Cmos
Cmos
net/publication/295256513
CITATIONS READS
5 3,264
1 author:
R. Jacob Baker
University of Nevada, Las Vegas
350 PUBLICATIONS 5,172 CITATIONS
SEE PROFILE
All content following this page was uploaded by R. Jacob Baker on 20 February 2016.
30
Data Converter Modeling
Figure 30.1 Generating the sinewave digital code for DAC simulation with an ideal ADC.
2 CMOS Mixed-Signal Circuit Design
v Analog Digital v v
v 111
101
011
t 000
t t t
t
Anti-aliasing filter
In AAF S/H ADC DSP DAC RCF Out
Figure 30.2 Signals resulting from A/D and D/A conversion in a mixed-signal system.
Sampler output
1
t y(t) = x(t) ⋅ δ u (t − nT s ) Sampling impluses
Sampling impluses
(b)
δ u (t − nT s ) (a)
Figure 30.3 (a) Simple sampling gate and (b) SPICE implementation of a sampling gate.
Chapter 30 Data Converter Modeling 3
Noting that the frequency of the input is fin while the sampling frequency is f s (= 1/T s ) , the
spectrum of the input signal is seen in Fig. 30.4a. If we take the Fourier transform of the
input signal after sampling, that is, we look at the spectrum on the output of the sampler,
we get
Vp ∞
T s k =Σ− ∞
Y( f ) = ⋅ [δ( f − f in + kf s ) + δ( f + f in + kf s )] (30.2)
This is the familiar result that a sampled spectrum is repeated, at intervals of f s , as seen in
Fig. 30.4b (shown is the one-sided spectrum, which is what we will use throughout the
book). Note that if an ideal lowpass filter (LPF) is applied to the output spectrum of the
sampler (the output of the sampler is connected to an LPF) with a bandwidth greater than
f in (and lower than f n [the Nyquist frequency]), then the higher order frequency
components can be removed so that only f in remains (this is our smoothing or
reconstruction filter shown in Fig. 30.2).
Y( f ) (2 f s − f in )
fs
Volts fn = Volts
2 Vp fn fs 2 fs
Vp
Ts
f in f f in f
(a) Input spectrum (f s − f in ) (f s + f in ) (2f s + f in )
(b) Output spectrum after sampling
Figure 30.4 One-sided spectrum of a sinewave (a) before and (b) after sampling.
Example 30.1
A sampling gate is strobed with an impulse train running at a frequency of 100
MHz ( f s = 100 MHz and the time in between the impulses, Ts, is 10 ns). Sketch
the resulting output frequency spectrum if a 60 MHz sinewave is applied to the
sampler. Also, sketch the time domain input and output of the sampler.
The resulting frequency spectrum is shown in Fig. 30.5. Notice how connecting
the output of the sampler through an LPF, with an ideal abrupt cutoff frequency of
f n , results in an output sinewave 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 f s /2 (the Nyquist frequency, again, f n ). Reviewing
Fig. 30.2, we see that this is the purpose of the antialiasing filter (AAF). Notice
how, ideally, both the AAF and RCF (reconstruction filter) in Fig. 30.2 are both
ideal LPFs with a cutoff frequency equal to half the sample frequency (the Nyquist
frequency). Figure 30.6 shows the time domain sketch of the sampler's output. T
4 CMOS Mixed-Signal Circuit Design
f s − f in Original signal, f in f s + f in
It should be clear from the preceding discussion that (1) sampling a signal results
in a reproduction of the sampled signal's spectrum at DC, fs , 2fs , 3fs , etc., (2) the input
signal's spectrum should have no significant spectral content above fn in order to avoid
aliasing, (3) to avoid aliasing both filtering the input signal using an AAF and increasing
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
lowpass RCF should be used.
Note that our discussion illustrates the operation of a sampling gate driven with
impulse signals. As shown in Fig. 30.2, a practical system would have other building
blocks. We would rarely, if ever, sample a signal and then reconstruct it without
processing it first.
25 ns 50 ns 75 ns 100 ns
Figure 30.6 Time domain input and output for Ex. 30.1.
these filters must be analog by design. The ideal cutoff frequency for the filters is fn
(assuming the sampling rate on the input of the system is the same as the sampling rate on
the system's output) and the filters should ideally have linear phase. Let's discuss these two
ideal characteristics.
The ideal magnitude response, shown in Fig. 30.7a, passes all spectral content
below the Nyquist frequency while removing all signals above this frequency. The ideal
phase response, shown in Fig. 30.7b, provides a constant delay, to , to all signals below fn.
In other words, the filters remove all unwanted signals while not distorting the wanted
signals.
slope = −2πt o
f n = f s /2 f
(a) (b)
Figure 30.7 (a) Ideal magnitude and (b) phase responses for the AAF and RCF.
Example 30.2
Discuss why the ideal AAF filter will not introduce distortion into the desired
portion of an input signal.
If our input signal is called v in (t) and the desired spectral content of this signal
after filtering is called v in (t) (that is, v in (t) contains nonzero spectral content only
at frequencies below f n ), then the output of the AAF, v out (t) , will be a time-shifted
(with a constant delay of to ) and filtered version of the input, as seen in Fig. 30.8.
Note that linear phase is equivalent to saying "constant delay." If our input signal
is already bandlimited to f n , then the output of the AAF is simply a time-shifted
version of the input. T
In Out
v in (t) v out (t) = v in (t − t o )
AAF
In Out v in (t) v out (t)
0 to t
Example 30.3
Suppose that the circuit, shown in Fig. 30.9, is used as an AAF filter in a data
conversion system. If the inputs to the system are two sinewaves with frequencies
of 4 MHz and 40 MHz determine whether the waveforms coming out of the AAF
will be distorted. Using SPICE show the input and output signals of the AAF.
v out (t)
∠
v in (t) 0
45
500 MHz
90
The 4 MHz input doesn't see any attenuation. The gain, or amplitude response, of
the filter at 4 MHz is unity (0 dB). The filter attenuates the 40 MHz input by 0.779
(−2.17 dB).
The phase response of the simple RC filter is given, in degrees by
∠v out /v in = θ(f ) = −tan −1 (2π ⋅ RC ⋅ f )
The phase shift through the filter at 4 MHz is approximately zero (the 4 MHz input
doesn't see any delay while passing through the filter). This is the ideal phase
response of this filter, i.e., to = 0. Looking at Fig. 30.9 we can conclude that only
at frequencies below approximately 5 MHz will the filter not exhibit phase
distortion. The phase shift through the filter at 40 MHz is −39° (the negative sign
indicates that the output is lagging the input or, in other words, occurs later in time
than the corresponding point on the input). Since phase is related to delay by
% of period, T
to
θ(f ) = ⋅ 360 = t o ⋅ f ⋅ 360
T
Chapter 30 Data Converter Modeling 7
the delay the 40 MHz sinewave sees passing through the filter is 2.7 ns. The
SPICE simulation results are shown in Fig. 30.10 assuming each sinewave input is
centered at ground and has an amplitude of 1V.
Also note that this filter does a poor job attenuating frequencies above 50
MHz. For example, the attenuation at 500 MHz (one decade above 50 MHz) is
only −20 dB (0.1). It can be concluded that unless f s /2 (the Nyquist frequency) is
much larger than the cutoff frequency of the simple RC LPF aliasing will (possibly)
still occur in significant amounts. In fact, we could argue that because of the
inherent noise present in any electronic circuit, aliasing will always occur when
sampling a signal (the wideband noise gets aliased down into the base spectrum
[the spectrum below the Nyquist frequency]). The question then becomes, "How
much aliasing is OK?" T
Input
Example 30.4
Determine the transfer function of the filter made with a 5 ns long (= to) ideal
transmission line shown in Fig. 30.11. Simulate the filter's frequency and phase
response using SPICE. This filter is called a continuous-time comb filter.
In this analysis, we assume that 500-ohm resistors do not load the input or output
of the delay line so that
v in + v in ⋅ e j⋅2π f⋅(−t o )
v out =
2
8 CMOS Mixed-Signal Circuit Design
v in e j2πf (−to )
v in Z o = 50 Ω t o = 5 ns
50 500
v out
500
v in
Notice that the phase response of this filter is linear. The SPICE simulation results,
plotted on a linear frequency scale, are shown in Fig. 30.12. The reason this filter
is called a "comb filter" should be obvious (the response looks like a comb). Notice
how the delay line length is related to the points where the magnitude response
goes to zero. Also note that this filter could be useful to isolate channels in a
communication system and easily implemented on a PC board using a microstrip
transmission line. The SPICE netlist that generated this figure is given below.
* Figure 30.12 CMOS: Circuit Design for Mixed-Signal Systems *
.AC LIN 1000 1MEG 1000MEG
Vin Vin 0 DC 0 AC 1
Rtout Vtout 0 50
Rt1 Vtout Vout 500
Rt2 Vin Vout 500
Chapter 30 Data Converter Modeling 9
Finally, before leaving this example, consider the dB (magnitude) and phase
responses of this filter on a log frequency plot, Fig. 30.13. It would appear that the
magnitude of the transfer function at 100 MHz, 300 MHz, 500 MHz, etc., is
nonzero. However, as the equation for the magnitude response shows, this isn't the
case. At these frequencies v out /v in is zero indicating, in the plots shown in Fig.
30.13, that the lower limit is set by step size (number of points per decade) used in
AC SPICE simulation. Increasing the number from 1,000, which is what was used
to generate Fig. 30.13 to, say, 10,000 will give more accurate results. T
100 MHz
Figure 30.12 SPICE simulation results for the comb filter of Ex. 30.4.
Figure 30.13 SPICE simulation results, in dB, for the comb filter of Ex. 30.4.
We know 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 the transfer function of the system, we apply the unit impulse to the input of the
system (a very large amplitude, very short time duration pulse, see Fig. 30.15). We then
look at the system's output in the time domain followed by taking the Fourier transform of
this output to get the system's transfer function. Therefore (in the reverse order), to
determine the time domain response of the ideal RCF, given the transfer function, we take
10 ns 30 ns 50 ns 70 ns 90 ns
the inverse Fourier transform of the transfer function. The ideal RCF's transfer function
(Fig. 30.7) can be defined by
H( f ) = 1/f s for f < f n else H( f ) = 0 (30.3)
The time domain response is then given, remembering 2f n = f s , by
fn
1 ⋅ e j⋅2π⋅f⋅t ⋅ df = e j⋅2π⋅f n ⋅t − e −j⋅2π⋅f n ⋅t =
h(t) = ∫ fs j ⋅ 2π ⋅ f s ⋅ t
(30.4a)
−f n
sin πf s ⋅ t
= Sinc(πf s ⋅ t)
πf s ⋅ t
H(jω)
1
1 0.64 = −3. 9 dB
fs
0.13
f n = f s /2 f t
0
−3T s −2T s − T s Ts 2T s 3T s
(a) Ts
(b) 0.21 = −13.5 dB
2
Figure 30.16 (a) Ideal RCF frequency response and (b) impulse response (time).
12 CMOS Mixed-Signal Circuit Design
30.17 shows the individual response outputs of an ideal RCF with the impulse train of Fig.
30.14 as the input. The output of the RCF is the weighted sum of the individual responses,
of the form Sinc(x), from each of the weighted impulse inputs into the RCF (using
superposition). While this figure is "busy," the basic concept of reconstruction should be
obvious.
time
resistance, zero output resistance, etc., the SPICE model shown in Fig. 30.18 will be used.
Voltage-controlled voltage sources, with the E prefix, are used to model the ideal
operational amplifiers. The switch is modeled with a voltage-controlled switch, an S
device. The model is used in the following netlist:
* Figure 30.19 CMOS: Mixed-Signal Circuit Design *
.tran .1n 500n 0 .1n UIC
Vin Vin 0 DC 0 Sin 0.75 0.75 5MEG
Vclock Clock 0 DC 0 Pulse 0 1.5 0 0 0 100p 10n
Vtrip Vtrip 0 DC .75
Ebufin Vinb 0 Vin Vinb 100MEG
S1 Vinb Vins CLOCK VTRIP switmod
Rout Vins 0 10k
Ebufout Vout 0 Vins Vout 100MEG
.model switmod SW
.end
v out
v in
Sampling impluses Vout
Vins Ebufout
0
time domain signal (we will use the Hanning window [a.k.a. von Hann window] unless
otherwise indicated), (2) sampling the signal, and (3) taking the Fourier transform of the
signal. Windowing ensures that abrupt transitions do not occur at the beginning and end of
the signal to be transformed. It's important to realize that taking the DFT of a signal that
has already been sampled results in a spectrum with amplitude errors (more on this in a
moment.) To perform a DFT in SPICE we first ensure that the signal to be transformed
has a linear time step. To do this we use
linearize Vout Vin
where Vout and Vin are the signals we are interested in transforming. The linearize
command with no arguments will linearize all of the variables available in the simulation.
Next, we use the spec command (spectral analysis command) in SPICE
spec 0 200MEG 2MEG Vout Vin
This command takes the DFT of Vout and Vin over the range of DC to 200 MHz with a
resolution of 2 MHz. The minimum resolution allowed when using the spec command
(DFT) is set by the transient simulation time, or
The spectrums, as an example, on the input and output of an impulse sampler sampling a
10 MHz sinewave at 100 MHz are shown in Fig. 30.20. Note that 1 µV was added to
both signals so as to set the noise floor in the display to −120 dB. The DC portion of the
input signal is 0.75 V while the peak voltage of the 10 MHz sinewave is also 0.75 V (0.75
V = −2.5 dB). Note that in Fig. 30.20b, because of the double sampling mentioned above,
the amplitude of the signals in the output is different than that predicted (see Fig. 30.4).
We can estimate the baseline reduction, resulting from taking the DFT of an impulse
sampled signal (the signal has nonzero values only during the sampling times) using
2 ⋅ (step size)
Baseline reduction (or duty cycle) ≈ (30.7)
Ts
For Fig. 30.20b the maximum stepsize specified in the transient simulation was 100 ps (for
each cycle of Ts one point is defined as having a total deviation, after being linearized, of
approximately 200 ps including rise and fall times, hence the factor of 2 in Eq. (30.7). The
baseline, using Eq. (30.7), is 200 ps/10 ns or 0.02 (−34 dB). The 10 MHz signal in Fig.
30.20b is −2.5 dB below the −34 dB baseline. Note how the DC signal is aliased up to the
sampling frequency (100 MHz). At DC, with reference to the baseline, the signal
amplitude is also −2.5 dB (−36.5 dB). However, when it is aliased up to the 100 MHz (the
sampling frequency) it is doubled (+6 dB). The doubling comes from adding the images
f s + 0 and f s − 0 and results in an amplitude of −36.5 dB + 6 dB or −30.5 dB.
The step size used in a transient simulation, as we saw above, is an important
parameter that needs specification when performing a spectral analysis using SPICE. In
the netlist that generates Figs. 30.19 and 30.20, we used 100 ps, but made no comment on
why this value was selected. Poor selection of the step sizes can give erroneous results if
the values are too large or cause the simulation to last a long time if the values are too
small. The transient simulation characteristics in a SPICE netlist are specified using
.tran print-step stop-time delay-time maximum-stepsize <UIC>
The step size, for a general simulation with nonideal components, can be set using
step size = 1 % ⋅ T s or with ideal components 10% ⋅ T s (30.8)
If our sample frequency, fs , is 100 MHz then we would set our step size to 100 ps using
.tran 100p 2000u 0 100p UIC
The term UIC forces the simulation to start with initial conditions (use initial conditions),
such as an initial voltage across a capacitor. The simulation always starts at zero time.
However, specifying a delay time in the simulation will make SPICE start saving data at
16 CMOS Mixed-Signal Circuit Design
2.5 dB (0.75V)
2.5 dB (0.75V) below baseline or 36.5 dB
−30.5 dB
−34 dB
Volts
10 MHz
10 MHz Input to the sampler Output of the sampler
(a) (b)
Figure 30.20 Spectrums of the signals shown in Fig. 30.19.
the time specified by the delay-time parameter. This parameter is useful in removing, from
a spectral response, for example, the start-up transients in a simulation or keeping the size
of a raw output file from getting too large.
Representing the Impulse Sampler's Output in the z-Domain
Consider the output of an impulse sampler, y(t) , with an input of x(t) shown in Fig. 30.21.
The sampler output can be written as
∞
y(t) = x(t) ⋅ Σ δ u (t − kTs )
k = −∞
(30.9)
y(t) Ts
Ts time
Y( f ) = ... + x(−1)e (1)⋅j2π⋅f⋅Ts + x(0)e (0)⋅j2π⋅f⋅Ts + x(1)e (−1)⋅j2π⋅f⋅T s + x(2)e (−2)⋅j2π⋅f⋅T s + ...
(30.11)
where the term e j2π⋅f⋅T s corresponds to a phase shift of 2π ⋅ f ⋅ T s (radians) when the output
of the sampler, Y( f ) , is evaluated at the frequency f. In other words, each consecutive
sample coming out of the impulse sampler is shifted in the time domain by Ts (which is
simply saying, in words, what Fig. 30.21 shows). If we define
f
j2π⋅ f
z ≡ e j2π⋅f⋅T s = e s (30.12)
then the output of our sampler can be written as
Y(z) = ... + x(−1)z 1 + x(0)z 0 + x(1)z −1 + x(2)z −2 + ... (30.13)
or
∞
Y(z) = Σ x(k) ⋅ z −k
k = −∞
(30.14)
Example 30.5
Determine the output of an impulse sampler in the z-domain if its input, x(t), is a
unit step. What is the sampler's impulse response H(z) [= Y(z)/X(z)] ?
The unit step, u(t), is defined by
u(t) = 1 for t ≥ 0
u(t) = 0 for t < 0
The time domain output of the sampler, with u(t) as an input, is given by
∞
y(t) = Σ u(t − kT s)
k=0
The time domain signals used in this example are shown in Fig. 30.22. The
sampler's impulse response is simply unity since the z-transform of the input
(assuming the input was passed through an ideal impulse sampler so that we can
take the z-transform of the signal) and output of the sampler are identical, that is,
H(z) = 1 . T
18 CMOS Mixed-Signal Circuit Design
Ts
Sampler output
Ideal impulse 1
In sampler Out 0
1 Sampler input
f s Sampling 0
clock time
Example 30.6
What is the effect of multiplying H(z), in Ex. 30.5 (or any z-domain transfer
function), by z−1?
Multiplying any z-domain transfer function by z−1 is equivalent to shifting the
system's output later in time by Ts. The result of changing the ideal sampler's
transfer function from unity to z−1 is shown in Fig. 30.23. Multiplying by z−L shifts
the output of the system later in time by L ⋅ T s T
Ts
Sampler output
Out 1
Ideal impulse z −1
In sampler 0
1 Sampler input
f s Sampling 0
clock time
An Important Note
It's important to note that our impulse sampler quantizes1 the input signal in time but not
amplitude (unlike an analog-to-digital converter which quantizes the input in both time and
amplitude). The amplitude out of the ideal impulse sampler is exactly the same as the
amplitude input to sampler at the sampling impulse time. We'll find that the z-transform
can be used to describe systems using both quantization in time as well as in amplitude. In
other words, whether we are discussing digital words, in a binary format or
sampled-analog waveforms with amplitudes of volts, amps, or coulombs, we can use the
z-transform to represent the discrete-time systems that process the signals.
1
Quantize: to limit the possible values of a quantity to a discrete set of values. Quantizing in time,
for example, means that the output amplitude is only defined at certain discrete times (such 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 discussed in
the next section).
Chapter 30 Data Converter Modeling 19
The switches S1 and S2 in the netlist above sample the input using the input clock. Note
that both switches can be closed, momentarily in Fig. 30.24, at the same time. The time
V out
V inbuf V ins V out1
V in
1e-16
Clock
Clock 1e-10
Figure 30.24 SPICE model of the ideal sample and hold (S/H).
20 CMOS Mixed-Signal Circuit Design
that the switches are closed is approximately equal to the transient step time. The charge
sharing between the capacitors is affected by having both switches closed at the same
time. Values given in this figure were selected so that a million-to-one ratio existed
between the two capacitors (120 dB range.) Because both switches are closed at the same
time, the difference between the two capacitors can be made smaller without affecting the
circuit's operation. Also note that over a time set by GMIN (remember a resistor with a
value of 1/GMIN is placed across every pn-junction in a SPICE simulation and GMIN's
default value is 1e-12 or 1 GΩ) and the capacitor values the charge on the capacitors will
leak off causing droop. For the 0.1f capacitor the associated RC time because of GMIN is
100 µs (increasing this capacitor to 10f won't affect the sampling operation and pushes the
RC time up to 10 ms).
The accuracy of the S/H is ultimately limited by the tolerances, that is, RELTOL,
ABSTOL, and VNTOL of the simulation. For an accurate simulation we may add
.options RELTOL=1u VNTOL=1u ABSTOL=1p
to the netlist. The accuracy of a simulation will be discussed in greater detail later.
S/H Spectral Response
Consider the application of a sinewave, at a frequency fin , to the ideal S/H shown in Fig.
30.26. To make the discussion as general as possible assume that the output of the S/H
can return-to-zero (RZ) as shown in Fig. 30.27 (which shows coarse time quantization for
Chapter 30 Data Converter Modeling 21
v in
f in
time
time
In Sample and Out
hold (S/H) y(t)
clock
a simpler figure and illustration of the concept of RZ). Note that as T approaches Ts we
get the operation of the S/H in Fig. 30.26. The output of the ideal S/H is given by
v in (t)
∞
h(t)
y(t) = Σ V p sin (2πf in ⋅ nT s ) ⋅ [u(t − nT) − u(t − (n + 1)T )] (30.15)
n=−∞
Note that the sine term is only defined at discrete sampling instances so that its spectrum
is given by Eq. (30.2). The spectrum of the sampling pulse, H( f ) , because of the duality
of the Fourier transform, is given by reviewing Fig. 30.16 or calculated using
Ts
Fourier[u(t − nT) − u(t − (n + 1)T] = ∫0 [u(t − nT) − u(t − (n + 1)T]e −j⋅2π⋅f⋅t ⋅ dt (30.16)
which is evaluated as
phase magnitude
−j⋅2π⋅f⋅T
H( f ) = e − 1 = e −j⋅π⋅f⋅T ⋅ e j⋅π⋅f⋅T − e −j⋅π⋅f⋅T = e −j⋅π⋅f⋅T ⋅ T ⋅ Sinc(π ⋅ f ⋅ T) (30.17)
−j ⋅ 2π ⋅ f j ⋅ 2π ⋅ f
Sinewave in
S/H out
time
Ts
The magnitude of Eq. (30.17), H( f ) , is plotted in Fig. 30.28. The phase response
corresponds to a shift in time of T/2 so, to simplify the math below, we will only concern
ourselves with the magnitude response of H( f ).
h(t) H( f )
u(t − nT) − u(t − (n + 1)T)
1 T
T Ts time
f
(a) 0
1 2 3
T T T
(b)
or
Ideal impluse sampler response
Weighting from S/H, H(f)
∞
Vp
Y( f ) = T ⋅ Sinc(π ⋅ T ⋅ f ) ⋅ ⋅ Σ [δ( f − f in + kf s ) + δ( f + f in + kf s )] (30.19)
Ts k = − ∞
As T → 0 (h(t) → δ u (t) ), the frequency response of the sample-and-hold approaches the
ideal impulse sampler of Sec. 30.1.1. Also, note that using an RZ format (making T < T s )
can reduce the amount of attenuation introduced by the S/H ( H( f ) doesn't roll off as
fast.)
For most circuit designs, T = T s so that, as Eq. (30.19) shows, the
sample-and-hold operation weights the amplitude of the ideal impulse sampler's frequency
πf πf
response by Sinc or Sinc . Note that at the sampling frequency (fs=1/Ts) the
fs 2f n
output of the ideal S/H goes to zero. Let's illustrate the frequency response of an ideal S/H
using an example.
Example 30.7
Using the ideal S/H SPICE model show and discuss the spectrum resulting from
sampling a 3 MHz sinewave at 100 Msamples/s.
Chapter 30 Data Converter Modeling 23
The results of passing a 0.75 V (peak) sinewave centered at 0.75 V (−2.5 dB)
through the ideal S/H are shown in Fig. 30.29. We have also plotted the response
of the S/H, H( f ) in this figure. The attenuation the 97 MHz image sees is
−13.5 dB
−3.9 dB
Volts, peak
S/H response
Noise floor
fs
50 MHz
Figure 30.29 Output of a S/H after sampling a 3 MHz sinewave at 100 MHz.
24 CMOS Mixed-Signal Circuit Design
repetitively sampling and holding a signal results in only one S/H attenuation hit
(assuming the timing is such that a sampling operation is not occurring when the
previous S/H stage's output is changing). This means that topologies that use
several S/H operations on an input signal, such as a pipeline ADC, only attenuate
the signal by Sinc(πf/f s ) once. T
The output of the S/H (assuming T = T s ) should be passed through a two-stage
reconstruction filter, to recover the input signal. One of the stages will have the frequency
response of the ideal RCF of Fig. 30.16. The other stage will have a frequency response
given by
1 πf
H RCFSH ( f ) = = (30.20)
Sinc fs 2f n sin 2f n
πf πf
to compensate for the attenuation of the S/H Sinc response. The shape of the ideal
reconstruction filter is shown in Fig. 30.30. Again, increasing the sampling frequency
relative to the input frequency will ease the requirements placed on the reconstruction
filter. Note how using the RZ format modifies the requirements placed on the
reconstruction filter to the point, when using impulse sampling, of having the brick wall
ideal RCF of Fig. 30.7.
H RCFSH (f )
0 dB
3.9 dB (1.56)
1 1
Overall response of RCF
for a S/H
f fn f
fs 2f s
Before we leave this section, let's answer the question: "What sets the value of the
noise floor in SPICE (Fig. 30.29)?" We can limit the noise floor, in Fig. 30.29 for
example, by adding 1 µV to voltages in the circuit. However, the SPICE-simulated
spectrum's noise floor, which is set by simulation variations, is limited by the RELTOL
parameter. Also, the length of the simulation can be important. ABSTOL, which defaults
to 1 pA, and VNTOL, which defaults to 1 µV, signify when a current or voltage has
converged in a SPICE simulation. If the step change in the simulation, for all currents and
voltages at a given time, is within ABSTOL (for currents) or VNTOL (for voltages), then
SPICE moves on to the next step in time (for a transient simulation). The parameter
RELTOL was added to SPICE so that simulations involving large currents and voltages
Chapter 30 Data Converter Modeling 25
were not forced to use ABSTOL and VNTOL to signify convergence. In other words, if a
current is approximately 10 A, we won't force the SPICE number for the current to be
10.000000000001. Instead we use 10.01 (the product of RELTOL [assuming = 0.001]
and 10 A) to signify convergence. To signify that a current has converged, we use the
larger of
ABSTOL or RELTOL ⋅ I simulated (30.21)
while for a voltage we use the larger of
VNTOL or RELTOL ⋅ V simulated (30.22)
For the simulation shown in Fig. 30.29, we set RELTOL to 10 −6 so that our 1 V level
signals simulate to within 1 µV of their "actual" values. This keeps the simulation noise
from setting our noise floor. The practical problem of reducing RELTOL is convergence
when nonideal components (e.g. MOSFETs) are added to the simulation. Trade-offs must
be made between simulation noise and convergence when using both ideal and nonideal
components in a simulation.
Circuit Concerns for Implementing the S/H
Figure 30.31 shows a single-ended input and output S/H implementation using either an
op-amp or an OTA (operational transconductance amplifier). At the time t0 , the φ 1 and φ 2
switches are closed while the φ 3 switches are open. During the time between t1 and t2 the
input charges the hold capacitor CH . The input is connected to the left side, or bottom
plate (the plate closest to the substrate), of CH , while because of the op-amp, the right side
(top plate) is connected to ground (or a common mode voltage, VCM). At t1 the φ 1 switch
opens and for a very short time (set by t 3 − t 1 ) the op-amp operates open loop (no
feedback). As the top plate is always at ground (or VCM) at t1 , the charge injection and
capacitive feedthrough resulting from the φ 3 switches turning off are independent of the
input signal. When the φ 2 switch turns off, the charge injection will, ideally, flow into the
low-impedance input, vin , since the impedance looking into the right of the φ 2 switch is
large. This, again ideally, leaves the voltage across the hold capacitor unaffected by the
charge injection resulting from turning off the switches. This sequence of turning off the
φ3
to t1 t2 t3
φ1 φ1
φ2 CH
v in φ2
φ3
v out
φ3
Bottom plate
switch to the right of CH followed by turning off the switch connecting vin to CH is often,
confusingly, called bottom plate sampling. Bottom plate sampling is illustrated in its
simplest form in Fig. 30.32. In this figure the switch connected to the bottom plate of the
capacitor is turned off first. When the φ 2 switch turns off, the charge can be injected into
the low-impedance node, the input vin , or into the series combination of CH and the off
switch. The charge takes the lowest impedance path to ground and thus most of the
charge injection resulting from the φ 2 switch turning off flows through vin , leaving the
voltage across the hold capacitor unaffected. We should see why the name "bottom plate
sampling" is confusing. Reviewing Fig. 30.31, we see that the top plate of the hold
capacitor is connected to the φ 1 switch while, in Fig. 30.32, the bottom plate of the hold
capacitor is connected to the φ 1 switch.
CH
v in
Turns off first
φ1
Returning to the discussion of the operation of the S/H of Fig. 30.31 we see that at
t3 the φ 3 switches turn on and the op-amp behaves as a voltage follower holding the
sampled input voltage. The sampling instant occurs between t1 and t3 (which should be
short to keep the op-amp output from drifting toward VDD or ground.)
Consider the ideal transfer characteristics of a 3-bit DAC shown in Fig. 30.33. (For
a detailed review of general DAC characteristics, see Ch. 28.) Notice in this figure that we
have drawn two reference voltages, V REF+ and V REF− , and are assuming that
V REF+ > V REF− . When a digital input of 000 is applied to the DAC, the output voltage
becomes V REF− . When the input code is increased to 001, the output of the DAC (an
analog voltage defined at discrete amplitude levels) increases by one least significant bit
(LSB). If the DAC has an input code with a number of bits, N, then we can define an LSB
as
VREF+ − V REF−
1 LSB = = V LSB (30.23)
2N
If, for example, V REF+ = 1.25 V and V REF− = 0.25 V and N = 3 , then our LSB, the vertical
distance between adjacent points in Fig. 30.33, is 0.125 V. Note that in our discussion of
V OUT − VREF−
V REF+ − V REF−
V REF+
8/8
V REF+ − 1LSB
7/8
6/8
5/8
1 LSB
4/8
3/8
2/8
1/8
V REF− Digital
0 input code, b 2 b 1 b 0
000 001 010 011 100 101 110 111
V REF+
b2
Ideal
b1 V OUT
3-bit DAC
b0
V REF−
an ideal DAC we are assuming that the output of the DAC ranges from V REF− up to
V REF+ − 1 LSB . We could just as easily have assumed that the output ranged from
V REF− + 1 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 V REF+ = VDD = 1.5 V and
V REF− = VSS = 0 V . In Ch. 33 we discuss a submicron CMOS process using these power
supply voltages, 1.5 V and 0 V. Selection of the power supply rails, which are noise free
in a SPICE 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). If we need more resolution
when using our ideal DAC, we will simply increase the number of bits, N, used and hence
decrease the value of the DAC's LSB.
SPICE Modeling Approach
We can write the output of the ideal DAC in terms of the reference voltages and digital
input codes b N (which are logic "0" or "1"), and assuming that an input code of all zeroes
results in an output voltage of V REF− , as
Example 30.8
Write the nonlinear dependent source statement for an ideal 12-bit DAC.
The statement follows:
Bout Vout 0 V=((v(vrefp)-v(vrefm))/4096)*
+(v(B11L)*2048)+v(B10L)*1024+v(B9L)*512+v(B8L)*256
+v(B7L)*128+v(B6L)*64+v(B5L)*32+v(B4L)*16+v(B3L)*8+
+v(B2L)*4+v(B1L)*2+v(B0L))+v(vrefm)
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. T
The next thing we need to concern ourselves with is the digital logic levels. We
want to use our ideal DAC with nonideal (real) circuits where the logic voltage levels may
not be well defined. We need to determine and use a switching-point voltage based on the
Chapter 30 Data Converter Modeling 29
power-supply voltage VDD. We will assume the input logic code is a valid logic "1" if its
amplitude is greater than VDD/2 and a logic "0" if its amplitude is less than VDD/2. We
can implement the VDD/2 switching point, or trip voltage, using the following SPICE lines
*Generate Logic switching point, or trip, voltage
R1 VDD trip 100MEG
R2 trip 0 100MEG
The solid logic levels can be generated using the following subcircuit SPICE code. The
switch implementation is shown in Fig. 30.34.
.subckt Bitlogic trip BX BXL
Vone one 0 DC 1
SH one BXL BX trip Switmod
SL 0 BXL trip BX Switmod
.model switmod SW
.ends
Using the above code, the subcircuit definition for an ideal 8-bit DAC can be
written, as shown in Fig. 30.35. Using this subcircuit in the following netlist, we can show
the operation of an ideal 8-bit DAC:
VDD VDD 0 DC 1.5
VREFP VREFP 0 DC 1.5
VREFM VREFM 0 DC 0.0
.ends
In this netlist we are assuming V REF+ = 1.5 V and V REF− = 0 . The pulse sources step the
DAC through all possible codes, i.e. from 00000000 (= 0 V) all the way up to 11111111
(= 1.5 V − 1 LSB) in increments of 1.5/256 or 5.859 mV (= 1 LSB.) The simulation
results are shown in Fig. 30.36. It should be very easy to see how to implement any
resolution of ideal DAC at this point using SPICE.
Before leaving the ideal DAC let's discuss how to shift the ideal output
characteristics up by 1 LSB. The DAC in Fig. 30.35 has an output range of 0 V (V REF− ) to
VDD − 1 LSB (V REF+ − 1 LSB) . We can rewrite Eq. (30.25) as
1 LSB
V OUT = (V REF+ − V REF− ) ⋅ 1N ⋅(b N−1 2 N−1 + b N−2 2 N−2 + ... + b 1 ⋅ 2 1 + b 0 ) + V REF−
2
(30.26)
Chapter 30 Data Converter Modeling 31
DAC output
DAC inputs
b7
DAC inputs (level shifted for easy viewing)
b0
To shift the output up by 1 LSB (so the output of the ideal DAC ranges from 1 LSB
above V REF− to V REF+ ) we simply add one to the binary-weighted term in the parentheses
1LSB
V OUT = (V REF+ − V REF− ) ⋅ 1N ⋅(b N−1 2 N−1 + b N−2 2 N−2 + ... + b 1 ⋅ 2 1 + b 0 + 1) + V REF−
2
(30.27)
This equation is trivial to implement in our ideal DAC by adding two characters to our
nonlinear dependent source, i.e., "+1."
30.2.2 The Ideal ADC
The characteristics of our ADC are shown in Fig. 30.37. (Again, a complete discussion of
ADC characteristics was given in Ch. 28.) Notice how in this figure the transfer curve is
shifted to the left. If we were to flip the curve on it's side and mark, with black dots, the
intersection of the analog input voltage with the ADC transfer curve, we would have the
DAC transfer curve of Fig. 30.33. Again 1 LSB is given by Eq. (30.23). Notice how
converting a (normalized) input voltage of 0.1 V will result in an output code of 000
which is the same output code resulting from converting 0 V. Unlike the ideal DAC, the
ideal ADC quantizes its input with the practical result of adding noise to the input signal.
This noise is often called quantization noise.
The implementation of the ideal ADC consists of an ideal S/H followed by passing
the output of the S/H (the held signal) through an algorithm to generate the output bits.
The algorithm we use is based on a pipeline ADC and follows:
1. The input signal is sampled and held.
2. This held signal is input to a comparator that compares the input value to a
reference voltage.
3. If the input signal is greater than the reference voltage, the output bit is set to a
high, and the reference signal is subtracted from the input. The difference is multiplied by
two and passed to the output of stage.
