 Implementing the dual summing block with a single capacitor
results in sensitivity to the top plate parasitic capacitance.
7.4 Show the spectrums (modulator input, digital output, and analog output after
7.25 In your own words, h
filtering) of the signals in question 7.3. Discuss what the spectrums indicate.
the effects of a jittery
7.5 If an extra delay, Z~l, was added to the forward path of the modulator in Fig. 7.2
7.26 Determine the transfe
7.7 For the NS modulator shown in Fig. 7.5 used for digital to analog conversion,
what component serves as the ADC? What component serves as the DAC?
7.9 What are we assuming about an input signal if the modulation noise follows Eq.
stability of the resultil
(7.5)?
7.11 What is the difference between quantization noise and modulation noise?
7.13 Is the statement on page 238 that "every doubling in the oversampling ratio results
7.34 Comment, in your (
7.17 In your own words, describe ripple in the output of a digital filter connected to an
NS modulator.
7.38 What transfer functio
7.18 Does adding a dither signal to the input of a NS modulator help reduce the
peak-to-peak ripple in the digital filter output? Does it help to break up tones in
X(z)
7.20 Repeat Ex. 7.3 if the integrator's gain is set to 0.5.
7.21 Estimate the range of Gc for the quantizer seen in Fig. 7.16. How does this
compare to the range of G, for the I-bit quantizer seen in Fig. 7.15? Name two
F
benefits of the I-bit quantizer over multi-bit quantizers.
7.23 In your own words, and without equations, describe integrator leakage. How
7.40 Derive the transfer n
would you relate integrator leakage, found in integrators that use an active
derivation). What is
7.24 Would large parasitic op-amp input capacitance affect the settling time of a DAI?
circuit's output (relm
Verify your answer using simulations with ideal op-amps (infinite open-loop gain)
good idea, now that
and non-ideal op-amps (open-loop gains around the oversampling ratio, K).
are tied to ground or
, and analog output after 7.25 In your own words, how does oversampling affect input-referred offset/noise and
: spectrums indicate . the effects of a jittery clock on an NS data converter?
. the modulator in Fig. 7.2 7.26 Determine the transfer function of the DAr shown in 7.20 .
.?
7.27 Derive Eq. (7.51).
7.28 Sketch the implementation of the full-differential second-order NS modulator.
;ital to analog conversion,
7.29 Derive Eq. (7.61)
:erves as the DAC?
7.30 Sketch the fully-differential equivalent of Fig. 7.33.
7.31 Resimulate the modulator in Ex. 7.4 if the gains are set to one. Comment on the
dulation noise follows Eq.
stability of the resulting circuit.
7.32 Resimulate the modulator in Ex. 7.4 if the input is only 50 mY. Comment on the
ency)?
stability of the resulting circuit.
flOdulation noise?
7.33 Regenerate Fig. 7.40 by selecting integrator gains so that the maximum output
swing of any op-amp is 800 m V peak-to-peak.
oversampling ratio results 7.34 Comment, in your own words, on why the actual SNR of a NS-based data
I? Explain. converter can be worse than the ideal values calculated in the chapter.
)w? 7.35 Derive Eq. (7.75). Make sure each step of the derivation includes comments.
7.36 Resimulate Fig. 7.44 using two-bit ADC and DAC.
;tADC? 7.37 Sketch a possible implementation of a quantizer for the crror feedback modulator
shown in Fig. 7.48.
gital filter connected to an
7.38 What transfer function does the following block diagram implement?
10dulator help reduce the
help to break up tones in
shift left
X(z) Y(z)
Bandpass
Figure 7.61 Circuit used in question 7.40.
7.41 Repeat question 7.40 for the op-amp circuit seen in Fig. 7.62.
Chapter
8
Bandpass Data Converters
.40.
;.7.62.
The data converter topologies we've covered in the last few chapters have performed
analog-to-digital conversion on a range of frequencies extending from DC to some
frequency, B. These topologies are sometimes called lowpass data converters. In this
chapter we develop the idea that we can perform data conversion on a bandwidth of
frequencies that doesn't include DC. The topologies developed in this chapter are called
bandpass data converters. Bandpass data converters are useful in communication circuits,
+ especially wireless communications, where the information is modulated up to some
higher frequency and thus contained in some fixed bandwidth in the frequency spectrum.
Before developing the idea of a bandpass data converter, let's review, on pages 4
41. through 6, why in-phase (1) and quadrature (Q) signals are used in communication
systems. In simple terms, IIQ signals are used because we can send more information
without increasing the bandwidth of the transmission channel. Note that other schemes
that use several phase-shifted sinewaves (e.g., 0, 30, 60, and 90 degrees) may also be
employed to further increase the information transmitted without increasing the channel
bandwidth. In this chapter, however, we focus only on 0 and 90 degree phase-shifted
(IIQ) sinewaves.
As seen in Fig. 1.6 and Eq. (1.11), summing I and Q components results in the I1Q
signal. Figure 8.1 shows time-domain, equal amplitude, I and Q signals and their sum (the
I1Q signal). The phase shift of the I1Q signal leads the Q component by 45 degrees while
lagging the I component by 45 degrees when equal amplitude I and Q signals are used. By
varying the amplitude of either (or both) the I or Q signals we vary the phase shift and
amplitude of the IIQ signal (this variation in amplitude and phase shift represent the
information we are transmitting). Note that this modulation scheme is often called
quadrature amplitude modulation (QAA-1).
In order to recover the information from the I and Q components in the IIQ signal
let's first write, see Eq. (1.10),
SIQ(t) =At· cos2rr.j~ . t+AQ' sin2rr.j~· t (8.1)
where j; represents the frequency of a carrier signal while AI and AQ represent the
information we are transmitting. Both AI and AQ vary with time and thus have some
-~~------~-----------------------------
8.1 Continuous-Time Be
aOOml"..,-~--~-==':'-----~--=':~'------,----,-'--"-.,------.
I/Q
100mV When we developed the (Iowpa:
.OOmV
SOOml'
used a capacitor to sum (sigma)
400mV .- .. .
300mI' .... . I
signals. The result was an abser
200mI'
Q DC where the impedance of the
100mV
Oml'
-SOOml'
fo 1I2n!IC .
-60DmV
-700mY
8.1.1 Passive-Component B
On$ Figure 8.2a shows the impleme
modulator using this approach.
Figure 8.1 In-phase (I), quadrature (Q) signals, and IJQ signals. use Fig. 8.2b. We ean write
topologies)
spectral representation, AI(!) and AQ(f). The AI component amplitude modulates the
cosine term while the AQ component amplitude modulates the sine term.
In order to recover the AI and AQ in the receiver we can multiply the received
signal, ideally a sealed version of SIQ(t) , by cosine and sine signals. For Ai' and
[A +B]) (8.3)
5[A +B]) (8.4)
sin2n2[c' t (8.5)
lowpass filter
(a) Circuit implementation of a bandpass modulator.
through a lowpass filter to
ed the maximum frequency
VQe(f)
:iply by a sinewave or Vin - Vinl
[A + BD (8.7)
sin2n2[c' t (8.8)
(b) B lock diagram
>wpass filter
or
I +s2LC
vOU! = vm . - - : - ' . ' . - - + VQe(f)· - Vin!' - - - - - - ' ' - - - (8.11 )
+s2LC 1 +s~+s2LC 1 +s~+s2LC
and tinally (see Fig. 3.39 and then Eq. [6.12] for the analogy with the lowpass passive
modulator)
(8.12)
The STF shows a bandpass response while the NTF shows a band reject response, Fig.
8.3. Like the lowpass modulator we have the extra noise/distortion term that causes the Figure 8.4 Simulati,
modulator to deviate from the ideal behavior. We know, at this point, that we can drive
V in , to zero by using an active element (amplifier). This is discussed in the next section.
Figure 8.3 Modulation noise spectral density for a bandpass modulator. ·140dB+----i-
20.0MHz 21.4MHz 2
Example 8.1
Figure 8.5 SPI
Simulate the operation of the NS modulator seen in Fig. 8.2 when R is lk, Cis 10
pF, L is 4.06 I-LH, and the clocking frequency is 100 MHz. Comment on the An Important Note
resulting simulation results and the operationllimitations of the circuit.