4. If the input signal is less than the reference voltage, the output bit is set low.
The input signal is multiplied by two and passed to the output of the stage.
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 V REF+ and V REF− followed by subtracting V REF− so that
the V CM is referenced to 0 V. This can be written as
V REF+ + V REF− V + V REF− V − V REF−
V CM = → V CM0 = REF+ − V REF− = REF+ (30.28)
2 2 2
We also want to level shift the input signal so that it is referenced to 0 V. In addition, we
want to shift the transfer curves to the left by 1/2 LSB as seen in Fig. 30.37. To do this we
Chapter 30 Data Converter Modeling 33
Digital
output code, b 2 b 1 b 0 1 LSB
111
110
Analog input
101
100
011
010
001
V IN − V REF−
000
0 1/8 2/8 3/8 4/8 5/8 6/8 7/8 8/8 V REF+ − V REF−
Analog input voltage
V REF−
V REF+ − 1 LSB
V REF+
b2
Ideal
V IN b1
3-bit ADC
f clk b0
V REF−
use the following SPICE statement (for an 8-bit ADC where V(OUTSH) is the output
voltage of the ideal S/H [the input to the pipeline algorithm above])
* Level shift by VREFM and 1/2LSB
BPIP PIPIN 0 V=V(OUTSH)-V(VREFM)+((V(VREFP)-V(VREFM))/2^9)
* Pipeline stage
.SUBCKT ADCBIT VDD VTRIP VCM VIN BITOUT VOUT
S1 VDD BITOUT VIN VCM switmod
S2 0 BITOUT VCM VIN switmod
Eouth Vinh 0 VIN VCM 2
Eoutl Vinl 0 VIN 0 2
S3 Vinh VOUT BITOUT VTRIP switmod
S4 Vinl VOUT VTRIP BITOUT switmod
.model switmod SW
.ends
*** END ADC Subcircuit *************************************
Example 30.9
Modify the SPICE code of Fig. 30.38 so that the subcircuit simulates an ideal
12-bit ADC.
We can change the level-shift statement (change 9 to 13) to
* Level shift by VREFM and 1/2LSB
BPIP PIPIN 0 V=V(OUTSH)-V(VREFM)+((V(VREFP)-V(VREFM))/2^13)
where the last statement is a modification, in the 8-bit ideal ADC, of the existing
statement for X7. T
We can simulate the operation of our ideal 8-bit ADC in several ways. Let's begin
by simply applying a ramp from V REF− to V REF+ (0 to 1.5 V) to the ADC while clocking
the ADC at 100 MHz. The results are shown in Fig. 30.39. Additional simulations using
the ideal ADC will be left as an exercise for the reader. We are now in a position to put
our ideal ADC and DAC together so that we can look at the spectral response and
limitations resulting from quantization noise.
Summary
It's important to realize the usefulness of the simulation models we have just developed. In
any mixed-signal simulation using SPICE we can use our ideal ADC to generate a digital
signal, most often a sinewave, as an input source. We can use the DAC to convert a digital
word into an analog waveform. We can then take the discrete Fourier transform of the
resulting analog waveform, using the SPICE "spec" command, and view the digital data's
spectrum.
Note that in this chapter we are only discussing the use of the offset binary format
(see Ch. 29) for our digital words (0000... corresponds to V REF− and 1111... corresponds
to V REF+ − 1 LSB ). It should be clear that we can modify our ideal data converters to
work with any data format. We could also add digital logic to our converter subcircuit for
the format conversion and continue to use the ideal ADC/DAC developed in this chapter.
ADC Input
b0
Clock
Figure 30.40 Passing a signal through an ADC and then through a DAC.
Signal plus noise, S+N, volts peak 7 MHz, 0.75 V peak (−2.5 dB)
ADC Input
fn f clk = f s
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 V REF+ and
V REF− 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.
2. The ADCs LSB is much smaller than the input signal amplitude. When this isn't
the case, the output of the ADC can appear squarewave like (when converted back into an
analog waveform) and result in a spectrum, once again, that contains spikes or spurs. We'll
see later in the book that adding or subtracting a fed-back signal (from the output based
on the expected or past quantization noise) to the input modifies this 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 of Fig. 30.41 1 LSB = 5.86 mV and T s = 10 ns so that the
input must change at least 5.86 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
make the requirement on the input of being busy practical in an actual circuit.
Chapter 30 Data Converter Modeling 39
We use these assumptions (Bennett's criteria) in the following discussion unless otherwise
indicated.
Figure 30.43 Taking the difference between the S/H input and output.
An Important Note
It's important to note that simply sampling an input waveform, using a S/H, does not result
in quantization noise, as seen in Fig. 30.29. The amplitude into the ideal S/H, at the
sampling instant, is exactly the same as the amplitude out of the ideal S/H. To understand
why this is important, consider the test setup shown in Fig. 30.43. If we input the 3 MHz
sinewave of Ex. 30.7 into this circuit, we get the outputs shown in Fig. 30.44. Clearly
there is a difference between the S/H's input and its output. However, this difference has
nothing to do with noise, an unwanted signal, since passing the output of the S/H, VOUTSH ,
V OUT
Difference output
Difference spectrum
Volts, peak
Input spectrum
Figure 30.44 (a) Time domain difference between S/H input and output and (b) spectrum.
40 CMOS Mixed-Signal Circuit Design
through the ideal reconstruction filter of Fig. 30.30 results in an exact replica of the S/H
input VIN .
RMS Quantization Noise Voltage
If we were to set up a test configuration similar to that shown in Fig. 30.43 (see Fig.
30.45), where the input to the ADC is subtracted from the DAC output, the resulting
output waveform would have little to do, in every case, with the quantization noise. This
is especially true when the input to the ADC contains a broad frequency spectrum
extending from DC to the Nyquist frequency, f n = f s /2 . However, if we simply apply a
slow linear ramp to the input of this test setup (to limit the input frequency spectrum), see
Fig. 30.45, we can (1) see the resulting quantization noise over a wide frequency spectrum
and (2) observe that the transfer curve, in the time domain, is similar to Fig. 30.37. Note
that this input violates Bennett's criteria (which, as we'll see, means the noise power
spectral density is flat from DC to the Nyquist frequency).
Slow
V IN Ramp
Ideal Ideal V OUT V OUTD
f clk ADC DAC
Figure 30.45 Taking the difference an ADC input and the DAC output.
A section of the input and output, using the test setup of Fig. 30.45, is shown in
Fig. 30.46a. 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 758.79 mV, in this figure, the ADC output code (input to the DAC)
changes. The ADC output code can be calculated as 755.9 mV/1 LSB ( 1 LSB = 1.5/256
= 5.86 mV for the present simulation) or 129 when the input is slightly below 758.79 mV
and 130 when the input is slightly above 758.79 mV. Looking at the transfer curves in this
figure it appears as though the output changes when the ADC code is 129.5 or 758.79
mV/1 LSB. This, as seen in Fig. 30.46b and discussed below, results in centering the
quantization error around the input (and is the reason we shifted the ADC transfer curves
by 1/2 LSB when we developed our ideal ADC model).
The difference output, between the two signals of Fig. 30.46a, is shown in Fig.
30.46b. Some points to note about this sawtooth waveform are that 1) its average value is
zero, 2) the waveform contains an abrupt transition (and so we expect a wideband output
spectrum similar to that which occurs after sampling a waveform), and 3) its peak-to-peak
amplitude is 1 LSB. Like a sinewave, which also has zero average value, we can
characterize this quantization error waveform by looking at its root-mean-square (RMS)
Chapter 30 Data Converter Modeling 41
ADC input
1 LSB = 5.86 mV
DAC output
764.65 mV
758.79 mV
1/2 LSB
−1/2 LSB
Note the frequency of this waveform is 10 MHz
Figure 30.46 (b) Difference between ADC input and DAC output.
42 CMOS Mixed-Signal Circuit Design
1/2LSB
Probability density function, ρ
∫ ρ ⋅ dQe = 1
−1/2LSB
1 1 LSB= V LSB
V LSB
Qe
−1/2 LSB 1/2 LSB
Figure 30.47 Probability density function for the quantization error in an ADC
assuming Bennett's criteria hold.
The quantization error noise power is the variance of the probability density
function. The RMS quantization error voltage is the square root of the quantization noise
power. The variance of the probability density function (the quantization noise power, PQe)
is given, knowing the average of the quantization error, Qe , is zero, by
1/2LSB
V 2LSB
P Qe = ∫ ρ ⋅ (Qe) 2 ⋅ dQe =
12
(30.31)
−1/2LSB
noise it is not because of the sampling process used in the ADC (and so quantization noise
doesn't experience aliasing). Quantization noise is added to the signal after the sampling
process during the analog-to-digital conversion process. To qualitatively understand why
the quantization error spectrum is white, in Fig. 30.42, we remember that there are abrupt
transitions in the DAC output, and if the quantization error is truly random, the times
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
quantization errors are further randomized and the resulting error spectrum will be flat.
Calculating RMS Quantization Noise Voltage from a Spectrum
The voltage spectrums for the quantization noise and the input signal (Fig. 30.46a and b)
are shown in Fig. 30.48. Note that the harmonics of the noise are, as we would expect,
spaced by 10 MHz (= 1/T ). Also note, the sampling frequency doesn't affect the value of
the RMS quantization noise voltage. The peak voltage of the fundamental tone in the
quantization noise voltage spectrum is approximately −55 dB or −58 dB RMS (peak
voltages [magnitudes] are used in the spectrum plots shown in this chapter unless
otherwise indicated). To relate the RMS noise voltage calculated above, i.e., −55.4 dB, to
the values shown in Fig. 30.48 we would: (1) convert the peak voltages to RMS values by
subtracting 3 dB from each value, (2) square each result to get the mean-squared voltage,
(3) sum the mean squared values, and (4) take the square root of this sum to get the RMS
quantization noise voltage. Looking at the peak values of the first three tones in the
Voltage, peak
Figure 30.48 Input and quantization noise spectrums for the signals of Fig. 30.46.
44 CMOS Mixed-Signal Circuit Design
quantization noise spectrum, −55 dB, −60 dB, and −65 dB we convert these values to
RMS voltages, 1.26 mV, 0.708 mV, and 0.398 mV. The quantization noise, calculated
using only the first three tones, is then (1.26) 2 + (0.708) 2 + (0.398) 2 = 1.5 mV, RMS or
−56.5 dB. Clearly, increasing the number of tones used in this calculation will cause the
result to approach Eq. (30.30) (1.69 mV).
To calculate the RMS quantization noise voltage from a DAC output spectrum we
sum the mean-squared contribution from each component (after removing the desired
tones from the spectrum) and then take the square root of the result (as mentioned above.)
See VDFT( f ) in Fig. 30.42 as an example. If the resolution of the DFT is fRES , then we can
write this as
M−1
V Qe,RMS = 1 ⋅
2
Σ V 2DFT ( k ⋅ f RES )
k=0
where M = #DFTpoints (30.33)
The factor of root two comes from changing the peak values in a spectrum to RMS
quantities. Note that the term "bin" is often used to describe the fact that the output of the
DFT is only valid at discrete frequencies (the bins.) The number of bins is also known as
the number of points in an DFT output vector (labeled #DFTpoints or M in Eq. [30.33]).
This is seen in Fig. 30.48 (also shown in Fig. 30.48 is a DFT problem known as the
"picket fence" effect, which will be discussed in a moment). If the DFT resolution in a
simulation is 1 MHz then the DFT output, assuming the starting frequency is DC (0), will
have nonzero values at DC, 1 MHz, 2 MHz, and so forth. If the stop frequency is 200
MHz, then the total number of points in a DFT output vector is 201.
Note that if Bennett's criteria hold, Eq. (30.33) will equal V LSB / 12 . If it doesn't
hold, then the V Qe,RMS calculated using Eq. (30.33) will be different from V LSB / 12 . An
input high-frequency sinewave violates Bennett's criteria. For example, consider sampling
a 25 MHz sinewave at 100 MHz. If the sample points occur at the zero crossing points on
the sinewave and at the peak/valley points, the resulting DAC output will be a rectangular
waveform with a well-defined spectrum.
After a simulation the length of the DFT output vectors can be determined using
display
These commands show, in the WinSPICE command window, the length of the vectors and
the type, e.g., complex, real, dB, etc. (for the display command) or the length of a
particular vector (for the "print length" command).
If we want to set a component of the DFT to a value, say zero, we may want to
use a command sequence like
let m=mag(voutd)
let m[7]=0
Chapter 30 Data Converter Modeling 45
This sequence of commands sets the eighth element of a vector to zero (since we start at
element zero). This is often done to remove a tone in an output spectrum resulting from
the input signal or some other distortion.
Example 30.10
Using WinSPICE calculate the RMS quantization noise voltage from the spectrum
of Fig. 30.48.
We begin by running the simulation that generates this figure (running the netlist
file Fig30_48.cir). After the simulation is completed we type, in the WinSPICE
command window,
display
We see from this that the length of the DFT is 401. Note that we could have used
the length command, as we'll do below, to determine the length of the DFT instead
of the display command.
To calculate the RMS quantization noise voltage we can use the following:
let m=mag(voutd)
let qnoise=0.707*sqrt(mean(m*m)*length(m))
print qnoise
which gives a result of 2.08 mV, a value larger than the 1.69 mV calculated for
V Qe,RMS earlier. Before we discuss the discrepancy between the two RMS voltages,
notice that we took the average (mean) of the mean-squared value of voutd and
then multiplied the result by its length to sum the mean-squared voltages as
specified by Eq. (30.33). T
Now we need to discuss the difference between the SPICE simulated and the
calculated RMS quantization noise voltages above. While the implementation of a discrete
Fourier transform is outside the subject matter of this book, we can comment here on two
DFT problems and how to reduce their effects; namely, the picket-fence effect and
spectral leakage.
The picket-fence effect, and the visual reason for its name, is shown in the insert
figure in Fig. 30.48. Coherent sampling (synchronizing the quantization error, Fig. 30.46,
with the sampling clock) was used to magnify the effect. In our discussion above we
assumed, for the first tone at 10 MHz, that the contribution to V Qe,RMS was − 55 dB. On
closer inspection, we see that there are also contributions, − 61 dB, to the quantization
noise at 9.5 MHz and 10.5 MHz. At these two side frequencies, the amplitude
46 CMOS Mixed-Signal Circuit Design
1
1
0
DFT points 0 1 2 0 1 2
Example 30.11
Repeat Example 30.10 if Eq. 30.34 is used to set the DFT resolution and
simulation length.
In Example 30.10 the simulation time was 2,000 ns. We could increase the
simulation time to 4,000 ns or reduce the DFT resolution from 500 kHz to 1 MHz.
In order to keep the simulation time reasonably short (and to avoid spectral
leakage discussed next) we will decrease the DFT resolution and leave the
simulation time at 2,000 ns. Figure 30.50 shows the resulting output spectrum with
the decreased DFT resolution (now 1 MHz). The RMS quantization noise voltage
calculated by SPICE, from this spectrum, is 1.71 mV RMS. T
To understand what is meant by "spectral leakage," consider the sinewave with
infinite duration shown in Fig. 30.51a. When a DFT is 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 is finite and set by the
simulation time or transient stop time, Tstop . We can think of taking the infinite duration
sinewave of Fig. 30.51a and multiplying it by the rectangular waveform of Fig. 30.51b to
obtain the waveform used in the simulation, Fig. 30.51c. This multiplication 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
Chapter 30 Data Converter Modeling 47
Figure 30.50 Eliminating the picket-fence effect from the simulation in Fig. 30.48.
1 1
t t
T stop
(a) (b)
1 1
T stop t f in f
f f in f
f in Linear amplitude
1/T stop
(e) Frequency spectrum of (c) (f), Von Hann (Hanning) window
Figure 30.51 Showing how spectral leakage, resulting from a DFT, affects a waveform.
48 CMOS Mixed-Signal Circuit Design
frequency domain. The result is that instead of the sinewave spectral response being an
impulse function, as seen in Fig. 30.51d, it is a weighted Sinc waveform, Fig. 30.51e. Note
how the DFT spectral response of the sinewave, Fig. 30.51e, 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 minimize these
sidelobes, other windowing functions are used. The one we are using in this chapter is the
von Hann (a.k.a. Hanning or Cosine) window shown, without the sidelobes, in Fig. 30.51f.
The response is shown on both linear and log amplitude scales and the width of the
window is 2/Tstop at its base (= 1 MHz if Tstop = 2,000 ns).
Example 30.12
Using SPICE, show the spectrum of a 1 V (peak) sinewave at 10 MHz over a
spectral range of DC to 200 MHz with an DFT resolution of 1 MHz and a
simulation time of 2,000 ns (windowed frequency spread of 1 MHz, Fig. 30.51e).
The results are shown in Fig. 30.52. Note how the only point in these figures that
has a nonzero value occurs at 10 MHz. The plotting lines are used to connect the
five DFT output points shown in each of these figures. T
Figure 30.52 Output spectrum of 10 MHz sinewave showing the window's effect.
Example 30.13
Repeat Ex. 30.12 if the sinewave frequency is changed to 10.4 MHz.
The DFT output is shown in Fig. 30.53. Note that although the input frequency is
at 10.4 MHz the peak in the DFT response still occurs at 10 MHz (a DFT output
point). Also notice how the spectrum of the sinewave is effectively wider than the
sinewave of Ex. 30.12. The 10.4 MHz sinewave is within the DFT resolution of
both the 10 MHz and 11 MHz points. The result is effective spectral content at
these frequencies. Sinewaves that do not fall exactly at the DFT calculation points
are smeared in the DFT output spectrum. This smearing can spread across the
Chapter 30 Data Converter Modeling 49
spectrum and affect spectral content at other frequencies. Consider the following
example. T
Figure 30.53 Magnitude and spectral response for the 10.4 MHz sinewave of Ex. 30.13.
Example 30.14
Using SPICE, plot the output spectrum resulting from adding the 10 MHz and
10.4 MHz sinewaves from the previous two examples.
The results are shown in Fig. 30.54. Note how, even though the 10 MHz sinewave
has an amplitude of 1 V, the resulting output spectrum reports an amplitude of
approximately 600 mV at 10 MHz. This is a result of contributions from the 10.4
MHz signal, after windowing, subtracting from the 10 MHz signal calculation
point. T
Figure 30.54 Magnitude and spectral response for the sum of the 10 and 10.4 MHz sinewaves.
It's important to understand that by increasing the simulation time, the window
length increases, causing the width of the Sinc spectrum, see Fig. 30.51e, to decrease.
However, increasing the simulation time without making a corresponding change in the
DFT resolution can actually be harmful to the results. For example, if 1/Tstop is 100 kHz
(simulation time of 10,000 ns) and the DFT resolution is 1 MHz then an input sinewave at
10.5 MHz will have no effect on the resulting output spectrum. In general, it's important
to make
Also note that in a general simulation, which includes MOSFETs, we can set the step size
used in a transient simulation with Eq. (30.8). However when using ideal components, as
in this chapter, the step size can be increased to speed up the simulation time.
Example 30.15
Determine the RMS quantization noise voltage from the DAC output spectrum
shown in Fig. 30.42.
Figure 30.42 was generated with a DFT resolution of 1 MHz and a simulation time
of 1,000 ns. We will increase the simulation time to 2,000 ns. The resimulated
spectrum of the DAC output noise is shown in Fig. 30.55. Notice in this spectrum
that there is a signal at DC and 7 MHz (from the input signal.) Also, the aliased
signals are present in the output spectrum at 93 MHz, 107 MHz, and 193 MHz.
To calculate the quantization noise we would have to first zero these components
out. We can use the following WinSPICE commands to calculate the quantization
noise:
let m=mag(vout)
let m[0]=0
let m[7]=0
let m[93]=0
let m[107]=0
let m[193]=0
let qnoise=0.707*sqrt(mean(m*m)*length(m))
print qnoise
Figure 30.55 Simulating the circuit shown in Fig. 30.40 for 2,000 ns.
Chapter 30 Data Converter Modeling 51
Note that the simulations we have shown in this chapter generate spectral
responses out to twice the clocking frequency or 200 MHz when using a 100 MHz clock.
To reduce simulation time we may limit the spectral response to the Nyquist frequency.
Also, an important component of the simulations can be the addition of
.options RELTOL=1u
to a netlist. Not including this statement or one similar (e.g., RELTOL = 10u) in a netlist
may limit the simulated noise floor to a significant voltage.
WinSPICE can also be useful if measured data is available. The data from a
spectrum analyzer can be written to a text file and loaded into a WinSPICE vector using
the load command. See the WinSPICE online manual.
The DFT's Relationship to the Continuous Time Fourier Transform
Before we leave this section, let's comment on how the discrete Fourier transform is
related to the continuous time Fourier transform. We can write the continuous time
Fourier transform of a time-varying function, v(t), using
∞
V( f ) = ∫−∞ v(t) ⋅ e −j2πf⋅t ⋅ dt (30.36)
To approximate this continuous time Fourier transform with discrete variables, we will use
the following
dt = ∆t and t = k ⋅ ∆t where k = 0, 1, ... N (30.38)
The variable ∆t is the transient step time (the time difference between points in the DFT
algorithm) and N is the number of time steps used in the algorithm (T stop = N ⋅ ∆t) . The
frequencies where the DFT is calculated, assuming Eq. (30.35) is valid, are given by
N
V( f ) f=n⋅f res ≈ Σ v(k ⋅ ∆t) ⋅ e −j(4π/N)nk
k=0
⋅ ∆t (30.40)
(noting that 4π is used in the exponent because our DFT resolution was twice the
reciprocal of the simulation time), or
52 CMOS Mixed-Signal Circuit Design
Analog Analog
Digital Digital
V IN Ideal V IN
f clk 8-bit ADC
V Qe ( f )
After looking at Eq. (30.43) we might think that by simply increasing the sampling
frequency we can reduce the amount of quantization noise an ADC introduces into an
analog input signal. While increasing the sampling frequency spreads the quantization
noise spectral density out over a wider range of frequencies (see Fig. 30.57) with a
corresponding reduction in amplitude, the sampling frequency doesn't affect the total RMS
quantization noise voltage. However, bandlimiting the spectrum using a filter reduces the
amount of quantization noise introduced into an input signal. In the simplest case a
lowpass filter, which we will think of as an averager, can be used on the digital outputs of
Chapter 30 Data Converter Modeling 53
V Qe ( f )
VLSB
12f s
f n = f s /2 f
the ADC to reduce the amount of quantization noise introduced into the signal. We can
write the amount of noise introduced into an input signal over a range of frequencies using
fH
Again the factor of 2 is used to account for the contributions to V Qe,RMS in the negative
frequency spectrum (the DFT routine, discussed in the previous section, uses a one-sided
spectrum so the factor of 2 is not necessary when making calculations using the SPICE
simulation output data). Because the output of the ADC is a digital word, we would
require a digital filter to bandlimit the output spectrum of the ADC. We will discuss digital
filtering in the next chapter. For now let's show that the sampling frequency indeed doesn't
affect the quantization noise, assuming Bennett's criteria are valid, and then let's discuss
the concept of averaging to reduce quantization noise.
Example 30.16
Repeat Ex. 30.11 if the sampling frequency is increased from 100 MHz to 200
MHz.
Doubling the sampling frequency has no effect on the output quantization noise. It
remains at 1.69 mV RMS. T
Example 30.17
Repeat Ex. 30.15 if the sampling frequency is increased from 100 MHz to 200
MHz.
Again, the RMS quantization noise voltage remains essentially unchanged, i.e.,
1.68 mV RMS. Recall that the circuit shown in Fig. 30.40 is used in this example
and Ex. 30.15 with a 7 MHz input. It's instructive to show the time domain output
of Fig. 30.40 when clocked at 200 MHz, Fig. 30.58, and compare it to Fig. 30.41
(the output of the circuit of Fig. 30.40 clocked at 100 MHz). Note how the DAC
output voltage step size has decreased in Fig. 30.58 when compared to Fig. 30.41,
yet the quantization noise remains unchanged. This shows, once again, that we
must look at the spectrum of a signal to determine the quantization noise voltage
and that the "coarseness" of an output signal has nothing to do with quantization
noise. T
54 CMOS Mixed-Signal Circuit Design
Figure 30.58 Output of the circuit shown in Fig. 30.40 with 7 MHz input and 200 MHz
sampling clock. This figure should be compared to Fig. 30.41.
Analog
8 V OUTA Analog
V IN Ideal Ideal
f clk = 100 MHz 8-bit ADC 8-bit DAC
Digital V OUT
8
Ideal Ideal
8-bit ADC 8-bit DAC V OUTB
Note that this configuration effectively samples the input at 200 MHz (200 Msamples/s
[2 ⋅ f s ] ), as was accomplished in Fig. 30.58 except that now we are averaging consecutive
samples. If we input a 7 MHz sinewave into this configuration, the same signal used in
Fig. 30.41 or Fig. 30.58, we get the output shown in Fig. 30.60. Note the resemblance to
Fig. 30.58. Also note the additional phase shift. The RMS quantization noise voltage now,
however, has dropped from 1.68 mV to approximately 1.18 mV.
Figure 30.60 Output of the circuit of Fig. 30.59 with a 7 MHz input sinewave.
or
V
V Qe,RMS = 1 ⋅ LSB (30.47)
2 12
56 CMOS Mixed-Signal Circuit Design
f n = f s /2 fs f
missing code (an input difference between two inputs at consecutive sampling times of
1-LSB results in the same output), then the averaging does nothing. If the data converter
is nonmonotonic (an increase in the data converter's input doesn't result in an 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 changing for
the ADC) or a value that isn't "busy" (not changing by at least 1 LSB in between sampling
instances) will not benefit from averaging. These topics are discussed further in the next
chapter.
Practical Implementation of Averaging in ADCs
The averaging topology shown in Fig. 30.59 is not practical in most situations. The silicon
area required to implement the extra ADC and DAC generally costs more than is gained
by the reduction in quantization noise. Also, as we'll see later, there are other techniques
for averaging that provide a more efficient way to reduce quantization noise. Having said
this and knowing that there are better ways, we will answer the question "How do we
implement an ADC using averaging?"
Figure 30.63 shows how we can add a digital averaging filter to the output of the
ADC to reduce quantization noise. The ADC and digital averaging filter are clocked at a
rate of fclk . If K = 2, for example, then the filter will sum its previous two inputs and
output the result at a rate of fclk . This filter could be implemented with a register and an
adder. Note that the output word size increases when using the digital filter (it had better
if we are reducing the quantization noise!). For example, if the output of the ADC is an
8-bit word, then the running sum coming out of the filter, again assuming K = 2, will be
9-bits.
Analog Digital
Digital
V IN Ideal Digital
averaging
f clk ADC filter
We might, at this point, assume that we can use a low-resolution ADC, say 6-bits,
with a significant amount of averaging 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 on the bandwidth of the signal coming into the ADC because,
unlike Fig. 30.59, we haven't increased the sampling rate of the signals. Therefore, the
amplitude of the spectral density remains unchanged. For an averaging of two, we would
have to limit our desired input signal bandwidth to fs /4. If this wasn't the case, then an
input sinewave at fs /2 would average to zero. Again, these topics will be discussed in
greater detail in Ch. 31.
58 CMOS Mixed-Signal Circuit Design
REFERENCES
[1] R. J. Baker, H. W. Li, and D. E. Boyce, CMOS: Circuit Design, Layout, and
Simulation, Wiley-IEEE, 1998. ISBN 0-7803-3416-7
[2] L. W. Couch, Modern Communication Systems: Principles and Applications,
Prentice-Hall, 1995. ISBN 0-02325286-3
[3] S. Haykin, An Introduction to Analog and Digital Communications, John Wiley
and Sons, 1989. ISBN 0-471-85978-8
[4] P. A. Lynn and W. Fuerst, Introductory Digital Signal Processing, Second
Edition, John Wiley and Sons, 1998. ISBN 0-471-97631-8
[5] E. P. Cunningham, Digital Filtering: An Introduction, John Wiley and Sons, 1995.
ISBN 0-471-12475-3
[6] R. K. Hester, Introduction to Oversampled Data Conversion, Notes from a
tutorial at the 1995 International Solid-State Circuits Conference (ISSCC-95).
[7] W. R. Bennett, "Spectra of Quantized Signals," Bell System Technical Journal,
Vol. 27, pp. 446-472, July 1948.
[8] J. C. Candy and G. C. Temes (eds.), Oversampling Delta-Sigma Data Converters,
IEEE Press, 1992. ISBN 0-87942-285-8
[9] S. R. Norsworthy, R. Schreier, and G. C. Temes (eds.), Delta-Sigma Data
Converters: Theory, Design, and Simulation, IEEE Press, 1996. ISBN
0-7803-1045-4
LIST OF SYMBOLS/ACRONYMS
AAF - Antialiasing Filter
ADC - Analog-to-Digital Converter
C H - Hold capacitor in a S/H
DAC - Digital-to-Analog Converter
DFT - Discrete Fourier Transform
DSP - Digital Signal Processing
∆t - Time difference between points used in a DFT
φ - Clock signal
fclk - Frequency of the input clock signal
f in - Input sinewave frequency
f n - Nyquist frequency ( f n = f s /2) which is 50 MHz in this chapter. Sometimes also called
the folding frequency
Chapter 30 Data Converter Modeling 59
VOUTD - Difference between an analog input and a digitized output, see Fig. 30.45.
VOUTdB - Output signal in decibels
VINdB - Input signal in decibels
V OUTSH - Output voltage of a S/H
V Qe ( f ) - Quantization Error Spectral Density, V/ Hz
v in (t) - Time domain input voltage
V inbuf - Input signal after buffering
V ins - Input signal after sampling
V LSB - See LSB
Vout - Output voltage
V p - Peak sinewave amplitude
V Qe,RMS - RMS quantization noise voltage
V REF+ - Positive reference used in an ADC or DAC, which is 1.5 V in this chapter
V REF− - Negative reference used in an ADC or DAC which is ground in this chapter
VSS - negative power supply voltage which is 0 V in this chapter
f
j2π f
z = e j2π⋅f⋅Ts = e s
QUESTIONS
30.1 Qualitatively, using figures, show how impulse sampling a sinewave can result in
an alias of the sampled sinewave at a different frequency.
30.2 What does linear phase indicate?
30.3 What does multiplying a signal by e j⋅2πf⋅(−td ) indicate? How does the magnitude of
the resulting signal change? How does the phase change?
30.4 Show, in the time domain, the input/output of the transmission line, and output of
the comb filter in Fig. 30.11 if the input signal is a sinewave with a peak amplitude
of 1 V and a frequency of 100 MHz. Show the two 500 Ω resistors average the
input signal and the output signal of the delay line (transmission line).
30.5 Regenerate Fig. 30.19 if the switches are closed for 5 ns instead of 100 ps.
30.6 What sets the minimum resolution of a DFT in a SPICE spectral analysis?
30.7 Explain why the sinewave in Fig. 30.19 is "double sampled."
30.8 Explain why z is used in signal processing. What does multiplying a signal by z−1 do
to the signal?
Chapter 30 Data Converter Modeling 61
30.9 Sketch the implementation of a circuit that will multiply a digital signal by z−1.
30.10 Sketch the time domain representation of the five signals shown in Fig. 30.29 on
different plots. Regenerate Fig. 30.29 if the input signal is a 1 V peak sinewave at
5 MHz and zero offset. Explain the resulting plot.
30.11 Sketch the input and output spectrum for the following block diagram. Assume the
DC component of the input is 0.75 V while the AC component is a sinewave at 4
MHz with a peak amplitude of 1 V. Assume the clock frequency is 100 MHz.
clock
30.12 Using the models developed in the chapter design a SPICE model for the S/H of
Fig. 30.31. Use the model to regenerate Fig. 30.29.
30.13 If VREF+ = 1.5 V and VREF− = 0 regenerate Fig. 30.33 using SPICE. (Design a 3-bit
ideal DAC model in SPICE.) The y-axis will be voltages in decimal form.
30.14 If, again, VREF+ = 1.5 V and VREF− = 0, sketch Fig. 30.33 for a 1-bit DAC. Note that
the digital input code will either be a 0 or a 1 and the analog voltage out of the
DAC will be either 0 or 1.5 V. Using Eq. (30.23) what is the voltage value of 1
LSB? How does this compare to the value of 1 LSB we get from the sketch? Is
Eq. (30.23) valid for a 1-bit DAC? Why? The 1-bit DAC will be a ubiquitous
component in our noise-shaping modulators in Ch. 32 (see Fig. 32.28).
30.15 Using SPICE, implement an ideal 4-bit DAC and regenerate Fig. 30.36.
30.16 Why do the transfer curves of Fig. 30.37 show a shift of 1/2 LSB to the left? How
do we implement this shift in SPICE?
30.17 Repeat question 30.16 for an ADC.
30.18 Using the models developed in questions 30.15 and 30.17 with a clock frequency
of 100 MHz apply an input sinewave that has an amplitude of 750 mV peak
centered around 750 mV DC and a frequency of 5 MHz to the input of the 4-bit
ADC. If the ideal 4-bit DAC is connected to the digital outputs of the ADC, also
show the DAC's analog output.
30.19 Using SPICE generate the spectrums of the input and output of the signals in
question 30.18.
62 CMOS Mixed-Signal Circuit Design
30.20 Does an ideal S/H introduce amplitude quantization noise into an input waveform?
Why or why not?
30.21 Why are the amplitudes of the mirror images decreasing with an increase in
frequency in Fig. 30.44b?
30.22 Show the details of the integration in Eq. (30.30).
30.23 How are voltage spectral density, power spectral density (PSD), average power,
and RMS voltage related for a random signal? What are the units of each? Provide
answers for both continuous signals and signals that are only defined at discrete
frequencies.
30.24 How would we convert the voltage spectral density of Fig. 30.48 into a power
spectral density plot? What term in Eq. (30.33) is the PSD? How would we
rewrite Eq. (30.33) to give the average power of the quantization error?
30.25 Repeat Ex. 30.10 if we want to determine the quantization noise power. Show the
simulation results and the commands used to determine this power.
30.26 Derive Eq. (30.43).
30.27 What term is the PSD in Eqs. (30.42) and (30.44)? What are its units?
30.28 Verify, with simulations, Ex. 30.16.
30.29 Verify, using simple circuit analysis, that resistors can be used to implement
averaging as seen in Fig. 30.59 and Eq. (30.45).
30.30 How does averaging K samples of a random voltage variable reduce its RMS
value? How does the power contained in the same variable get reduced by
averaging?
Chapter
31
Data Converter SNR
In the last chapter we developed the idea of treating an analog-to-digital converter (ADC)
as a noisy circuit block where the output of the ADC is the sum of quantization noise and
the input signal. Logically, the next step in our development of concepts is to characterize
a system using ADCs and DACs in terms of the signal-to-noise ratio (SNR). The ideal
SNR for a data converter was developed back in Ch. 28 and is repeated here for
convenience.
If we apply a sinewave with an amplitude of Vp (and thus an RMS value of
V p / 2 ) to an ADC input then, knowing the RMS quantization noise added to a busy
ADC input signal is V LSB / 12 (see Eqs. [30.30] and [30.32]), the resulting SNR for the
ADC is given by
Vp / 2
SNR ideal = 20 ⋅ log (31.1)
V LSB / 12
If we remember that
V REF+ − V REF−
V LSB = 1 LSB = (31.2)
2N
and we assume that the largest possible amplitude sinewave is the ADC input (to
maximize the SNR), that is,
2V p = V REF+ − V REF− (31.3)
then Eq. (31.1) can be rewritten as
2 N 12
SNR ideal = 20 ⋅ log = 6.02N + 1.76 (in dB) (31.4)
2 2
64 CMOS Mixed-Signal Circuit Design
Our goal in this chapter is to discuss how to determine the actual SNR of a data
conversion system and to present topologies for improving data conversion system SNR
(e.g., averaging, noise shaping, and others).
Example 31.1
Determine the effective number of bits for an ADC with V REF+ = 1.5 , V REF− = 0 ,
and a measured V Qe,RMS of 2 mV.
If we assume that the input peak amplitude, Vp , is 0.5 ⋅ (V REF+ − V REF− ) or 0.75 V,
then the measured SNR is given by
0.75/ 2
SNR meas = = 265 = 48.5 dB
2 mV
The effective number of bits, Neff , is (from Eq. (31.5]) 7.76 bits. T
Normally, the measured SNR (SNR meas ) is determined from the RMS quantization
noise, which is determined using Eq. (30.33) with measured data. The amplitude and
frequency of the input sinewave can be selected to maximize the SNR.
Example 31.2
Using the ideal 8-bit ADC and DAC shown in Fig. 31.1, which were developed in
the last chapter, and a sampling frequency of 100 MHz (= f s ) show, using SPICE,
that applying a full-scale sinewave at 24 MHz to this configuration will cause the
resulting SNR to approach the ideal value given by Eq. (31.4).
Let's begin by calculating SNR ideal . From Eq. (31.4), SNRideal is roughly 50 dB, as
the data converters have 8-bit resolution.
The time domain input and output of the circuit shown in Fig. 31.1, and the
corresponding DAC output spectrum, are shown in Fig. 31.2. The input to the
Chapter 31 Data Converter SNR 65
Figure 31.1 Test setup used in Ex. 31.2 to show deviation from ideal SNR.
ADC in Fig. 31.1 is a 24 MHz sinewave with a peak amplitude of 0.75 V centered
on a DC voltage of 0.75 V (the peak-to-peak voltage of the input waveform is 1.5
V). The VQe,RMS measured with SPICE, remembering to zero out the wanted signals
at DC and 24 MHz and the images at 76 MHz, 124 MHz, and 176 MHz before
calculating the noise (see Ex. 30.15), is 1.497 mV. The simulated SNR is
(0.75/ 2 )/1.497 mV or 354 (51 dB), which is very close to the value calculated at
the beginning of the example. T
It's important to understand that poor selection of the input frequency (or
windowing function) can result in an SNR that is different from the ideal value calculated
using Eq. (31.4). Selecting an input sinewave frequency, fin , such that fs /fin is a whole
number creates a condition where an integral number of input sinewave cycles fit perfectly
into the sampling window (coherent sampling). This results in an output spectrum that
contains isolated tones (helping to reduce the effects of spectral leakage on the SNR).
ADC DAC
Input output
Signal plus Noise
(a) (b)
Figure 31.2 Example 31.2 (a) time domain input and output, and (b) spectrum of DAC output.
66 CMOS Mixed-Signal Circuit Design
An additional effect to consider occurs when fin is comparable to fs. The input tone
at fin undergoes amplitude attenuation (−3.9 dB at fn , Fig. 30.29).
Example 31.3
Repeat Ex. 31.2 if the input sinewave frequency is increased to 45 MHz.
The results of increasing the ADC input sinewave frequency in Fig. 31.1 to 45
MHz are shown in Fig. 31.3. Note how the DAC output contains tones at 5 MHz,
10 MHz, 15 MHz, etc., in addition to the desired tone at 45 MHz. The simulated
VQe,RMS is 2.26 mV. The input amplitude of the 45 MHz sinewave is 0.75 V (−2.5
dB). The simulated peak output amplitude at 45 MHz is 0.53 V (−5.5 dB). The
simulated SNR can be calculated
0.53/ 2
SNR= 20 log = 44.4 dB
2.26 mV
or 5.6 dB below the ideal value of 50 dB calculated using Eq. (31.4) for a data
converter with a resolution of 8 bits. T
Signal plus Noise
(a) (b)
Figure 31.3 (a) DAC output with 45 MHz input and (b) the DAC output spectrum.
we calculate the RMS quantization noise voltage using Eq. (30.33) and nonideal
components, we are actually calculating the noise plus the distortion in the spectrum. Up
to this point we have only used ideal components, so that distortion in the output
spectrums was absent. We can rewrite Eq. (30.33) to indicate that when it is used with a
measured spectrum, both noise and distortion are included in the result as
M−1
V Qe+D,RMS = 1
2
Σ V 2DFT (k ⋅ f RES )
k=0
where M = #DFTpoints (31.6)
Example 31.4
Suppose that the test setup shown in Fig. 31.1 is used with an input sinewave
having a frequency of 7 MHz, a peak amplitude of 0.75 V, and centered around
0.75 V (so that, once again, the sinewave swings from 0 V to 1.5 V.) Using
SPICE simulation, determine the SNDR if there is a gain error in the ideal ADC in
Fig. 31.1 (it's no longer ideal) so that each stage in the pipeline algorithm used to
implement the ideal SPICE ADC has a gain of 2.1 instead of the ideal 2.0.
The resulting DAC output spectrum is shown in Fig. 31.4. The RMS noise plus
distortion voltage, V Qe+D,RMS , is calculated to be 16.9 mV, using SPICE and
remembering to zero out the desired terms at DC and 7 MHz as well as the
undesired images at 93 MHz, 107 MHz, and 193 MHz. The SNDR is then
0.75/ 2
SNDR = 20 log = 30 dB
16.9 mV
and the effective number of bits is 4.7. In other words, a 5% gain error in the ADC
amplifiers results in an effective resolution of almost half the ideal, 8-bit value. T
Measuring SNDR requires a spectrum analyzer, when looking at the output of a
DAC, or loading digital data (most often in decimal form) into a program that can perform
a DFT (such as WinSPICE [utilizing the load command] or Matlab), when looking at the
output of an ADC. Trying to measure SNDR using a time domain instrument, such as an
oscilloscope, is usually a waste of time because the dynamic range of the instrument is
comparable to the dynamic range of the data converter under test. Spectrum analyzers
utilize narrow band filtering on their input to reduce the inherent noise measured in a
circuit and can have dynamic ranges in excess of 120 dB over a very wide frequency
spectrum. Also note that the SNDR is sometimes abbreviated as SINAD (signal-to-noise
and distortion.)
68 CMOS Mixed-Signal Circuit Design
Figure 31.4 Output spectrum with ADC gain error (see Ex. 31.4).
Another way to specify DR is as the ratio of the RMS full-scale input sinusoid
amplitude, Vp/ 2 , to the input sinusoid amplitude (RMS) that results in an SNDR of 0
dB. (The RMS amplitude of the input signal is equal to the RMS quantization noise plus
distortion, V Qe+N,RMS , when the SNDR is 0 dB.) This is nothing more than saying that the
SNDR can be used to specify DR.
Example 31.5
Determine the DR for the ideal ADC in Ex. 31.2 using Eq. (31.10). Compare the
result to the SNDR calculated in Ex. 31.4.
Using Eq. (31.10), the DR is 48.16 dB (the ideal value). The SNDR calculated in
Ex. 31.4 was 30 dB. Clearly, the SNDR is a better indication of DR than is the
value obtained using Eq. (31.10). T
Specifying SNR and SNDR
The SNR and the SNDR are usually specified as a function of input sinewave amplitude at
a fixed frequency, Fig. 31.5. The x-axis in Fig. 31.5 is normalized so that an input
sinewave with a peak-to-peak amplitude of V REF+ − V REF− 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. [31.6] and [30.33]). We continue to calculate
SNDR using a data converter output spectrum, remembering to zero out the desired tones
and images, and Eq. (31.6) as was done in Ex. 31.4. When we calculate the SNR, we
follow the same procedure except that now we also zero out any spikes or spurs (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. 31.5 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).
Signal-to-noise plus distortion, dB
SNR
SNDR
60
40 V REF+ − V REF−
20
0
-80 -60 -40 -20 0
Normalized input signal amplitude (dB)
∆T s = peak-to-peak jitter
Ts
V p sin 2πf n t = V p sin πf s t
Error in sampling, ∆V s
t
Ideal
sampling point
Aperture
uncertainty/jitter
∆V s
= πf s V p , or ∆V s = ∆T s ⋅ πf s V p (31.12)
∆T s
If we require the uncertainty in the sampled voltage, ∆V s , to be at most 0.5 LSB = (VREF+
−V REF− )/2 N+1 and we remember V p = (V REF+ − V REF− )/2 , then our maximum allowable
peak-to-peak clock jitter can be determined for a particular data converter using
∆T s ≤ 1N ⋅ 1 (31.13)
2 πf s
or in terms of the sampling clock stability
∆T s
Stability, ppm = ∆T s ⋅ f s = = 1N (31.14)
Ts π⋅2
Table 31.1 relates the stability requirements placed on a sampling clock for a data
converter resolution, N, if less than 0.5 LSBs aperture error or sampling voltage
uncertainty is required of the data converter.
Example 31.6
Suppose a phase-locked loop (PLL) is used to generate a clock signal for a data
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 jitter
allowed in the output of the PLL. Assume that the maximum sampling error
allowed is 0.5 LSB and that the data converter is sampling a sinewave with a
frequency of 100 MHz.
Because the sinewave being sampled has a frequency below the Nyquist value, Eq.
(31.13) cannot be used directly. Instead, after reviewing the derivation of this
equation, we can rewrite it in terms of any input signal frequency, fin , as
∆T s ≤ 1N ⋅ 1 (31.15)
2 2π ⋅ f in
72 CMOS Mixed-Signal Circuit Design
noting that when fin = fn = fs /2, Eq. (31.15) reduces to Eq. (31.13). Using Eq.
(31.13) with the numbers in this problem results in a peak-to-peak jitter of 1.56 ps!
The reader familiar with PLL design will recognize that this is a very challenging
requirement when designing a PLL (that is, to design a PLL with an output
frequency of 900 MHz and an output jitter of 1.56 ps). T
We're now in a position to answer how, given the peak-to-peak clock jitter ∆T s ,
the SNR of a data converter is degraded from the ideal value (given in Eq. [31.4]) when
the input sampling clock isn't ideal. Rewriting Eq. (31.15) and assuming
∆T s ≥ 1N ⋅ 1 (31.16)
2 2π ⋅ f in
(a resolution loss ≥ 0.5 LSB), we get
∆T s = 1 ⋅ 1 (31.17)
2 N−N Loss 2π ⋅ f in
where fin is, once again, the frequency of the input sinewave, and NLoss is the number of bits
lost due to the excess jitter. Assuming Eq. (31.16) is valid, then when NLoss is zero, the loss
in resolution is 0.5 LSB and Eq. (31.17) reduces to the equality condition in Eqs. (31.15)
or (31.16). The ideal data converter's SNR, assuming the only nonideal factor in the
system is clock jitter, can be written as
effective bits, N eff
Example 31.7
For an ideal 8-bit ADC clocked at 100 MHz, determine the SNR of the data
converter with 100 ps of peak-to-peak jitter in the input sampling clock, ∆T s ,
assuming the ADC's input is a full-scale sinewave at 25 MHz.
We can write the number of bits lost by solving Eq. (31.17) as a function of
peak-to-peak jitter as
In other words, we are getting at least 2K samples for every cycle of the input sinewave. If
we were sampling at twice the Nyquist frequency ( fs ), where K = 1, then we would get
two samples for every cycle of the input signal. Notice that Eq. (31.15) gives the
maximum jitter specification for a given input frequency and data converter resolution, but
it doesn't specify the sampling frequency, fs , or the sampling frequency period, Ts .
For a given maximum jitter, ∆T s , we can reduce the requirements placed on the
stability of the oscillator by increasing the sampling frequency. This can be written as
Stability(new), ppm = [stability(old), ppm] ⋅ K (31.21)
If we were sampling at 1 MHz and the stability required was 10 ppm, then the jitter in the
sampling clock would be at most 10 ps, peak-to-peak. Increasing the sampling rate to 100
MHz, with 10 ps jitter would require an oscillator stability of 1,000 ppm. If we were to
increase the sampling clock frequency to 1 GHz, then the stability of the clock would be at
least 10,000 ppm (the period of the sampling signal is 1 ns and the jitter is 10 ps or 1%
[10,000 ppm] of the sampling period).
Note that the oversampling factor symbol, K, is the same symbol that indicates the
number of samples averaged on the output of a data converter to reduce RMS
quantization noise voltage (see Sec. 30.3.2). The choice of variable was made for a reason
(which will be discussed further in the next section).
Example 31.8
In Table 31.1 we saw that the 16-bit data converter clocked at 44.1 kHz could
have at most 111 ps peak-to-peak jitter to limit the sampling uncertainty to 0.5
LSB. We saw that the stability required of the oscillator under these circumstances
was 5 ppm at 44.1 kHz. What would happen to the stability requirements of the
oscillator generating the sampling clock if we increased the sampling clock
frequency to 128 ⋅ 44.1 kHz= 5.645 MHz ?
We know that the input bandwidth, prior to increasing the sampling frequency, is
limited to 44.1 kHz/2 or 22.05 kHz (= B, the bandwidth of the input signal). We
then assume the maximum input frequency, fin , remains at or below 22.05 kHz
even after we increase the sampling frequency. We can define the oversampling
factor, in this example, as
f s /2 2.822 × 10 6
K= = = 128
f in 22.05 × 10 3
The jitter requirement remains 111 ps whether we use a sampling frequency of
44.1 kHz or 5.645 MHz. However, now that the clock frequency has increased to
5.645 MHz, the stability required of the oscillator has gone from approximately 5
ppm to 640 ppm. T
It's important to note that the oversampling ratio, K, is given by
f s /2 f n
K= = for f in ≤ B (31.22)
B B
74 CMOS Mixed-Signal Circuit Design
If we desire less than 0.5 LSBs aperture error, and we are using oversampling, then we
can use Eqs. (31.13) and (31.22) to write
∆T s ≤ 1N ⋅ K = 1N ⋅ 1 (31.23)
2 πf s 2 2πB
where, once again, B is the bandwidth of the input signal and K is the oversampling ratio.
As shown by this equation and in Eq. (31.21), using oversampling reduces the
requirements placed on the stability of the sampling clock.
A Practical Note
We need to point out that the effects of clock jitter are possible even if the clock is
perfectly stable because of the clock's finite transition times (rise and fall times). If the rise
time of the clock signal in Fig. 31.6 is finite, say 50 ps, then the same derivations and
discussions concerning jitter in the previous 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 input signal frequency increases.
Modeling Clock Jitter with SPICE
It's useful in many situations to determine how clock jitter affects a data converter's
performance. Consider the block diagram shown in Fig. 31.7. This is our basic
configuration, used previously, to show data converter operation. Now, however, we have
changed the sampling clock from an ideal pulse source to a source that contains jitter. The
questions we want to answer in this section are "How do we use SPICE to model a jittery
clock source?" and "How does the jitter affect the SNDR of the data converter?"
Jittering clock
Figure 31.7 Simulating ideal data converters using a sampling clock with jitter.
To begin, let's consider the single frequency frequency modulation (SFFM) source
available in SPICE. This source generates a frequency modulated (FM) sinewave using the
following syntax
SFFM(VO VA FC MDI FS)
The modulation index (MDI) will set the peak-to-peak jitter time in the waveform while
the signal frequency (FS) describes the rate at which this jitter varies. The carrier
frequency (FC) will set the clock frequency (FC = 100 MHz in Fig. 31.7). The term VO is
available to add a DC offset to the signal (which we will assume is zero in our discussion),
and the value VA sets the amplitude of the frequency modulated sinewave. The question
we need to answer now is, "How do we convert the FM sinewave generated using Eq.
(31.24) into a squarewave suitable for driving our ADC?"
Figure 31.8 shows that a switch and the SFFM source can be used to generate the
sampling clock with jitter. When the FM source transitions above zero, the top switch
closes and the clock output goes high. When the source transitions below zero, the bottom
switch closes and the clock output goes low.
VDD = 1.5 V
Figure 31.8 Generating a clock with jitter using a switch and an SFFM source.
The average period of the sampling clock generated by the SFFM source is given
by
T s = 1 or f s = FC (31.25)
FC
The peak phase excursion of the clock signal is set by MDI. The peak phase excursion can
be related to the sampling frequency using
∆T s
2⋅MDI = 2π ⋅ (31.26)
Ts
or the peak-to-peak jitter, ∆Ts, is given by
Example 31.9
Using SPICE, generate the output spectrum of an oscillator assuming the oscillator
frequency is 100 MHz and the peak-to-peak jitter is 100 ps. Assume FS = 1 MEG.
We can begin this example by noting FC = 100 MHz and the modulation index is
π ⋅ 100ps⋅100MEG = 0.0314 . The SPICE netlist and the resulting spectrum are
shown as follows and in Fig. 31.9, respectively.
76 CMOS Mixed-Signal Circuit Design
.model switmod SW
.end
Note that the resolution of the DFT was 200 kHz (set by Eq. [30.35] and that the
signal frequency, FS, was 1 MEG (which sets the spacing between the tones in Fig.
31.9). The period of the 100 MHz clock, in this example, varies sinusoidally from
9.95 ns to 10.05 ns over a time frame of 1 µs (1/FS.) While this spectrum is
interesting, it isn't representative of an actual oscillator where the noise is random.
Peak Signal plus Noise Voltage, dB
While our simple model will never be capable of generating truly random noise, we
can, for a given simulation time, make the simulated spectrum of the oscillator
approach something that looks more realistic (and thus more random over a given
simulation time). Toward this goal, let's attempt to make the oscillator spectrum
more continuous. We can do this by requiring
FS = 1 (31.28)
simulation time
We apply this result in the following example. T
Example 31.10
Repeat Ex. 31.9 if the FS is set using Eq. (31.28).
Using Eq. (31.28) and a simulation time of 10,000 ns, we get an FS = 100 kHz.
Resimulating, using this value of FS, gives the results shown in Fig. 31.10. The
amplitude of the square wave varied from 0 to 1.5 V (varied from 0 to A).
Remembering that the harmonics of the clock (a square wave) have a value of A/2
at DC and a value of 2A/nπ , where n = 1, 3, 5, ... at the other harmonics, we can
calculate the peak amplitude of the fundamental tone in Fig. 31.10 (or Fig. 31.9) as
(2 ⋅ 1.5)/π = 0.955 V or −0.4 dB. T
Peak signal plus noise voltage, dB
Figure 31.10 Oscillator spectrum using Eq. (31.28) to set phase variation time.
Note that a measured spectrum with anomalous spikes would generally indicate
that the noise (the spikes) is not random and could be the result of coupling from an
adjacent circuit. The coupling could be through the substrate or power connections, or it
78 CMOS Mixed-Signal Circuit Design
could be capacitive. If the oscillator jitter is only due to MOSFET noise, in general, no
unwanted spikes will exist in the oscillator's output spectrum.
In the frequency domain the jitter (which is called phase noise) is usually specified
at some offset to the fundamental carrier and taken with reference to the carrier. For
example, at a 1 MHz offset in Fig. 31.10 (at 99 MHz or 101 MHz), the spectrum has a
peak amplitude of approximately −50 dB. The phase noise at a 1 MHz offset would then
be given with reference to the carrier as −50 dB − (−0.4 dB) (the carrier amplitude) =
−49.6 dBc. However, when talking about phase noise, we are generally referring to a
single-tone sinusoid (not a square wave with odd harmonics).
Using Our SPICE Jitter Model
The model we've just developed is difficult to use in a practical simulation because of the
finite step time used by SPICE. For example, if we are trying to model the effects of 100
ps of clock jitter on a data converter's performance, then our step size in the simulation
should be much smaller than 100 ps. This requirement can lead to very long simulation
times or, if the step size is comparable to the simulated jitter, questionable results (see the
example below). Nevertheless, the model is useful in many situations.
Example 31.11
Suppose a 100 MHz sampling clock has 500 ps of jitter. Determine how the SNR
of an ideal data converter will be affected when clocked with this signal. Assume
the topology and input signal of Fig. 31.1 are used.
We begin by using Eq. (31.19) to calculate the number of bits lost
N Loss = 8 + 3.33 log(2π ⋅ 25MEG⋅500p) = 4.3 bits
The SNR of the 8-bit system is determined using Eq. (31.18)
SNR = 6.02(8 − 4.3 − 0.5) + 1.76 = 21 dB
The next factor that we need to determine is the modulation index, MDI,
which simulates 500 ps of jitter. Using Eq. (31.27) we get
MDI = 2π ⋅ 100MEG⋅500p = 0.314
To keep the simulation time relatively short, we'll simulate for 2,000 ns with a step
size of 100 ps. The resolution of the DFT is 1 MHz (set by Eq. [30.35]). The rate
at which the jitter varies is set by Eq. (31.27) and is 500 kHz. The DAC output
spectrum is shown in Fig. 31.11. The simulated V Qe+N,RMS is 93 mV. The SNR is
then given by 20 ⋅ log 0.75/ 2 /93mV = 15 dB (practically worthless). We see
the simulated SNR is fully 6 dB below the calculated SNR. We may speculate that
this is due to both the jitter we purposely introduced into the clock and the jitter
introduced by the varying step size in the SPICE simulation. When we used a pulse
source to clock our ADC in previous simulations, the jitter was absent because of
the exact timing of the pulse statement and the fact that the rise and fall times were
set by the simulation step size. T
Chapter 31 Data Converter SNR 79
or knowing
The spectrum of the average value of a function can be found by taking the Fourier
transform of the autocorrelation function. The result is called the power spectral density
function (PSD) and is given by
∞
units, V 2 /Hz or V 2 ⋅ s (31.35)
P in ( f ) = ∫ R in (t) ⋅ e −j⋅2πf⋅t ⋅ dt
−∞
The power spectral density function of Eq. (31.29) is then, with the help of Eq. (31.34),
V 2p
P in ( f ) = ⋅ [δ( f + f in ) + δ( f − f in )] (units, V2/Hz) (31.36)
4
This is simply two impulses in the frequency spectrum located at ± fin with an amplitude of
V 2p /4 (V 2 /Hz) . The total average power of this signal is given by
∞ ∞
P AVG =
−∞
∫ P in ( f ) ⋅ df = 2 ⋅ ∫ P in ( f ) ⋅ df (units, V 2/Ω or watts) (31.37)
0
The RMS value of Eq. (31.29) is simply, as one would expect for a sinewave, V p / 2 .
Note the similarity between Eq. (31.38) and Eq. (31.6). The factor of root 2 in Eq. (31.6)
is used because VDFT ( f ), the output of WinSPICE, is the peak voltage at a given
(one-sided spectrum) frequency (and so dividing VDFT [ f ] by 2 results in RMS voltages
as a function of frequency).
Chapter 31 Data Converter SNR 81
Example 31.12
Determine the ACF, PSD, average power, and RMS value of a signal V(t) made up
of three sine waves with peak amplitudes of V1 , V2 , and V3 with frequencies of f1 ,
f2 , and f3 .
Using Eqs. (31.30) and (31.34), the ACF is
V 21 V2 V2
R(t) = cos 2πf 1 t + 2 cos 2πf 2 t + 3 cos 2πf 3 t (units, V 2 )
2 2 2
The PSD (positive frequencies) is determined using Eqs. (31.35) and (31.36)
V 21 V2 V2
P( f ) = ⋅ δ( f − f 1 ) + 2 ⋅ δ( f − f 2 ) + 3 ⋅ δ( f − f 3 ) (units, V 2 /Hz)
4 4 4
The average power, using Eq. (31.37), is
V 21 + V 22 + V23
P AVG = (units, watts)
2
Finally, the RMS value of the signal is given by
V 21 + V 22 + V 23
V RMS = (units, V)
2
Note that if we added phase shifts to our signals the results would be the same; the
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. T
Next, suppose that the sinewave specified by Eq. (31.29) is sampled at a rate of fs
V in (nT s ) = V p sin (2πf in ⋅ nT s ) (31.39)
The ACF for a sampled signal can be written as
N
R in (nT s ) = lim
N→∞
1
(2N + 1) Σ V in (kT s) ⋅ Vin (kTs + nTs )
k=−N
(31.40)
which results in
V 2p
R in (nT s ) = cos 2πf in ⋅ nT s (units, V 2 ) (31.41)
2
The PSD is the Fourier transform of this equation (see Eq. [30.2] in the last chapter),
V 2p ∞
P in ( f ) =
4T s Σ [δ( f − f in + kf s ) + δ( f + f in + kf s )]
k=−∞
(31.42)
The RMS value of the sampled sinewave, Eq. (31.39), assuming we have passed the signal
through an ideal reconstruction filter (RCF) with a bandwidth of fs/2, is simply, once
again, V p / 2 . The PSD of the signal, after passing through the RCF, has an amplitude of
V 2p /4 at frequencies of ± fin.
82 CMOS Mixed-Signal Circuit Design
Example 31.13
Determine the average value of the jitter with the PDF shown in Fig. 31.12.
Trace
1 Ideal clock edge position ρ(t)
Probability density function, PDF
2 Edge too early
5 1 3 4
2
Edge too late 1
3
∆T s
−∆T s 0 ∆T s time
5 Edge close to boundary
2 2
Peak-to-peak jitter, ∆T s
Figure 31.12 Clock jitter assuming the edge falls with the same probability
anywhere within the peak-to-peak limits.
Chapter 31 Data Converter SNR 83
We can use Eq. (31.43) to determine the average value of any PDF. Applying this
equation to the PDF shown in Fig. 31.12 results in
∆T s /2
Average value, y, = ∫ t ⋅ 1 ⋅ dt = 0
−∆T 2
∆T s
s
This somewhat obvious result means that the average position of the clock rising
edge is the ideal position indicated by trace 1 in Fig. 31.12. Any PDF that is
symmetrical about some center point will have an average equal to the center
point. T
The variance of the PDF is defined as the average of the square of the signal's
departure from its average value. For a random signal this can be written as
∞
∫ (y − y)
2 2
σ 2 = (y − y) = ⋅ ρ(y) ⋅ dy (31.44)
−∞
where σ is the standard deviation of the PDF (the square root of Eq. [31.44]). For our
purposes, in this book, we can think of variance as the average power of a random
(voltage) signal and the standard deviation as the RMS value of the signal (see Eqs.
[31.37] and [31.38]). Example random signals include the time difference between the
actual edge of a clock and the ideal edge location (jitter), the voltage difference between
the input of an ADC and the ADC's reconstructed output (quantization noise), and the
random fluctuations of electrons due to thermal motion in a resistor (thermal noise).
Example 31.14
Determine the RMS value of the jitter when the jitter has a probability density
function, PDF, as shown in Fig. 31.12.
Using Eq. (31.44) the variance of the jitter PDF is
∆T s /2
(∆T s ) 2
σ2 = ∫ t 2 ⋅ 1 ⋅ dt = 1 ⋅ t 3 ∆T s /2
−∆T s /2 = (seconds2)
−∆T /2
∆T s 3 ⋅ ∆T s 12
s
ρ(t)
Probability density function, PDF
∆T s
RMS jitter = σ ≈
6
ρ(t) = 1 ⋅ exp − t 2
Peak-to-peak jitter ≈ ∆T s σ 2π 2σ 2
3σ
σ 2σ Time
0
6σ ≈ ∆T s
where δT s (t) is a random variable indicating the jitter in the sampling clock at a given
time. (The variable δT s (t) is the time difference between the actual clock transition time
and the expected transition times that are spaced by Ts [see Fig. 31.12].) The peak-to-peak
value of δT s (t) is ∆T s , while its average value is zero. Again, we assume that the jitter
probability distribution function is Gaussian, as seen in Fig. 31.13.
Rewriting Eq. (31.45) using a discrete time step nTs, the sampling error can be
written as
Sampling error amplitude Carrier term
Sampling error amplitude spectrum Data converter output spectral content resulting from jitter
f 0 f in f
0
Example 31.15
Repeat Ex. 31.7 assuming the clock jitter has a Gaussian PDF.
In this example the peak amplitude of the input signal, Vp , is 0.75 V, the input
frequency, fin , is 25 MHz, and the peak-to-peak jitter is 100 ps. The average
power in the sampling error amplitude spectrum is
2 2
(V p ⋅ 2πf in ) ∆T 2 (V p ⋅ 2πf in )
P AVG,jitter = σ 2 ⋅ = s ⋅ (31.47)
2 6 2
or
100 ps 2 (0.75 ⋅ 2π ⋅ 25 MHz) 2
P AVG,jitter = ⋅ = 1.93 × 10 −6 V 2
6 2
while the RMS voltage associated with this error is 1.39 mV. The quantization
noise associated with this 8-bit data converter is
V LSB V REF+ − V REF−
V Qe,RMS = = = 1.69 mV
12 2 N 12
The RMS noise voltage due to clock jitter and quantization effects is then given by
1.39 2 + 1.69 2 mV = 2.1 mV
We can calculate the SNR using
0.75/ 2
SNR= 20 ⋅ log = 48.1 dB
2.1 mV
giving an effective number of bits, from Eq. (31.5), equal to 7.7. Note that this is a
significant improvement over what was calculated in Ex. 31.7, where the jitter
variation was always the peak-to-peak value. T
The PSD of the sampling error amplitude, described by Eq. (31.46), can be
determined with the help of Eq. (31.37)
2 ∞
(V p ⋅ 2πf in )
σ2 ⋅ = 2 ∫ P jitter ( f ) ⋅ df (31.48)
2 0
If the spectrum of the phase noise due to jitter is narrow, as seen in Fig. 31.14, then the
spectral density of the sampling error, Pjitter( f ), 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. [31.48] from DC to
fs/2), we can write
2
2 (V p ⋅ 2πf in )
P jitter ( f ) = σ ⋅ (31.49)
fs 2
The power spectral density of the sampling error voltage, assuming even distribution of
the noise throughout the base spectrum, is shown in Fig. 31.15.
86 CMOS Mixed-Signal Circuit Design
P jitter ( f ), V 2 /Hz
2
σ 2 ⋅ (Vp ⋅ 2π ⋅ f in )
fs 2
0 f s /2 f
Figure 31.15 Sampling amplitude error PSD assuming sampling error spectrum is white.
P osc ( f ), V 2 /Hz
0 fs f
f L1
f H1
Figure 31.16 Measured oscillator spectrum.
Chapter 31 Data Converter SNR 87
5.00
3.33
1.67
0
1 10 100 1k K Number of points averaged
where P Qe ( f ) is the quantization noise power spectral density. Also, from the last chapter
we can write
2
P Qe ( f ) = [V Qe ( f )] (31.55)
We might think that if the quantization noise is white (Bennett's criteria hold, so
there is no correlation from one data converter output sample to the next) then the
spectral content of the noise is spread evenly in frequency from zero to infinity (P Qe [ f ] is
a constant with frequency). This would also mean that P Qe ( f ) approaches zero, from Eq.
(31.54), in order to make the average power, that is, the variance (σ 2 ), of the quantization
noise equal to (V LSB ) 2 /12 .
Before we address this concern (our spectral density approaching zero), let's
review how we calculated the RMS quantization noise voltage, V Qe,RMS , from the
spectrum given in Fig. 30.48 back in Ch. 30. In this figure we looked at the entire
spectrum (or most of the spectrum, up to 200 MHz or 2fs , where significant spectral
content is found) to determine V Qe,RMS (see Ex. 30.11 and the discussion concerning the
figure). We know that the quantization noise doesn't experience aliasing since quantization
occurs after sampling. So while it is correct to look at a wide spectrum to calculate noise,
it would be more useful to limit our view of the spectrum to frequencies up to the Nyquist
frequency (= f n = f s /2) , where our desired signal spectrum should reside. We can do this
by assuming the entire quantization noise power lies in the base spectrum or
fs /2
V 2LSB
12
=2 ∫ P Qe ( f ) ⋅ df (31.56)
0
or
2
V
P Qe ( f ) = 1 ⋅ LSB (31.57)
f s 12
The PSD of the quantization noise is plotted in Fig. 31.18. Note the similarity to Fig.
30.57 (the voltage spectral density of the quantization error).
Consider the result of adding two consecutive ADC outputs as shown in Fig.
31.19. A simple sum will be considered the average of the two consecutive ADC output
signals. The finite digital output word length, in this case 8 bits, can limit the resolution of
the resulting sum. In the cases where we do need to do a division by two we could simply
use the top eight bits of the sum (a shift-right operation). In most of the discussions
related to digital words in this book, averaging will be equivalent to addition. The current
Chapter 31 Data Converter SNR 89
P Qe ( f ), V 2 /Hz
V 2LSB
12 ⋅ f s
f n = f s /2 f
time sample coming out of the ADC is labeled x(nT s ) , while the previous ADC output is
x[(n − 1)T s ] . The output of the simple digital averager is
y(nT s ) = x(nT s ) + x[(n − 1)T s ] (31.58)
remembering that Bennett's criteria must be valid for averaging to effectively reduce the
quantization noise. For example, applying a DC input signal to the circuit of Fig. 31.19
will not result in higher accuracy (the output of the averager will remain the same as the
output of the ADC [actually the averaged output is twice the ADC output]). We'll discuss
this restriction in more detail in a moment when we discuss adding a dither or pseudo
random noise signal to the input to randomize the quantization noise (make its spectrum
white). Also note that there are restrictions on the allowable range of input frequencies
when using this configuration to avoid amplitude distortion. For example, if f in is 50 MHz
(with f s = 100 MHz ), then it's easy to show that the resulting digital averager output is
zero (see Fig. 31.20).
Analog
Simple digital averager.
V IN = V p sin (2πf in ⋅ t) 8
Ideal x[nT s ]
f s = 100 MHz 8-bit ADC
y[nT s ]
9
Digital Digital Output
(averaged)
8 Eight 8
latches x[(n − 1)T s ] Adder
Let's show that the digital averager of Fig. 31.19 can be thought of as a filter and
look at how passing the ADC output through the averager affects the ADC's signal plus
quantization noise and distortion output spectrum. This will also tell us how we have to
restrict the input frequencies applied to the ADC to avoid amplitude distortion or
something similar to what's shown in Fig. 31.20.
90 CMOS Mixed-Signal Circuit Design
time
20 ns 40ns
Figure 31.20 The limitations placed on ADC input frequency when using averaging.
Consider the redrawn (Fig. 31.21) z-domain version of the digital averaging filter
of Fig. 31.19. The filter's transfer function can be found directly from Fig. 31.21 or by
taking the z-transform of Eq. (31.58) as
Y(z)
H(z) = = 1 + z −1 (31.59)
X(z)
f
j2π f
Remembering from Eq. (30.12) that z = e s we can write
real imaginary
f in
−j2π f f in f in
H(z) = 1 + e s = 1 + cos −2π + j⋅sin −2π (31.60)
fs fs
Taking the magnitude of this equation results in
2 2
f in f in
H(z) = 1 + cos 2π + sin 2π (31.61)
fs fs
f
H( f ) = 2 1 + cos 2π (31.62)
fs
and the phase is given by
−sin 2π f
fs
−1
∠H( f ) = tan (31.63)
1 + cos 2π f
f s
X(z)
Y(z)
ADC output Simple digital filter Output
z −1 X(z)
z −1
Referring to Ex. 30.4 of the last chapter the phase can be written as
f
∠H( f ) = −π ⋅ (units, radians) for f < f s /2 (31.64)
fs
The magnitude and phase responses of this simple digital filter are shown in Fig. 31.22.
Note that this is the discrete version of the comb filter discussed in Ex. 30.4. Also note
that (1) the phase response is linear, (2) the response is periodic (as is the response of any
digital filter), and (3) at an input frequency of half the Nyquist frequency, f s /4 , the
magnitude response is 2 (3 dB down from the DC gain of two).
∠H( f )
H( f ) 2 degrees
90
2 f in (Hz)
-90
f s /4 f s /2 fs 3f s /2 f in (Hz) f s /2 3f s /2
Figure 31.22 Magnitude and phase response for the simple digital filter of Fig. 31.21.
We can also see that (1) averaging results in an attenuation of many of the input
signal frequencies (as shown in Fig. 31.22) and (2) indeed the average of the input signal
goes to zero, as was shown in Fig. 31.20, when the input signal frequency is f s /2.
If we assume the output quantization noise power spectral density, P Qe ( f ) , for the
ADC shown in Fig. 31.19 is white then the output of the simple digital filter has the PSD
shown in Fig. 31.23 (the product of the filter response squared with the noise PSD). The
average power contained in this PSD is
f s /2
V2LSB f
P AVG = 2 ∫0 12 ⋅ f s
⋅ 2 1 + cos 2π df
fs
(31.65)
P Qe ( f ), V 2 /Hz
V 2LSB
3 ⋅ fs
f n = f s /2 f
Figure 31.23 Quantization noise power spectral density after averaging two samples.