Notice that the value of the indm
The conversion is centered around, from Fig. 8.3, 25 MHz. What this means is if modulator. If one understands th
our input signal is a sinew ave at 25 MHz we should be able recover an exact replaced with an active circuit, I
replica of this input after removing the modulation noise. If a DC signal is applied of frequencies (the range we w,
to the modulator we should get out no signal (actually the modulator should representation).
output a sequence, like 101010101 that averages to VCM or, for a mixed-signal
system, no signal). 8.1.2 Active-Component Bal
Figure 8.4 shows the simulation results. Note the ripple in the output, after To remove the extra noise/distor
filtering, amplitude. This variation can be removed by more filtering (increasing 6.2.4, an active circuit. Figure
the Q of the simple filter used in the simulation by increasing the resistor) to variation in vim is zero when tht
remove noise. It is useful, when learning, to vary the input frequency in the this topology as
simulation and look at how the output of the modulator changes (keeping in mind
the frequency response, Fig. 3.39, of the LC tank circuit). Finally, Fig. 8.5 shows
the spectrum of the modulator's output when a 25 MHz input tone is applied.
Again, as in the passive noise-shaping lowpass topology, the extra, unwanted,
term in (8.12) limits the SNR. •
\1ixed-Signal Circuit Design Chapter 8 Bandpass Data Converters 289
Modulator output
(8.12)
Input signal
>
f
mdpass modulator.
:.2 whenR is lk, Cis 10 Figure 8.5 Spectrum of the modulator's output in Ex. 8.1.
MHz. Comment on the An Important Note
If the circuit.
Notice that the value of the inductor we used in Ex. 8.1 is too big to be integrated with the
'z. What this means is if modulator. If one understands the concepts in this section we see that the LC tank can be
e able recover an exact replaced with an active circuit, like a gm-C filter, that has a high-impedance over a range
If a DC signal is applied of frequencies (the range we want to convert from an analog representation to a digital
'I the modulator should
representation).
lor, for a mixed-signal
8.1.2 Active-Component Bandpass Modulators
}ple in the output, after To remove the extra noise/distortion tenn in Eq. (8.12) we can use, as we did back in Sec.
lore filtering (increasing 6.2.4, an active circuit. Figure 8.6 shows one such implementation. Noting that the
~reasing the resistor) to variation in V inl is zero when the op-amp's gain is infinite we can rewrite Eq. (8.12) for
input frequency in the this topology as
langes (keeping in mind STflf) lvTF(f)
Finally, Fig. 8.5 shows
l input tone is applied. --~-
Vout ,VQe(f) (8.13)
y, the extra, unwanted,
290 CMOS Mixed-Signal Circuit Design Chapter 8 Bandpass Data C
vln-;-_ _ _ _- J
Analog
R
R
~ I :
ADC (comparator)
+ Digital
R~~
Vinl
V out
VCM +
Op-amp j,
I
DAC
-Vout
Simulations demonstrating the operation of this topology are found at CMOSedu.com. 8.1.3 Modulators for C(
/
L L
R
ADC (comparator) vin----N, --.-_, I \ I \ I
CMf>f,
-vr I
igital
Vout >--+-V:...;2'--i +
VCM +
f, VCM
VC_M_~
Antenna VI(Z) ~
L
V2(Z),_ _ _ _ _ _ __
ADC
L -_ _ _ _~~------~+
>---<>---- Digital
LNA is VI (z)
VOUI(Z) = ---
Figure 8.9 Design of a bandpass modulator for data conversion at RF.
VeM
~c
>---'0--- Digital
" - - - - - - - - - - - - - - - - < - 1 r-------.J
I conversion at RF.
(V Input
OV-=-~'Th.J
Note the inversion in the signal
-(V
V2 = VCM signal see Fig. 3.6 and, for a
-1IV ••.••.. : ••••••. ; .•......:.......•••••..•..•.. ; ............••••......•.
, , 1 + z-2 , has the shape seen in F
"
" ,,
·12V ,
..,, , , ,
.." we move away from is 14. Note
- - - - -_. - - - - - - - j. - - . - .'. - - -
,
- -". - - -_ •• - -' - - - -~- - - - - - - - .' - - - - - - - -
·16V
.
"
"
---~
, , ,
Vout(Z) =
VI (Z) . Z-1I2
I
- - - - --, - --- - - j. - - - ---' -- - - - - - - _. - - - - - - '. - -- - - .'. - - - - _.'- - - - - - - - - - --
Example 8.3
Suggest a modification to the resonator in Fig. 8.10 to eliminate the need for the
gain of -1 block. Verify your circuit modification with SPICE.
Reviewing Fig. 2.56 we see that if we connect the fed back output signal to the
top plate of the 2CF capacitor, Fig. 8.12, then the signal is inverted when it's Figure 8.13 Block
transferred to the output of the resonator. Simulation files for this circuit verifying
correct operation are available at CMOSedu.com. Note that we'll use the topology Example 8.4
seen in Fig. 8.10 since it is more tolerant of charge injection errors on the 2CF Simulate the operation of tJ-:
capacitor than the topology seen in Fig. 8.12 .• modulator's output spectrurr
out of the modulator at a r:
input frequency of 25.1 MH
an ideal digital bandpass fil
<1>1 <1>2 then estimate the final SNR h
,
.,,, , ,
'
''
-20dB -----T----------,----------,---_·
, , , -
·30dB
Figure 8.12 Implementing a resonator for use in a bandpass modulator.
·40dB
8.2.2 Second-Order Modulators
-50dB ................. ..
Figure 8.13 shows the implementation of a second-order modulator based on the
first-order lowpass topology seen in Fig. 7.1. Remembering, from page 290, that a ·60dS+------1---j---i----'
0.1 104Hz 7.1104Hz 14.1104Hz 21.1Io4Hz 2B
second-order bandpass modulator's SNR is similar to (calculated in the same way as) the
first-order lowpass modulator's SNR we can write Figure 8.14 Spectr
Mixed-Signal Circuit Design Chapter 8 Bandpass Data Converters 295
(8.20)
~Output
or, assuming G 1 = I (as discussed in Sec. 7.1.5),
STF NTF
.--Input (8.21 )
Note the inversion in the signal transfer function. This is trivial to remove; for an analog
signal see Fig. 3.6 and, for a digital signal, see Fig. 6.3. The noise transfer function,
1+ ,has the shape seen in Fig. 8.7. Atfsl4 the modulation noise is zero increasing as
VI (z) . z-1I2 we move away from j~ 14. Note that this topology is second-order because the number of
1 +z-I poles in the NTF is two (a first-order lowpass modulator has one pole in its NTF).
Vout(Z) Out
landpass modulator.
(8.23)
Again, as discussed in Sec. 8.1 the fourth-order bandpass modulator's SNR can be
estimated using the same approach that we used for the second-order lowpass modulator
discussed in Sec. 7.2.
Figure 8.17 A fourth-ord
STF
(8.23)
is modulator's SXR can be
ld-order lowpass modulator
Extracting the JIQ components digitally eliminates the J and Q channel phase response
mismatch (difference in delay through each path) encountered in typical baseband
demodulators. Figure 8.19 shows the simulation results showing how the 100 kHz (offset
sired output tone from 25 MHz) signal seen in Fig. 8.16 is demodulated to baseband. Note that the input in
25.1 MHz. the simulation is simply a 25.1 MHz sinewave, not an JIQ generated signal (see
simulation example at CMOSedu.com using an JIQ signal).
lIodulation noise After looking at Fig. 8.18 we might wonder how we multiply the modulator
output, a I-bit word, by 1, 0, and -1. As seen in 6.3, a 1 coming out of the modulator
°
corresponds to 1 (+ 1) in two's complement while a
MUltiplying these outputs by ° ° corresponds to 11 (-1).
results in an output of 00 (0) in two's complement.
Multiplication by +1 or -1 resultsinOI x(+I) 01,01 x (-1) 11,11 x(+I)=ll,and
°
11 x (-1) = 1. A data selector and a multiplexer with some logic for output selection can
be used to implement the multiplier (see the simulation example used to generate Fig.
8.19 at CMOSedu.com for one example).
cussed in Ex. 8.5.
J
298 CMOS Mixed-Signal Circuit Design Chapter 8 Bandpass Data Con'
i""d"
IF Extraction at 20 MI
C is pp. 1920-1934, Decem
4
[5] S. R. Norsworthy, R
I---~ I output Converters: Theory, J
978-0780310452
In
Multipliers Digital filters [6] B.-S. Song, "A 4th-C
Op-Amps," IEEE Soh
I---~ Q output 1995
Is [7] B.-S. Song, "A Fourt
Numbers of Op-Amp,
sin 2rc· ~ . nTs = sin n~ = 0, 1,0, -1... pp. 1309-315, Deceme
150rnV
-50rnV
1 QUESTIONS
-100rnV
-150rnV
[2] A. I. Hussein and W.B. Kuhn, "Bandpass Lf1 Modulator Employing 8.6 Sketch the implement
Undersampling of RF Signals for Wireless Communication," IEEE Transactions but using a multi-bit q
on Circuits and Systems 11: Analog and Digital Signal Processing, Vol. 47, No.7,
8.7 Show the details of he
pp. 614 - 620, July 2000
8.8 Derive the transfer ful
[3] A. Jayaraman, P.F. Chen, G. Hanington, L. Larson, and P. Asbeck, "Linear
High-Efficiency Microwave Power Amplifiers using Bandpass Delta-Sigma 8.9 Using the modulator
Modulators," IEEE Microwave and Guided Wave Letters, Vol. 8, No.3, pp. 121 sinusoid to the modul
123 March 1998 digital data through a
input and output sigm
ixed-Signal Circuit Design Chapter 8 Bandpass Data Converters 299
[4] A. K. Ong and B. A. Wooley, "A Two-Path Bandpass SD Modulator for Digital
0,-1,0...