92 CMOS Mixed-Signal Circuit Design
or
f = f /2
V 2LSB V 2LSB f s V2
P AVG = + sin 2π = LSB (31.66)
6 6π fs f = 0 6
The power in an input sinewave before averaging is V 2p /2 (the RMS voltage of the
sinewave is V p / 2 ). Averaging (adding) two samples results in an increase in the desired
signal amplitude by two and so the power increases to 2V 2p (the RMS voltage increases to
(2V p )/ 2 ). This is important because now the SNR, on the output of the digital averaging
filter, is
2V p / 2 Vp / 2
SNR= 20 log = 20 log (31.67)
V LSB / 6 V LSB / 24
or in terms of a generic averaging constant K (see Eqs. [31.51] or [30.48]), the effective
RMS quantization noise voltage is
1 V LSB
V Qe,RMS = ⋅ (31.68).
K 12
Without rederiving the equations presented at the beginning of this section, we should see
how averaging affects a data converter's SNR.
An Important Observation
Equation (31.68) assumes the averaging filter does not attenuate the input signal. If, for
example, the input frequency were f s /4 , then the RMS amplitude of the desired signal
would change from (2V p )/ 2 to V p ⋅ 2 / 2 or simply V p (because of the root two
gain at f s /4 , as shown in Fig. 31.22) and the SNR would be the same as the output of the
ADC in the nonaveraged circuit. If the input frequency were greater than f s /4 , then the
SNR would actually be worse than the nonaveraged SNR! Therefore, we have to restrict
the input frequency bandwidth, B, to frequencies less than f s /4 when averaging two terms
in order to avoid degrading the data converter's SNR. In general, for an arbitrary number
of averages K, we can write the restrictions on the input bandwidth using
f s /2 f n
B= = and f in ≤ B (31.69)
K K
We have already presented this equation (Eq. [31.22]) when discussing how oversampling
affects sampling clock jitter stability requirements. The averaging factor K is commonly
called the oversampling ratio to denote the ratio of the Nyquist frequency to the input
signal bandwidth. This can sometimes be confusing since, as we showed in the last
chapter, oversampling an input waveform alone, without averaging, does not lower the
amount of quantization noise in a data converter's output spectrum. Nevertheless, stating
that a data converter is using oversampling is synonymous with stating the data converter
employs an averaging filter. The averaging filter used on the output of an ADC is called a
Chapter 31 Data Converter SNR 93
decimating filter while the reverse averaging filter used on the input of a DAC is called an
interpolating filter. We will discuss these filters in detail in the next sections.
Example 31.16
Suppose the input sinewave in Fig. 31.19 has a peak amplitude of 0.5 V and a
frequency of 20 MHz. Determine the peak amplitude of the averager output and
the delay through the circuit. Comment on any assumptions made.
Using Eq. (31.62) we get
H( f ) = 2 1 + cos 2π 20 = 1.62
100
and so the peak amplitude of the output sinewave is 0.5 ⋅ 1.62 = 809 mV . Ideally,
the amplitude out of the averager is twice the input or, in this case, 1 V.
The delay through the filter is determined using Eq. (31.64) and knowing
Phase shift
f
2π ⋅ f ⋅ ∆t = −π
fs
The constant delay (knowing the minus sign indicates the output of the filter
occurs after, or later in time than, the input signal) can then be written as
T
∆t = 1 = s (31.70)
2f s 2
and so, for this example, ∆t = 5 ns .
Note that we are not discussing the effects of quantization noise, that is, the
fundamental minimum voltage that can be resolved. We are assuming continuous
amplitude signals throughout the system in order to simplify the filter calculations.
This assumption falls apart if, for example, the peak-to-peak amplitude of the input
sinewave is reduced to a value below one least significant bit. This will cause the
circuit to function as if the input were a DC signal. T
Jitter and Averaging
We can apply the averaging discussion just developed directly to the jitter discussion
presented earlier in the chapter and answer the question, "How does averaging affect the
sampling amplitude error power (resulting from jitter) in a data conversion system?" If we
assume that the jitter has a Gaussian PDF, then the average power in the sampling error
amplitude, from Ex. 31.15, is
2
Vp
P AVG,jitter = σ ⋅ ⋅ 2πf in (31.71)
2
where σ is the standard deviation of the jitter (see Fig. 31.13). It may be helpful to rewrite
Eq. (31.68) in terms of the quantization error power as
94 CMOS Mixed-Signal Circuit Design
2
V
P Qe,AVG = (V Qe,RMS ) = 1 ⋅ LSB
2
(31.72)
K 12
and apply the same derivation to Eq. (31.71) to give
2
Vp
P AVG,jitter = 1 ⋅ σ ⋅ ⋅ 2πf in (31.73)
K 2
This equation shows that the sampling error amplitude power, PAVG,jitter, introduced into the
data converter's output spectrum decreases with averaging. Averaging two samples causes
the sampling error amplitude power to decrease by 3 dB. This effectively reduces the jitter
requirements placed on the sampling clock. While this may not appear to be very
significant at first glance, consider what happens if, for example, 256 samples are averaged
(K = 256 ). The sampling error power decreases by 24 dB, making clock jitter, when using
a reasonably stable oscillator, almost not an issue. Also note that a doubling in the jitter's
standard deviation, σ , results in a 6 dB increase in sampling error amplitude power.
Relaxed Requirements Placed on the Antialiasing Filter
The use of averaging will also lead to relaxed requirements of the antialiasing filter (AAF).
Figure 31.24a shows the requirements placed on the AAF without averaging. As we saw
in the last chapter, ideally, the transition from the 3 dB frequency to the "stop frequency"
or Nyquist frequency should be infinitely sharp (the filter should abruptly change from a
gain of unity to a gain of zero [something small]). When using averaging, Fig. 31.24b, we
have to limit our desired input signal bandwidth to B; see Eq. (31.69). The rolloff of the
filter in part (b) of the figure can be much more gradual and in many cases a simple, single
pole, RC filter is all that's needed for an AAF. Also, our averaging filter will attenuate the
ADC output spectrum, as seen in Fig. 31.22, and help to remove input signal power above
f s /2K . The 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 (as
discussed in the next section). Of course, the penalty for the relaxed requirements of the
AAF is reduced signal bandwidth for a fixed sampling frequency.
H( f ) H( f )
f 3dB B = f s /(2K)
1 1
f n = f s /2 f f n = f s /2 f
(a) (b)
Figure 31.24 (a) AAF requirements without averaging, and (b) AAF requirements with averaging.
Chapter 31 Data Converter SNR 95
ADC output 1
This is especially true when a resolution greater than 10 bits is desired with INL and DNL
less than ± 0.5 LSBs. Later in the chapter, and in the next chapter, we will look at
feedback topologies that may relax the accuracy requirements placed on the ADC and
allow averaging to more effectively remove quantization noise.
Example 31.17
To illustrate the requirements placed on the accuracy of the original ADC in more
detail, consider averaging 16 consecutive ADC output samples: 15 at a digital
code of zero and one at a digital code of 1 LSB. Determine the accuracy required
of the ADC, the size of the word coming out of the averaging filter assuming the
ADC is 8 bits, and the final word size after considering the increase in resolution.
The ideal output of the digital filter will be 1/16 of the original ADC's LSB. We
can write this as
16 consecutive ADC outputs
Example 31.18
Specify the accuracy required of an 8-bit ADC if it is to be used with oversampling
to attain 12 bits with INL and DNL of ± 0.5 LSBs.
The increase in the number of bits, NInc , is 4. The accuracy required of the 8-bit
ADC, from Eq. (31.75), is ± (1/32) of an LSB. If VREF+=1.5 V and VREF−= 0 then
the LSB of the ADC is 5.86 mV. The output of the ADC must be within ± 183 µV
of the ideal ADC output levels in order to arrive at a final, after averaging,
resolution of 12 bits (with a 12-bit accuracy of ± 0.5 LSBs). Also, according to
Eq. (31.53), we will have to average 256 consecutive ADC outputs to get a 12-bit
output. T
Adding a Noise Dither to the ADC Input
Our assumption, when discussing the benefits of averaging or calculating the spectral
density of the quantization noise, falls apart for DC or slow-moving signals (the ADC
input is not "busy"). To solve this problem consider adding a noise signal to the ADC
input that has a frequency content that falls within the range
fs fs
≤ f < (31.76)
2K 2
so that it can be filtered out with the averaging filter. This noise is often called dither (a
state of indecision or agitation) because it helps to randomize the spectral content of the
quantization noise making it white (a flat spectrum, see Fig. 31.18).
Figure 31.26 shows the basic idea. In part (a) a DC signal is applied to the ADC
that falls halfway between two ADC transition codes spaced apart by 1 LSB. The output
code of the ADC remains unchanged with time. In part (b) a noise signal is added to the
DC input which has two benefits: (1) the quantization noise (the difference between the
input signal and the reconstructed ADC output code) changes with time, and (2) the
output of the ADC has some variation which makes it possible to determine the DC
voltage after averaging.
DC Input signal
(a) (b)
Figure 31.26 (a) DC input signal and (b) DC input signal with dither added.
98 CMOS Mixed-Signal Circuit Design
We can add the noise signal to our desired input signal with a circuit similar to
what's shown in Fig. 31.27. Simple resistors add and reduce the noise signal 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 (= 1.5 V in this chapter).
In this figure note that we have indicated that the dither signal amplitude should be
approximately 0.5 LSB RMS (remembering the signal is, ideally, random and bandlimited
as specified by Eq. [31.76]). This number, 0.5 LSB RMS, is subjective, and no exact rules
as to its selection can be given other than the desire that the peak-to-peak amplitude be
greater than 1 LSB. One disadvantage of adding the dither is that the 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).
5,000
Dither Approximately 0.5 LSB RMS dither Dither
noise 0 to 1.5V generating
source digital
circuit
Block diagram
Circuit implementation
Figure 31.27 Adding dither to an ADC input signal.
ρ(t)
Probability density function, PDF
1 ⋅ exp − (V in − V in )
2
ρ(t) =
RMS dither = σ = 0.5 LSB σ 2π 2σ 2
Amplitude variation
with time
σ = 0.5 LSB Volts
DC in, V in
Figure 31.28 Input to the ADC, dither and DC, with a Gaussian probability distribution.
Chapter 31 Data Converter SNR 99
would also mean that we can have some dither spectral content below f s /2K as long as
we average enough ADC output samples to make its contribution to the SNDR small. It
is generally a good idea to use Eq. (31.76) as a guide 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 VDD/2 before amplitude reduction). If not, an unknown DC offset
(the known DC offset is the VDD/2 attenuated by the resistive divider in Fig. 31.27) in the
data converter's (actually the filter's) output will result.
An example of an implementation of a dither noise source is shown in Fig. 31.29.
The outputs of the rows of inverters, which are tied together, will occur asyncronously
and fight against each other causing the amplitude of the dither signal to occupy levels
other than the normal logic levels of VDD and ground for significant amounts of time. The
dither signal can be made more random by adding more rows of inverters. The 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.
Dither
out
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 by looking at the
z-domain representation of the system. Toward this goal, consider the transfer function of
the simple digital averager depicted in Figs. 31.19 and 31.21 with a z-transform of
= 1 + z −1 = z +z 1
Y(z)
H(z) = (31.77)
X(z)
Y(z) is the system's output, while X(z) is the system's input. Note that the system is
discrete in time but not necessarily in amplitude. We can apply the z-transform to
switched-capacitor circuits with continuous-valued amplitudes as well as to data
converters with quantized values of amplitude. It is very useful, for an intuitive
100 CMOS Mixed-Signal Circuit Design
understanding of the frequency response of a discrete system, to plot the transfer function
in the z-plane (Fig. 31.30). Figure 31.30 also shows how Eq. (31.77) can be displayed on
the z-plane. A pole is located at z = 0 (at the location the denominator goes to zero and
the transfer function goes to infinity) and a zero is located at z = −1 (at the location where
the numerator goes to zero).
f Imaginary
j2π⋅ f
z=e s
z-plane
Indicates a pole 1
Indicates a zero
6
Indicates six poles at a location
2
Indicates two zeroes at a location Real
The z-plane is usually used to describe the frequency response of a discrete time
system, H( f ), by assuming the input to the system is a unit magnitude sinusoid with
f
j2π f
varying frequency, f. This input, 1 ⋅ e s ( = z) , evaluates the output of the system or
phase
magnitude
f
j2π f
H( f ) = H(z) evaluated when z = 1 ⋅ e s (31.78)
We should now see that the unit circle, shown in Fig. 31.30, indicates the relationship
between z and f when specified by Eq. (31.78). Therefore, to determine H( f ) from a plot
of H(z) on the z-plane, we simply evaluate H(z) along the unit circle. To show how this
transfer function evaluation is performed, consider Eq. (31.77) and the corresponding plot
of its pole and zero shown in Fig. 31.30 along with the magnitude of Eq. (31.77) or Eq.
(31.59) plotted against the frequency in Fig. 31.22. At DC ( f = 0 and z = 1 ⋅ e 0 = 1∠0)
point A in Fig. 31.31, the gain of the circuit is two and is calculated using
z-plane
f
j2π⋅ f
z=e s
Point B, f = f s /4, 5f s /4 ...
Point A, DC or f = 0, f s , 2f s ...
Figure 31.31 The z-plane pole and zero for Eq. (31.77).
which, as seen in Fig. 31.22, results in a phase angle of zero. Next consider evaluating the
H(z) at f s /4 f = f s /4 and z = 1 ⋅ e j 2 = 1∠90 , point B in Fig. 31.31. The distance from
π
the pole to point B is 1 while the distance from the zero is 2 resulting in a magnitude
2 . The angle from the pole along the x-axis to point B is 90°, while the angle from the
zero is 45° resulting in an overall phase response of −45° (verify with Fig. 31.22).
Also note that (1) any digital filter's or system's frequency response is periodic with
period f s (one complete revolution around the unit circle), (2) we normally are only
concerned with evaluating H(z) over the top half of the unit circle (from DC to f s /2 [the
Nyquist frequency, f n ]), and (3) a pole at the origin has no effect on the magnitude
response of H(z) but does affect the phase response (as shown in Ex. 30.6 in which
multiplying H(z) by z−1, adding a pole at the origin to H(z), simply shifts the output later in
time). Finally note that the number of poles in H(z) must be greater than or equal to the
number of zeroes if the digital filter/system is to be realizable in hardware (the output of
the system cannot occur before the system's input).
Example 31.19
Determine, using the graphical approach just discussed, the magnitude and phase
of the transfer function shown in Fig. 31.32 at a frequency of f s /4 .
If we label the length from a pole (zero) to the evaluation point p (z), then the
magnitude of the transfer function is given by
z
H(z) = p ⋅1p
1 2
Labeling the angles for the poles and zeroes as indicated in the figure, we can write
the phase response as
∠H(z) = θ 3 − θ 2 − θ 1 T
102 CMOS Mixed-Signal Circuit Design
z-plane z-plane
Evaluated here
z1 p1 θ1
p2 θ3
θ2
Figure 31.32 The z-plane pole and zero plot for Ex. 31.19.
Example 31.20
Determine the frequency response of a digital system with the time domain
response
y[nT s ] = x[nT s ] + y[(n − 1)T s ]
Sketch the hardware implementation of the system and its frequency response.
The z-domain transfer function for this system is
Y(z) = X(z) + Y(z) ⋅ z −1
or
H(z) = 1 = z (31.81)
1 − z −1 z − 1
The hardware implementation of the system is shown in Fig. 31.33 along with the
z-domain representation. Note that the size of the words used (the number of bits
coming out of the adder and the number of latches) depends on the application.
x[nT s ]
y[nT s ] X(z)
Y(z)
y[(n − 1)T s ] Y(z)z −1
z −1
Latches 1 = z
Clock H(z) =
1 − z −1 z − 1
Figure 31.34 shows the z-plane representation of this system along with the
magnitude and phase response of the system. This circuit is called a digital
integrator. To show why, let's determine the magnitude and phase responses, using
Eq. (31.81), and noting the z in the numerator is simply a phase shift
Phase shift
H( f ) = 1 = 1 (31.84)
2 2 f
−1 + cos 2π f + sin 2π f 2(1 − cos 2π f s )
fs fs
f
− π + π = 180 − 90 (degrees) for 0 < f < f s (31.85)
f f
∠H( f ) = 2π
fs fs 2 fs
At DC the phase contribution from the zero is 0°, while the phase contribution
from the pole, at a frequency just above DC, is 90°. The result is an overall phase
response of −90°. At fs /4 the phase contribution from the zero is 90°, while the
phase contribution from the pole is 135°, resulting in an overall phase response of
−45 °. T
1 = z H( f )
H(z) =
1 − z −1 z − 1 z-plane
0.5
f s /2 fs 3f s /2 f
∠H( f )
degrees
90
f
90
Figure 31.34 The z-plane representation along with magnitude and phase response
for a digital integrator.
104 CMOS Mixed-Signal Circuit Design
Consider the possible input to our digital integrator (Fig. 31.33) and the resulting
output shown in Fig. 31.35a and 31.25b respectively. In this figure we are using +1 to
indicate the peak positive input sinewave amplitude and −1 to indicate the peak negative
input amplitude. The frequency of the sinewave is f s /2 and so, according to our
magnitude plot in Fig. 31.34, the gain is 0.5 (the peak-to-peak amplitude of the sinewave
is reduced by one-half). Looking at Fig. 31.35, we should see that the initial state of the
register (the latches) used in the integrator will, together with the input, determine the DC
offset in the output waveform. For example, if the latches initially contained zero and the
first sample was +1, as seen in Fig. 31.35a, then we would get the waveform shown in Fig.
31.35b. If, instead, the first sample were −1, then the entire waveform in Fig. 31.35b
would shift downwards by +1. In any integrator, digital or analog, the "initial conditions"
will affect the output waveform.
The next important factor we should notice in Fig. 31.35 is that we picked the
peaks of the input sinusoid as our inputs to the digital integrator. Shifting our ADC output
sampling points by Ts/2 results in a signal of all zeroes being applied to the digital
integrator. The result is no change in the integrator's output. Shifting the sampling points
by Ts/3 results in an integrator input of ± 0.5 with a corresponding integrator peak-to-
peak output of ± 0.25 . In any case, at f s /2 , the output of the digital integrator is one-half
the input signal's amplitude.
x(t)
x(nT s )
+1
time
-1 (a)
y(t) 2T s 4T s
+1
time
y(nT s ) (b)
Figure 31.35 (a) Input and (b) output of the digital integrator of Fig. 31.33.
In the above discussion we used decimal numbers to represent the input and output
signals of the integrator. In a practical implementation we use binary numbers. Let's use
our ideal 8-bit data converters to illustrate the number system concerns. In these
converters V REF− = 0 and V REF+ = 1.5 V with V LSB = 5.859 mV . A code of all zeroes
corresponds to 0 V while a code of all ones corresponds to 1.494 V. The common mode
voltage, V CM , = (V REF+ + V REF− )/2 = 0.75 V . Again this number system is called offset
binary. Figure 31.36 illustrates the representation of a full-scale sinusoid using the offset
binary number format.
Chapter 31 Data Converter SNR 105
11111111 (1.494 V)
(255)
10000000 (0.75 V)
(128) time
00000000 (0 V)
(0)
Figure 31.36 Representing a sinusoid in binary offset format.
We should compare the binary offset numbers of Fig. 31.36 to the decimal
numbers of Fig. 31.35 and notice that if we apply the DC components of each signal to the
integrator, we get totally different results. In Fig. 31.35 the DC component of the input
has a decimal value of 0. Applying 0 to our integrator causes the output of the integrator
to remain unchanged. In Fig. 31.36 the DC component of the input is the common mode
voltage (halfway between the reference voltages) of 10000000. Applying this value to the
integrator results in the integrator's output increasing until the output changes from all
ones to all zeroes (the output rolls over). Clearly, the binary offset representation has
some practical limitations when used in a digital integrator. To avoid these problems, the
binary offset format is usually converted into the two's complement format prior to
application to the digital integrator.
In two's complement the left-most bit is the sign bit. A zero represents a positive
number (except for all zeroes, 00000000, or the common mode voltage) and a one
represents a negative number (see Fig. 31.37). A binary offset number can be translated
back and forth between a two's complement number by simply complementing the MSB of
the code (running the MSB through an inverter). For this reason, and others that will be
discussed later (ease of implementing subtraction and no overflow problems), two's
complement is the preferred format for data words in digital filtering.
01111111 (1.494 V)
V CM + 127 ⋅ V LSB
(+127)
00000000 (0.75 V) V CM = 0.75 time
11111111 (0.7441) V CM − V LSB
(−1)
V CM − 128 ⋅ V LSB
10000000 (0 V)
(−128)
Figure 31.37 Representing a sinusoid in two's complement format.
106 CMOS Mixed-Signal Circuit Design
which is nothing more than averaging K input samples. Before going any further, we
should make sure we understand this equation. Our decimation filter will take K input
samples, add them together, and then divide the result by K to obtain the average of the
input. If K = 16 and i = 1, then samples x[0] through x[15] are summed and divided by 16.
As we saw in Ex. 31.17, the actual division by K is dependent on the increase in
the number of bits in the output word. For example, if our input word is 8-bits and K = 16
then the output word, before dividing by K, is 12-bits (adding sixteen, 8-bit words, results
in a 12-bit word). If the ultimate increase in resolution is 2-bits, then the final output word
size is 10-bits and we throw the lower 2-bits away. This could effectively mean that
instead of dividing by 16, we divide by 4.
Chapter 31 Data Converter SNR 107
We can rewrite Eq. (31.87) in the z-domain (so that we have the z-domain
representation for the decimation filter) as
K−1
Y(z) 1
H(z) = =
X(z) K Σ z−n = K1 (1 + z −1 + z −2 + ... + z 1−K)
n=0
(31.88)
or
−1
H(z) = 1 ⋅ 1 − z −1 ⋅ (1 + z −1 + z −2 + ... + z 1−K ) (31.89)
K 1−z
or finally, the z-domain transfer function for the decimator (averager), is
−K
H(z) = 1 1 − z −1 (31.90)
K1−z
If K = 2, Eq. (31.90) becomes
(z + 1)(z − 1) 1
H(z) = 1 = (1 + z −1 ) (31.91)
K z(z − 1) K
noting that we have already discussed this case, Eq. (31.59), earlier. Note also that the
division by K may be ignored in this case since, as discussed earlier, the word size
increases by one bit when adding the two words. (However, our realized increase in
resolution is only 0.5 bits meaning the SNR increases by 3 dB.)
One circuit used to implement Eq. (31.90) is shown in Fig. 31.39 and is called an
accumulate-and-dump circuit. To understand the operation of this circuit let's assume the
bottom set of latches are reset. The sampling clock is used to clock this set of latches K
times until the sum of K inputs is accumulated. At this time the accumulated sum is
dumped into the output latches. Also, at this time, the bottom set of latches is reset to zero
to start the accumulation process for the next set of K input samples. Note the clock rate
on the input of the circuit is fs and the clock rate coming out of the circuit is fs /K.
Clock
Dump the sum into
this set of latches.
Clock, f s
Latches
Reset
Accumulate the sum in this f s,new = f s /K
set of latches
divide by K
Assuming K = 8
f f
(31.92)
K −j⋅2π⋅ fs K −j⋅2π⋅ fs
1−e 1−e
or, knowing 1 − e −jx = (1 − cos x) + j sin x = (1 − cos x) 2 + (sin x) 2 = 2(1 − cos x) ,
sin Kπ f s sinc Kπ fs
f f f
2(1 − cos K2π f s )
H( f ) = 1 ⋅ =1⋅ = (31.93)
sin π f s sinc π f s
K K
2 1 − cos 2π f s
f f f
fs Clock in H( f )
H( f )
z-plane
−2
K=2 H(z) = 1 − z −1
2 1−z
f s /2 3f s /2 f in (Hz)
B
4
−4
Note different scale! K=4 H(z) = 1 − z −1
1.08 1−z
3
f s /4 f s /2 3 f s /4
B
8
K=8 −8
Note different scale! H(z) = 1 − z −1
1−z
1.7 7
f s /8 f s /4 3f s /8
B
obvious reasons, Eq. (31.93), the frequency response of the accumulate-and-dump circuit,
or an averaging filter with the z-domain response given by Eq. (31.90), is sometimes
called a sinc filter.
Example 31.21
Determine the pole and zero locations (verify the z-plane plot in Fig. 31.40) for an
averaging filter that averages eight samples.
We can write the z-domain representation of the averager using Eq. (31.90)
without the scaling factor K as
−8
H(z) = 1 − z −1 = 7z − 1
8
1−z z (z − 1)
or
H( f )
K Main lobe
First sidelobe
The attenuation can be determined using Eq. (31.93) evaluated at 1.5(f s /K) as
Main lobe = K ⋅ sin 1.5π for K ≥ 3 (31.94)
First sidelobe K
This equation is plotted in Fig. 31.42 against averaging factor K. Note how the maximum
amount of attenuation, as the number of averages increases, approaches 13.5 dB. This is a
significant limitation and results in the need to cascade averaging filter stages to attain a
large amount of attenuation at frequencies above f s /K (more on this topic in a moment).
110 CMOS Mixed-Signal Circuit Design
dB
13.5 dB
13
Main lobe
11
First sidelobe
3 4 6 8 10 K
It's also of interest to determine how much droop the filter will introduce into the
signal frequencies of interest. Figure 31.43 shows the droop (attenuation) at the maximum
input bandwidth, B. We can calculate the amount of droop, again using Eq. (31.93) when
f = f s /(2K) = B , as
Droop = 1 (31.95)
K ⋅ sin 2K
π
H( f ) droop, dB
K -3.5
droop
-3.6
-3.8
B = 0.5( f s /K) f -4.0
3 4 6 8 10 K
Figure 31.43 Droop at edge of signal bandwidth when using an averaging filter.
In Out
Accumulate Accumulate Accumulate
f s clk and dump clk and dump clk and dump clk f s /K 3
f s /K f s /K 2
x(1) + x(2) + x(3) + x(4) + x(5) + x(6) + x(7) + x(8) + x(9) + ... (31.96)
We can get a much more efficient and practical filter if we average the input samples
without decimation (sample frequency reduction). The final output clocking frequency
(after decimation) can be set to 2B (Eq. [31.69]) and can occur in a later stage in the
filter's construction. The reduction in sampling frequency reduces power and circuit
complexity (for example, serial multipliers can be used in a digital filter).
Consider the running sum shown below
First output of averager
x(1) + x(2) + x(3) + x(4) + x(5) + x(6) + x(7) + x(8) + x(9) + ...
Second output of averager
x(1) + x(2) + x(3) + x(4) + x(5) + x(6) + x(7) + x(8) + x(9) + ...
Third output of averager
x(1) + x(2) + x(3) + x(4) + x(5) + x(6) + x(7) + x(8) + x(9) + ...
Fourth output of averager
x(1) + x(2) + x(3) + x(4) + x(5) + x(6) + x(7) + x(8) + x(9) + ... (31.97)
It should be obvious that the outputs of the averager occur at the same rate as the
averaging filter's input (no change in the sampling frequency). The z-domain
representation of the averager is the same as the accumulate-and-dump's transfer function
x[nT s ] + x[(n − 1)T s ] + x[(n − 2)T s ] + ... X(z)(1 + z −1 + z −2 + ... + z 1−K )
y(nT s ) = → Y(z) =
K K
(31.98)
or, reviewing Eqs. (31.88) and (31.89), results once again in
−K
H(z) = 1 1 − z −1 (31.99)
K1−z
where the division by the number of points averaged, K (K is four in Eq. [31.97]) is
performed by simply adjusting the final word size to the desired length (as discussed
112 CMOS Mixed-Signal Circuit Design
earlier and in Ex. 31.17). The transfer function of a cascade of L of these averaging filters
can be written as
−K L
H(z) = 1 1 − z −1 (31.100)
K 1 − z
or
L
sinc Kπ f
H( f ) =
fs
f
(31.101)
sinc π
fs
Before we discuss the implementation of the averaging filter (a.k.a. sinc filter), let's
borrow the results from Figs. 31.42 and 31.43 and notice the attenuation for a cascade of
L averaging filters is
L
Main lobe = K ⋅ sin 1.5π ≈ L ⋅ 13 dB for K ≥ 8 (31.102)
First sidelobe K
while the droop, at the maximum input frequency, B, is
−L
Droop = K ⋅ sin π ≈ L ⋅ (−3.9) dB for K ≥ 8 (31.103)
2K
Also note that Eq. (31.69) is still valid, which means that we don't have a significant
restriction on the maximum allowable input frequency.
To compare the cascade of averaging filters to the limitations imposed by a
cascade of accumulate-and-dumps as discussed earlier, let's once again assume we need 60
dB of attenuation through the averaging filter and our sampling frequency is
f s = 100 MHz . If K = 8, then we need to cascade five averaging filter stages as seen in
Fig. 31.45. The clock frequency coming out of the final stage is 100 MHz and needs to be
reduced to 12.5 MHz in the last stage by simply dividing the clock down before clocking a
final set of latches or by using an accumulate-and-dump for the final stage. The 12.5 MHz
output rate and 6.25 MHz input frequency bandwidth, B, should be compared to the 3
kHz output clock frequency and 1.5 kHz input frequency calculated earlier for the cascade
of accumulate-and-dumps. The droop B remains 19.5 dB and can be a serious concern in
many situations. Obviously, limiting the input bandwidth further reduces the droop at B.
Figure 31.46 shows the frequency response of a cascade of averaging filters (the frequency
response is given by Eq. [31.101]).
In
1 − z −8 1 − z −8 1 − z −8 1 − z −8 1 − z −8 Out
clk 1 − z −1 1 − z −1 1 − z −1 1 − z −1 1 − z −1 clk
H( f )
L ⋅ 3.9 dB
KL
B = 0.5( f s /K) L ⋅ 13 dB
f s /K 2( f s /K) 3( f s /K) f
f s /(2K)
Figure 31.46 General frequency response of an averaging filter, Eq. (31.101).
f s − f s /K f f s /(2K) f s − f s /K f
(a) Input spectrum to the ADC. (b) Output spectrum of the ADC with aliasing.
digital filter
response
f s /(2K) f s − f s /K f f s /(2K) f s − f s /K f
(c) Same as (b) except showing a continuous (d) Output of the digital filter after removing
spectrum. Thin line is digital filter response. aliased components.
Figure 31.47 How a digital (averaging) filter helps remove aliased spectral components.
114 CMOS Mixed-Signal Circuit Design
why the AAF has to limit the ADC input spectral content to f − f s /K . At f s /K the output
of the digital filter goes to zero. This assumes the first sidelobe in the digital filter is
sufficiently small so that the amount of aliasing past fs /K is negligible. In many systems
that employ a digital signal processor, an additional abrupt-cutoff lowpass digital filter is
used after the averaging and decimation process to provide additional alias signal removal.
Note, in Fig. 31.47d the output of the digital filter, as mentioned earlier, has a
periodic frequency response (and so the spectrum out of the digital filter will still have
aliased components but hopefully not in the base spectrum). As a further example of the
periodic nature of all digital filters, consider the K = 8 averaging filter frequency response
shown in Fig. 31.48 (see Fig. 31.40 for a comparison). Note that the number of points in
the frequency response between DC and f s that go to zero (the number of zeroes in the
transfer function) is seven (since the zero at DC is canceled by the pole at DC [z = 1 ]).
H( f )
1 − z −8
K=8 H(z) =
1 − z −1
8
f s /4 f s /2
f s /8 3f s /8 fs 2f s
B
X(z) Y(z)
z −1
Y(z) = [Y(z) + X(z)]z −1
y(nT s ) = y[(n − 1)T s ] + x[(n − 1)T s ] Y(z) −1
H(z) = = z
X(z) 1 − z −1
Figure 31.49 Alternate digital integrator.
Chapter 31 Data Converter SNR 115
f
H( f ) = 2 1 − cos 2π (31.107)
fs
and the phase response is
∠H( f )
H( f ) H(z) = (1 − z −1 ) = z −z 1 degrees
2
90
2 f (Hz)
-90
f s /4 f s /2 fs 3f s /2 f (Hz) f s /2 3f s /2
Next consider the digital comb filter (or differentiator over a range of frequencies)
circuit shown in Fig. 31.52. We should recognize this circuit's transfer function from Eq.
(31.104) as
H(z) = 1 − z −K = z −k 1
K
z
with a magnitude response given by
f
H( f ) = 2 1 − cos 2πK
fs
H(z) = 1 − z −K = z −k 1
clk clk clk K
Clock
z
fs Figure 31.52 Block diagram of a digital comb filter.
Figure 31.53 shows the z-plane and frequency responses for comb filters with
various values of comb filter delays K. Before we proceed with the implementation of the
averaging filter defined by Eq. (31.104), let's discuss, intuitively, how we take the basic
digital comb filter and make a lowpass, averaging filter. Remember that we evaluate
1 − z −K around the unit circle, in the z-plane, to determine H( f ) . If we look at Fig. 31.40,
we see that the only difference between a comb filter and a lowpass averaging filter is the
fact that we have added a pole to the transfer function at DC (i.e. z = 1) to cancel the zero
at DC. This is important as we will be able to make highpass and bandpass averaging
circuits by taking a comb filter and canceling the zeroes placed at other points on the unit
circle. We'll discuss this in more detail in the next two sections. Note that by using Eqs.
(31.79) and (31.80), we should be able to see why canceling a zero with a pole at DC
results in an a lowpass filter.
The comb filter of Fig. 31.52, or the differentiator of Fig. 31.50, is an example of a
finite impulse response (FIR) digital filter. Applying a unit amplitude impulse to the input
of the comb filter, and zeroes at all other times, causes the output of the comb filter to go
to a one at the moment the impulse is applied and KT s seconds later, and a zero at other
times. In other words, the output response of the filter has a finite duration.
The integrator (sometimes called an accumulator) shown in Fig. 31.34 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, 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.
Chapter 31 Data Converter SNR 117
H( f ) H(z) = 1 − z −K = z −k 1
K
z z-plane K=3
2 H(z) = 1 − z −3
3
f s /3 2f s /3 fs f (Hz)
H( f )
z-plane K=4
2 H(z) = 1 − z −4
4
f s /4 f s /2 3f s /4 fs f (Hz)
H( f )
z-plane K = 16
2 H(z) = 1 − z −16
16
f s /4 f s /2 3f s /4 f s f (Hz)
f s /16 15 f s /16
Figure 31.53 Frequency response and z-plane plots for various values of K in a comb filter.
Example 31.22
Consider an averaging filter (Eq. 31.104) using L = 3 and K = 8 with an input
word length of 8 bits. Determine the final number of bits coming out of the filter
and show that a constant input of 01110000 (+112 two's complement) results in
the correct output code. Also, discuss overflow concerns.
118 CMOS Mixed-Signal Circuit Design
From Eq. (31.53) we can calculate an increase in resolution of 1.5 bits per filter
stage so the final output word size should be 13 bits (the original 8-bits input plus
an additional 4.5 bits from averaging 83 [= KL ] samples).
The "gain" of the filter at DC is, from Fig. 31.46, K L = 512 . This means that
our 8-bit input of 112 (0111 0000) will be multiplied by 512 and result in an
output, prior to scaling, of 57344, or a binary code of 0 1110 0000 0000 0000 (17
bits in the general case). From Eq. (31.100) we would then divide this code by 512
(drop the lower nine bits.) However, this would result in an output word size equal
to the input word size (both 8-bits and no increase in resolution). So, for the
general input signal that is time-varying and to get the final 13 bits, we divide the
17-bit filter output by 16 (drop the lower four bits so our final output is 13 bits or
0 1110 0000 0000 [3,584 = 32 ⋅ 112 since 13-bits]). A block diagram of the filter
implementation is shown in Fig. 31.55. A MUX is needed in between each
integrator stage to adjust the two's complement word size up by log 2 K bits.