IF Extraction at 20 MHz," IEEE Journal afSolid-State Circuits, Vol. 32, No. 12,
pp. 1920-1934, December 1997
[5] S. R. Norsworthy, R. Schreier, and G. C. Ternes (eds.), Delta-Sigma Data
I output Converters: Theory, Design, and Simulation, Wiley-IEEE Press, 1997. ISBN
978-0780310452
filters [6] B.-S. Song, itA 4th-Order Bandpass ~z: Modulator with Reduced Number of
Op-Amps," IEEE Solid-State Circuits Conference, pp. 204-205, 367, February
Q output 1995
[7] B.-S. Song, "A Fourth-Order Bandpass Delta-Sigma Modulator with Reduced
Numbers of Op-Amps," IEEE Journal of Solid-State Circuits, Vol. 30, No. 12,
1. 0, 1... pp. 1309-315, December 1995
[8] D. B. Ribner, "Multistage Bandpass Delta Sigma Modulators," IEEE Circuits and
1. Systems II.' Analog and Digital Signal Processing, Vol. 41, No.6, pp. 402 405,
June 1994
[9] S.A. Jantzi, W.M. Snelgrove, and P.F. Ferguson Jr., "A Fourth-Order Bandpass
--'-+--I1Q Sigma-Delta Modulator," IEEE Journal of Solid-State Circu.its, Vol. 28, No.3,
pp. 282 291, March 1993
Q
[10] J. C. Candy and G. C. Ternes (eds.), Oversampling Delta-Sigma Data Converters,
Wiley-IEEE Press, 1992. ISBN 978-0879422851
I QUESTIONS
8.1 Show, using SPICE, how to adjust the phase and amplitude of the I and Q signals
discussed in the beginning of the chapter to modulate the amplitude and phase of
the resulting I1Q to construct a constellation diagram for 8-level rectangular QAM.
8.2 Suggest a topology for the bandpass passive-integrator NS modulator where the
input and fed back signals are currents. Derive a transfer function for your design.
n be extracted digitally. Does your topology have the extra noise/distortion term seen in Eq. (8.12)? Why
or why not? Simulate the operation of your design.
8.3 Show the details of deriving the transfer function for the modulator in Fig. 8.6.
8.4 Repeat question 8.3 for the modulator seen in Fig. 8.8.
a-Sigma Data Converters,
8.5 Derive the transfer function for the modulator seen in Fig. 8.9.
~ Modulator Employing 8.6 Sketch the implementation of a modulator, based on the topology seen in Fig. 8.9,
~ation,"IEEE Transactions but using a multi-bit quantizer and feedback DAC.
Processing, Vol. 47, No.7,
8.7 Show the details of how Eq. (8.18) is derived.
8.8 Derive the transfer function of the modulator seen in Fig. 8.12.
1, and P. Asbeck, "Linear
19 Bandpass Delta-Sigma 8.9 Using the modulator topology in Ex. 8.4, show that if we apply a 25 MHz input
~rs, Vol. 8, No.3, pp. 121 - sinusoid to the modulator we can recover this input signal by passing the output
digital data through a bandpass filter with a very small bandwidth (show that the
input and output signal amplitudes are equal).
300 CMOS Mixed-Signal Circuit Design
8.10 Derive the transfer function of the topology seen in Fig. 8.17. VerifY that the
topology is unstable by determining the location of the topology's poles.
8.11 Using a bandpass modulator and digital demodulation (sketch the schematic of
your design) show how to recover a 10 kHz sinewave that is amplitude modulated
with a carrier frequency of I MHz. Use SPICE to verifY the operation of your
design.
A High-Spe
The majority of the data converter topologies we've discussed up to this point have traded
off time for resolution (so the signal bandwidth is generally much smaller than the
clocking frequency). In other words, to get higher signal-to-noise ratios, SNRs, (wider
output word size, N) we've averaged (filtered) the output of a noise-shaping modulator to
remove the modulation noise. The averaging (again, filtering) has the undesired effect of
reducing the noise-shaping modulator's allowable signal bandwidth. In this chapter we
tum our attention towards high-speed topologies. Our approach is based on the material
covered in Sec. 6.2.3 where K-paths are used but with switched-capacitor, SC, circuits
used for the required feedback subtraction from the input (delta) and integration (sigma).
9.1.2 Implementation
Figure 9.4 shows how we woul
single integrator, I-sigma, wit
switched-capacitors, Fig. bu
has to respond very quickly, a
integration capacitor, 4C" is sele
in detail in the last chapter. Also
outputs from each path can be
summed to generate an output Cl
outputs are high). Summing the
function of
H(;
is,new Kpath'ls
9.1.2 Implementation
Figure 9.4 shows how we would implement a high-speed analog-to-digital topology, a
single integrator, I-sigma. with K feedback paths, K-deltas, (see Fig. 6.24) using
switched-capacitors, Fig. building blocks. The amplifier used in the active integrator
has to respond very quickly, as described in the last chapter. Further note that the
integration capacitor, 4Ci , is selected based on an integrator gain of 0,25, again, discussed
in detail in the last chapter. Also seen in this figure are the clock signals and the way the
5 =~ outputs from each path can be combined together. The outputs of the K paths can be
6 = $2 summed to generate an output code ranging from 0000 (aU outputs are low) to 1000 (all
$3 outputs are high). Summing the outputs together results in Sinc filtering with a transfer
function of
1 _Z-8 .
H(z) "" ~ With Is,new Kpa1h'ls (9.2)
as seen in 9.3. Note that this is exactly the same response that we get when using the
:avcd sampling and the
sample-and-hold, SIB, Fig. 2.17, clocked at/,. While there are other ways to combine the
path outputs let's do some examples.
ve around confused because Simulate the operation of the data converter seen in Fig. 9.4 if C; 100 tF and/, =
:h path in the converter (see 100 MHz. Show that an input frequency of 50 MHz results in, as seen in Figs. 9.3
Figure 9.5 shows the simulation results when the input is a 7 MHz sinusoid
Fig. 9,1 centered around 500 mV with an amplitude of 400 mY. Notice that the rate at
which the data comes out of the summing circuit is 800 MHz (the output changes
Fig. 2.37 every 1.25 ns). Also notice that the change in the output is the reference voltage,
Switched-capacitor Figure 9.6 shows the input and output if the frequency is increased to 50 MHz
charging!discharging (the Nyquist rate if we were to decimate down to a clocking frequency of 100
MHz). We expect the amplitude of the output to be 0.64·0.4+0.5 or 756 mY.
capacitor, The simulation shows an output of 750 mY (remember the quantization noise) .
•
304 CMOS Mixed-Signal Circuit Design Chapter 9 A High-Speed Data 0
I-sigma
4JI-l 4J2-1
VCM ~'
i: C ! :
----.~4J1-2~ 4J2-2
.:~~~------~~+---~
-~,
<Pl~
-L:
4J2-~: : : :r i r-1
J:'-+----+ l'~;;:L--j
4J1-3~
L
4J1-3 4J2-3
4J2-3~ I I "" I '
Figure 9.5 Simulating
a 7 MHz ir
4J 1-4 '~;;I~
r-:--:--:l:, r-1
L-.J
i
~.
Non-overlapping clocks. 720mV
IT '-+-------r--__+__ s
4J2-1 4JH :§l
IT: Ys
0
Summing circuit.
J'-+-----+-_+__
YI 1On. 20n.
Yz
o.avil·--d--··'it·--·+----··}··--·+
::: . .:·;--:·::.~~·:·::: ...:·:L:.·--i .=::r::::..
0.5V --.~---- .. C__ .
. ·~----· .. 7
.
.... - -------- ..... +h,.•. ; .....
.
::: ..:...L:::.. :::::.: ~.~
.::::.L· .;::::.:::·::::::.L.\
O.2V ~ -1 ___ :w. ______; --.".~ 1 f ",wf_·_·
O.lV ....... ( .... : . . . . +. . ~ .... ;........;.......; ....... ~.
o.ov , I I I I --i---i--"-i-----i
Ons 30ns SOns 90ns 120ns 150ns 180ns 210ns 2~Ons 270ns 300ns
Figure 9.5 Simulating the operation of the data converter in Fig. 9.4 with
a 7 MHz input signal.
gOOmV'.--~--~;?---.,---;,.,-,---~,---=~-.",.__.-__,
810m\!