1 − z −8
17 17 17 17 17
1 − z −8 1 − z −8
Use inverter for 17 1 Out, 13-bits
subtraction in two's Drop lower 4 bits
complement Carry
in (to change to binary
8 registers offset complement
17
D Q D Q D Q the MSB)
clk clk clk
Let's discuss overflow concerns. To keep the discussion simple, let's just
consider a single integrator and comb filter stage, that is, L = 1. A constant two's
complement, 8-bit, input of 0111 0000 (+112) into the integrator will result in an
output, assuming we start with all zeroes, of
Output (Sum0) 000 0000 0000 0 (or V CM [two's complement])
Input1 000 0111 0000 112
Sum1 000 0111 0000 112
Chapter 31 Data Converter SNR 119
These sums are applied to our comb filter. Since K = 8, we won't have a
meaningful comb filter output until our ninth integrator output. At this time the
output of the comb filter will be the difference between Sum1 and Sum9, that is,
Sum9 − Sum1. The ninth integrator output is 1008 while the first is 112. The
difference being 896 (011 1000 0000 or 8 ⋅ 112 with 11 bits). In fact, we can take
any difference between sums spaced eight clock cycles apart, even after overflow,
and get this result (looking only at the lower 11 bits.) For example, Sum18 −
Sum10 is
Sum18 111 1110 0000 (2016)
minus Sum10 100 0110 0000 (1120)
(896)
In two's complement, for subtraction, we complement, and add one (set the adder
carry-in bit high) to the number we are subtracting, which gives, for this example,
Sum18 111 1110 0000
Sum10(comp) 011 1001 1111
plus 1 (adder carry)
011 1000 0000 (896)
noting that we threw out the 12th bit in the sum. As a final example,
Sum19 000 0101 0000 (80) or Sum19 000 0101 0000
minus Sum11 100 1101 0000 (1232) or Sum11(comp) 011 0010 1111
difference (-1152 or 896) plus 1
011 1000 0000 (896)
While this last discussion focused on L = 1, we could use any number of stages as
long as the register size in our integrators can accommodate a binary number of at
least K ⋅ 2 N , where N is the number of bits in the input word. T
120 CMOS Mixed-Signal Circuit Design
We might notice from this example that the amount of hardware needed to
implement the averaging filter is significant. The main contributors to the final filter layout
size are the registers used in the comb filters (a total of twenty-four 17-bit registers are
used). It turns out, as discussed earlier, that we can use the reduction in clock frequency
(decimation) to reduce the number of registers used in the comb filter, Fig. 31.56. By
dividing the clock frequency down by K, we can reduce the number of registers used in
each comb filter to one. In either Fig. 31.54 or Fig. 31.56 the delay used in the comb filter
is KTs. Figure 31.56 is the preferable way to implement decimation/averaging filters
(however, see aliasing description below for practical implementation concerns).
In 1 1 1
1 − z −1 1 − z −1 1 − z −1
fs clock Note the use of one register
L comb filters
÷K 1 − z −1 1 − z −1 1 − z −1 Out
f s /K clock
L
1 − z −K In Out
K
1 − z −1
Transfer function of Decimate and average
decimating and averaging filter (schematic symbol)
Figure 31.56 Using a reduction in clock frequency to lower complexity in averaging filters.
31.47), we can use the averaging filter for additional aliased signal removal. While we may
only be interested in signal content up to B (Eq. [31.69]), we can still have unwanted
signal content between B and f s /K that will alias into the base spectrum. It is desirable to
eliminate this problem altogether. We can do this by resampling at 2(f s /K) . The
averaging/decimation filter shown in Fig. 31.56 is changed so that the divider divides by
(a)
f s /K f s /2 fs f
(b)
f s /8 3f s /8 5f s /8 7f s /8 f
f s /4 f s /2 3f s /4 fs
Output spectrum after resampling at 2( f s /K) = f s /(K/2) (= 4B = f s /4 here)
(c)
f s /4 f s /2 3f s /4 fs f
Figure 31.58 Showing signal spectrum (a) prior to decimation and (b) and (c) after.
122 CMOS Mixed-Signal Circuit Design
K/2 . By adding a register to the comb filter the comb filters are 1 − z −2 . The resulting
spectrum is shown in Fig. 31.58c. Note that aliasing is a very important concern when
designing the averaging filter.
If the first sidelobe amplitude isn't sufficiently small, as was assumed in Fig. 31.58,
a larger resampling frequency or decimation frequency can be used to minimize aliasing. A
common intermediate clocking frequency is 4 ⋅ (f s /K) . A sample averaging filter output
waveform is shown in Fig. 31.59a, where the sidelobe amplitude is no longer insignificant.
Figure 31.59b shows the spectrum, assuming K = 8 and the resampling frequency is
f s /2 (= 4 ⋅ [ f s /8]) (meaning the divider in Fig. 31.56 is f s /[K/4]) and each comb filter
stage uses four registers [1 − z −4 ] ). Note how the third side lobe is aliased into the base
spectrum in this example, while the first sidelobe was aliased into the base spectrum in Fig.
31.58c (although it was not shown in the figure). Figure 31.60 shows that the ratio of the
main lobe to the third sidelobe is approximately 20 dB (assuming K ≥ 8 ). The possible
large amount of baseband aliasing together with the droop at B may result in the desired
input bandwidth, B, being limited to frequencies below f s /(2K) with an the external
analog AAF or an additional digital filter (more on this in a moment). Finally, notice that
at DC (or very low frequencies) in Figs. 31.58b, 31.58c, or 31.59b there is essentially no
aliased signal. This is the result of the zeroes in the averaging filter transfer function, Eq.
(31.99), falling at multiples of the decimation frequency.
(a)
f s /8 f s /2 fs f
Third side
(b)
lobe aliased
in the base
spectrum
f s /8 f s /2 fs f
Figure 31.59 Showing signal spectrum with significant side lobes (a) prior to
decimation and (b) after decimation.
H( f )
KL
Droop
≈ L ⋅ 20 dB
L ⋅ 3.9 dB
f s /K 2( f s /K) 3( f s /K) f
B = 0.5( f s /K) (7/2)( f s /K)
Figure 31.60 Decimating at four times the Nyquist rate. Showing (aliased) third sidelobe.
0.5 is simply a shift-right operation). The output of the circuit in the time domain may be
written as
y[nT s ] = x[(n − 1)T s ] + a ⋅ y[(n − 1)T s ] (31.109)
or
y[nT s ] = x[(n − 1)T s ] + a ⋅ x[(n − 2)T s ] + a 2 ⋅ x[(n − 3)T s ] + a 3 ⋅ x[(n − 4)T s ] + ... , (31.110)
which will obviously blow up if a > 1 .
X(z) Y(z)
z −1
Y(z) = [aY(z) + X(z)]z −1
y(nT s ) = a ⋅ y[(n − 1)T s ] + x[(n − 1)T s ] a
Y(z) z −1
H(z) = =
X(z) 1 − a ⋅ z −1
H(z) = z −1 a (31.111)
Figure 31.62 shows the z-plane and magnitude plots specified by this equation. If a > 1
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
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
problem unless we start to try to round numbers to decrease hardware complexity
(performing divisions or multiplications) without being careful.
124 CMOS Mixed-Signal Circuit Design
H(z) = z −1 = 1
1 − az −1 z − a
z-plane
H( f )
1
a 1−a
1
1+a
f s /2 fs 3f s /2 f
Decimating Down to 2B
In many situations (for example, we want to transmit the modulator output) it is desirable
to reduce the clocking frequency down to twice the Nyquist frequency. Since the Nyquist
frequency for our input bandwidth B is f s /(2K) , our final output clocking frequency
would be f s /K = 2B . As we have just discussed this second stage decimation can result in
significant aliasing. To eliminate this aliasing, a digital filter is often used to limit the
bandwidth, after the first stage decimation, to f s /(2K) (= B). Before we proceed any
further, let's summarize the discussion so far for the different situations.
1. The input to the ADC is bandlimited to B [= f s /(2K)] . In this situation we can
use the topology of Fig. 31.56 directly, decimating the sampling clock from fs to fs /K in
one stage. The comb filters use one register. The smallest size averaging/decimation filter
results in this situation and the output clock rate is 2B.
2. The input, to the ADC is bandlimited to 2B (= f s /K). In this situation the
averaging filter provides some aliased signal removal, see Fig. 31.47. Assuming the aliased
signal content in the first side lobe is insignificant, we can use the topology of Fig. 31.56
with a divider of (K/2) and two registers in each comb filter; that is, each comb filter has a
transfer function of 1 − z −2 . The sampling clock gets reduced from fs to fs/(K/2). The
output clocking rate is now 4B.
3. The input to the ADC is bandlimited to the Nyquist frequency fs /2 (the general
situation where the analog AAF has the most relaxed requirements). In this situation the
averaging filter will again provide some aliased signal removal. We can use the topology
of Fig. 31.56 with a divider of (K/4) and four registers in each comb filter, that is, each
comb filter has a transfer function of 1 − z −4 . The sampling clock gets reduced from fs to
fs/(K/4) and the output clock rate is 8B.
Chapter 31 Data Converter SNR 125
4. Figure 31.63 shows the entire system for case 3 above, with the addition of a
digital filter on the output of the first-stage decimation for reducing the sampling, or
output clocking frequency, to 2B. The digital filter used in the second decimation stage is
generally a half-band digital filter (covered in most books on digital filtering). Half-band
filters are used because of the simplicity of their implementation (half of the filter's
coefficient are zero) and the fact that the filter's transition frequency is symmetric around
its clocking frequency divided by four (which we can use, with a divider or using two,
cascaded, half-band filters to set the filter's cutoff frequency to B when the clocking
frequency is 4[ f s /K] = 8B ).
Figure 31.64 shows the spectrum of the signals in Fig. 31.63 for the general
situation. Figure 31.64a is the AAF input, which we have drawn with an arbitrary shape.
In 31.64b we see that the AAF limits the input signal spectrum to f s /2 (and we should
see, once again, the relaxed requirements placed on the AAF when using oversampling).
Figure 31.64c shows the ADC's output spectrum resulting from sampling the input
waveform (actually the output of the AAF). After first stage decimation, Fig. 31.64d, the
signal is passed through a sinc (averaging) filter and then resampled at 4(f s /K) (= f s /2
when K = 8). Figure 31.64e shows the output of the half-band filter prior to
down-sampling (second stage decimation). The figure assumes the half-band filter is
clocked with an effective clock frequency of 4B (the actual clock frequency is 8B as
discussed above) or 4[ f s /(K/4)] . So the filter's cutoff frequency is B. The half-band filter's
implementation may also use the high frequency clock signal to simplify the filter's
implementation. Finally, Fig. 31.64f shows the spectrum resulting after final decimation.
The clock frequency out of the final stage is 2B, while the desired signal bandwidth is B.
It's important to remember that unless there is some reason to lower the clock
frequency (for example, we want to store the ADC/averaging filter's digital output in
memory), we can avoid the aliasing problems associated with decimation (see Fig. 31.58)
and the added complexity. Also note that the desired spectrum of Fig. 31.64e must
ultimately be reconstructed using a DAC and a reconstruction filter (RCF). Because of the
unwanted spectral content, directly adjacent to the desired content, the reconstruction
becomes more challenging when decimating down to 2B.
126 CMOS Mixed-Signal Circuit Design
f s /2 fs Assuming K = 8
B 8B = f s /2
Input signal spectrum
(a)
f
f
ADC output
(c)
f
After first stage decimation
(d)
f
4B
After passing through half-band filter
(e)
f
After second stage decimation down to 2B
(f)
B f
Desired signal spectrum
Figure 31.64 Spectrums of the resulting signals for the decimation scheme shown in Fig. 31.63.
fs Clock in
Interpolation Filter
Figure 31.65 Block diagram of a DAC that uses interpolation to increase effective DAC
resolution.
Chapter 31 Data Converter SNR 127
Also, as in the last section, Bennett's criteria must be valid. In particular, the digital
word must be busy and the DAC must be linear to the final desired resolution (that is, the
N-bit DAC must be linear to N + NInc).
The Dump and Interpolate
Figure 31.66 shows the basic idea of introducing digital words in between the words
coming into the interpolation circuit of Fig. 31.65. The inputs to the interpolating filter are
indicated by the thicker lines in the figure. The interpolator introduces additional samples
in between these inputs. If the frequency of the input samples is 2B then the frequency of
the samples coming out of the interpolator is
f s = K ⋅ 2B (31.112)
noting that, since we are using the same notation as used in the last section, the rate of
words being clocked into the DAC is the same ( f s ) as the rate at which the ADC was
clocked in the last section.
time
Interpolator inputs
Figure 31.66 How the interpolation circuit increases the sample rate while
introducing samples in between the existing samples.
If the inputs to the interpolator are x[Ki ⋅ T s ] and the outputs of the interpolator
are y[nT s ] , we can write
K⋅i−1
x[Ki ⋅ T s ] − x[K(i − 1) ⋅ T s ]
y[n ⋅ T s ] = x[K(i − 1) ⋅ T s ] + Σ
n=K(i−1)
[n − K(i − 1)] ⋅
K
(31.113)
to describe the operation of the interpolator. Rewriting this equation to show only the
change between adjacent outputs results in
x[Ki ⋅ T s ] − x[K(i − 1) ⋅ T s ]
y[n ⋅ T s ] − y[(n − 1) ⋅ T s ] = (31.114)
K
Taking the z-transform of this equation results in
−K
Y(z)(1 − z −1 ) = X(z) ⋅ 1 − z (31.115)
K
or
Y(z) 1 1 − z −K
H(z) = = (31.116)
X(z) K 1 − z −1
128 CMOS Mixed-Signal Circuit Design
which is the familiar transfer function for our averager presented in the last section. The
implementation of our interpolator, termed a Dump and Interpolate, is shown in Fig.
31.67. The input words are dumped into latches which serve two purposes: (1) to store
two consecutive, slow input words for generation of the incremental change in the fast
output samples, and (2) to pass the interpolator input words directly to the output,
through the multiplexer (MUX), every K clock cycles. The ÷ K is implemented simply by
removing the lower bits of the adders output word. As we saw with the
accumulate-and-dump circuit, this implementation has practical problems that result in the
need to use other implementations (which we'll discuss next).
÷K Latches
Input, x[Ki ⋅ T s ]
Latches Latches
MUX Latches
Clock, f s /K
divide by K Output, y[nT s ]
Clock, f s
Figure 31.67 A dump-and-interpolate circuit used for interpolation and reverse averaging.
Figure 31.68 Implementation of a single stage interpolation filter. See Fig. 31.55
for handling the word size increase in the integrators.
2. The input to the interpolator/DAC has the spectrum shown in Fig. 31.58c and is
clocked at 4B. The interpolating filter has the same topology shown in Fig. 31.68 except
that two registers are used in the comb filter and the divider is changed to K/2. Again, the
DAC is clocked at fs. The RCF together with the sinc filter used in the interpolator limits
the output spectral content.
3. The input to the interpolator/DAC has the spectrum shown in Fig. 31.59b and is
clocked at 8B. The interpolating filter has the same topology as the one shown in Fig.
31.68 except that four registers are used in the comb filter (see Fig. 31.63) and the divider
is changed to K/4. Again, the DAC is clocked at fs. The RCF together with the sinc filter
used in the interpolator limit the output spectral content.
4. Our input spectrum is shown in Fig. 31.64f and clocked into the interpolation
filter at 2B (the general situation that results in the most relaxed requirements on the
RCF). The basic, general interpolation structure is shown in Fig. 31.69. Figure 31.70
shows the spectrums at various points in this circuit. The input spectrum to the
interpolation circuit is shown in Fig. 31.70a. This input is connected to a set of latches
clocked at 2B. The output of these latches is connected to the digital filter, which clocks
the values in at 8B and has a response, as seen in Fig. 31.70b. We need to understand
what's happening at this point. If we look at the output register used in Fig. 31.63, we see
Analog
f s /K = 2B 4( f s /K) = 8B
÷4 ÷ K/4
Clock, f s
Figure 31.69 General interpolation and reverse averaging topology for an oversampled DAC.
130 CMOS Mixed-Signal Circuit Design
f s /2 = 8B here fs Assuming K = 8
Input Nyquist rate
(a)
B f
Input holding register sinc response
(b)
B 4B f
Nyquist rate after resampling
After passing through half-band filter
(c)
B 4B f
Sinc response After passing through second-stage interpolator filter
(d)
f
Desired signal spectrum. Nyquist rate after resampling
(e)
3B f
RCF response may leave some unwanted spectral content
RCF reponse eliminates all unwanted spectral content
Figure 31.70 Spectrums of the resulting signals for the interpolation scheme shown in Fig. 31.69.
that we were only saving one out of every four samples coming out of the digital filter. In
Fig. 31.69 we are "estimating" these samples by simply clocking each input value four
times into the digital filter. Figure 31.71 shows the situation in more detail. It is desirable
to determine how this input holding register affects the spectrum of the input signal. We
can relate the input of the register (a set of latches) to the register's output using
4⋅i−1
x[4(i − 1) ⋅ T s ]
y[nT s ] = Σ
n=4(i−1) 4
(31.117)
t nT s
Figure 31.71 Showing effects of digital filter input holding register on the input data.
1 − z −8
= 7z −
8 1
8 1+z −1 z (z + 1)
7
f s /2 DC
f s /4 f s /2 fs f
3f s /8
(31.120)
In 1 Out
1 − z −K
1 − 2 cos 2π f s ⋅ z −1 + z −2
f
1 − z −8 = z 8 − 1 f s /4
4 1 + z −2 z 6 (z 2 + 1)
6
DC
f s /8 3f s /8 fs f
f s /4 f s /2
Figure 31.74 A bandpass filter implementation using a comb filter and digital resonator.
We can determine the magnitude response of Eq. (31.122) following the same
procedure used to determine Eq. (31.93). The result, for the f s /4 resonator, is
At the center of the passband, that is f s /4 , H( f ) = K/2. The ratio of the main lobe to the
first side lobe, on either side, is plotted in Fig. 31.75 along with the lowpass sinc filter
response and is calculated using
Main lobe = K ⋅ sin 3π (31.124)
First side lobe 2 K
The cosine term in Eq. (31.121) can be set to ±1 when f = f s /6 or f s /3 resulting in
a bandpass filter that is easy to implement. 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 combinations of simple bandpass filters can be implemented
using these techniques.
134 CMOS Mixed-Signal Circuit Design
11
Main lobe
First sidelobe
9 Bandpass response when f = f s /6 or f s /3
7
Bandpass attenuation when f = f s /4
5
3
3 4 6 8 10 12 16 K
Figure 31.75 Lowpass and bandpass filter attenuation versus number of comb filer zeroes K.
The ratio of the main lobe to the first sidelobe for f = f s /6 or f s /3 is given,
assuming K = 12, 24, ... , by
K sin 2K
3π π 3π
sin 3 − 2K
Main lobe = = 1.15K sin 3π sin π − 3π (31.125)
First side lobe sin π3 2K 3 2K
which is approximately 13.5 dB for K = 24, 36, 48 ... and 10.15 dB for K = 12, see Fig.
31.75.
To increase the amount of attenuation between the main lobe and the first side lobe
in a bandpass filter implementation, we can cascade filter sections (as we did in the
lowpass filter implementations discussed earlier). For example, cascading five f s /4
bandpass filters with K = 8 will result in an attenuation of 57 dB. Also, note that by
changing the sampling, or filter clock frequency fs , 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 31.23
Sketch the block level circuit diagram for an f s /4 digital resonator.
From Eq. (31.121) the time domain representation of the f s /4 resonator can be
written as
y[nT s ] = x[nT s ] − y[(n − 2)T s ]
The implementation is shown in Fig. 31.76. T
Chapter 31 Data Converter SNR 135
x[nT s ] y[nT s ]
Q D Q D H(z) = 1
clk clk 1 + z −2
fs
f s /4 resonator
Digital resonators
f1
−1
f2
f1
Comb filter f3
In
1 − z −K Out
−1
f X−1
fX
Sinc responses
Desired filter response
f2 f4 f6
Figure 31.77 A frequency sampling filter.
136 CMOS Mixed-Signal Circuit Design
CF
φ1 φ2
φ1
V CM
CI φ2 v out
v1
v2
V REF+ + V REF−
Ts V CM =
2 Bottom
φ1 plate
(the plate closest
φ2 to the substrate)
t
n−1 n
n − 1/2
To begin, let's assume the output of the DAI is connected to the op-amp through
the φ 1 switch. When the φ 1 switches are closed ( φ 1 is high) at n − 1 (the instance when
the switches shut off), the charge stored on CI is
Q 1 = C I (V CM − v 1 [(n − 1)T s ]) (31.126)
and the output of the integrator is v out [(n − 1)T s ] . When the φ 2 switches turn on the
charge stored on CI becomes
Q 2 = C I (V CM − v 2 [(n − 1/2)T s ]) (31.127)
remembering that the op-amps holds its noninverting input terminal at VCM. The difference
in these charges, Q 2 − Q 1 , is transferred to the op-amp's feedback capacitor resulting in an
output voltage change. This change can be written as
(v out [nT s ] − v out [(n − 1)T s ])C F = C I (v 1 [(n − 1)T s ] − v 2 [(n − 1/2)]T s ) (31.128)
or writing this equation in the z-domain results in
CI
V out (z)(1 − z −1 ) = (V 1 (z) ⋅ z −1 − V 2 (z) ⋅ z −1/2 ) (31.129)
CF
The transfer function of the DAI with the output connected to the φ 1 switches is then
C I V 1 (z) ⋅ z −1 − V 2 (z) ⋅ z −1/2
V out (z) = ⋅ (31.130)
CF 1 − z −1
Similarly, if we connect the output through the φ 2 switches (the edges we label n in Fig.
31.78 shift in time by T s /2 ) we can write
138 CMOS Mixed-Signal Circuit Design
Example 31.24
Determine the transfer function of the DAI of Fig. 31.78 without the switches on
the output of the op-amp.
Reviewing Fig. 31.78 we see that charge is transferred to the feedback capacitor
only when the φ 2 switches are closed. Therefore, the output only changes states
during the time interval when the φ 2 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.
(31.134). Using the φ 1 switches simply adds a half clock cycle delay, z −1/2 , to the
integrator's transfer function (instead of the output changing with the rising edge
of φ 2 , the output changes one-half cycle later on the rising edge of φ 1 ). T
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. 31.79. The summation, gain, and
integrating blocks are implemented with a single switched-capacitor DAI having the
transfer function given by Eq. (31.134). The gain, G, of the DAI is set by the ratio of
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
ADC ) and DACs (V2[z] is connected to the output of a DAC).
It should be clear from both Fig. 31.79 and Table 31.2 that many different
combinations of discrete analog building blocks are possible. Figure 31.80 shows two
Chapter 31 Data Converter SNR 139
CI
G V 2 (z) G=
CF
Discrete analog integrator
other possibilities. In part (a) the capacitors used have the same parasitic capacitance on
each plate (see the lateral capacitor in Fig. 33.11 for example), so there is no benefit to
using a bottom-plate insensitive topology (Fig. 31.78). The transfer function of this DAI is
CI −1
V out (z) = ⋅ z ⋅ (V 2 (z) − V 1 (z)) (31.136)
C F 1 − z −1
noting each input signal sees the same delay, i.e., z−1 when the outputs are connected
through φ1 controlled switches and z−1/2 delay when no switches or φ2 controlled switches
are used. If the integrator inputs must see the same delay and the capacitors available have
asymmetric parasitic capacitance, the topology of Fig. 31.80b can be used. Its transfer
function is
φ1 φ2 CF
(a)
V 1 (z) φ1
V CM V out (z) =
CI
V 2 (z) CI z −1
⋅ ⋅ (V 2 (z) − V 1 (z))
C F 1 − z −1
CF
(b)
V CM C I1 φ1
V CM V out (z) =
V 1 (z)
z −1 ⋅ I1 ⋅ V 1 (z) + I2 ⋅ V 2 (z)
C C
1 − z −1 C F CF
C I2
V 2 (z)
31.3.2 Modulators
The basic topology of a feedback modulator or coder is shown in Fig. 31.81. Depending
on the circuit blocks used for A(z) and B(z) feedback modulators can be separated into
two categories: predictive modulators and noise-shaping modulators [10].
DAC B(z)
A(z) 1
Y(z) = ⋅X(z) + ⋅E(z) (31.138)
1 + A(z) ⋅ B(z) 1 + A(z)B(z)
In a predictive modulator the feedback filter, B(z), has a large gain so that, ideally, the fed
back signal equals the input signal. If A(z) = 1 (a wire), then both the STF (signal transfer
function) and the NTF (noise transfer function) have a value of, approximately, 1/B(z).
Recovering the input signal requires passing the output of the predictive modulator
through an analog filter with a transfer function of precisely B(z) (noting that B[z] is a
digital filter in the modulator of Fig. 31.82). The required precision of the analog filter (the
matching between the filter in the modulator and the filter in the demodulator) limits the
attainable resolution when using predictive modulators. Notice that both the input signal
and the quantization noise experience the same spectral shaping (spectral 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 output of the
summer to zero. If the prediction is perfect, the signal that is fed back exactly matches the
input signal.
z-domain representation
E(z) of the quantization error
ADC
In X(z) Y(z) Out
A(z)
DAC B(z)
In a noise-shaping modulator the gain of the forward path, A(z), is large in the
signal bandwidth so that the STF is approximately unity (assuming B[z] = 1). The NTF, on
the other hand, will approach zero, ideally, in the bandwidth of interest. Note that the
signal spectrum passes through the modulator essentially unchanged, while the
quantization noise spectrum is shaped (and thus the name noise-shaping). No precision
filter or analog components are required, as discussed earlier, except, perhaps, for the
DAC in the feedback path of the modulator. We'll see in the next chapter that if A(z) is an
integrator, the quantization noise is pushed to higher frequencies so that it can be removed
with the averaging filter. This is a very important concept, as a noise-shaping modulator
does not reduce the quantization noise to attain higher resolutions, but rather pushes the
noise to frequencies outside of the signal bandwidth of interest.
142 CMOS Mixed-Signal Circuit Design
REFERENCES
[1] R. J. Baker, H. W. Li, and D. E. Boyce, CMOS: Circuit Design, Layout, and
Simulation, Wiley-IEEE, 1998. ISBN 0-7803-3416-7
[2] L. W. Couch, Modern Communication Systems: Principles and Applications,
Prentice-Hall, 1995. ISBN 0-02325286-3
[3] S. Haykin, An Introduction to Analog and Digital Communications, John Wiley
and Sons, 1989. ISBN 0-471-85978-8
[4] P. A. Lynn and W. Fuerst, Introductory Digital Signal Processing, Second
Edition, John Wiley and Sons, 1998. ISBN 0-471-97631-8
[5] E. P. Cunningham, Digital Filtering: An Introduction, John Wiley and Sons, 1995.
ISBN 0-471-12475-3
[6] R. K. Hester, Introduction to Oversampled Data Conversion, Notes from a
tutorial at the 1995 International Solid-State Circuits Conference (ISSCC-95).
[7] W. R. Bennett, "Spectra of Quantized Signals," Bell System Technical Journal,
Vol. 27, pp. 446-472, July 1948.
[8] J. C. Candy and G. C. Temes (eds.), Oversampling Delta-Sigma Data Converters,
IEEE Press, 1992. ISBN 0-87942-285-8
[9] S. R. Norsworthy, R. Schreier, and G. C. Temes (eds.), Delta-Sigma Data
Converters: Theory, Design, and Simulation, IEEE Press, 1996. ISBN
0-7803-1045-4
[10] S. K. Tewksbury and R. W. Hallock, Oversampled, Linear Predictive and
Noise-Shaping Coders of Order N>1, IEEE Trans. Circuits and Sys., Vol.
CAS-25, pp. 436-447, July 1978.
LIST OF SYMBOLS/ACRONYMS
ACF - Autocorrelation Function, see Eq. (31.30)
ADC - Analog-to-Digital Converter
B - Bandwidth of the input signal, see Eq. (31.69)
DAC - Digital-to-Analog Converter
dBc - Decibels with respect to the carrier, see Eq. 31.9
DR - Dynamic Range, see Eq. (31.10)
∆T s - Peak-to-peak amount of clock jitter
∆V s - Uncertainty in the sampled voltage
fclk - Clock frequency. Also, sometimes called sampling frequency, fs
fin - Input sinewave frequency
Chapter 31 Data Converter SNR 143
fn - Nyquist frequency, which is fs /2. Sometimes also called the folding frequency.
fres - Resolution of a DFT
fs - Sampling frequency. Also, sometimes called clock frequency, fclk , or Nyquist rate.
H( f ) - Transfer function of a system
H(z) - Z-Domain representation of H( f )
K - Oversampling factor or number of points averaged. See Eq. (31.22) or Eq. (31.51)
L - Order of sinc averaging filter
LSB - Least Significant Bit, see Eq. (31.2)
M - Order of a noise-shaping modulator (see Ch. 32.)
N - Ideal data converter resolution (number of bits)
Neff - Effective number of bits, see Eq. (31.5)
NFinal - Final data converter resolution after averaging, see Eq. (31.74)
NInc - Increase in data converter resolution, see Eq. (31.53)
NLoss - Number of bits lost because of sampling jitter. see Eq. (31.17)
PDF - Probability Density Function, see also ρ(t)
PAVG - Total average power in a waveform
Pin( f ) - Power Spectral Density (PSD), see Eq. (31.35)
Pjitter( f ) - PSD of the sampling error voltage due to jitter, see Eq. (31.49)
Posc( f ) - PSD of the output of an oscillator, see Eq. (31.50)
PSD - Power Spectral Density
PQe( f ) - Quantization noise power spectral density
ppm - Parts per million, a multiplier of 10 −6
L - Number of sections used in an averaging filter
Rin(t) - Autocorrelation function, see Eq. (31.30)
ρ(t) - Probability Density Function, PDF
σ - Standard deviation or square root of the variance
σ 2 - Variance of a PDF, see Eq. (31.44)
Rin(t) - Autocorrelation function, see Eq. (31.30)
RMS - Root Mean Square
SFDR - Spurious Free Dynamic Range, see Eq. (31.9)
144 CMOS Mixed-Signal Circuit Design
31.7 Suppose a perfectly stable clock is available (∆Ts is zero in Eq. [31.12]). Would we
still have a finite aperture window if the clock has a finite rise time? Describe why
or why not?
31.8 How do the number of bits lost because of aperture jitter change with the
frequency of an ADC input sinewave? If the ADC input is a DC signal is aperture
jitter a concern? Why?
31.9 Show the time domain signal that generates the spectrum shown in Fig. 31.10.
Verify in the time domain that the signal's rising and falling edges do indeed vary
from their ideal positions.
31.10 Describe in your own words the problems with simulating clock jitter using
SPICE.
31.11 What does the autocorrelation function (ACF) tell us about a signal? What is the
ACF of a 1 V DC signal. Show the simple calculations leading up to your answer.
31.12 Plot the power spectral density of a sinewave. From this plot show how to
determine the average and RMS values of the sinewave. Show the procedure for
both one-sided and two-sided spectrums.
31.13 Sometimes the average power specified by Eq. (31.37) is termed total average
normalized power of a signal. Why?
31.14 When WinSPICE generates a plot from a DFT the units on the y-axis are volts
peak (the peak value of a sinewave at a given frequency). How do we change this
plot into RMS voltages, voltage spectral density, and power spectral density vs.
frequency?
31.15 Repeat Ex. 31.12 if the sinewaves are first sampled.
31.16 Suppose the jitter in a clock signal can be characterized using the PDF shown in
Fig. 31.12. Further if ∆Ts = 100 ps estimate the RMS value of clock jitter, standard
deviation, and variance of the jitter.
31.17 Suppose that a noise voltage has the PDF shown in Fig. 31.12. If the maximum
voltage deviation from the ideal value is 10 µV estimate the RMS value of the
noise (the standard deviation) and the noise power (the variance).
31.18 Repeat question 31.17 if the noise voltage has a Gaussian PDF as seen in Fig.
31.13.
31.19 Repeat Ex. 31.1 if we want to include an error from sampling jitter, PAVG,jitter of 1
µW.
31.20 If a DC signal is input to a data conversion system, is Eq. (31.51) valid? Name
three conditions on the input signal in order for this equation to be valid.
146 CMOS Mixed-Signal Circuit Design
31.21 Suppose the standard deviation of the quantization noise in a data conversion
system is 1 mV. Using Eq. (31.56) plot the PSD of the quantization noise.
Comment on the assumption that the noise power is limited to the Nyquist
frequency. Does this result in an over- or underestimate for the actual noise power
in the spectrum of interest?
31.22 Show why averaging two 8-bit words, as seen in Fig. 31.19, must result in a 9-bit
word. (Why isn't the sum of the two words divided by two [the average] another
8-bit word?)
31.23 Why must Bennett's criteria be valid for the averaging to reduce the quantization
noise in Fig. 31.19? Give an example where averaging will not reduce quantization
noise.
31.24 Show a figure similar to Fig. 31.20 where an input sinewave with a frequency of
fs/4 is sampled at fs. Show that the magnitude of the resulting fundamental
(indicating that the sinewave lies in the frequency range of DC to fn) sinewave is
2 , as indicated in Fig. 31.22.
31.25 Assuming Eq. (31.68) is valid rederive Eq. (31.4) including the effects of
averaging K ADC output samples. Is Eq. (31.4) or the equation derived here valid
for a slow or DC input signal? Comment on why or why not.
31.26 Assuming Bennett's criteria are valid, does averaging ADC outputs (or DAC
inputs) put any restrictions on the bandwidth of the input signal? Why? Give an
example.
31.27 Comment on the statement "The factor of 2 in the magnitude response of Fig.
31.22 at low-frequencies simply indicates that the digital word length increases by
one bit."
31.28 What is the magnitude response of z −2 + z −3 .
31.29 Repeat Ex. 31.15 if 16 ADC outputs are averaged, that is, K = 16.
31.30 How accurate does an 8-bit ADC have to be in order to use a digital filter to
average 16 output samples for a final output resolution of 10-bits (see Eq.
[31.53]). Assume the ideal LSB of the 8-bit converter is 10 mV. Your answer
should be given in both mV and % of the full-scale.
31.31 If a DC signal is applied to a data converter can a digital averaging filter be used to
increase the system's resolution? What about if a dither signal is added to the DC
input? Use simple time domain drawings to illustrate your answers.
31.32 Name three characteristics of all digital filters.
31.33 Plot Eq. (31.59) on a z-plane. Using this plot show how to graphically determine
the magnitude and phase responses shown in Fig. 31.22.
31.34 The magnitude response shown in Fig. 31.34 becomes infinite as the input signal
approaches DC. Since the filter is digital, what is the maximum output of the filter?
Chapter 31 Data Converter SNR 147
31.35 Show that the peak (+127) and valley (−128) amplitudes of the two's complement
signals in Fig. 31.37 sum to −1.
31.36 Summarize the method of changing a number from binary offset to two's
complement. Demonstrate addition and subtraction using two's complement
numbers. Show how, in two's complement, 8, 33, and 111 sum to 152. Assume a
10-bit word size.
31.37 Suppose a digital filter sums 16 inputs and then outputs the total. If the filter is
clocked at 100 MHz, plot the magnitude response of the filter.
31.38 Comment on the benefits and drawbacks of using an averaging filter with and
without decimation.
31.39 Verify the z-domain function specified by Eq. (31.100) has a frequency response
given by Eq. (31.101). How are the typical input and output signals in the time
domain related for this filter?
3
31.40 What is the magnitude response of (1 − z −1 ) . Sketch a block diagram
implementation for this filter.
31.41 Resketch Fig. 31.53 if, in each transfer function H(z), a pole is added at DC.
31.42 Show the problem with not using a MUX at the input of the adders in Fig. 31.55.
31.43 Is it possible for the accumulate-and-dump circuit to output a spectrum with
aliasing if the input signal is bandlimited to fs? Why or why not?
31.44 In the discussions in this chapter we assumed the digital signals are much larger
than an LSB of a data converter. What happens if this is not the situation for the
sinc averaging filter?
31.45 Is it possible to decimate a digital waveform down to 2B and then later, with some
other hardware or software digital filter, remove all of the aliased signals from the
desired signal?