~on-overlapping clocks.
360mV
270mV
Summing circuit.
180mV
9UmV
OmV+--~-~-i--i---i---4-~-~-~-~
Ons 20llS 3Dns 4Dns 50ns 60ns 80ns 90n. lOOn.
Figure 9.6 Again, simulating the operation of the data converter in Fig. 9.4 but
now with a 50 MHz input signal.
Example 9.2
Repeat Ex. 9.1 if the output is decimated down to the 100 MHz clock rate.
Figure 9.7a shows that adding a register in series with the K outputs can be used
to decimate the word rate down to Is. A K-bit register can be placed in front of the
summation circuit or a log2 K wide register can be placed on the output of the
summing circuit (not shown in Fig. 9.7). Timing mistakes in this digital part of the
circuit are an important practical concern. The digital signals are moving quickly
the eight outputs together.
so eliminating the skew introduced by the summing circuit, by placing the
'ath filter, 1 - z-8 on the input side (Fig. 9.7a), is the approach we take here.
I
The simulation results are seen in 9.8. Notice that, when the input signal
~ers
frequency is at the Nyquist rate in Fig. 9.8b, that the sampling instances repeat.
Avoiding decimation (down-sampling), as can be seen when comparing Figs. 9.5,
ng mixed-signal techniques,
9.6, and 9.8, and discussed in detail earlier in the book, helps keep the original
signal intact.
J
306 CMOS Mixed-Signal Circuit Design Chapter 9 A High-Speed Dau
From To summing
To summing
circuit.
Ys
9.1.3 Filtering
Thinking about what we've done in this section we might come to the conclusion that the
topology seen in Fig. 9.4 is nothing more than a flash ADC implemented with The output of the summing
8-comparators. We are only getting, effectively, 3-bits out of the topology (see previous 1000 (8 in decimal). We pa
examples) so why not simply use a 3-bit flash ADC? The simple answer to this is that word size by 3-bits each fOJ
averaging (read filtering) has little, Fig. 5.31, to no (for a DC input) effect on the centered around the comme
resolution of a Nyquist-rate ADC while averaging can be used in a noise-shaping by 2 and center the word (a
topology to increase the resolution, Fig. 7.7. We are repeating material presented earlier more than required. The iS1
in the book so let's just do some examples to provide additional discussions.
Mixed-Signal Circuit Design Chapter 9 A High-Speed Data Converter 307
~ .D
N
~
elk.
T $2-1
To summing
circuit.
-~""-=0.993 (9.3)
)me to the conclusion that the
Ish ADC implemented with The output of the summing circuit is a 4-bit word that ranges from 0000 (0 in decimal) to
of the topology (see previous 1000 (8 in decimal). We pass the 4-bit output through two Sinc filters which increase the
simple answer to this is that word size by 3-bits each for a total output word size of 10-bits. However, the word is not
r a DC input) effect on the centered around the common-mode voltage so we can shift the word left to do a multiply
be used in a noise-shaping by 2 and center the word (as discussed earlier). This reduces the word size to 9-bits, still
jng material presented earlier more than required. The issue is that we also reduce the signal bandwidth. For our 6.25
mal discussions.
308 CMOS Mixed-Signal Circuit Design Chapter 9 A High-Speed Data Co
Thermometer code ~fs @f, 0000 1000 Multiply output by 2, shift to 8is
\ \ &+-1\0 gHOfrOm IOt09b~~tS'
~@f\ J \
4
In~+\ 4 1-z- 8
7 l-Z=~ 9 Out
I Fig. 9.4 + DQ
elk
Analog '------' Lf I - Z-I I - z . Digital
8fs
.. is 100 MHz
ts:::.~'9dB=0'64
~ I
i elk
_ _ __
i -7.8 dB = 0.41 @ 6.25 MHz
50 MHz
l~f
8is 4/s
. ~f Figure 9.12 D,
50 100
6.25'12.5 MHz
Figure 9.1 0 Decimating and filtering the output of the K-path modulator, an
approach that won't work for Nyquist-rate conversion.
810mV
72UmV
MHz input signal, as seen in Fig. 9.10, this means that we attenuate the signal by OA1. 630mV
The total attenuation of our 6.25 MHz input signal is then OA07 so the peak amplitude 540mV
out of the filter is 663 mV (ideally it's 900 mY). Figure 9.11 shows simulation results
verifying our hand calculations. Note the delay through the filter when starting.
Studying Fig. 9.10, we see that the decimation right at the beginning of the digital
filter is limiting our bandwidth. Consider spreading the decimation out through the filter
as seen in Fig. 9.12 (see also Sec. 4.2.5). Here we've put a register on the output of the
summing circuit instead of on its input, Fig. 9.7a. For ideal signals we may not need this OmV·-j-L-.,.----t-
Ups 50ns 1 OOn~
register; however, as mentioned earlier, eliminating skew and timing errors is important
for proper data converter operation. Simulation results are seen in Fig. 9.13. The input
signal undergoes less attenuation but the final resolution is 7-bits instead of 9-bits Figure 9.13 Simul
(indicating that there is less filtering in Fig. 9.12 than 9.10). In order to move towards our
design requirements, an 8-bit converter with a signal bandwidth of 50 MHz that is Figure 9.14 shows our ill
clocked at 100 MHz, our final output clock rate will be increased to 200 MWords!s results, with a 6.25 MHz input, a
(instead of 100 MWords/s as in Figs. 9.10 and 9.12). Fig. 9 A we are still attenuating 01
of the adder is changing at 200
place in the filter. The chain of a\
size by I-bit per stage. The
(0.707)5 :::: 0.177, see Fig. 1.17.
results in an overall attenuation
signal frequency and simulate to '
Figure 9.16 shows yet ane
MHz input should see the -3.9 (
9.10. In addition, as seen in Fig.
the first zero point occurs at
additional, minor, attenuation, Fi
remove the modulation noise in
fiNo, the filter doesn't remove th
Figure 9.11 Simulating Fig. 9.10 with a 6.25 MHz input signal. goals? Further, how do we detem
lixed-Signal Circuit Design Chapter 9 A High-Speed Data Converter 309
'l ~
1-
~~@f,
Z-1
Out
Digital
@8j,
J \
I~ Fig. 9.4~1--4T--1~tF
t:<.='
~
: ~f
@ 50 Mlh
Multiply
Sls 50 100 output
by 2,
7
/ Out
I dB = 0,41 @6.25 MHz
4j~
:-palh modulator. an
onversion.
dwidth of 50 MHz that is Figure 9.14 shows our next attempt at filtering and decimation. The simulation
ncreased to 200 MW ords/s results, with a 6.25 MHz input, are seen in Fig. 9.15. By adding together the 8 outputs in
Fig. 9A we are still attenuating our input signal with a Sine response. Note that the output
of the adder is changing at 200 MWords/s. We don't employ decimation at any other
place in the filter. The chain of averaging filters, on the adder's output, increases the word
size by I-bit per stage. The attenuation from this digital filter at 50 MHz is
(0,707)5 0.177, see 1.17. This, combined with the attenuation from the addition,
results in an overall attenuation for a 50 MHz input signal of 0.113 (change the input
signal frequency and simulate to verify).
Figure 9.16 shows yet another filter. This filter doesn't employ decimation so a 50
MHz input should see the -3.9 dB attenuation from adding the 8 outputs together,
9.10, In addition, as seen in Fig. 1.17 (but with frequencies on the x-axis scaled by 8 so
the first zero point occurs at 400 MHz), the averaging stages will provide some
additional, minor, attenuation, Fig. 9.17. The question we need to ask is "Does this filter
DOns ~50ns 500no
remove the modulation noise in addition to increasing the word size?" If the answer is
"No, the filter doesn't remove the modulation noise," then how do we meet our design
[z input signal.
goals? Further, how do we detelmine if the filter removes the modulation noise?
310 CMOS Mixed-Signal Circuit Design Chapter 9 A High-Speed Dat,
~~
~
8fs
~
remov
word ~
1.011
.. > 0.911
O.BV
so/is Is 21s'200 0,711
2 0.611
Figure 9.14 Another filter; however, this topology has problems removing 0.511
the modulation noise.
0.411
0.311
0.211
O.W
D.OV
On. SOns
:::fKT
0.811-+-(K
On. 1DOns 150n8 200ns 250ns 300n. 350n. 400n. 450n. 500ns
o. 7V4 1""\ ~
Figure 9.15 Simulating Fig. 9.14 with a 6.25 MHz input signal. Note the
the nonlinearity in the output.
-ti ...
0 6v
D,5114--·,\"
+.
In order to move towards answering these questions, notice the nonlinearity in
0.41'1. ....,\.
both Figs. 9.15 and 9.17. We implemented the filter by focusing on only increasing the ~.3V+ ..... A\
,
0.2V~--"'"''
output word size to 8-bits. We didn't concern ourselves with removing the modulation
noise and thus we see noise and nonlinearity in our output signal. Reviewing Sec. 7.1.3, 0.1V4 ..·....;·
we can remove the modulation noise by using the filter seen in Fig. 9.1 0 but without the o.ov-r---i-
Dns 10m
decimation (noting that if L 2 then only the path filter and one Sinc filter are used for a
7-bit output). This also helps us move closer towards the design goals. Figure 9.18 shows
the simulation results when a 50 MHz input is applied to the topology seen in Fig. 9.l0 Figure 9.18 Using th
due to 3·
without decimation. The input signal undergoes 3-Sinc filter responses (0.64)3 0.262.
[ixed-Signal Circuit Design Chapter 9 A High-Speed Data Converter 3 II
8f,
MUltiply
Multiply output
output by 2.
8 8 by 2.
~Out
'_L_.~ . Diigital
8f,
@2!s
Figure 9.16 Filtering without decimation. This filter won't do a good job
removing the modulation noise but it does increase the output
word size to 8-bits .
problems removing
Figure 9.17 Simulating Fig. 9.16 with a 6.25 MHz input signal. Again, note the
the nonlinearity in the output.
Direction
In order to move towards our design goals we can use the filter seen in Fig. 9.12 but with ""1iiW
~l~
an additional averaging, I + z I , stage on the filter'S output. This increases the word size
to 8-bits. The penalty is an extra 3 dB attenuation at 50 MHz (so the total attenuation at
50 MHz is 10.S dB or 0.2S8). The practical issue with this topology, for very high-speed
operation, is avoiding timing errors in the stages clocked at Sis. For a general design, we'll
use the topology seen in 9.10. Timing errors are straightforward to minimize. The
~ H(z)
~3 -'
drawback is the reduction in signal bandwidth. ••
•
An area that should be investigated further is the use of biquad filters for
removing the modulation noise, as discussed in Sec. 4.3.4.
,_I~
9.1.4 Discussion
If the NS modulator seen in Fig. 7.2 is clocked at 100 MHz we get the noise transfer
function, NTF, shape seen in Fig. 9.l9a (out to 800 MHz or Sis). While we are used to
/
H(z) = Vin(Z)'
focusing on the spectral content between DC andls (= 100 MHz here), we know that any
discrete-time system's frequency response repeats with the sampling frequency, Is. The
signal bandwidth, as seen in the figure, is 6.25 MHz, assuming Kavg = S. Note that we are
not showing the frequency response of the digital filter used for removing the modulation
noise (which is where Kavg is used, Eq. [6.47] and Fig. 7.8). Figure 9.19b shows the
assumed spectrum of the quantization noise while Fig. 9.19c shows the single path
clocked at 100 MHz (noting we are not showing the non-overlapping clock signals).
INTF(f) I2
200
: 50 100 200 400 500 600 700 800 MHz (c) Quantization noise for aJ
6.25 (a) NTF ofa noise-shaping topology clocked at 100 MHz, 7.6.
visB i . V;n(Z) i M d It
-'-Iou a or.
H()
Z
Vin(Z)' + (l-Z-l) ,VQe(f)
12fs~ , elk ~ Desired spectrurr
50 100 MHz ~IOOMHz
(b) Quantization noise PSD, Fig. 5.12. (c) Modulator clocked at 100 MHz. 50 100 200
Next, examine the S-path topology of modulators seen in 9.20 (modified from quantization noise out over a
Fig. 2.37 for this discussion). In (a) of the figure each modulator is clocked at!" the same noise added to our signal in t
as in Fig. 9 .19c, but with time-shifted clock signals. In (b) we show the equivalent circuit, our desired signal bandwi.
see discussion on page 224 for details, noting the equivalent circuit's output changes with improvement in SNR. Note
a frequency of Kporh ·Is. By using an effectively higher clock frequency we spread the SOOMHz.
\fixed-Signal Circuit Design Chapter 9 A High-Speed Data Converter 313
~~~.--------~~
~~ ••
•
e use of biquad filters for
~.~~)
~8 8 ~l
.:.
Kpath 8
Ca) 8 modulators in parallel
fz v.e get the noise transfer H(z) Vin(Z)·z-l+ (l-z-I). VQe(f)
Ir ~f,). While we are used to
fHz here), we know that any
sampling frequency,!,. The ~lK JL..h--.SLS
ng Kavg 8. Note that we are ~2K~
for removing the modulation ~
.8). Figure 9.l9b shows the TsIKpafh
> 100 200 300 400 500 600 700 800 MHz
500 700 800 MHz (c) Quantization noise for an 8-path topology clocked 100 MHz, compare to Fig. 9.19b.
)cked at 100 MHz, Fig. 7.6.
2
NTF(f)
~
INTF(f) I • 8. i21s
ineZ) . Z-l + (1- z-l) . VQe(f)
n in Fig. 9.20 (modified from quantization noise out over a wider range of frequencies, (c). This reduces the modulation
ator is clocked at!" the same noise added to our signal in the bandwidth from DC to 6.25 MHz as seen in Fig. 9.20d. If
e show the equivalent circuit, our desired signal bandwidth remains at 6.25 MHz then we will 'clearly get an
circuit's output changes with improvement in SNR. Note that the output of the topology is I-bit changing at a rate of
)ck frequency we spread the 800 MHz.
314 CMOS Mixed-Signal Circuit Design Chapter 9 A High-Speed Dat
Suppose we want to use this topology for high-speed data conversion where we 9.1.5 Understanding the
use the fact that our effective output clock frequency is 800 MHz (and so the Nyquist
frequency is 400 MHz)? Is this possible or does it make sense? This is a good point to Notice that in Fig. 7.2 we us
remember the fundamentals. As mentioned on page 201, using noise-shaping does not our high-speed topology sef
result in a reduction in quantization noise. Rather, noise-shaping topologies push the strobe the comparator. For e)
quantization noise to frequencies outside the region of interest. Reviewing Fig. 9.20d we are clocked with <P1-I whill
see that (all of) our quantization noise is still present in the spectrum from DC to 400 captured on C, at the falling
the comparator. Since neal
MHz (see example below). It's impossible to reduce this noise with a digital filter without
effecting the desired signal in this same frequency range. associated with transferring
When the comparator makes
Example 9.3 with minimal delay, so that
Estimate the RMS value of the quantization noise for the 8-path topology clocked path's decision. The result is
at 100 MHz seen in Fig. 9.20d. topology clocked at Kpa'h ·Is
V~e,RMs == I
4/,
INTF(f) I
2
• 8~~~Is . df ==
2
I
4(,
4· sin
2
(f)
11:
V2
1; . 8. ~~Is . df
In a practical imp lei
integrator and comparator
experiences a delay. The re~
this instability. In this exarr
== ViSB .J[I-COS(211:.L)].df == ViSB mV here, and the clock sign
4·12Is 0 Is 12 used for both the input pair
when using ideal componen'
or VQe,RMS == VLSB / Jl2 , no reduction in quantization noise. While this result was the input signal; however, v
derived for the case Kpolh == 8 the reader should see that the result is valid for any
shown, the oscillation is m
value of K po1h ' •
signals to minimize the df
always occurs at a high free
Example 9.4
the output of the modulator
Estimate SNR;deal for the topology seen in Fig. 9.20a.
Another example, bt
The procedure for calculating SNR was given in Sec. 5.2. Since we just calculated
that the average of the outp
the quantization noise, all that's left is to calculate the desired output signal power.
the clock signals applied to
If we apply a sinusoid to the topology with a peak amplitude of Vp , then adding
grows and shrinks. In tr
the output signal powers from each path results in Kpalh . V')I2. We can then write
comparator or through the ~
~.Vp/J2 into account when selectir
SNRideal == 20 . log ~ concern is the increase in
VLSB/" 12 (which can be reduced by ir
or (see Fig. 5.31)
1.0V I ,i I
SNRideal == 6.02N + l.76 + 10 log Kpalh (9.4) o·9Vi-- --~-t
For every doubling in the number of paths we get a 3-dB (0.5 bits) increase in the O.BVi· "r'
SNR. Such a modest increase in SNR is, generally, too insignificant for the
hardware costs so this approach isn't used .•
O.N
O.6V .... r'1"". ·
O.5V-f--r'-"-
The virtues of the topology seen in Fig. 9.4 should be clearer. Using a single O.4V
integrator (I-sigma) with K feedback paths (K-deltas) we can attain conversion O.3V
bandwidths approachingf)2 with reasonable resolutions. Since a single integrator is used,
common information is applied to the input of each quantizer. This allows the O.1V
quantization noise to be pushed to very high frequencies. In other terms, we get fast O.OV+-J--;--
Dns 1Dns
sampling by using K feedback paths while we are able to push the quantization noise to
higher frequencies using a single integrator common to all quantizers.