31.46 Suppose the waveform shown in Fig. 31.66 is the input to a decimator. If K = 8,
what would the output of the decimator look like? Use integers to illustrate your
understanding.
31.47 Suppose a digital word is clocked into a hold register and held there for eight
clock cycles before another word is clocked into the hold register. Is this similar to
the analog sample-and-hold? If the sampling rate (clock frequency) is increased by
a factor of 8 after the hold register what kind of digital filter can we think of the
hold register as being?
31.48 Show that the digital resonator of Fig. 31.76 can be modified if we add a multiplier
to the circuit, so that Eq. (31.120) can be implemented.
148 CMOS Mixed-Signal Circuit Design
31.49 It is more correct to write our DAI continuous time input signals in Fig. 31.78 as
v 1 (t) + V CM and v 2 (t) + V CM
Knowing this rederive Eq. (31.130).
31.50 Repeat question 31.49 for Eq. (31.134).
31.51 Using the results from question 31.49, derive the transfer function, Eq. (32.139),
for the circuit shown in Fig. 32.92 (in the next chapter).
31.52 Show the detailed derivation of Eq. (31.138).
31.53 Summarize the advantages and disadvantages of predictive and noise-shaping data
converters.
Chapter
32
Noise-Shaping Data Converters
In this chapter we discuss the design of noise-shaping (NS) data converters. Our approach
will be to develop NS theory along with SPICE behavioral models to illustrate, using
simulations, the operation of NS ADCs and DACs. Our goals are to discuss the
fundamentals of NS data converter design and to put a framework together for SPICE
simulations. Having a simulation framework available will allow us to (1) perform a
behavioral simulation using nearly ideal components to determine fundamental
performance limitations of a particular NS converter topology, and (2) replace behavioral
models with actual, MOSFET-based circuits in steps to design and simulate the operation
of a NS data converter in stages. While we can replace the behavioral models with
MOSFET-based circuits, we will delay this discussion (MOSFET-based circuit design in a
submicron process) until the next chapter. This chapter will focus on NS theory and
examples, using simulations, to illustrate the use of the theory.
In the material that follows we attempt to develop models that are robust, including
fundamental limitations of the circuits used, while at the same time attempting to generate
simple models for fast simulations.
Nonoverlapping Clock Generation and Switches
In this chapter, as we did in Chs. 30 and 31, we assume VDD = 1.5 V (the positive power
supply voltage), VSS = 0 V (the negative power supply voltage), VREF+ = 1.5 V (the
positive data converter reference voltage), V REF− = 0 V (the negative data converter
reference voltage), and f s = f clk = 100 MHz (the sampling, or clock frequency, of the data
converter). The SPICE pulse statements used to generate two 100 MHz nonoverlapping
clocks can be written as
Vphi1 phi1 0 DC 0 Pulse 0 1.5 0 200p 200p 4n 10n
Vphi2 phi2 0 DC 0 Pulse 0 1.5 5n 200p 200p 4n 10n
R2 phi1 0 1MEG
R3 phi2 0 1MEG
where the resistors ensure that the clocks are not floating (not the only elements
connected to the nodes phi1 and phi2 as the clocks may be used exclusively to control
switches in a simulation). The statements used to set up the power supply voltages,
reference voltages (if used), common-mode voltage, and switch trip points can be written
as
VDD VDD 0 DC 1.5
Vtrip Vtrip 0 DC 0.75
VCM VCM 0 DC 0.75
VREFP VREFP 0 DC 1.5
VREFMVREFM0 DC 0
The trip voltage is used in simulating the operation of the switches to indicate when the
switch should be opened or closed. Figure 32.1a shows the use of the basic switch in
SPICE. When phi1 (φ 1 ) is above the trip voltage (0.75 V here), the S1 switch is closed.
When the node phi2 is above the trip voltage, S2 is closed. The SPICE statements
specifying the operation of these switches, in the manner described, are
S1 1 2 phi1 Vtrip switmod
S2 2 3 phi2 Vtrip switmod
.model switmod SW RON=1k
The parameter RON can be used to model the switches' on resistance as shown in Fig.
32.1b. This may be useful when simulating finite settling time effects in a data converter.
φ1 φ2 Node numbers
RON
1 2 3
S1 S2
(a) (b)
Op-Amp Modeling
Behavioral modeling of op-amps could take up an entire chapter by itself. Here we
introduce a trivial model that is easily modified to account for real op-amp imperfections.
Figure 32.2 shows the basic op-amp symbol and a voltage-controlled-voltage-source used
to simulate the operation of an op-amp. The SPICE statement that specifies the op-amp is
Ein 3 0 2 1 100MEG
1 3
3 1
2 E1
2
Example 32.1
Determine and simulate the gain of the circuit shown in Fig. 32.3.
This circuit is our discrete analog integrator (DAI) shown in Fig. 31.78 with the v2
input connected to VCM. The transfer function of this circuit in the z-domain is
H(z) = z −1 (32.1)
1 − z −1
From Fig. 31.34 and Eq. (31.84) the magnitude of Eq. (32.1), in the frequency
domain, is (noting the z −1 in the numerator of Eq. [32.1] is a delay of Ts that adds
to the phase of the integrator but doesn't affect the magnitude response)
C F = 1p
φ1 φ2
φ1
V CM
V out
1p
V in C I = 1p
V CM = 0.75 V f clk = f s = 100 MHz
H( f ) = 1 = 3.2 (32.2)
2 1 − cos 2π 100
5
*Input Signal
Vin Vin 0 DC 0 Sin 0.75 50m 5MEG
*Clock Signals
Vphi1 phi1 0 DC 0 Pulse 0 1.5 0 200p 200p 4n 10n
Vphi2 phi2 0 DC 0 Pulse 0 1.5 5n 200p 200p 4n 10n
R2 phi1 0 1MEG
R3 phi2 0 1MEG
.end
T
Chapter 32 Noise-Shaping Data Converters 153
Output
Input
VDD
S1
Vinm
Inputs
A(z) 1
Y(z) = ⋅X(z) + ⋅E(z) (32.3)
1 + A(z) 1 + A(z)
where STF( f ) is the signal's transfer function and NTF( f ) is the noise's transfer function.
E(z)
Delta Sigma ADC
In X(z)
A(z) = z −1 Y(z) Out
1 − z −1
Digital
Integrator
Analog
DAC
φ1 φ2 φ1
1p
V CM
V CM V out
1p V CM
Clocked comparator
V in
f s = 100 MHz
digital word representing the analog input voltage. For a detailed discussion of the
requirements placed on the anti-aliasing filter (AAF) and the digital filter (or
digital-decimation filter if decimation is used), see Ch. 31.
1-bit
Analog input Digital
AAF NS modulator Digital Output
Filter
fs
Clock input Decimation filter
We might wonder if we can use an analog filter, instead of a digital filter in the
topology of Fig. 32.9, to remove the high-frequency quantization noise. The output of the
resulting circuit will be analog, so it can't be used as an ADC. While this may not be of
practical use at the moment, it does help in understanding how the NS modulator
functions. Passing the modulator output of Fig. 32.8 through a simple RC lowpass filter,
with a time constant of 100 ns, results in the waveform shown in Fig. 32.10. Increasing the
time constant results in a smoother output signal. However, increasing the time constant
too much can affect the amplitude of the desired signal. Also, note the phase shift through
the modulator and filter.
Chapter 32 Noise-Shaping Data Converters 157
Output after RC
filtering
Input
Figure 32.10 Using a simple RC lowpass filter on the output of the NS modulator of Fig. 32.7.
1-bit
Digital input Digital Analog output
NS demodulator RCF
Filter
fs
Clock input
Interpolation filter
E(z)
Quantizer
In X(z) z −1 Out
1 − z −1
Digital
Accumulator
Digital
(a)
V CM = 00000.... = 0.75 V (here)
Accumulator Quantizer
In MSB Out
D Q
clk
011111...
1
fs 100000... 0 MUX
(b)
Figure 32.12 Block diagram of (a) a NS demodulator and, (b) a more detailed
implementation for use in a DAC.
where N is the number of bits used in the low-resolution ADC/DAC in the modulator.
Using a single-bit ADC/DAC in a NS modulator, N = 1, results in V LSB = 1.5 V (see Prob.
30.14 for a discussion of when Eq. [32.9] isn't valid). This again shows that we are not
reducing the quantization noise, but are rather pushing it to higher frequencies so that it
can be filtered out. Using Eq. (31.107), we can write the PSD of the NTF (the PSD of the
first-order modulator's modulation noise) as
V 2LSB f units, V 2 /Hz
NTF( f ) 2 ⋅ V Qe ( f ) 2
= ⋅ 2 1 − cos 2π (32.10)
12f s fs
Figure 32.13 shows the PSD of the first-order NS modulation noise for V LSB = 1.5 V and
f s = 100 MHz . Note how now we are discussing modulation noise instead of quantization
noise. The modulation noise is the quantization noise after being differentiated by the NS
modulator. The modulation noise is the unwanted signal added to the input signal. After
reviewing Fig. 32.13 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. Figure 32.14 shows the PSD of the noise if we limit our view to 1 MHz. The point
here is that by restricting the bandwidth of the modulation noise we can, theoretically,
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 by
utilizing a 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.
V 2 /Hz, × 10 −9
f, Hz, × 10 6
f n = f s /2 = 50 MHz
Figure 32.13 Modulation noise for a first-order NS modulator.
160 CMOS Mixed-Signal Circuit Design
V 2 /Hz, × 10 −12
f, Hz, × 10 6
Example 32.2
Using SPICE, show the modulation noise spectrum associated with the NS
modulator of Fig. 32.7. Compare the simulation results to the theoretical results
shown in Fig. 32.13.
Following the procedure given back in Sec. 30.3.1, Fig. 30.46, to determine a data
converter's quantization noise spectrum, we apply a slowly moving voltage ramp
to the input of the modulator. Then we look at the difference between the input
and output of the modulator (the modulation noise). The simulation results are
shown in Fig. 32.15a and 32.15b. We used the following WinSPICE commands to
generate these plots (added directly into the netlist)
*#plot Vout Vin
*#let Vqev=Vout-Vin
*#linearize Vqev
*#spec 0 100MEG 200k Vqev
*#let Vqedb=db(Vqev)
*#plot Vqedb
The first command is used to generate Fig. 32.15a. The second command is used
to generate the difference between the modulator's input and output (the
modulation noise). Notice how, in Fig. 31.15a, the output of the modulator stays
low most of the time, when the input to the modulator is close to ground, while the
output stays high most of the time when the input is close to VDD. The third and
Chapter 32 Noise-Shaping Data Converters 161
Modulator output
Input ramp
(a)
Volts
Voltage, peak
(b)
Figure 32.15 (a) Input and output of the NS modulator of Fig. 32.7, and (b)
modulation noise output spectrum.
162 CMOS Mixed-Signal Circuit Design
fourth commands in the above list generate the spectrum of the modulation noise
(the spectrally shaped quantization noise). The last commands are used to plot,
Fig. 32.15b, the spectrum of the modulation noise (units of Volts). It may be
helpful, at this point, to review Fig. 30.48 and the associated discussion. To
change this plot (Fig. 31.15b) into a power spectral density (units of V 2 /Hz ) we
can square the magnitude of the modulation noise and then divide the result by the
resolution of the Fourier transform ( f res = 200 kHz in Fig. 32.15). The list of
commands used to generate a power spectral density from Fig. 32.15b would be
*#let mrms=mag(Vqev)/1.414
*#let Vqepsd=10*log10(mrms*mrms/200k)
*#plot Vqepsd
Using this sequence of commands results in a spectrum with amplitude values that
are similar to the values given in Fig. 32.13 (and have the same units). However,
the shape would remain essentially unchanged from Fig. 32.15b.
Notice that the shape of the modulation noise shown in Fig. 32.15b (in
decibels) matches fairly well with the spectrum shown in Fig. 32.13. This is the
case even though the quantization noise spectral density, E( f ), is not flat (is not
white), as was assumed in Eq. (32.7). The important thing to notice in Fig. 31.15b
is that the modulation noise spectrum decreases with decreasing frequency (at 20
MHz and below), as was predicted using Eq. (32.10) and shown in Fig. 32.13. T
RMS Quantization Noise in a First-Order Modulator
If we limit the range of frequencies we look at to calculate the quantization noise to values
below fs , then we can rewrite Eq. (32.10) as
f
units, V/ Hz
V LSB
NTF( f ) ⋅ V Qe ( f ) = ⋅ 2 sin π (32.11)
12f s f s
The RMS quantization noise present in a bandwidth, B, can be calculated, see Eq. (30.44),
using
B B
V 2LSB f
V 2Qe,RMS = 2 ∫ NTF( f ) 2 VQe ( f ) 2 ⋅ df = 2 ⋅ ⋅ 4 ⋅ ∫ sin 2 π ⋅ df (32.12)
0
12f s 0
fs
Remembering that the maximum bandwidth of our input signal is related to the sampling
frequency, fs , and the oversampling ratio, K, by
fs
B= (32.13)
2K
and, for small values of x,
sin x ≈ x (32.14)
then
Chapter 32 Noise-Shaping Data Converters 163
V LSB π
V Qe,RMS ≈ ⋅ ⋅ 13/2 (32.15)
12 3 K
This equation should be compared to Eq. (31.51). Further we can describe the ideal data
converter SNR using first-order NS, see Eqs. (31.1) - (31.4) as
6.66
0
1 10 100 1k 10k K , Oversampling ratio
Examples 32.3
Determine the ideal signal-to-noise ratio and the maximum signal bandwidth
allowed, B, for the first-order NS modulator of Fig. 32.7 if 16 of its output
samples are averaged (K = 16).
Because the sampling frequency, fs, is 100 MHz, we can use Eq. (32.13) to
determine the maximum input signal bandwidth, B, is 3.125 MHz. Using Eq.
(32.17) we can solve for the SNRideal as (knowing that the NS modulator of Fig.
32.7 uses a 1-bit quantizer) 38.73 dB. This corresponds to an equivalent data
converter (ADC) resolution, using Eq. (31.4), of 6.14 bits (number of bits added is
164 CMOS Mixed-Signal Circuit Design
5.14). Note that the ADC is made with the components, modulator, and digital
filter, shown in Fig. 32.9. T
Decimating and Filtering the Output of a NS Modulator
It's important to note that Eq. (32.17) was derived assuming the output of the modulator
was passed through a perfect lowpass filter with a bandwidth of B. Passing the output
through a sinc averaging filter, see Fig. 31.41, 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. [31.104]), of sinc averaging filter
should be used in the digital filter on the output of the NS modulator, and (2) assuming we
use only this filter (no half-band filter or 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
bits, Ninc, as
30 log K − 5.17
N inc = (32.18)
6.02
If our NS modulator uses a 1-bit ADC, then the final, after the digital filter, resolution of
the resulting data converter is Ninc + 1 bits. (An NS modulator using a 5-bit ADC [often
called a multibit NS modulator] would ideally have an output resolution of Ninc + 5 bits.)
Further, we saw in Ex. 31.22 and Fig. 31.55 that the word size increased by log 2 K bits in
each stage so that we can require
30 log K − 5.17
L ⋅ log 2 K ≥ (32.19)
6.02
Solving this equation results in L being greater than or equal to 2. In general, we can write
L = 1+M (32.20)
where M is the order of the modulator. For a first-order modulator we use two stages in
the averaging filter, or,
−K 2
H(z) = 1 1 − z −1 (32.21)
K 1 − z
In the next section we discuss second-order NS modulators (M = 2). For these modulators
we use a sinc averaging filter with L = 3.
Example 32.4
Sketch the implementation of the digital decimation filter for the modulator
described in Ex. 32.3. Assume the final output clocking frequency is 100 MHz/16
or 6.25 MHz. Do not be concerned with aliasing (use only the averaging filter).
The transfer function of the digital filter is (see Eq. [31.93])
Chapter 32 Noise-Shaping Data Converters 165
2
sin 16 ⋅ π f s
f
− −16 2
H(z) = = H( f ) =
1 z 1
16 ⋅
1 − z −1 sin π f s
f
The block diagram of the filter is shown in Fig. 32.17. The increase in resolution
through each accumulator (integrator) stage is log 2 16 = 4 bits. The resolution
calculated in Ex. 32.3 was 6.14 bits, which we round up to 7-bits. Because the
output of the digital filter is 9-bits, we drop the lower two bits (divide by 4) to get
our final 7-bit resolution. T
In 5 59 9 9 9 Out
1 1 1 − z −1
00001 select −1
1 − z −1
1 1−z −1 1−z Digital
11111 0 MUX
fs Add MUX f s /K
÷K Drop the lower
Clock 2 bits so that
Decimation filter output is 7 bits.
Figure 32.17 Sinc filter used for decimating the output of the NS modulator of Fig. 32.7.
Next let's answer how filtering with a sinc filter affects the SNR of the data
converter. Remember the SNRideal was calculated in Eq. (32.17) assuming the modulation
noise was strictly bandlimited to B. Figure 32.18 shows the PSD of the NTF 2 ( f )
⋅ V Qe ( f ) 2 (the modulation noise) of the first order NS modulator. Also shown in this
figure is the shape of the averaging filter's magnitude response squared (see Eqs. [32.22],
[31.90] and [31.93]). Here we are showing the shape of a filter with L = 2 (set by Eq.
[32.20] for a first-order modulator) and K = 16. We limit our range to fs /2.
4
f
1 sin Kπ f s
H( f ) 2
= ⋅
K sin π f
fs
V 2LSB f
NTF( f ) 2 ⋅ V Qe ( f ) 2 = ⋅ 4 sin 2 π
12f s fs
f s /K f s /2
B, Ideal maximum input frequency
Figure 32.18 Showing modulation noise and filter response.
166 CMOS Mixed-Signal Circuit Design
We can calculate the RMS quantization noise resulting from a cascade of a first-
order modulator and an averaging filter using
f s /2
sin 4 Kπ fs
f
f s /2
V2
V 2Qe,RMS = LSB ⋅ 84 ⋅ ∫0 ⋅ df (32.23)
sin 2 π f s
12f s K f
f
If we let θ = π , then this equation can be written as
fs
= K⋅ π4
V2 sin 4 (Kθ)
2
8 fs
V 2Qe,RMS = LSB ⋅ 4 ⋅ π ⋅
12f s K ∫0 sin 2 θ
dθ (32.24)
and finally,
V LSB 2
V Qe,RMS = ⋅ 3/2 (32.25)
12 K
This equation should be compared to Eq. (32.15), which was derived assuming the digital
filter was ideal with a bandwidth of B. The SNR resulting from using a first-order (M = 1)
NS modulator and a second-order (L = 2) sinc averaging filter is
SNR sinc = 6.02N + 1.76 − 3.01 + 30 log K (in dB) (32.26)
Comparing this to SNRideal given in Eq. (32.17), we see that using a sinc filter for
averaging results in only a 2.16 dB difference (increase) in SNR over the ideal filter. If we
remember that using a sinc filter results in a droop in the desired signal, see Fig. 31.46, the
SNR will be lower than what is predicted by Eq. (32.26). (Note that an analysis of higher
order modulators using sinc averaging filters would show that as long as Eq. [32.20] is
valid the deviation from SNRideal is negligible.)
We have not talked about the effects of sample rate reduction (decimation) on the
SNR. The modulation noise aliased into the base spectrum is a concern when decimating
the output of the modulator. The major difference between filtering the outputs of the data
converters in the last chapter and filtering the output of an NS modulator is the spectral
characteristics of the noise. In the last two chapters we assumed the quantization noise
spectral density was white (see Fig. 30.57). As we've seen in this chapter the modulation
noise increases with increasing frequency (see Fig. 32.18). This can mean that the amount
of modulation noise aliased into the base, or desired, spectrum can be more of a concern.
The major concern when decimating, in most situations, is the input signals that reside at
frequencies > B and the resulting aliasing degradation of the SNR (see, for example, Fig.
31.59).
Chapter 32 Noise-Shaping Data Converters 167
(32.27)
1−z
This equation shows that adding a second accumulate-and-dump results in averaging K2
samples while not providing a reduction in the first sidelobe's amplitude. Because the
bandwidth of the filter is reduced using this topology, we get a corresponding reduction in
quantization noise on the output of the filter (and corresponding increase in the
attenuation when looking at a fixed frequency in the stop band). While we could modify
the cascade of accumulate-and-dumps to operate properly, with a final output clocking
frequency of fs /K, the simplicity and ease of designing the averaging filter using the
topology of Figs. 32.17, 31.54, or 31.56 makes them the topology of choice, in most
situations, for a NS modulator averaging filter.
Accumulate-and-dumps
In Out
1 − z −K 1 − z −K
2
The time interval we average the output of the modulator over is related to the
order of sinc averaging filter as shown in the following two examples. A time interval
longer than KTs is used when L ≥ 2 .
Example 32.5
Determine the time domain impulse response of a first-order averaging filter (L =
1) with K = 8. Assume decimation is not employed in the filter.
The transfer function of the filter is given, after reviewing Eqs. (31.89) and
(31.90), by
168 CMOS Mixed-Signal Circuit Design
−8
H(z) = 1 − z −1 = 1 + z −1 + z −2 + z −3 + z −4 + z −5 + z −6 + z −7
1−z
The time domain relationship between the input and the output is then
y[nT s ] = x[nT s ] + x[(n − 1)T s ] + x[(n − 2)T s ] + ... + x[(n − 7)T s ]
The time-domain impulse response of the first-order averaging filter is shown in
Fig. 32.20. Note the rectangular shape. T
x[nT s ]
1 Impulse input
time, n/T s
0
y[nT s ]
1 Output
time, n/T s
0 1 2 3 4 5 6 7 8 9 10 11
Example 32.6
Repeat Ex. 32.5 if a second-order averaging filter is used.
The transfer function of the filter is
−8 2
H(z) = 1 − z −1 = 1 + 2z −1 + 3z −2 + 4z −3 + 5z −4 + ... + 3z −12 + 2z −13 + z −14
1 − z
The time domain relationship is
y[nT s ] = x[nT s ] + 2x[(n − 1)T s ] + 3x[(n − 2)T s ] + ... + 2x[(n − 13)T s ] + x[(n − 14)T s ]
The impulse response of the second-order averaging filter is shown in Fig. 32.21.
Note the triangular shape of the curve and how the impulse response of the
second-order filter lasts twice as long as the first-order's response. T
y[nT s ]
8
1
time, n/T s
0 1 2 3 4 5 6 7 8 9 10 1112 13 14 15 16
Out H(z) = 1 − z −K
In
1
KT s
Voltage controlled voltage source
used for summing.
Transmission line delay of KT s
(a)
Out
In
1 Ts
Transmission line
1
H(z) = 1
Used as an 1 − z −1
inverting buffer (b)
transfer function equivalent to Eq. (32.28) (see Eqs. [31.100] and [31.101]), we can
cascade L of these sections. The resistors used on the output of the transmission lines, in
Fig. 32.22 are used to terminate the transmission lines. The following example
demonstrates the implementation of a sinc averaging filter in SPICE.
Example 32.7
Generate a SPICE model for the digital filter used with the first-order modulator
of Fig. 32.7. Assume, as in Ex. 32.4, that K = 16. Specify the droop in the output
when an input sinewave has a frequency of B.
In this modulator the clocking frequency is 100 MHz (Ts = 10 ns). The delay in the
comb filters is 160 ns, while the delay in the integrators is 10 ns. The SPICE netlist
is shown below. Note how the input to the filter is isolated using a buffer, and the
integrators are isolated by the comb filters. The output is scaled at the end of the
netlist to 1/K2 to normalize the filter's gain (see Fig. 31.46).
The simulation results are shown in Fig. 32.23. The input to the filter is a 1-V
sinewave that is swept from DC to 50 MHz (= fs /2). The droop at B (= 100
MHz/(2⋅ 16) = 3.125 MHz) is 7.8 dB (which matches what was predicted in Fig.
31.46).
* Figure 32.23 CMOS: Mixed-Signal Circuit Design *
Vin Vin 0 DC 0 AC 1
*Input buffer
Ebuf1 Vobuf 0 Vin 0 1
*Comb filter 1
EC1 Vo1 0 Vobuf Vf1 1
TC1 Vobuf 0 Vf1 0 ZO=50 TD=160n
RC1 Vf1 0 50
*Integrator filter 1
EI1 Vo2 0 Vo1 Vb1 1
TI1 Vo2 0 Vf2 0 ZO=50 TD=10n
RI1 Vf2 0 50
EB1 Vb1 0 0 Vf2 1
*Comb filter 2
EC2 Vo3 0 Vo2 Vf3 1
TC2 Vo2 0 Vf3 0 ZO=50 TD=160n
RC2 Vf3 0 50
Chapter 32 Noise-Shaping Data Converters 171
*Integrator filter 2
EI2 Vo4 0 Vo3 Vb2 1
TI2 Vo4 0 Vf4 0 ZO=50 TD=10n
RI2 Vf4 0 50
EB2 Vb2 0 0 Vf4 1
.end
T
= −1 (V in − V out )
iF
V OI = (32.30)
jωC jωRC
where ω = 2πf and noting how we switched the inverting and noninverting inputs of the
comparator to compensate for the inverting gain of the integrator. Knowing, from Eq.
(31.82), that
z −1 = 1 (32.31)
1 − z −1 −1 + cos 2π f + j sin 2π f
fs fs
f
or, for f << f s (which must be valid for any oversampling converter) where cos 2π fs ≈ 1
f f
and sin 2π fs ≈ 2π f s , then
z −1 ≈ 1 (32.32)
1 − z −1 j2π f
fs
Finally, we see that the topology of Fig. 32.24 behaves like a modulator with a block
diagram of Fig. 32.6 when
fs = 1 (32.33)
RC
If this equation doesn't hold, the topology of Fig. 32.24 may still function correctly as a
NS modulator, except that the above analysis would include an integrator gain. (Which,
R φ Clocked at f s
C V CM
V in R
V out
V CM
−V out
combined with the high gain of the comparator, may still result in an overall forward path
gain of one. We will discuss component gains in the modulator later.)
Analog integrator-based implementations of modulators can be simpler and easier
to breadboard and test on the bench, lower power, and less susceptible to clocking noise
(capacitive feedthrough and charge injection). The two main drawbacks are the difficulty
in setting the integrator gain to a precise value (in integrated versions) and the integrator's
susceptibility to the fed back pulse shape (a problem also encountered using the analog
sinc filters of the last section).
Using a DAI, the gain of the integrator is set by a ratio of capacitors, see Fig.
31.79 (and Ch. 27). Variations in the absolute oxide capacitance for a given process run
don't affect the integrator's gain. Using the analog integrator in a purely monolithic form,
however, can result in RC time constant variations of 50% or more.
Figure 32.25 shows how the shape of the pulse affects the output of the analog
integrator. 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 the finite rise time and fall time can affect the actual
area under the curve and thus the output of the integrator. To minimize these unwanted
effects we can use wider pulses as shown in parts (c) and (d), which means we run the
modulator at a slower clocking frequency. Increasing the width of the pulses minimizes the
percentage of the area affected by the transition times. Note that the feedback signal
directly subtracts from the input signal so that any noise or unwanted variation in the fed
back signal, such as an amplitude variation, can be considered as adding noise to the input
(and thus degrading the modulator's SNR). This is important! We will discuss the fed back
signal, and how to isolate/implement the actual voltage fed back to the integrator, again in
the next section.
Ideal area
digital output, doesn't limit the rate the switched capacitor is charged but rather the
limiting factor is the op-amp (since the top plate is tied to the op-amp's inverting input and
charged from the op-amp's output through the feedback capacitor).
The Feedback DAC
Up until this point we have been feeding the output of the comparator directly back to the
integrator. This works fine as long as the logic "1" (VDD) or logic "0" (ground) voltages
are clean (have no noise on them). Noise on these fed back voltages directly adds or
subtracts from the input signal and thus decreases the modulator's SNR. In any practical
mixed-signal integrated circuit the digital supply and return are commonly noisy with
variations in the hundreds of mV. As discussed in Ch. 28, it is common to separate the
analog and digital power supplies on-chip (and so the modulator should be powered with
the analog supply). Because of this, and the desire to set the fed back voltage V REF+ and
V REF− independent of VDD and ground, the output of the comparator is often connected to
the simple 1-bit DAC circuit shown in Fig. 32.26. It's important to remember that the
output of the DAC must be able to charge the integrator's input switching capacitance (see
Fig. 32.7) to within the final desired resolution of the converter in half a clock cycle
(before the φ 2 switches open).
To integrator To comparator
Understanding Averaging and the Use of Digital Filtering with the Modulator
In the following discussion we assume, as before, that V REF+ = VDD = 1.5 V and V REF− =
ground (1 LSB = 1.5 V) . We also assume the input voltage to the modulator falls within
VDD and ground so that the output of the modulator doesn't saturate in a string of ones or
zeroes (more on this below). In this section we want to discuss, intuitively, the operation
of the digital filter used on the output of the modulator (see Fig. 32.9).
Before discussing the digital filter let's remember that we can recover the analog
input voltage using an analog filter as shown in Fig. 32.10. This means that if the output of
the modulator is a continuous string of zeroes, then the output of the analog filter will be
zero volts. Other possible modulator outputs and their averages are shown in Fig. 32.27.
Chapter 32 Noise-Shaping Data Converters 175
1.5
Average =
1.5
4
= 0.375
0
1.5
Average =
1.5
2
= 0.75
0
Modulator output
1.5 (31)⋅(1.5)
(Repeats every 32) Average = 32
0
1.5 (7)⋅(1.5)
Average = 8
0
1.5 1.5
Average = 8
0
0 Average = 0 V
8 16 24 32 40
Example 32.8
Plot the ideal I/O transfer curve for the 1-bit DAC. Also plot the non-ideal transfer
curve if V REF+ = 1.45 V (instead of 1.5 V). Comment on how the output of the
NS-modulator/decimator (the data converter) will be affected.
The transfer curves are shown in Fig. 32.28. The offset in the positive reference
voltage results in a gain error in the data converter (but no nonlinearity). T
Ideal
1.5
Output, V
Nonideal
0
0 0.75 1.5 Comparator input, V
Figure 32.28 Ideal and nonideal transfer curves for the 1-bit DAC and comparator cascade.
Next let's consider the case where the modulator's input is a DC signal and our
data converter's resolution is only limited by the number of samples we can take in a given
time, KTs . Equation (32.13) shows that as K approaches infinity, B approaches zero (DC).
Figure 32.14 shows that at DC the spectral density of the modulation noise is zero.
Feeding the output of our first-order modulator to a single accumulate-and-dump, see
Figs. 31.39 and 31.41, can provide the needed digital filtering. As seen in Fig. 31.41,
176 CMOS Mixed-Signal Circuit Design
increasing K causes the filter's amplitude response at DC to increase (the number of bits
coming out of the counter increases) and the sidelobes move toward DC. Also, as seen in
Fig. 31.58b, we don't have aliasing at DC, so the reduction in output clocking frequency
(decimation) to f s /K can be accomplished with the single accumulate-and-dump stage.
However, there may still be aliasing from higher frequencies. Note that we have just
described a first-order NS modulator driving a counter where the counter is reset and
read-out every KTs clock cycles and clocked at a rate of fs . Practically, for large K, the
limiting factor in the resolution of the NS modulator is the noise inherent in the circuit
(mainly thermal and flicker noise sources from the MOSFETs) and, more importantly, the
finite gain and linearity of the op-amp (more on this later). Note that in our perfect
modulator a single output going high out of one-million outputs would correspond to an
average input voltage of only 15 µV . This would also mean that we would need to
average at least one million and one (1,000,001) modulator outputs with our simple
counter for a constant output.
In our cascade of two accumulate-and-dumps, Eq. (32.27) and Fig. 32.19, we
average K2 samples. The cascade behaves, from a frequency response point of view, like a
single accumulate-and-dump. For the cascade of L sinc filters, however, the number of
modulator outputs averaged is
Number of modulator outputs averaged = L ⋅ K − 1 (32.34)
As we saw in Ex. 32.6, the cascade of sinc filters results in a weighted average of the
filter's inputs. For our first-order modulator of Fig. 32.7 and the digital filter of Fig. 32.17
with a transfer function, once again, of
−16 2
H(z) = 1 − z −1 (32.35)
1−z
we perform a weighted average on 31 of the modulator's outputs. The time domain
impulse response of this filter (without decimation), again see Ex. 32.6, is given by
y[nT s ] = x[nT s ] + 2x[(n − 1)T s ] + 3x[(n − 2)T s ] + ... + 15x[(n − 14)T s ] + 16[(n − 15)T s ] +
15x[(n − 16)T s ] + 14x[(n − 17)T s ] + ... + 2x[(n − 29)T s ] + x[(n − 30)T s ] (32.36)
For a continuous filter input of "1" the output of the filter is 256. Note that this is the same
result we get with the cascade of two accumulate-and-dumps and is the "gain" ( KL ) of the
filter at DC. For our current discussion the minimum resolution we can represent with the
maximum output value of 256 (realizing the minimum output value, which corresponds to
a continuous modulator output of all zeroes, is 0) is (1.5)/256 = 5.86 mV . We want this
value to be less than the resolution calculated in Ex. 32.3, which was 6.14 bits (1.5/26.14 =
21.27 mV) so that our modulation noise, for a given bandwidth, limits the data converter's
resolution and not the digital filter. This is why adding an additional sinc filter stage (say,
L = 3) will not increase the data converter's resolution. The fundamental way to increase
the first-order modulator's resolution is to increase the number of samples averaged (the
oversampling ratio), K. Note also, that increasing L will have the undesirable effect of
increasing droop in the bandwidth of interest.
Chapter 32 Noise-Shaping Data Converters 177
Example 32.9
If the desired input to an ideal first-order NS modulator is a DC signal, would it be
better to use a single sinc filter (a counter or accumulate-and-dump) or a cascade
of two sinc filters of the form given by Eq. (32.35)?
Equation (32.20) was derived assuming we wanted to maximize the input signal
bandwidth, B. If we are measuring a DC signal with, ideally, zero-bandwidth, then
we want to minimize the digital filter's bandwidth to remove unwanted noise that
may corrupt the DC signal. This means, for higher resolution with correspondingly
longer conversion time, the single-stage filter is the best choice. T
In Fig. 32.17 we indicated that the word size coming out of the filter, after
dropping the lower two-bits, is seven-bits. This means, in two's complement, that outputs
of 011 1111 (+63), 100 0000 (−64), and 000 0000 (0) correspond to the maximum input
(V REF+ − 1 LSB ), minimum input (V REF− ), and common-mode voltage (VCM = 0.75 V)
respectively, see Fig. 31.37. A continuous modulator output of "1" would correspond to
an input of VREF+ , which would be outside the possible digital filter output words and
result in the incorrect filter output code of 100 0000. Note that here
1 LSB= (V REF+ − V REF− )/2 7 = 11.719 mV , which is, again, below the 21.27 mV
fundamental RMS noise limit of the modulator in a bandwidth, B.
Next, let's consider the situation where one out of every 64 modulator outputs is a
logic 1 (and the sequence repeats indefinitely). As we saw in Fig. 32.27, averaging this
output results in an analog voltage of 1.5/64 or 23.44 mV. Using Eq. (32.36), we can
write the sequence of digital filter outputs (in decimal form and assuming our single-pulse
input to the filter marks the beginning of the output) without decimation as
1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0, (followed
by 32 more zeroes and then the sequence repeats itself).
If decimation is employed, as in Fig. 32.17, then the outputs of the filter look like
1,15,0,0 (summed = 16)
noting that shifting in time doesn't change the sum of the sequence if the samples output
by the decimation process are spaced apart by 16 (K). For example, the sequence
4,12,0,0 (summed = 16)
also has the same sum of 16. Figure 32.29 shows plots of our modulator input and the
reconstructed filter output. Our modulator input is a constant DC signal of 23.44 mV
while at the same time the digital filter output is a repeating sequence. What we are seeing
is the ripple associated with passing the output of the modulator through the filter.
Additional filtering, including the RCF, will reduce the ripple (because of the reduction in
bandwidth). We should point out that this ripple represents a major difference between
Nyquist-rate data converters and oversampling converters. Note how averaging 4 output
samples after decimation or 64 samples before decimation results in the digital
reconstructed output value matching the input value.