Figure 9.21 Sl
ixed-Signal Circuit Design Chapter 9 A High-Speed Data Converter 315
Since we just calculated Another example, but with an AC input signal, is seen in Fig. 9.22. Notice, again,
red output signal power. that the average of the output does equal the input signal. It can be instructive to change
tude of Vp , then adding the clock signals applied to the comparators in the modulator to show how this oscillation
];12. We can then write grows and shrinks. In the practical implementation the finite delay through the
comparator or through the switches charging or discharging the capacitors must be taken
0. into account when selecting the clock phases for the comparators. Another practical
concern is the increase in the output swing of the integrator because of this instability
(which can be reduced by increasing the size of the integrator's feedback capacitor).
(9.4)
:0.5 bits) increase in the
)0 insignificant for the
1.0V-,,--...,.,....-n-~=------,r-T---,-·~-~-····--+--'-=---:-
D.SY
D.7V
O.SV
D.SY
OAV
O.3V
O.ZV
Figure 9.24 Ring oscill:
O.W
O.OV+----+---jL-L-.lj---+-.JL--i--···-r----r----'---'t--'------r--'-----'-'l
Ons JOns 60ns 90ns 120ns 150no 180ns 210ns 240n5 270ns JOOns process. Note that the frequen
inverters in the delay stage. A
Figure 9.22 Showing built-in instability with a sinewave input signal. 200 MHz (lower for a simu
sampled at a rate offs.nev.. or 1.t
then decimated, 9.7a.
9.2 Practical Implementation
:~lffi
In this section we'll discuss the practical implementation of the high-speed topology
proposed in this book. The goal of this section is to provide discussions and provoke
thought that should prove helpful when designing a converter using this topology. The
:~~i~·:::nf:
goal is not to provide definitive solutions for specific applications. This endeavor is left
for discovery by the engineers and researchers doing mixed-signal circuit design.
12V -- •. -,-'-
9.2.1 Generating the Clock Signals lOV .-- ---- :- ••
8V -- .-.--- ---
Generating the 16 clock signals needed for the topology seen in Fig. 9.4 can be 6V ----.-.
challenging. We could use a delay-locked loop (DLL) that takes an input clock signal and
generates the 16 clock signals (but that adds complexity). Here we use a ring oscillator
that runs asynchronously with an external clock signal. Figure 9.23 shows the basic delay
01f?E
-zv i i
stage schematic and icon used in thc oscillator. Figure 9.24 shows the complete ring Dns 30s
oscillator while Fig. 9.25 shows some simulation results in a 500 nm, 5-V, CMOS
Figure 9.25
Figure 9.24 oscillator, schematic and icon, for use with the data converter.
process. Note that the frequency of the oscillator can be adjusted by adding or removing
inverters in the delay stage. Also notice that the simulated oscillation frequency is around
\ ave input 200 MHz (lower for a simulation with layout parasitics). The analog input signal is
sampled at a rate ofIs.new or 1.6 GHz. The 8 path outputs are added together, Fig. 9.4, and
then decimated, Fig. 9.7a.
Figure 9.26 shows how the external (synchronous) and internal (asynchronous)
clocks can be interfaced using a synchronizer. We are assuming that the internal clock is
running faster than the external clock (so the output is decimated). Note that the
synchronizer doesn't introduce aperture jitter as discussed in Sec. 5.2.1 since it only
proeesses digital signals. After reviewing this section, however, we might wonder if the
ring oscillator is a practical choice for providing the clock signals to the data converter.
Ring oscillators are certainly not as stable as crystal-controlled oscillators. However,
notice that by combining the Kpalh outputs together we reduce the variance of the aperture
jitter by K,""lh' see, for example, Eq. (5.30). Further filtering reduces the effects of a jittery
clock signal. Slow variations in VDD, ground, or temperature will also have essentially
no affect on the data converter's performance (via the ring oscillator) since these changes
simply vary the sampling rate (the internal clock frequency). Note that for large
differences in the internal and external cloek frequencies aliasing concerns, when
decimating, should be taken into consideration in the synchronizer (as should
:hematic and icon. metastability concerns, e.g., the external clock (not) going low just after the internal clock
goes high resulting in a glitch on the output of the AND in Fig. 9.26).
318 CMOS Mixed-Signal Circuit Design
1, Chapter 9 A High-Speed Data
Synchronizer
In Digital
Analog Filter ~.DQ Out
T
External clock
Internal clock,
clocked with ring oscillator, Fig. 9.24. External clock (not)
Internal clock
Irm=:
Figure 9.26 Synchronizing the external and internal clock signals.
Ou(rn
:
Finally, the size of the capacitors used in this circuit, as discussed on page 307, is set by
thermal noise considerations.
2.55V
i
m
2.53\1 ....... ~ ....... ~
Synchronizer
Out Inpf~tP
i
External clock
Inffl~tm
Amplifier inputs
Figure 9.29 Simulating the operation ofthe amplifier in Fig. 9.28 with 400 if loads.
changes states (only) on the rising edge of the input clock signal. The outputs of the
vices (which interface nicely
NAND gates are buffered up to ensure that they can drive the switched-capacitors. If
Fig. 9.29. The non-inverting
these outputs don't fully charge the switched-capacitors or the VDD/ground connected to
; pulsed from 2.45 to 2.55 V
these gates droops or bounces, the feedback signal will be affected causing noise and
.8 (2 V change) so the gain is
nonlinearity. Finally, the inputs to the comparator should be above the threshold voltage
ts of the amplifier are loaded,
of the NMOS device but not so large that the input devices are pushed into deep triode.
:c. 7.1.8, is that we don't see
Note, as discussed already, that we need to be careful to ensure minimal forward delay
e larger than required for our
elk
Inp~ t.c""",
InN _ . bar
Figul
elk
1
~bar
on.
..
• T· - - - - - - ~ - - - - - - -..,
•....... , ...... .
-: ------- ~- ------ Output synchronize the output data to an external clock. This means that the output will change
at the rate of the ring oscillator (roughly every 5 ns or 200 MHz). An input signal
..
clock frequency of (roughly) 200 MHz/16 or 12.5 MHz will be attenuated by -3.9 dB. Figure
9.34 shows the simulation results for the ADC (modulator and filter) when the input
signal is 10 MHz (same as Fig. 9.33 but showing the output of the digital filter). These
8
1 ............. ..
.;.......L....
. .
+ input
+--.
conversion rates are comparable to any high-speed converter topology implemented in a
.i....... i........ - mput
500 nm CMOS process (even faster if we use less decimation early in the digital filter) .
-,-------,-------
S.OV
itor in Fig. 9.30. 4.SV
4.0V
3.SV
3.0V
minus the digital filter) we 2.SV
2.0V
:l VDD of 5 Y. The resistors
1.SV
lod of summing the digital 1.0V
g up the simulation, they are O.SV
The finite transition times, O.OV
riewing analog data. Figure -O.svr+---i---i---i---i---i---i---i---i--';--
Ons 30n5 60ns 90ns 120n5 150n5 180ns 210ns 240n5 270n5 JOOns
input signal.
ile ADC, we need to add the Figure 9.33 Simulating the ADC in Fig. 9.32 with a 10 MHz input signal.
,igma topology seen in Fig.
ldulator clocked at Is.new Figure 9.35 shows a simulation where the ADC's input is a ramp signal that
:tively connect each path's transitions from 0 to VDD in 1 )..ls. The delay through the filter is 50 ns so the input signal
is). Let's use the decimator should be shifted 50 ns later in time when comparing the ADC's input to its output. Of
3, L = 2 (the adder and only course, the ADC's output will then be valid only after this delay (so we neglect the
ffective I-bit output by the start-up transient). We can estimate the minimum step size in the output of the ADC as
the Sinc filter. We won't VDD!2 6 or 78 mY. We use 26 instead of2 7 (the exponent is the number of bits coming out
of the filter) because in order for the output signal to swing rail-to-rail we have to
...~.;:1·';~'
..
,,-;
~;
~.-;1
~.".
1i
--- - - - - -
l
322 CMOS Mixed-Signal Circuit Design i Chapter 9 A High-Speed Data C
I QUESTIONS
II
~.OV
9.1 What is a time-interlea\
3.5V
different from the K-I
implementation of a
3.0V
Delta-Sigma modulators
2.5V
2.0V
1.5V
I 9.2 Using the modulator fro
the K- Delta-I-Sigma top
1.0V ------+-- 9.3 Repeat question 9.2 but
0.5V integrator isn't critical
O.OV-t---i---i---i---i---i---;---i---;---;------1 converter using differenl
Dns 30n5 60ns 90ns 120n5 150n5 180ns 210ns 2040n5 270n5 300n5
5.0V,--_-,----_...,..-'-'VI=oU"'IJ-,---_-,---_-,---_--,--_--,--:-VI"-vi""nJ'-;-_--,------::?l
,
.. .
~.5V -- - -- - - j--. -----;"- ----. --~ ------. ~ ------- -;- --. ----~ -------j-------;
" ,
Explain how this filter
outputs. Does this filter
9.3 Conclusion
This chapter has proposed a new topology for high-speed analog-to-digital conversion
using the mixed-signal circuit design techniques presented in this book. The topology
should prove useful when designing high-speed ADCs in nanometer CMOS. Future work
where z = e j2 1[jT, . Usin
can be focused in many areas including, but not limited to: fully-differential signal paths,
paths in parallel, Fig
the integrator (amplifier) design, higher-order (perhaps passive, Sec. 6.2.1) topologies,
topology's outputs.
bandpass converters, segmenting feedback paths, digital calibration (e.g. for offset, Sec.
7.1.9) and the design of the digital filter (including the decimating stage).
Mixed-Signal Circuit Design Chapter 9 A High-Speed Data Converter 323
QUESTIONS
9.1 What is a time-interleaved data converter? Why is a time-interleaved converter
different from the K-Delta-l-Sigma converter seen in Fig. 9.4? Sketch the
implementation of a time-interleaved data converter implemented with
Delta-Sigma modulators. Also sketch the clock signals used in the topology.
9.2 Using the modulator from Ex. 9.1 show that capacitor matching isn't important in
the K-Delta-l-Sigma topology.
9.3 Repeat question 9.2 but show that thc opcn-Ioop gain of the amplifier used in the
integrator isn't critical (compare, using simulations, the performance of the
converter using different gains).
to.ns 270ns 300n5
9.4 Show the details of (derive from the time-domain outputs of the K-Delta-l-Sigma
.he output of the digital filter, modulator) how the path filter seen in Fig. 9.4 has a z-domain transfer function of
1 Z-8
I
Explain how this filter performs a moving-average filtering of the modulator's
outputs. Does this filter decimate the K-Delta-l-Sigma outputs? Why or why not?
9.5 Sketch the decimate by Ki4 topology similar to the topologies seen in 9.7.
Ensure the proper clock signals are used in your sketch.
9.6 Explain, in your own words, why oversampling (averaging the outputs) using a
3-bit Flash converter (eight comparators), won't result in as significant
improvement in SNR as the K-Delta-I-Sigma topology.
9.7 What is the frequency response (an equation) of the filter seen in Fig. 9.1 0'1
9.8 What is the frequency response (an equation) of the filter seen in Fig. 9.12'1
p input signal. Shift
re to output. 9.9 What is the frequency response (an equation) ofthe filter seen in Fig. 9.14?
9.10 What is the frequency response (an equation) of the filter seen in Fig. 9.16?
to display the digital data as
9.11 Show how the switches on the inputs and outputs of the 8 modulators in parallel
seen in Fig. 9.20 can be described using the unit matrix and delays or
)reasing resolution) requires
.ignal conversion bandwidth 1 0 0 0 0 0 0 0
ely following the modulator o1 0 0 0 0 0 0 Z-118
:e Fig. 9.7b for the decimate 001 0 0 0 0 0 z·218
o0 0 0 0 I 0 0
analog-to-digital conversion 0000001 0
in this book. The topology o0 0 0 0 0 0 I
lometer CMOS. Future work
lIly-differential signal paths, where z = . Using these relationships show how to relate the inputs of the K
sive, Sec. 6.2.1) topologies, paths in parallel, Fig. 9.20, to the transfer function H(z) and the resulting
,bration (e.g. for offset, Sec. topology's outputs.
iting stage).
324 CMOS Mixed-Signal Circuit Design
9.12 In the previous question we showed that we can think of the switches on the input
or output of the K-path topology as multiplying the signals by I or O. Let's take
this a step further and ask "What can we accomplish if we multiply the signals by
I, 0, or -I (the analog multiplication by -I is implemented by swapping the
fully-differential amplifier's inputs)?" Reviewing Fig. 8.19 from the last chapter
Index
we see that proper selection and sequence of these multiplying terms can result in
modulating or demodulating our signals. Remembering that our (desired) analog
input spectrum extends from DC to some frequency B (lowpass data conversion)
we see that this movement of the spectrum is undesirable since it moves our
desired input signal spectrum directly up into the larger-amplitude modulation A
noise after it's passed through the modulator (again assuming a lowpass modulator
AAF. See Antialiasing filter.
is used). This makes the signal difficult to recover and lowers the topology's SNR. Accumulate-and-dump circuit, 129,
Perform a literature search and comment on these problems ensuring that your Accumulator, 19, 129, 134,235-236
understanding of the issues is clear and that you've referenced the literature used Accuracy, 195,200,269
9.13 The effective sampling frequency of the K-Delta-l-Sigma ADC discussed in Sec. 109, 112, 11 6
9.2 is roughly 1.6 GHz. Can any component of the ADC operate at, or be clocked Adaptive filtering, 93
at, 1.6 GHz? Verify your answers using SPICE and the 500 nm, 5 V, CMOS
Adder, 94, 126, 145, 147, 151, 156,
models used to generate these figures. What is the most critical component then, 309,320,322
from a timing perspective, (the DFF used to capture the eight bits coming out of Aliasing, 21, 27-72, 93, 127-129, 1~
the modulators) and what is critical in that component (the DFF's setup and hold 306,317
times)? What happens if an error is made in the most critical component? (The quantization noise, 169
wrong count is captured. For example, we should capture 6 logic Is but we Analog RC filter, 80, 82-83, 87-88,
actually capture 5 or 7 logic Is. Since we have a significant amount of averaging 103-106,109,112,116
in the digital filter the effects should be small. If an equal number of positive and Analog Sinc averaging filter, 10-11
negative errors are made the errors average to zero and don't affect the converter's Analog-to-digital eonverter (ADC),
performance. ) 164-224,225-284
I-bit, 236
bandpass, 134,287-300
gain, 246
high-speed, 301-324
pipeline, 38
Antenna. 292
198-199,203,205-206,302
315
Averaging,
decimation, 209-211
234,239-241,266-267
190-194
Averaging (continued),
interpolating filter, 43-50, 71, 139-142, 199,
235, 273
B
A B, bandwidth of the input signal, 129-130, 133-134,
AAF. See Antialiasing filter. 182, 198-199, 208-209, 212, 220, 229,
Accumulate-and-dump circuit, 129, 132 237-241, 248, 250, 267, 285, 290, 296, 324
Accumulator, 19, 129, 134, 235-236 B device. See Nonlinear-dependent source.
Accuracy, 195, 200, 269 Bandpass filter, 19, 96, 103-111, 129, 134-139, 160
Active-RC Integrator, 73, 75, Bandpass modulators, 263, 285-298, 301, 324
filter, 80, 82-83, 87-88, 97-98, 101, 103-106, Bennett’s criteria, 165-166
109, 112, 116 Bilinear transfer function, 95
Adaptive filtering, 93 Binary offset, 123-125, 204-205, 274
ADC. See Analog-to-digital converter. Biquad filters, 73, 99-101
Adder, 94, 126, 145, 147, 151, 156, 158-159, 273, Biquadratic transfer function, 99
309, 320, 322 Bottom plate of a capacitor, 67
Aliasing, 21, 27-72, 93, 127-129, 140, 143-144, Bottom plate sampling, 39-40, 54, 219
306, 317
quantization noise, 169
Amplitude modulation, 187, 207, 285, 288, 315 C
Analog RC filter, 80, 82-83, 87-88, 97-98, 101,
103-106, 109, 112, 116 CAI, See continuous-time analog integrator.
Analog Sinc averaging filter, 10-11 Calibration, 322
Analog-to-digital converter (ADC), 21, 27, 94, Canonic digital filter, See Digital filter.