178 CMOS Mixed-Signal Circuit Design
Figure 32.29 Ripple in the output of a digital filter. Note how the
average of the filter output is equal to the input voltage.
Figure 32.30 shows the output of an RC filter, with a time constant of 1 µs, when
connected to the output of the modulator of Fig. 32.7. The modulator input voltage is
23.44 mV. The analog RC filter, like the digital filter, will not totally filter out the higher
frequency components of the modulation noise and thus there will be some ripple on the
output signal. Again, as in the digital filter, reducing the bandwidth of the filter (increasing
the time constant) will reduce the peak-to-peak amount of ripple.
Figure 32.30 Showing how we have ripple on the output of an analog filter connected to
a modulator with a DC input.
Example 32.10
What are the 7-bit, two's complement, representations of the numbers in Fig. 32.29
assuming they are originally represented using 9-bit, two's complement, words as
in Fig. 32.17? What is the four-sample average of these words?
Chapter 32 Noise-Shaping Data Converters 179
Figure 32.31 Filter ripple when input is the common-mode voltage of 0.75 V.
180 CMOS Mixed-Signal Circuit Design
outside our base spectrum (which, from Ex. 32.3, is from DC to 3.125 MHz) it will not, in
a significant way, affect the SNR. (In the digital filter a zero in the digital filter's transfer
function will most likely fall at half the clocking frequency eliminating the ripple altogether
and resulting in a constant filter output value.) The frequency of the ripple in Fig. 32.30,
however, is 1/640 ns or 1.564 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 becomes,
"How do we minimize the possibility of unwanted tones appearing in the data converter's
output spectrum?"
If we look at the digital filter output data shown in Fig. 32.29 we see that the
"ripple" amplitude of the digital data is fully 87.9 mV peak-to-peak or significantly above
the data converter's LSB value (noting that after the RCF/half-band filter this ripple value
will be reduced). Looking at this figure, we see that it would be better to spread or flatten
the data out over all four cycles of the repeating waveform. Although we may still have a
tone, or repeating sequence, at a frequency in the base spectrum, the amplitude of the tone
will be well below the LSB of the data converter (and so it won't affect the SFDR of the
data converter). To accomplish this spread or randomization we can add a noise dither
source (see the last chapter) to our basic NS modulator, as seen in Fig. 32.32. By applying
the dither to the input of the comparator (quantizer) the dither will be noise-shaped like
the quantization noise (the spectral content of the dither, Eq. [31.76], is less important).
Dither V CM
Source
φ2 φ1
φ1
1p
V CM
V out
V in 1p V CM
As a final example, let's consider how a tone can occur in a modulator output that
has heavy transition densities (numerous one-zero transitions). If the input to the
modulator is the common mode voltage, VCM , of 0.75 V, then the output of the modulator
is an alternating sequence of ones and zeroes. Changing the input voltage upwards by a
small amount will result in the output of the modulator staying high once in a while instead
of going low (resulting in two consecutive logic one outputs). An example is seen in Fig.
32.34 where the input to the modulator was increased to 0.77 V. As the double ones are
spaced apart by approximately 350 ns, we can estimate a tone in the resulting output
spectrum at a frequency of 1/350 ns or 2.86 MHz.
Finally, note that unwanted tones are usually not a problem if the input signal is
busy and random (not DC as discussed in this section). Later in the chapter, we discuss
second-order modulators that utilize two integrators. The second integration helps to
spread the repeating sequences out over a longer period of time so that, hopefully,
negligible unwanted tone energy is present in the base spectrum.
Figure 32.34 Modulator output showing how tones can occur with higher transition density.
182 CMOS Mixed-Signal Circuit Design
0.75
E(z)
ADC
In X(z) z−1 Y(z) Out
GI Gc
1 − z −1
GF = GI ⋅ Gc
GF = GI ⋅ Gc (32.38)
We can rewrite Eq. (32.4) using this gain as
z −1 ⋅ G F 1 − z −1
Y(z) = −1 (G
⋅ X(z) + ⋅ E(z) (32.39)
1+z F − 1) 1 + z −1 (G F − 1)
If GF approaches zero (the integrator saturates while the comparator gain stays finite),
then the output of the modulator is the sum of the integrated input and the quantization
noise. (This is bad.) Since the quantization noise is not spectrally shaped it will be difficult
to filter the modulator's output to recover the input signal. If the forward gain is greater
than two, then, as seen in Fig. 31.62 and the associated discussion, the poles of the
transfer function reside outside the unit circle and the modulator will be unstable. We can
restrict the values of the forward gain to
0 ≤ GF ≤ 2 (32.40)
Ideally, however, the gain is one.
Example 32.11
Show, using SPICE simulations and the modulator of Fig. 32.7, that an integrator
gain of 0.4 will result in an op-amp output range well within the power supply
range.
Figure 32.37a shows a schematic of the modulator with GI = 0.4. Figure 32.37b
shows the output of the integrator (the output of the op-amp) in the modulator of
part (a) with the input sinewave shown in Fig. 32.8. The output swing is limited to
roughly 80% of the supply range. For general design it is desirable to set our
integrator gain to 0.4. This ensures our integrator doesn't saturate unless the input
to the modulator goes outside the supply voltage range.
It's interesting to note that in both modulators, Fig. 32.7 and Fig. 32.37, the
forward gain is unity. This is a result of the effective gain of the comparator
changing forcing the forward gain, controlled by the fed back signal, to unity.
What this means is that our modulator functions as expected with a signal gain of
one (Eq. [32.4] is valid) whether GI is 1 or 0.4. We discuss how this change in
comparator gain occurs next. T
184 CMOS Mixed-Signal Circuit Design
φ1 φ2 φ1
1p
V CM
V CM V out
V CM
(b)
Figure 32.37 (a) First-order NS modulator with an integrator gain of 0.4, and (b)
the output of the op-amp.
Figure 32.38 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
rails (1.5 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 1.
We can write this using equations as
Integrator gain, G I Comparator gain, G c GF
1.5
0
0 0.75 1.5 Comparator input, V
Input signal
If the modulator is functioning properly, then the average value of the modulator output
will be equal to the modulator input and thus GF = 1. It's interesting to note that this result
(precise integrator gain isn't important) will apply to any integrator that is directly
followed by an ADC.
Before leaving this section, let's point out a couple of problems with a noise-
shaping modulator that uses a multibit ADC, Fig. 32.39. Since the output of the integrator
is the input signal to the ADC, the limited integrator output swing will directly affect the
range of ADC output codes. Limiting the range of ADC output codes will then limit the
allowable range of modulator inputs unless scaling is used (shifting the output codes or
sizing of capacitors in the DAI). Next, notice in Fig. 32.39 how the variation in the gain of
the ADC, with input signal, is more limited than the gains attainable with the simple
comparator of Fig. 32.38. Limiting the range of ADC gains can result in modulator
Digital
output code
111
110
101
100 Dashed lines indicate ADC gain
011
010
001
000
0 1/8 2/8 3/8 4/8 5/8 6/8 7/8 8/8
Analog input voltage (integrator output)
Figure 32.39 A 3-bit ADC.
186 CMOS Mixed-Signal Circuit Design
forward gains that are not exactly unity. This is especially true at high input frequencies
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.
Comparator Gain, Offset, Noise, and Hysteresis
It's of interest to determine how the performance of the comparator affects the operation
of the modulator. Both the comparator's offset and input-referred noise, Fig. 32.40a, can
be referred back to the modulator's input, Fig. 32.40b. By doing so we can determine how
they effectively change the input signal seen by the modulator. As seen in Fig. 32.40, the
high gain of the integrator, A( f ), reduces the effect of the comparator's 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 mV, then the input-referred offset is only 50 µV.
Comparator's input-referred
noise and offset, V n,comp ( f ) Input-referred noise and offset, V n,comp ( f )/A( f )
In Out In Out
A( f ) A( f )
Comparator Comparator
Integrator Integrator
(a) (b)
Figure 32.40 (a) Referring the comparator offset and noise to (b) the input of the modulator.
Using this result in Eq. (32.3) and, as discussed in the last section, assuming the forward
gain of the modulator, GF , is one gives
C
z −1 1 + C FI A 1
OL ( f)
− z −1
Y(z) = C
⋅ X(z) + C
⋅ E(z) (32.45)
1 + CI A 1
1 + CI A 1
F OL ( f ) F OL ( f )
If we assume the contribution to the noise from the error term squared, ε 2gain , is
small, which is valid for op-amp gain
A OL ( f ) > K (the oversampling ratio) (32.50)
over the frequency range of DC to B, then we can rewrite Eq. (32.16), to include the
effects of finite op-amp gain, as
SNR gerr = 6.02N + 1.76 − 20 log π + 20 log K 3/2 − 10 log (1 + ε gain ) (32.51)
3
The largest degradation in the SNR, resulting from integrator leakage, can be estimated as
0.5 dB if K ≥ 8 (ε gain ≈ 1/8 neglecting GI ). The minimum gain⋅ bandwidth product of the
op-amp is estimated as
Op-amp unity gain frequency, f u = K ⋅ B = f s /2 (32.52)
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
oversampling ratio, K.
To illustrate typical op-amp requirements, let's consider the modulator of Fig.
32.37 with K = 16 and B = 3.125 MHz (see Ex. 32.3). The fu of the op-amp is estimated,
using Eq. (32.51), as 50 MHz. If the open-loop response of the op-amp starts to roll off at
10 kHz, then the DC gain of the op-amp must be at least 5,000. However, we could also
use an op-amp with a DC gain of 100 (remembering low integrator gain increases the
undesirable effects [noise and offset] of the comparator on the performance of the
modulator) that rolls off at 500 kHz.
Op-Amp Settling Time
Equation (32.52) can be used, for the moment, to provide an estimate for the settling time
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 (assuming a
dominant pole compensated op-amp, see Eqs. [27.37] and [27.38]) as
× 100% = 1 − exp − π ⋅
v out CF
× 100% (32.55)
V outfinal 2 CF + CI
Chapter 32 Noise-Shaping Data Converters 189
CF
β ⋅ v out = v f = v out
CI + CF CF
vf
CI v out
The output will only reach 67% of its ideal final value in the modulator of Fig. 32.37 when
the op-amp used has a unity gain frequency of fs /2.
In deriving Eq. (32.55) we used an op-amp unity gain frequency specified by Eq.
(32.52) 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
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
Settling gain error, G s
C V (z) ⋅ z −1/2 − V 2 (z)
V out (z) = I ⋅ (1 − e −πβ⋅(f u /f s ) ) ⋅ 1 (32.56)
CF 1 − z −1
Full, or complete, settling requires that the op-amp's unity gain frequency, fu , be much
larger than the sampling frequency, fs (in other words we can't use Eq. [32.52] to specify
the required bandwidth of the op-amp if settling time is important). The constant gain
error, resulting from incomplete settling, can be tolerated in the first-order modulator
because the integrator is directly followed by a comparator, as discussed earlier. In some
of the modulator topologies, though, the integrator is not followed by a comparator so
settling time becomes more important. 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 modulator is then
evaluated to determine the allowable values of Gs for the application.
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. (32.56), will
not be a constant and will introduce distortion into the modulator's output spectrum
(whether a comparator follows the integrator or not).
Op-Amp Offset
The operation of the DAI is subject to the op-amp's offset. It can be shown that this offset
will effectively add (or subtract) from the common-mode voltage, VCM, and thus effectively
shift the input signals upwards or downwards. The resulting modulator output will then
show an offset equal to the op-amp's offset. To circumvent this problem, offset storage
can be used in the integrator.
190 CMOS Mixed-Signal Circuit Design
In Out
A( f )
Comparator
Integrator
Figure 32.42 The modulator's input-referred noise contributions from both the
comparator and the integrator.
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. (32.4), the
modulator's input, and thus its input-referred noise, pass through the modulator with a
delay of z-1. If we assume the modulator's input-referred noise is white and bandlimited to
fs /2 such that
Vn
V n,ckt ( f ) = for f < f s /2 (32.57)
fs
then passing the output of the modulator through an ideal lowpass filter with a bandwidth
of B ( = fs /[2K] ) results in
B
V 2n V
V ckt,RMS = 2 ⋅ ∫ ⋅ df = n (32.58)
0
fs K
Noting that not passing the output of the modulator through a lowpass filter results in an
RMS output noise of Vn, we see that the averaging filter (the lowpass filter) reduces the
noise by the root of K. We could also think of the filtering as reducing the PSD of the
modulator's input-referred noise by K. Remembering the jitter discussion from the last
chapter, we see a direct parallel in the derivations of how averaging affects the RMS value
of a random signal (noise or jitter).
Finally, as used in Ex. 31.15, we can estimate the finite SNR of a data converter
from quantization noise, jitter, and circuit noise using
and
Vp / 2
SNR= 20 ⋅ log (32.60)
V n,RMS
where Vp is the peak amplitude of an input sinewave, see Eq. (31.1), and
v 2+ CF
φ1 φ2
v 1+
CI
v out+
V CM
v out−
CI
v 1−
v 2− CF
C I V 1 (z) ⋅ z −1/2 − V 2 (z)
V out (z) = ⋅
CF 1 − z −1
Figure 32.43 Fully-differential discrete-analog integrator (DAI) implementation.
It's important to understand the signal levels in the fully-differential DAI. Let's
assume VCM = 0.75 V and the input voltages can range in amplitude from 0 to 1.5 V.
Assuming the input is balanced correctly if v 1+ = 0.85 V , then v 1− must equal 0.65 V. The
maximum input voltage is v1max= 1.5 − 0 = 1.5 V. The minimum input signal, on the other
hand, is v1min = 0 − 1.5 = −1.5 V. The range of inputs, or outputs, is then 3 V or twice the
range of the single-ended DAI.
192 CMOS Mixed-Signal Circuit Design
V out
φ1 φ2 CF
V in+ φ1
CI
V CM V out
CI
V in−
CF
−V out
v o+
v− E1
V CM
v− v o+
v+ v o−
v+ E2
v o−
Figure 32.45 SPICE modeling a differential input/output op-amp with common-mode voltage.
Figure 32.46 shows the simulation results for the outputs of the modulator of Fig.
32.44 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, because the gain of the integrators is 0.4, the
integrator output's swing is at most 80% of the supply rails. Also note that the input
common-mode voltage of the op-amp remains at 0.75 V. This is important as the design
of the op-amp becomes more 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 modulator. Because many input signals will not be fully-differential, we
briefly discuss differential modulator design with single-ended inputs next.
Chapter 32 Noise-Shaping Data Converters 193
(a) (b)
Figure 32.47 (a) Filtered modulator outputs with full-scale (1.5 V peak-to-peak) single-
ended input, and (b) how the input common-mode voltage of the op-amp changes.
194 CMOS Mixed-Signal Circuit Design
If we restrict our frequency range to frequencies less than fs /2, then we can rewrite
Eq. (32.63) as
V LSB f
NTF( f ) ⋅ V Qe ( f ) = ⋅ 4 sin 2 π (32.64)
12f s f s
16.6
13.3
First-order noise-shaping, M = 1
10.0
6.66
0
1 10 100 1k 10k K , Oversampling ratio
NTF(z) = 1 = (1 − z −1 )
2
(32.69)
1 + A(z)B(z)
E(z)
ADC
In X(z) Y(z) Out
A(z)
B(z)
A(z) = z −1 (32.70)
2
(1 − z −1 )
and
B(z) = 2 − z −1 (32.71)
The second-order modulator can be implemented using the topology shown in Fig.
32.51a. The output of B(z) is the sum of the modulator output and the differentiated,
(1 − z −1 ) , modulator output. We can redraw the block diagram of Fig. 32.51a, as shown in
Fig. 32.51b, resulting in the implementation of a second-order NS modulator shown in
Fig. 32.51c.
E(z)
Comparator
A(z)
In X(z) 1 z −1 Y(z) Out
1 − z −1 1 − z −1
1 − z −1
B(z)
(a) E(z)
1
1 − z −1
1 − z −1
(b)
E(z)
(c)
The second-order (de) modulator topology of Fig. 32.51c can be used directly to
implement a NS DAC (see Figs. 32.11 and 32.12). However, this topology doesn't lend
itself directly to implementation using the DAI. 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 32.52a shows how we can add an integrator gain to the block diagram of
Fig. 32.51c without changing the system's transfer function. Figure 32.52b shows pushing
the gain, 1/GI , through the second summer so that it is directly preceding the second
integrator. Notice how in Fig. 32.52b this (the second integrator's gain) is in series with
the comparator's gain (not shown; see Fig. 32.36 and the associated discussion). This
means we can arbitrarily change the second integrator's gain because of how the
E(z)
Comparator
X(z) Y(z)
In 1 1 z −1 Out
GI
1 − z −1 GI 1 − z −1
(a)
X(z) Y(z)
In 1 1 z −1 Out
GI
1 − z −1 GI 1 − z −1
GI
(b)
E(z)
X(z) ⋅ z −1/2
X(z) 1 z −1 Y(z)
z −1/2 G1 G2 Out
1 − z −1 1 − z −1
(c)
comparator gain changes to force the loop gain to unity (see Fig. 32.38). Figure 32.52c
shows the resulting configuration where the second integrator has the a gain of G2 and the
first integrator has a gain of G1. Also notice how we have added a delay in series with the
input signal. This delay was added to show how using a DAI results in an added delay in
series with the input signal. The delay doesn't affect the magnitude of modulator's transfer
function but rather indicates the input signal arrives half a clock cycle later.
Figure 32.53 shows the DAI implementation of the second-order modulator of Fig.
32.52c. Note how the output of the modulator is fed back and immediately passes through
the first integrator and is applied to the second integrator (no delay as seen in Fig. 32.52c).
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 this
point it should be trivial to sketch the circuit implementation of the fully-differential,
second-order modulator (see Fig. 32.44).
C F2
V CM C F1 V CM φ 1 φ2 V CM φ 1
φ2 φ1
C I2 V out
C I1
V CM
V CM
V in
C I1 C I2
f s = 100 MHz G1 = G2 =
C F1 C F2
Figure 32.53 Implementation of the second-order modulator of Fig. 32.52c.
Integrator Gain
As we showed in Eq. (32.41) for the first-order modulator, the forward gain of a second-
order modulator will be unity when the modulator is functioning properly. We now need
to discuss how to select the integrator gains to avoid harmful integrator saturation. If
noise and offsets were not a concern, as shown in Fig. 32.40 and the associated
discussions, then we could make our integrator gains very small (ultimately limited by
imperfections in the switches such as clock feedthrough and charge injection). In a
practical modulator, integrator saturation (the integrator's gain going to zero) can also
lead to modulator instability, as shown in Eq. (32.40), and the associated discussion.
Figure 32.54 shows the integrator outputs for the modulator of Fig. 32.53 if both
integrator gains are set to 0.4. Notice how both outputs go outside the supply voltage
range. If we replace the ideal op-amps in the simulation with transistor-based op-amps, the
integrator outputs will saturate at some voltage within the supply range. This saturation
can be thought of as noise and ultimately limits the data converter's SNR. Integrator
saturation can be avoided by limiting the input signal range, designing with small
integrator gain, and using op-amps that have a wide output swing.
Chapter 32 Noise-Shaping Data Converters 199
Output1
Output2 Supply voltage range
Example 32.12
Using SPICE simulations, show how an ideal second-order NS modulator can
become unstable if the integrator gain is too low.
Because the second integrator is directly followed by a comparator, its gain is
more tolerant to variations allowing it (the gain) to be made small. The first
integrator's gain, however, is isolated from the comparator by the second
integrator restricting its values. Figure 32.55a shows the (unstable) output of the
modulator in Fig. 32.53 if G1 = 0.01, while Fig. 32.55b shows the integrator
outputs. T
(a) (b)
Figure 32.55 (a) Modulator output, and (b) integrator outputs ifG 1 = 0.01
200 CMOS Mixed-Signal Circuit Design
(32.72)
The poles of this transfer function are located at
z p1,p2 = (1 − G F ) ± (1 − G F) 2 − (1 − G F ) (32.73)
We know that for the modulator to remain stable the poles must reside within the unit
circle. This means that our values of forward gain are restricted to
0 ≤ G F ≤ 1.333 (32.74)
Again, if the modulator is functioning properly, GF = 1 (because of the comparator's gain
variation as seen in Fig. 32.28 and the associated discussion).
E(z)
2 − z −1
We should make some observations at this point. Reviewing Eq. (32.40), 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 (GI doesn't approach zero), stability for either modulator is easy to attain. An
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
modulator can have a forward gain of at most 1.15. Finally, notice that the input signal
range is more restricted for the second-order modulator, in order to avoid integrator
saturation, as seen in Fig. 32.54. We'll discuss methods to attain wider input signal range
and more robust stability criteria by adjusting the feedback gains later in this section.
Chapter 32 Noise-Shaping Data Converters 201
Notice that we are treating our modulator as a linear system even though it isn't
linear; the comparator gain is a nonlinear variable. The linear approximation is useful 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 are added to
determine the average comparator gain, Gc. Figure 32.57 shows this schematically.
Assuming we know GI (the gain coefficient of the integrators), we can then look at the
stability and forward gain of the modulator for varying DC input signal voltages.
DC input
v LPF LPF
G c = voutc
inc
v inc v outc
Figure 32.57 Simulating the gain of the comparator.
Next consider the more generic block diagram of the second-order NS modulator
shown in Fig. 32.58. In a moment we'll discuss how to implement the feedback gain, G3,
using the DAI. The transfer function of this topology can be written as
G 1 G 2 G c ⋅ z −1 X(z) + (1 − z −1 ) 2 ⋅ E(z)
Y(z) = (32.75)
1 + z −1 ⋅ (G 1 G 2 G c + G 2 G 3 G c − 2) + z −2 ⋅ (1 − G 2 G 3 G c )
Notice that if G 1 = G 2 = G 3 = G c = 1 (where G 1 G 2 G c = G F ), then this equation reduces
to Eq. (32.62). The poles of this equation are located at
2 − G 1 G 2 G c − G 2 G 3 G c ± (2 − G 1 G 2 G c − G 2 G 3 G c ) 2 − 4(1 − G 2 G 3 G c )
z p1,p2 =
2
(32.76)
When the modulator is functioning properly we require the (linearized) coefficient of the
input, X(z) in Eq. (32.75), to be unity
G1G2Gc
=1 (32.77)
(z − z p1 )(z − z p2 )
Again, if we set G 1 = G 2 = G 3 = 1 (and Gc = 1), then the poles are located at DC, that is,
z p1,p2 = 0 (32.78)
Equation (32.76) is useful to estimate the modulator's stability when scaling amplitudes by
adjusting the integrator gain coefficients, G1, G2, and G3.
202 CMOS Mixed-Signal Circuit Design
E(z)
X(z) G1 G 2 ⋅ z −1 Y(z)
Gc
In 1 − z −1 1 − z −1 Out
G3
GF = G1G2Gc
C F2
φ1 φ2
V CM
C I2 V CM v out
v1
C I3
v2
V 1 (z) C I2 −1/2
⋅z
1 V out (z)
C F2 1 − z −1
C I3 V 2 (z)
C F2
DAI of Fig. 32.59
(a)
G3 =
C I3 G3 V 2 (z)
C I2
DAI of Fig. 32.59
(b)
Example 32.13
Sketch the circuit implementation of a second-order NS modulator based on the
topology of Fig. 32.58, where G 1 = G 2 = G 3 = 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. 32.61. We could dissect Eq.
(32.76) 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.
φ1 φ2 1p φ1
φ2 φ1 1p
V CM V CM
V CM 0.4p V out
0.4p
V CM
V in V CM
0. 16p
Figure 32.62 shows the output swing of the integrators. This figure should be
compared with Fig. 32.54. 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.
Output2
Figure 32.62 Integrator output signals in the modulator shown in Fig. 32.61.
Let's attempt to get an idea for the stability of the modulator by adding LPFs,
as seen in Fig. 32.57, to the simulation (with a DC input) to measure Gc. Figure
32.63 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. 32.61 both
the comparator output and integrator outputs are ideal voltage sources, so we
don't need the isolation (and therefore we can add the RC LPF directly into the
simulation).
Out
In 10k
1 100 pF RC >> 1/f s
Figure 32.64 shows the comparator input and output, after lowpass filtering,
for the modulator of Fig. 32.61 when the input signal is 0.1 V (DC). Longer
simulation times reveal the average comparator input is 0.4 V. The resulting
comparator gain is then only 0.25. Using Eq. (32.76) to calculate the location of
the poles results in z p1.p2 = 0.96 ± j ⋅ 0.195 . These poles are very close to the unit
circle. Small shifts in the DAI gains can result in an unstable modulator. Increasing
the input signal amplitude makes the modulator more stable. Increasing G3 also
increases the modulator's stability.
Figure 32.64 Average comparator input and output when using the modulator
of Fig. 32.61 with an input signal of 0.1 V.
The simulation that generated Fig. 32.64 can be very useful in understanding
basic second-order modulator's stability criteria. Changing the simulation variables
and looking at the resulting simulation outputs can be very instructional. Note that
increasing the simulation time in the netlist that generated Fig. 32.64 would reveal
that the comparator input actually has small amplitude oscillations. Also note how
Fig. 32.62 shows the output of the second integrator going way outside the power
supply limits when transitioning negative (going well below 0 V) while staying
bounded to the power-supply rail when transitioning positive (above 1.5 V). This
is related to the stability of the modulator being a function of the input voltage. T
Using Two Delaying Integrators to Implement the Second-Order Modulator
Consider the second-order modulator topology shown in Fig. 32.65. This topology can be
implemented using the circuits of Figs. 32.53 or 32.61 by simply switching the phases of
206 CMOS Mixed-Signal Circuit Design
E(z)
X(z) G 1 ⋅ z −1 G 2 ⋅ z −1 Y(z)
Gc
In 1 − z −1 1 − z −1 Out
G3
the clocks in the first integrator (by making both integrators delaying). The transfer
function of this topology is
G 1 G 2 G c ⋅ z −2 X(z) + (1 − z −1 ) 2 ⋅ E(z)
Y(z) = (32.82)
1 + z −1 ⋅ (G 2 G 3 G c − 2) + z −2 ⋅ (1 − G 2 G 3 G c + G 1 G 2 G c )
The poles are located at
2 − G 2 G 3 G c ± (2 − G 2 G 3 G c ) 2 − 4(1 − G 2 G 3 G c + G 1 G 2 G c )
z p1,p2 = (32.83)
2
This equation should be compared to Eq. (32.76). Remembering that for a stable
modulator the poles must be inside the unit circle, we see that using two delaying
integrators will not result in a modulator that has as robust stability criteria as the general
implementation of Fig. 32.58. Figure 32.66 shows the implementation of a second-order
NS modulator using two delaying integrators. One advantage of this topology over the
topology of Fig. 32.58 is the reduced slew-rate requirements of the op-amps since neither
op-amp in Fig. 32.66 has to drive both the feedback capacitance and the switched input
capacitance of the next stage during the same clock phase.
C F2
V CM C F1 V CM φ 1 φ2 V CM φ 1
φ1 φ2
C I2 V out
C I1
V CM
V CM
V in
C I1 C I2
f s = 100 MHz G1 = G2 = G3 = 1
C F1 C F2
2 − G 1 G 2 G c − G 2 G c ± (2 − G 1 G 2 G c − G 2 G c ) 2 − 4(1 − G 2 G c )
z p1,p2 = (32.84)
2
Keeping in mind that the reason we are not setting all gains to one is to avoid integrator
saturation, we can look at Eq. (32.84) as a guide to determine how we can reduce G1 and
G2. Since G2 is directly followed by the comparator, we can set its gain to 0.4 as discussed
earlier. Practically then, we can reduce the value of G1 to a very small number and still
have a stable modulator (see Ex. 32.12). At the same time using small G1 avoids integrator
saturation. The practical problem with small G1, as discussed earlier, is the increase in the
input-referred noise. Again trade-offs must be made for given design criteria. Figure 32.67
shows the integrator outputs for the modulator of Fig. 32.58 when G1 = 0.2, G2 = 0.4, and
G3 = 1. Note how, when compared to Figs. 32.54 and 32.62, the outputs are very well
behaved. We don't have the abnormal transitions above the power-supply rails indicating
that the modulator stability is becoming marginal with input signal values close to the
power-supply rails.
G 1 = 0.2
G 2 = 0.4
G3 = 1
Figure 32.67 Integrator outputs for a modulator with first integrator gain of 0.2.
208 CMOS Mixed-Signal Circuit Design
G = 0.4
Volts
V in = 0.75 + 0.5 sin (2π ⋅ 500kHz ⋅ t)
Voltage, peak
We know that for the modulator to be useful its output must be passed through a
lowpass filter to remove the modulation noise. In simulations we can approximate a
lowpass filter with a bandwidth B by limiting the spectral analysis range. To estimate the
SNR from the simulations of Figs. 32.68 or 32.69, with K = 16, we perform the spectral
analysis up to 50 MHz/16 or 3.125 MHz. The quantization noise plus distortion is
calculated as discussed in the last chapter and shown in the SPICE netlists.
To demonstrate the calculation of a modulator's SNR let's use the second-order
modulator simulation used to generate Fig. 32.69. Using Eq. (32.65) with VLSB = 1.5 V
and K = 16 results in an RMS quantization noise of 1.86 mV. The SNRideal is calculated,
using Eq. (32.66) as 55 dB. However, Eq. (32.66) was derived assuming Eq. (32.9) was
valid. For the 1-bit ADC/DAC it is not. For the 1-bit DAC/ADC VLSB is twice the value
given by Eq. (32.66) or VREF+ − VREF−. The doubling in VLSB results in a subtraction of 6 dB
from Eq. (32.66). The SNRideal is 49 dB. To discuss this further consider a sinewave with a
peak-to-peak amplitude of VLSB (= 1.5 here). We can write the SNR of the modulator as
Chapter 32 Noise-Shaping Data Converters 209
Volts
0.75 DC
0.5 V peak input sinewave
Second harmonic
Third harmonic
Figure 32.70 Same as Fig. 32.69 but with limited spectral range.
The RMS noise in a bandwidth, B, can be written, see Eqs. (32.25) and (32.65), as
V Qe,RMS =
V LSB
⋅ πM ⋅ M1+ 1/2 (32.88)
12 2M + 1 K
Chapter 32 Noise-Shaping Data Converters 211
A(z) = z −1 (32.93)
M
(1 − z −1 )
and a feedback filter transfer function of
M
1 − (1 − z −1 )
B(z) = (32.94)
z −1
The block diagram of an Mth-order NS modulator is shown in Fig. 32.71.
E(z)
Nondelaying integrators Delaying integrator
X(z) 1 1 z −1 Y(z)
In 1 − z −1 1 − z −1 1 − z −1 Out
th
Figure 32.71 Generic block diagram of an M -order NS modulator.
212 CMOS Mixed-Signal Circuit Design
sin2(M+1) Kπ fs
f
fs /2
V2 2(M+1)
V 2Qe,RMS = 2 ⋅ LSB ⋅ 2 2M ⋅ 1 ⋅ ∫0 ⋅ df (32.98)
K
sin2M π fs
12f s f
f
If we let θ = π , then we get
fs
M
= K2 π⋅ Π
m=1
2m−1
2m
The change in SNR, when using the sinc averaging filter decimator instead of the ideal
filter with bandwidth, B, is given by looking at the ratio of Eq. (32.88) to Eq. (32.100)
M
2M + 1
Increase in SNR = −20 log 2 M+1/2 ⋅ Π 2m − 1 ⋅ (32.101)
m=1 2m πM
Chapter 32 Noise-Shaping Data Converters 213
For first-, second-, and third-order modulators, the difference in the SNRs is 2.16, 6.35,
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,
Figs. 31.43 or 31.46. To avoid the droop, as discussed earlier, the desired signal content is
often limited to frequencies well below fs/2K (= B). When the droop (reduction in the
desired signal amplitude) is taken under consideration, the SNR, when using the sinc
averaging filter, is worse than the ideal filter with bandwidth B.
Implementing Higher Order, Single-Stage Modulators
The single-stage, higher order modulator of Fig. 32.71 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. 32.65 and Eqs. (32.82) and (32.83), the
stability criteria of a modulator using only delaying integrators is poorer than the criteria
of the topology shown in Fig. 32.71 (where only the last integrator is delaying). While we
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
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
too long (because of the large number of integrators), the signal fed back may add to the
input signal instead of subtracting from it.
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. Towards this goal, consider the modified NS topology for higher
order modulators shown in Fig. 32.72. The forward and feedback transfer functions can be
written as
M−i+1
a 1 ⋅ z −M a 2 ⋅ z −(M−1) a ⋅ z −(M−2) a ⋅ z −1 M −1
A(z) = + + 3 + ... + M −1 = Σ a i ⋅ z −1
(1 − z −1
)
M
(1 − z −1 )
M−1
(1 − z −1 )
M−2 1−z i=1 1−z
(32.102)
or
A(z) = (z − 1) −M ⋅ a 1 + a 2 (z − 1) 1 + a 3 (z − 1) 2 + ... + a M (z − 1) M−1 (32.103)
and
b 1 ⋅ z −M b 2 ⋅ z −(M−1) b 3 ⋅ z −(M−2) b M ⋅ z −1 M 1
M−i+1
−A(z)B(z) = M
+ M−1
+ M−2
+ ... + =
1 − z −1 i=1
Σ b i ⋅
z −1
(1 − z −1 ) (1 − z −1 ) (1 − z −1 )
(32.104)
or
−A(z)B(z) = (z − 1) −M ⋅ b 1 + b 2 (z − 1) 1 + b 3 (z − 1) 2 + ... + b M (z − 1) M−1 (32.105)
214 CMOS Mixed-Signal Circuit Design
E(z)
In
X(z) a1 a2 aM
z −1 z −1 z −1 Y(z)
1 − z −1 1 − z −1 1 − z −1 Out
bM
b2
b1
th
Figure 32.72 Block diagram of a modified M -order NS modulator.
Before going any further, let's explain what we are trying to do with the modified,
higher-order, NS topology of Fig. 32.72. We know that the NTF(z), for a general
modulator, is of the form (1 − z −1 ) M with a shape seen in Fig. 32.73. At high frequencies
the modulation noise will get very large. At fs /4, for example, the magnitude of the noise
M
transfer function, NTF( f ) , is 2 (see Fig. 31.51). For the modified NS modulator
we will try to reduce the modulation noise at higher frequencies by changing the shape of
the NTF(z). Our modified NTF(z) will be of the form
M
NTF(z) = LPF(z) ⋅ (1 − z −1 ) = LPF(z) ⋅ z −z 1
M
= HPF(z) (32.106)
where LPF(z) [HPF(z)] is a lowpass [highpass] filter implemented with the feedback
coefficients bx. The goal is to flatten out the higher frequency modulation noise (keep the
noise from getting too large) thereby reducing the NTF( f ) at high frequencies and
keeping the modulator stable. One drawback of using this technique is that the signal no
longer sees just a delay in its transfer function but rather it sees the lowpass response. The
modified STF will be of the form
M
STF(z) = NTF(z) ⋅ A(z) = LPF(z) ⋅ Σ a i ⋅ (z − 1) i−1 (32.107)
i=1
NTF( f )
An example shape of a
higherorder NTF.
(1 − z −1 ) M
LPF(z) ⋅ (1 − z −1 ) M
f
B
Figure 32.73 Showing the change in the NTF in a higher order modulator.
Chapter 32 Noise-Shaping Data Converters 215
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
NTF(z) = 1 (32.108)
1 + A(z)B(z)
or
i=1
The coefficients, bx, are selected for a highpass response. Note also that our coefficients
are positive since the feedback paths, as seen in Fig. 32.72, are subtracting. The design of
the modulator at this point is to determine the feed-forward and feedback coefficients
using basic digital-signal processing filter design (and, to keep the algebra simple, a
computer program of some sort), then to simulate the design to see if it exceeds
specifications. One challenge, among others, is to meet a given SNR without causing
harmful integrator saturation.
Other topologies have been developed to implement higher-order NS modulators.