164-224, 225-284 Capacitance,
1-bit, 236 bottom plate, 54
bandpass, 134, 287-300 input, 86, 93, 247, 282, 319
gain, 246 load, 88-89
high-speed, 301-324 noise calculation, 307
Nyquist-rate, 43, 63, 220-225, 241, 306-308 parasitic, 66, 69, 83-84, 219, 279-281
pipeline, 38 Capacitive feedthrough, 39, 251, 257
SPICE models for, 119-123, 160, 176 Cascade of integrators and combs filters (CIC
Antenna, 292 filters), 142
Antialiasing filter (AAF), 27-40, 73, 93, 139, 144, Cascaded modulators, 233, 264, 275-280
198-199, 203, 205-206, 302 Charge injection, 39-40, 251, 257, 294
Aperture jitter, 179, 202, 317 Charge pump, 120
Autocorrelation function (ACF), 183 Chopper stabilization, 71
Autozeroing, 57-58, 62, 71 CIC filters, see Cascade of integrators and combs
Average power, 183-187, 198 filters
Average value, 44, 168, 183-187, 200, 212, 244, Clock generation, 53, 316-318
315 Clock jitter, 178-189
Averaging, Coherent sampling, 175-176, 202
bandpass filter, 19, 134-139, 160, 290-292 Comb filter, 8-21, 25, 129-131, 133-143, 224
clock jitter, 198 Common-mode feedback (CMFB), 60-61, 88-89
decimation, 209-211 Common-mode rejection ratio (CMRR), 87, 252
filter, 13-14, 126, 131, 133-134, 201, 220, 223, Common-mode voltage, VCM = 500 mV in this book
234, 239-241, 266-267 Comparator,
highpass filter, 19, 129, 134 clocked, 163, 319-320
improving signal-to-noise ratio (SNR), 13, gain, 244-246, 259, 262-263, 295
190-194 hysteresis, 246-247
326 Index
Comparator (continued),
LSB, 204 E
noise, 163, 246-247, 250
offset, 246-247 Effective number of bits, Neff , 173, 176, 179,
placement, 261 181-182, 188, 202
Constant delay, 10, 31, 145, 171 Error feedback, 235, 271-275, 283
Continuous-time analog integrator (CAI), 73, 75, 95 Expected value, 186
Continuous-time comb filter, 10-11
Continuous-time Fourier transform, 23
Convolution, 35, 175 F
Correlated double sampling (CDS), 59, 66, 71
Cosine window. See Hanning window. Feedback gains in a DAI, 260
Counter, 45, 126-129, 130, 270-271 Filters,
active-RC, 75
accumulate and dump, 129
D bandpass, 19, 96, 103-111, 129, 134-139, 160
DAC. See Digital-to-analog converter. bilinear, 95
DAI. See Discrete analog integrator. biquad, 73, 99-101
dBc (decibels with respect to the carrier), 177, 189 comb filter, 8-21, 25, 129-131, 133-143, 224
Dead zones, 215, 217, 232 decimating filters, 143-144, 157, 203, 239
Decimation, 33-34, 127-129, 132, 222-223, digital. See Digital filter.
306-312, 321-322 exact frequency response of discrete-time filter,
bandwidth limitations, 34 94
decimating filter, 143-144, 157, 203, 239 finite impulse response (FIR), 133-134,
Delta modulation, 199 145-146, 161
Delta-sigma modulation, 199, 233-234 See gm-C filters. See transconductor-C filter.
noise-shaping. highpass, 19, 129, 134
Demodulator, 199-200, 235-236, 271-272 in a higher-order modulator, 267
Diff-amp, 61, 86, 89, 222 infinite impulse response (IIR), 134
Differentiator, 14-15, 119, 275-276 integrating filters, 73-118
Digital averaging filter, 192, 220 interpolating filters, 139-142
Digital comb filter, 11-18 ladder filter, 117-118
Digital filter, 119-162 lowpass, 75
biquad, 155-159 MOSFET-C, 83
canonic, 151-155 Q peaking and instability, 112
finite impulse response (FIR), 133-134, Sinc, 133
145-146, 161 SNR, 82-83, 87
infinite impulse response (IIR), 134 stability in digital filters, 146
stability, 146-147 switched-capacitor, 90-95
Digital integrator, 19-21, 235 transconductor-C filter, 97
Digital I/Q extraction to baseband, 297-298 tuning, 80, 83, 86, 93, 98
Digital resonator, 134-139, 160 Finite impulse response (FIR) filter, 133-134,
Digital signal processing (DSP), 27, 32, 203 145-146, 161
Digital-to-analog converter (DAC), 69, 119-123, Finite op-amp gain-bandwidth product, 55-57,
204-205, 213-214, 235-236, 272-275 78-79
Dirac delta impulse, 23, 28 First-order noise shaping, 208, 210, 224, 233, 238,
Discrete analog integrator (DAI), 66-70, 90-93, 254
233, 251, 260 FIR. See Finite impulse response.
Discrete-time integrator, 90-95 Forward modulator gain, 243
Distortion, 35, 37, 49, 65, 71, 176-178, 187, Fourier series, 23
206-208, 213-214, 225-226, 249, 251-252, 269, Fourier transform, 24
273, 288-289
Dither, 166, 195-198, 242, 282
Droop, 37, 39-41, 47-48, 127, 129, 132, 142, 157, G
209, 212, 223, 241, 267
DSP. See Digital signal processing. Gain error term, 176-177, 204, 213-214, 219,
Dynamic range, 83, 177, 251, 276, 279 247-249, 280
gm-C (transconductor-C) integrators, 86-90
Gaussian probability distribution, 187, 197
Index 327
H M
Hanning window, 175 MASH modulator. See Noise-shaping
Harmonics, Matching, 53, 55, 94, 200, 211, 212, 230, 279, 323
square wave, 189 Matlab, 171
High-frequency transconductor, 89-90 Mean, 186
Highpass filter, 19, 129, 134 Miller integrator, 75
Higher order modulators, 267 Mixed-signal system block diagram, 27
Hold register, 46-50 Modulation noise, 206-208, 212, 237
Hysteresis, 246-247 Modulator,
bandpass, 287-297
I cascaded, 275
first-order, 233
I, in phase or real component, 4-6, 285-286, gain, 243
297-299 higher-order, 265-269
IIR. See Infinite impulse response filter. interpolative, 271-272
Impulse sampling, 28-31 multibit, 269
Infinite impulse response (IIR) filter, 134 noise-shaping, 199-201, 233
Integrator, predictive, 200
analog. See Continuous-time analog integrator. second-order, 253
digital, 19-21, 235 MOSFET-C Integrators, 83-85
discrete analog integrator (DAI), 66-70, 90-93, Multibit modulators, 269
233, 251 Multiplexer (MUX), 236, 297
gain, 113, 244-246, 248, 255, 257-260, 264, Multipliers, 34, 64, 94, 291, 297-298
276, 303 Multistage noise shaping modulator (MASH), 275
leakage, 247-248, 282 MUX. See Multiplexer.
Interpolation, 32, 43-46, 48-50, 71, 139-142, 235,
273
Interpolative modulators, 271-272 N
Inverse Fourier transform, 23
Noise,
averaging, 190
J clock jitter,
correlated double sampling (CDS), 59, 66, 71
Jitter. See Clock jitter. modulation, 206-208, 212
quantization noise, See Quantization noise.
Noise transfer function (NTF), 200-201, 205-208,
K 224, 233, 253-254
Noise-shaping (NS), 199-201, 233, 253
K, oversampling ratio, averaging factor, or Non-overlapping clock generation, 51, 53, 304, 317
interpolation increase, 33, 43-50 Nonrecursive filters, 133
clock jitter, 181-182 Number representation, 123-126
K-Delta-1-Sigma, xvi, 222, 303-304, 320-324 Nyquist,
K-path, Kpath, 50-53, 220-224, 228, 230, 301-302, differences between noise-shaping converters,
317 241
frequency, 17, 21, 27, 29, 31, 37-40
rate ADC, 43, 63, 220-225, 241, 306-308
L
Ladder filter, 117-118 O
Least significant bit (LSB), 119, 204 Offset,
Limit cycle oscillations, 241, 315 autozero, 57-59, 62
Linear phase, 25, 30-31, 145 comparator, 221-222, 224, 246
Linearity, 212, 214 DAC, 214
Linearity requirements for averaging, 191-195 op-amp, 54-55, 58-62, 66, 72, 212, 247-250
Lowpass filters, 73-74 Offset binary format, 123-125, 204-205, 274
LSB. See Least significant bit.
328 Index
T
Taylor series, 6
T/H, See Track-and-hold.
Thermometer code, 308
Time-domain description of reconstruction, 31-33
Time-interleaved ADC, 220-222, 232, 313, 323
Track-and-hold, 41-43
Transconductor-C filters, 86
bilinear, 97
biquad, 114
Triangular response, 133
Tuning, 80, 83, 86
orthogonal tuning, 98
Two’s complement, 123-126, 131, 137, 147,
204-205
subtraction, 126
U
Unit step function. See u(t).
u(t), 35-36
V
Variance, 168, 186
VCM , (VDD/2 or 0.5 V in this book)
VDD, (1.0 V in this book)
Von Hann window. See Hanning window.
W
White noise, 169, 172, 269
Windowing, 175-176
Wireless, 285, 291
I/Q extraction, 297-298