The reader is referred to Chs. 4 and 5 of [2] for further information.
32.2.2 Multi-Bit Modulators
Throughout this chapter we have assumed N = 1; that is, we have used a comparator for
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 1-bit feedback DAC.
Feedback DAC linearity is important because the output of the DAC is directly subtracted
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 SNR.
The benefits of using a multibit (N > 1) quantizer in a NS modulator are increased SNR
(see Eq. 32.89), better stability (the modulator behaves closer to the linearized theory
developed in this chapter), fewer spectral tones, and simpler digital-decimation filter. The
drawbacks of using multibit topologies, are the increase in ADC complexity (the ADC
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 32.74 shows a circuit-level implementation of a first-order, multibit, NS modulator
using a 4-bit ADC and DAC. Figure 32.75 shows the SPICE simulation outputs of this
modulator in the time and frequency domains with the same input sinewave used in
generating Fig. 32.68. Comparing Fig. 32.75 to Fig. 32.68 the decrease in modulation
noise is obvious. Note that (a) of the figure shows both the input to the modulator and the
output of the ideal DAC while (b) is the DAC's output spectrum. Looking at the output of
the DAC avoids the need for a Fourier transform on the modulator's output digital data.
216 CMOS Mixed-Signal Circuit Design
φ1 φ2
V REF+ = VDD
1p φ1
Clk
V CM 4-bit Out V out
V CM ADC
In 4-bits
0.4p V REF− = 0
V in Out 4-bit In
f s = 100 MHz DAC
(a) (b)
Figure 32.75 Output of the (a) multbit modulator and (b) its spectrum.
Most of the design effort, when developing multi-bit modulators, goes into the
design of the feedback DAC. Because it is nearly impossible to design highly accurate
DACs without trimming, or some sort of error correction, methods have been developed
that attempt to randomize DAC errors. If the 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 32.76 shows one possible implementation of a DAC that utilizes resistive
unit elements. While this figure is busy, let's attempt to explain how the DAC functions.
The DAC is based on unit-element (equal value) resistors. In one case we connect VDD to
one corner of the resistor 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 (1/8)VDD, to (2/8)VDD, ... up
to (8/8)VDD. Depending on the output of the decoder, one tap from each side is fed to the
analog output. Because there are two sides, the outputs from each side are combined and
effectively averaged.
The purpose of the counter is to vary the connections of VDD and ground around
the outside of the resistive divider to randomize variations in the output voltage due to
resistor mismatch. To understand this in more detail consider a constant DAC output
Chapter 32 Noise-Shaping Data Converters 217
b0
Decoder
8 Analog
b1 10,2 Output
b2
2,10 11,3
Counter
Clock 16
1,9 Switch
VDD bus
0,8 15,7 14,6 13,5 12,4
Ground bus
Connected to VDD when counter output is 0
and connected to ground when counter output is 8
voltage of VDD/2. As the counter changes output values, so do the connections to VDD
and ground in the resistor string. To keep a constant output voltage of VDD/2 the
switches in the center of the DAC move accordingly based on the output of the counter
and the input to the decoder. In this way variations in the resistors, hopefully, average out
to a constant value.
Multibit Demodulator (Used in a NS DAC) Implementation (Error Feedback)
The NS topologies we've discussed so far are sometimes called interpolative modulators
since the signal fed back is the average of the input signal interpolated between known
values of the modulator output (the average of the modulator outputs should be the input
signal). However, NS modulators were first introduced (see C. C. Cutler, "Transmission
systems employing quantization," 1960, U.S. Patent No. 2,927,962 [filed 1954]) using the
error feedback topology shown in Fig. 32.77. Error feedback topologies are not used in
analog input modulators because errors in the analog subtraction directly add to the input
signal. We can use this topology, however, in the implementation of a digital input
demodulator (sometimes also called a modulator), as the subtraction is digital.
218 CMOS Mixed-Signal Circuit Design
E(z)
Quantizer
F(z)
E(z)
z −1
Looking at Fig. 32.77 we note that by definition the difference between the input
and the output of the quantizer is the quantization noise, E(z). This noise is subtracted
from the input after a delay (for a first-order modulator) resulting in
Y(z) = X(z) − F(z) ⋅ E(z) + E(z) = X(z) + E(z)[1 − F(z)] (32.110)
Note that the signal transfer function for an error feedback-based modulator is simply one;
that is, STF( f ) = 1. For a first-order NS modulator we set F( z ) = z−1 (a register), which
results in
Y(z) = X(z) + E(z) ⋅ (1 − z −1 ) (32.111)
2
A second-order modulator with a NTF( f ) = (1 − z −1 ) would use a feedback filter,
noticing from Eq. (32.110) that NTF( f ) = 1 − F(z) , of
2
F(z) = 1 − (1 − z −1 ) = z −1 ⋅ (2 − z −1 ) (32.112)
Implementation of a second-order NS modulator is shown in Fig. 32.78. Note that when
trying to implement higher-order modulators using error feedback we run into the same
problem we encountered when using an interpolative modulator, namely, instability
resulting from a NTF that is too large at higher-frequencies. As with interpolative
modulators, we can design the NTF to be a highpass response.
E(z)
shift-left
Quantizer
×2
z −1
z −1
We've introduced the error feedback topology with the idea that it can be used in a
modulator (demodulator) that performs digital-to-analog conversion. We first introduced
a modulator for use in a DAC back in Fig. 32.12. At this point we need to answer the
question, "Why is the NS topology of Fig. 32.77 a better choice for DAC implementation,
in general, then the topologies of Figs. 32.12 and 32.71?" The answer to this comes from
the realization that the quantizer and difference block in Fig. 32.77 can be implemented by
simply removing lower bits from the digital input words. This is illustrated in Fig. 32.79.
The resulting error feedback modulator will be simpler to implement than the modulators
based on interpolative topologies. Figure 32.80 shows Fig. 32.77 redrawn to show the
simpler implementation.
E(z)
Figure 32.79 Showing how quantizer and difference block are implemented.
The number of bits used in the modulator, N, is selected to avoid overflow when
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.
Figure 32.81 shows the block diagram of an NS-based DAC. As we saw in Fig.
32.11, if a 1-bit output is used, the modulator can be connected directly to the
reconstruction filter (RCF). The 1-bit DAC is perfectly linear so distortion concerns are
reduced. Using a 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.
220 CMOS Mixed-Signal Circuit Design
Interpolation filter
Digital input Digital Analog output
NS (de)modulator DAC RCF
filter
Let's comment on how to estimate the quantization noise added to the signal from
the error feedback quantization process. Assuming we are using two's complement
numbers we know
1 LSB
N bits
V REF+ − V REF−
0111111... = V REF+ − (32.113)
2N
and
1000000... = V REF− (32.114)
For the DAC to function properly we must change the numbers back to binary offset
(complement the left-most or most-significant bit) unless the DAC input uses two's
complement format. This is easy to see if the output of the modulator is a single bit
(N − F = 1) since an MSB of 1 = V REF+ and a 0 = V REF− where V LSB = V REF+ − V REF− . By
dropping F bits the voltage weighting of an LSB in the modulator output can be written as
V REF+ − V REF−
V LSB = if N − F > 1 (32.115)
2 N−F
(See question 30.14 or Ex. 35.21 for further discussions.) This result is used in Eq. (32.8)
to estimate the quantization noise spectrum in a NS modulator.
Implementation Concerns
We know from our discussions in the last chapter that most digital additions and
subtractions utilize two's complement numbers because of the simplicity (see Fig. 31.55
and the associated discussion) in implementing the hardware. However, consider the two's
complement N-bit input in Fig. 32.79. 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.
To circumvent these types of problems, the topology shown in Fig. 32.82 can be
used. The input to the quantizer/subtractor is changed from two's complement format into
binary offset format. (See Figs. 31.36 and 31.37 for a comparison of the formats.)
Quantization is then performed; the lower bits are dropped from the output and fed back.
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. 31.55).
Chapter 32 Noise-Shaping Data Converters 221
MSB
N bits N bits N − F bits Y(z)
In
Two's complement Binary offset Binary offset
select MSB
Two's complement 1 0000... F bits
MUX
0 1111... Binary offset
E 1 (z)
− E 1 (z) z −1 Y 2 (z)
1 − z −1
1 − z −1
Differentiator
see Fig. 31.51
where, ideally, the output quantization noise of the first modulator, E1out(z) , is E1(z). If the
modulator is functioning properly, then GF = 1 independent of GI as discussed earlier.
Equation (32.120) can then be written as
E 1out (z) = (1 − G I1 ) ⋅ z −1 ⋅ X(z) + [1 − (1 − G I1 ) ⋅ z −1 ] ⋅ E 1 (z) (32.121)
Using this equation in Eq. (32.117) while assuming the second modulator uses an
integrator scaling factor, GI2, and GF2 is one results in (rewriting Eq. [32.118])
2
Y(z) = z −2 X(z) + z −1 (1 − z −1 )E 1 (z) − z −1 (1 − z −1 )E 1out (z) + (1 − z −1 ) E 2 (z)
Desired output Unwanted term
−1 2
= z X(z) + (1 − z ) E 2 (z) + [E 1 (z) − X(z)] ⋅ z −2 (1 − z −1 ) ⋅ (1 − G I1 )
−2
(32.122)
While we can set the second modulator's integrator gain coefficient, GI2, to 0.4 to avoid
integrator saturation, as discussed earlier, we must set GI1 as close to unity as possible.
Using a unity gain coefficient results in a reduction in the modulator's overall dynamic
range (see Fig. 32.35 and the associated discussion). Note how the input signal appears in
the unwanted term in Eq. (32.122). 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 how the number of bits in the 1-1 modulator's
output will be more than one bit (two bits if comparators are used in each first-order
modulator).
Third-Order (1-1-1) Modulators
By adding a third first-order modulator to our 1-1 modulator of Fig. 32.83 we get a 1-1-1
or third order modulator, Fig. 32.84. The output of the added third modulator can be
written as
Y 3 (z) = −z −1 E 2 (z) + (1 − z −1 )E 3 (z) (32.123)
while the ideal output of the 1-1-1 cascade is given by
3
Y(z) = Y 1 (z) + Y 2 (z) + Y 3 (z) = z −3 X(z) + (1 − z −1 ) E 3 (z) (32.124)
Again, as we saw in Eq. (32.122), noise from the first modulator can leak through to the
output and spoil the overall cascade's SNR. Indeed, if the leakage from the first modulator
is large enough, we get no benefit from adding the third modulator. Notice, in Eq.
(32.122), that the unwanted term exhibits first-order differentiation, (1 − z −1 ) . We might
expect better overall performance, that is, less leakage if the first modulator is second
order. The unwanted term would then exhibit second-order differentiation.
Third-Order (2-1) Modulators
A third-order modulator formed by using a second-order modulator followed by a first-
order modulator is shown in Fig. 32.85. The output of the first modulator is given by
2
Y 1 (z) = z −1 X(z) + (1 − z −1 ) E 1 (z) (32.125)
224 CMOS Mixed-Signal Circuit Design
Out
Y(z)
E 2 (z)
− E 1 (z)
z −1 Y 2 (z)
1 − z −1 z −1 (1 − z −1 )
E 3 (z)
− E 2 (z)
z −1 Y 3 (z) 2
(1 − z −1 )
1 − z −1
E 1 (z)
In 1 z −1 O 1 (z) Y 1 (z)
z −1
X(z) 1 − z −1 1 − z −1
Out
Y(z)
E 2 (z)
− E 1 (z)
z −1 Y 2 (z) 2
(1 − z −1 )
1 − z −1
3
Y(z) = z −2 X(z) + (1 − z −1 ) E 2 (z) +
Undesired term
−1 −1 2
(1 − G 1 G 2 ) ⋅ z −1 X(z) + [(1 − G 2 ) − (G 2 G c − G 2 )z −1 ] ⋅ z −1 E 1 (z)
z (1 − z )
1 + z −1 ⋅ (G 2 G c − 1) + z −2 (1 − G 2 G c )
(32.131)
When this equation is compared to Eq. (32.122), we see that the undesired term is second-
order differentiated. Also, we have more control over the integrator gains. Third-order
modulators using the 2-1 topology are much more robust than the 1-1-1-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 1 = G 2 = 1 .
One of the interesting uses of the 2-1 modulator is the configuration where the first
(second-order) modulator utilizes a 1-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, and
potentially ubiquitous, data converters are discussed in greater detail in Ch. 7 of [2].
226 CMOS Mixed-Signal Circuit Design
Figure 32.86 Showing implementation of the dual summing block as a single block.
One way to implement the extra subtracting input and the integrator is shown in
Fig. 32.87. This DAI is a modification of the DAI shown in Fig. 32.59. The output of this
integrator is related to the inputs by
C I2 z −1/2 C 1 C 1
V out (z) = O 1 (z) ⋅ ⋅ − Y 2 (z) ⋅ I2 ⋅ − Y 1 (z) ⋅ I22 ⋅ (32.132)
C F2 1 − z −1 C F2 1 − z −1 C F2 1 − z −1
If we set CI2 = CI22 = CF2 and we realize that the comparator in the second modulator,
assuming it is clocked with the rising edge of φ 1 (or the falling edge of φ 2 ), adds a
half-clock cycle delay in series with the Y1(z) input and a full clock cycle delay in series
with O1(z) and Y2(z), then we can write
−1
V out (z) = [O 1 (z) − Y 1 (z) − Y 2 (z)] ⋅ z −1 (32.133)
1−z
Figure 32.88 shows the implementation of a 2-1 modulator.
C F2
φ1 φ2
V CM
C I2 V CM V out (z)
O 1 (z)
V CM C I22 Y 2 (z)
Y 1 (z)
Figure 32.87 Implementing the dual summing block for a cascaded modulator.
Chapter 32 Noise-Shaping Data Converters 227
φ1 φ2 φ1
φ2 φ1
Y 1 (z)
V in
φ1 φ2 φ1
Y 2 (z)
O 1 (z)
Y 1 (z)
We could also use the topology shown in Fig. 32.89 to implement the summing
block of Fig. 32.86. This topology has the benefit of using a single capacitor for a simpler
circuit and no matching differences between CI2 and CI22 . Unfortunately, as discussed in
Ch. 27, the topology is no longer insensitive to the parasitic capacitance on the top plate
of the switched capacitor. In the parasitic insensitive topologies, Fig. 32.87 for example,
the top plate of the capacitor is always held at the common-mode voltage, VCM. In the
topology of Fig. 32.89 the top plate is charged to y1(t) when the φ 1 switches are closed
and discharged to VCM when the φ 2 switches are closed. The difference between these
voltages combined with the value of the unwanted parasitic capacitance to ground on the
top plate causes unwanted charge to transfer to the feedback capacitor and a gain error.
This by itself isn't too bad. However, the unwanted capacitance can have a large depletion
capacitance component, resulting in a voltage-dependent capacitance and thus nonlinear
gain. Nevertheless, in some applications this topology may still prove useful.
φ1 φ2
Y 1 (z)
V out (z)
V CM
O 1 (z)
Y 2 (z)
Figure 32.89 Implementing the dual summing block with a single capacitor
results in sensitivity to the top plate parasitic capacitance.
228 CMOS Mixed-Signal Circuit Design
E(z)
X(z) 1 −z −2 Y(z)
G1 G2 Out
1 + z −2 1 + z −2
Figure 32.91 shows the modulation noise spectrum of a fourth-order modulator, Eq.
(32.127). Notice how, at fs /4, the modulation noise goes to zero. The oversampling ratio
is, once again, defined as
fs
K= (32.138)
2B
If fs /4 = 25 MHz and the desired bandwidth is 50 kHz, then K = 1,000.
V 2 /Hz
f s /4 f s = 100 MHz
Figure 32.91 Modulation noise in a bandpass modulator.
CF
φ1 φ2
V CM
CI V out (z)
V CM
V 1 (z)
V 2 (z)
V CM
2C F V CM
−V out (z)
1
In a bandpass modulators 1/f noise isn't a concern since it is filtered out with the
digital filter. If the modulator is used in a wireless application where the intermediate
frequency (IF) is fs /4, extracting the real (I, or in-phase component) and imaginary (Q, or
quadrature component) becomes trivial (see Fig. 32.93). Since a single modulator is used,
and the I/Q extraction is digital, the channel mismatch encountered in typical baseband
demodulators is absent.
We might wonder, after looking at this figure, how we multiply the modulator
output, a 1-bit word, by 1, 0, and −1. We know that a 1 coming out of the modulator
corresponds to VREF+ and a 0 corresponds to VREF−. After reviewing Fig. 31.37 in the last
LPF I output
IF Bandpass
Multipliers
In modulator
LPF Q output
f s = 4 ⋅ f IF
chapter, we can convert these 1-bit outputs to 2-bit words in two's complement format. A
1 output is changed to 01 (+1) and a modulator output of 0 becomes 11 (−1) in two's
complement prior to multiplication. Multiplication results in 01 × 1 = 01, 01 × 0 = 00, 01
× −1 = 11, 11 × 1 = 11, 11 × 0 = 00, or 11 × −1 = 01. A simple MUX with some logic for
output selection can be used to implement the multiplier.
REFERENCES
[1] J. C. Candy and G. C. Temes (eds.), Oversampling Delta-Sigma Data Converters,
IEEE Press, 1992. ISBN 0-87942-285-8
[2] S. R. Norsworthy, R. Schreier, and G. C. Temes (eds.), Delta-Sigma Data
Converters: Theory, Design, and Simulation, IEEE Press, 1996. ISBN
0-7803-1045-4
[3] R. K. Hester, Introduction to Oversampled Data Conversion, Notes from a
tutorial at the 1995 International Solid-State Circuits Conference (ISSCC-95).
[4] P. A. Lynn and W. Fuerst, Introductory Digital Signal Processing, Second
Edition, John Wiley and Sons, 1998. ISBN 0-471-97631-8
[5] E. P. Cunningham, Digital Filtering: An Introduction, John Wiley and Sons, 1995.
ISBN 0-471-12475-3
[6] A. K. Ong, and B. A. Wooley, "A Two-Path Bandpass SD Modulator for Digital
IF Extraction at 20 MHz," IEEE Journal of Solid-State Circuits, Vol. 32, No. 12,
pp. 1920-1934, December 1997.
QUESTIONS
32.1 Show the details and assumptions leading to Eq. (32.1).
32.2 Would it be possible to operate the DAI of Fig. 32.3 without a 0.75 V supply?
Give an example. Show simulation results with the output initially at 0.75 V and
the same input used to generate Fig. 32.4. Are the DAI outputs the same?
32.3 Show the derivation of Eq. (32.4) from the block diagram shown in Fig. 32.6.
32.4 In the basic NS modulator shown in Fig. 32.7, what component serves as the
ADC? What component serves as the DAC?
32.5 Show, using timing diagrams, how Eq. (32.5) is correct.
32.6 Show, using SPICE simulations, how increasing the RC circuit's time constant in
Fig. 32.10 removes additional modulation noise making the output smoother.
What happens to the amplitude of the desired signal?
32.7 Show the spectrums (modulator input, output, and output after filtering) of the
signals in question 32.6. Discuss what the spectrums indicate.
32.8 Explain how the quantizer in Fig. 32.12 functions.
232 CMOS Mixed-Signal Circuit Design
32.9 What are we assuming about an input signal if the modulation noise follows Eq.
(32.7)?
32.10 What is the magnitude of Eq. (32.7)?
32.11 What is the difference between quantization noise and modulation noise?
32.12 Show the steps and assumptions leading to Eq. (32.15)?
32.13 Is the statement that on page 163 that "every doubling in the oversampling ratio
results in 1.5 bits increase in resolution" really true if K is small (say 8 or 16)?
Explain.
32.14 Does noise-shaping work for DC input signals? If so, how?
32.15 Show the steps leading up to Eq. (32.25).
32.16 What is the impulse response of the following z-domain transfer function?
−16 2
H(z) = 1 − z −1
1−z
R φ
R V CM
V in V out
C
−V out
shift left
×2
X(z) z −2 Y(z)
z −1
32.40 In Fig. 32.84 sketch the block diagram implementation of the circuit in series with
the Y2(z) output.
2
32.41 Sketch the block diagram implementation of the transfer function (1 − z −2 ) . What
kind of filter does this transfer function implement?
234 CMOS Mixed-Signal Circuit Design
33
Submicron CMOS Circuit Design
In this chapter we turn our attention toward transistor-level circuit design using a
submicron CMOS process, that is, a CMOS process with a minimum channel length, Lmin,
less than 1 µm. We divide the chapter up into three sections. The first section covers basic
submicron CMOS processes and device models. The second and third sections provide
digital and analog circuit design examples and discussions, respectively. We assume
throughout the chapter, that the reader is well grounded in the material presented in
CMOS: Circuit Design, Layout, and Simulation [1] (the first CMOS book).
Before getting too far into this chapter, let's discuss how we distinguish the
material presented here from the material presented in the first CMOS book. If we recall,
in the first CMOS book, an older CMOS process was used to illustrate the fundamental
design ideas, methods, and procedures and to provide practical models for comparing
hand-calculations and simulation results. The older CMOS devices followed the
"square-law" MOSFET model. While one can argue that the older CMOS processes will
never be obsolete because of their inherent higher voltage handling capability (a modern
CMOS process generally limits VDD, with VSS = 0, to between 1 and 3.3 V), this isn't the
main reason for using them to illustrate design fundamentals. The main reason comes from
the fact that hand-calculation parameters (transconductance and output resistance, for
example) are derived from the SPICE level 1 square-law model. An older MOSFET that
can then be modeled relatively well using the level 1 model yields simulation results that
match hand-calculations and provides immediate feedback to the designer that "I know
what I'm doing." The level 1 model can accurately model devices with an Lmin > 5 µm.
Reasonable hand-calculation accuracy, say within 20%, can be achieved in a process with
Lmin between 1 and 5 µm (in any case, though, the most accurate SPICE MOSFET model
available should be used). However, if hand-calculations, based on the level 1 square-law
model, are used in a submicron CMOS process, the error is generally well above 100%!
As an example, Fig. 33.1a shows an IV plot of a typical 0.5 µm (= Lmin ) NMOS transistor,
while Fig. 33.1b shows the level 1 SPICE results.
236 CMOS Mixed-Signal Circuit Design
ID ID
V GS = 3 V V GS = 3 V
(a)
V DS (b) V DS
L = 0.5 µm and W = 3 µm
Figure 33.1 (a) IV characteristics of a submicron MOSFET, and (b) its level 1
SPICE representation.
Looking at Fig. 33.1, we should make several key observations. To begin, notice
how the drain current, ID , varies as the square of the gate-source voltage, VGS , in the level
1 model (b), while the variation in the actual curves (a) is more or less linear with VGS.
Next, notice how the point the MOSFET enters the saturation region, VDS,sat , is defined by
VGS − VTHN (the gate-source voltage minus the NMOS threshold voltage) in the level 1
model (b). This, as was discussed in Chs. 5 and 6, results in an overestimate of VDS,sat and is
one reason body-effect is neglected so often in DC hand-calculations. Finally, notice how
the slope of the curves, when the MOSFET is operating in the saturation region, is larger
in (b) then in (a). We know that the small-signal output resistance of the MOSFET is the
reciprocal of this slope. Further then, we would expect the level 1 model to provide lower
values of gain in simulations than a model that more exactly matches part (a).
Part (a) starts with a p-type wafer or a p+ wafer with a p− epitaxial layer as
discussed in Ch. 2. The active areas are patterned by first depositing a thin oxide and a
nitride. A photoresist is deposited and patterned on the top of the nitride. The exposed
areas of nitride are removed
Part (b) shows the result of etching down into the silicon areas that are exposed,
that is, those not covered with the photoresist (or nitride after the nitride is etched off).
Part (c) shows the cross-section of the wafer after a thick oxide has been deposited
over the entire wafer and the top of the wafer has undergone polishing using chemical-
mechanical polishing (CMP). Notice how the top of the wafer is flat. Also notice the
shallow trenches are filled with a chemical vapor deposited (CVD) oxide (called STI).
STI is used in place of LOCOS because of the ability to define smaller openings in the
top of the wafer (smaller active area windows). The effective encroachment on the
devices' width is reduced and the MOSFETs can be placed closer together.
Part (d) shows the implant used for making the body of the PMOS transistors (the
n-well) and the implants used to adjust the threshold voltage.
Part (e), at the top of the next page, shows the result of patterning the polysilicon
gates on the top of the wafer.
Nitride
Pad oxide
Photoresist (after patterning)
p-type (a)
After etching
p-type (b)
Undoped polysilicon
n+ p+
STI n+ n+ STI p+ p+ STI
n+/p+ implants
p-type (g)
silicide p+
n+
STI STI STI
Add silicide
NMOS
PMOS (h)
p-type
In part (f) light, and shallow, implants are shown which are used in lightly doped
drain, or LDD, MOSFET formation.
Part (g) shows formation of the lateral oxide spacer adjacent to the gate poly (used
for LDD MOSFET formation) and the results of using implants to heavily dope the
gates, sources and drains. Implanting the gate polysilicon is important. The p+ poly
used in the PMOS formation here results in a surface device (conduction between the
drain and source occurs along the oxide/semiconductor interface) rather than a buried
device (conduction occurs through a buried channel). The threshold voltage of the
PMOS is easier to set precisely (because we don't have to counter dope the channel)
and the short channel effects become less severe. The drawbacks of switching from a
buried-channel PMOS to a surface-channel PMOS are the reduction in mobility and the
increase in flicker (1/f ) noise.
Finally, part (h) shows the cross-sectional view of the resulting NMOS and PMOS
devices after a silicide (combination of silicon and a refractory metal such as tungsten)
has been deposited. The addition of the silicide complicates the process but produces
devices that have significantly less parasitic series gate and source/drain resistance.
Chapter 33 Submicron CMOS Circuit Design 239
Cg
ε ox
C ox = Cg
t ox
Accumulation Strong inversion
Depletion
VX 400 mV = V THN VG
poly
n+ active
n-well
(a) Layout view
n+
STI n+ n+ STI
n-well
p-type
Cg
ε ox Cg
C ox =
t ox
100 mV VG
but rather assume that the capacitor and resistor time constant is so large compared to the
signal frequencies of interest that little charge, ideally zero on average, flows in the big
resistor. An increase in voltage on A, in the figure, will cause an accumulation of charge
under the left MOSFET's gate oxide. At the same time an equal and opposite charge will
appear under the right MOSFET's gate oxide. The result is that the charge accumulated on
B will be equal and opposite to the charge accumulated on A. Since the capacitors are in
series the overall capacitance seen between A and B, CAB is the series connection of each
MOSFET's own gate-oxide capacitance as seen in the figure.
VA
VDD A B
C AB = C ox /2 C AB = C ox WL/2 = C ox /2
Big (assuming each MOSFET
C AB < C ox /2 resistor is the same size)
VX
VB
VX VDD
Figure 33.5 A floating MOS capacitor.
Chapter 33 Submicron CMOS Circuit Design 241
To implement the big resistor in Fig. 33.5, a topology like the one seen in Fig. 33.6
can be used [6]. The long L PMOS device is used to generate a very small current. This
current is mirrored across and used to hold the n-well and p+ active areas at ground (on
average). Because the current is so small, the actual voltages seen in the n-well and active
areas can be considerably different than ground over short periods of time allowing the
capacitor action discussed above to occur. Finally, notice that the MOSFETs used in the
capacitor operate in accumulation as long as the voltage on either side of the capacitor is
greater than VX. For the PMOS equivalent of Fig. 33.3, VX is positive and VTHP is negative
with the source/drain and body grounded.
VDD
Figure 33.6 Implementing a large resistor for the floating MOS capacitor.
Metal Capacitors
Probably the most common method of making capacitors in a submicron CMOS process is
using the metal layers. Consider the cross-sectional view of a parallel plate capacitor
shown in Fig. 33.7. If the plate capacitance between the metal1 and metal2 dominates
because the metals have a large layout area (that is, the fringe capacitance contribution is
small), then the capacitance can be estimated using
C 12 =Area ⋅ (capacitance per area) (33.1)
If the capacitance per area is 50 aF/µm2, then it would take an area of 100 µm by 200 µm
to implement a 1 pF capacitor. While large area is a problem, it isn't the main problem with
a metal parallel-plate capacitor. The main problem occurs from the extremely large bottom
Metal2
Metal1 Insulator
Insulator
Substrate (p-type)
plate parasitic capacitance, that is, the capacitance from metal1 to substrate. This parasitic
capacitance can be anywhere from 80 to 100% of the desired capacitance. Further it
usually slows the circuit response and results in a waste of power.
To help decrease the bottom plate's percentage of the desired capacitor value,
consider the cross-sectional view shown in Fig. 33.8, where four layers of metal
implement a capacitor. The capacitance of this structure can be estimated using
C = C 12 + C 23 + C 34 (33.2)
2
If plate capacitance between each metal layer is, again, 50 aF/µm , then the area required
to implement a 1 pF capacitor is 100 µm by 66 µm. The area needed is reduced by
one-third of the area used in the metal1/metal2-only capacitor. While we used the same
plate capacitance value in between each level, we know that this will vary because of the
differing thickness in between the metals. The absolute value of the capacitors, in most
situations, isn't important but rather the ratio of capacitors is the important parameter, as
seen in the last chapter. Also notice how, in a modern CMOS process, the thickness of the
metals (made most often now with copper) increases as we move away from the substrate.
Metal4
Metal3
Tungsten plugs
Substrate (p-type)
The value of the capacitors in Figs. 33.7 and 33.8 was set by the areas of the
metals and the corresponding plate capacitance. We assumed the perimeter of the metals
and the resulting fringe capacitance was a small contributor. Figure 33.9 shows typical
minimum sizes and distances between pieces of metal1 where the fringe capacitance
dominates. We can make a capacitor using the two pieces of metal1 shown in this figure.
A typical value of capacitance per length is 25 aF/µm. The parasitic bottom plate
capacitance is half of this value or 12.5 aF/µm. Since, as seen in Fig. 33.9, the electric
fields can terminate on the close adjacent metal, the bottom component is a smaller
percentage than it was when the plate capacitance dominated.
Chapter 33 Submicron CMOS Circuit Design 243
Layout view
Cross-sectional view with field lines
Metal1
Example 33.1
Estimate the size of a metal1 only 1 pF capacitor. Also estimate the bottom
parasitic capacitance.
If we look at the layout of the capacitor shown in Fig. 33.10, we can estimate the
capacitance of a 1 µm by 2 µm section as 25 aF/µm2. This would mean that we
need an area of metal1 that measures 200 µm by 200 µm to implement a 1 pF
capacitor. The bottom plate capacitance can be estimated as 0.5 pF.
Note that while this capacitor results in twice the area af the metal1/metal2
capacitor of Fig. 33.7 and the associated discussion, it only uses one layer of metal
and the bottom parasitic is smaller. One might wonder if further benefits can be
achieved by using the fringe capacitance and multilevels of metal. T
Consider the use of metal2 and via1 in the implementation of a capacitor shown in
Fig. 33.11. While the fringe capacitance is still a major component in this capacitor
because of the addition of the via between the metals, it is sometimes called a lateral
capacitor (there exists a "plate" capacitance between the vias). A typical value of
25 aF 25 aF
1 µm
2 µm
Figure 33.11 Using two layers of metal and the via to implement a lateral capacitor.
capacitance for this structure is 200 aF/µm. The bottom plate capacitance remains
approximately 15 aF/µm. Using additional vias and metal layers will increase the
capacitance but generally not linearly. The higher levels of metal, e.g. metal4 or metal5,
generally have larger spacing and width design rules than do the lower levels of metal.
Nonetheless, using the lateral capacitor with several layers of metal and vias is the most
common way to implement a capacitor in a mixed-signal circuit.
Example 33.2
Repeat Ex. 33.1 using the lateral capacitor of Fig. 33.11.
Since the capacitance per length is now 200 aF/µm, the area required for a 1 pF
capacitor, keeping in mind that both metal1 and metal2 are used, is 50 µm by 100
µm. The bottom plate capacitance becomes 0.15 pF.
It's interesting to note that if we used four layers of metal with three vias and a
lateral capacitance value of 500 aF/µm, the area drops to 50 µm by 40 µm. T
An Important Note
While we've concentrated on the bottom plate parasitic, it is also possible to have a top
plate parasitic. Often, to avoid coupling noise into the relatively large area occupied by the
capacitor, a ground plate is placed above the capacitor. This would allow noisy digital
signals to be routed above the capacitor, as seen in Fig. 33.12.
Metal3
Resistors
Using a large number of capacitors in a circuit can result in large layout area, as just
discussed. Because of this, resistors are used whenever and wherever possible. For
example, a DAC may have been implemented using a charge redistribution topology in the
past. Now, however, it can be implemented using an R-2R topology and in considerably
less space. Table 33.1 shows the various characteristics of resistors available in a
submicron CMOS process. In the following discussion we assume that we are concerned
with implementing an R-2R DAC (see Chs. 29 and 34).
At first glance after reviewing Table 33.1, it may appear as though the n-well
offers the best choice for a resistor in a data converter since it has < 0.1% matching
characteristic. However, after reviewing the voltage coefficient specification for the
n-well, i.e., 0.008/V, we see a problem. If the voltage across one resistor in the R-2R
string is 2 V, while the voltage across another resistor is 0 V, we will see a mismatch
between the two resistors of 1.6%. Such a large voltage-varying mismatch can cause
severe linearity problems. It should be mentioned in passing that the origin of the n-well
voltage coefficient comes from the extension of the depletion region into the n-type
material used in the n-well (a problem not found in polysilicon resistors).
The polysilicon resistors available depend on the process steps. In one scenario the
poly is doped in situ while it's being deposited. Since the poly is silicided (called a
polycide) the possible pn-junction between the p+ and n+ poly isn't a concern. If the poly
is doped with an implant after it has been deposited, then a nitride layer above the gate
oxide is required to keep the implant from penetrating into the MOSFET's channel. In
either case, a silicide blocking mask is generally available to block out the siliciding of poly
(or active.)
246 CMOS Mixed-Signal Circuit Design
The p+ and n+ diffusions (the source and drain areas) can also be used for resistor
implementation. Again, the silicide block can be used to keep from siliciding the diffusions
(a silicided diffusion is called a salicide). Since the matching characteristics, temperature
behavior, and voltage coefficient are, overall, better for the polysilicon resistors, they are
generally preferred in the implementation of precision circuits such as data converters.
In general, the resistor’s width and length should be at least 10 and 100 times the
minimum feature size of the process, respectively. For example, if Lmin is 0.15 µm, then the
minimum width of the resistor should be 1.5 µm. Requiring minimum widths and lengths
for the resistors is important both for matching and to ensure that the self-heating, which
occurs because of the different current densities flowing in the R-2R resistors, doesn't
cause any noticeable differences in DAC linearity. In simple terms, the larger resistor area
dissipates heat better than the same valued resistor in a smaller area.
Figure 33.13a shows the conceptual layout of an R-2R resistor string in a minimum
area. Figure 33.13b shows the actual layout of the resistors having large width and length
along with a large number of contacts to reduce metal/resistive material contact resistance.
Figure 33.13c shows the problem of laying out metal over the resistive material, that is,
resistor conductivity modulation. The figure shows what happens when a metal, having a
potential greater than the potential of the underlying resistor is laid out directly over a
resistor. Electrons are attracted towards the surface of the resistor causing spots of lower
resistivity. The solutions to avoiding or reducing conductivity modulation are (1) avoiding
running metal over the resistors, (2) using higher levels of metal to route the resistive
signals so as to increase the distance between the resistor and the overlaying metal
(remembering vias and contacts must be plentiful to avoid adding unwanted series
resistance), or (3) inserting a conducting “shield” connected to analog ground and made
with metal1 between the resistors and the routing wires above the R-2R resistor array.
Finally, to conclude this subsection, we ask, “What is the best method of laying out
the resistors in an R-2R string to avoid process gradients and achieve good matching?”
While there are no absolute answers, we will discuss a possibility where layout area is a
concern. In other words, we won't discuss methods that use a large amount of layout area
to average out process variations but will limit our averaging to at most twice the layout
area of the R-2R string shown in Fig. 33.13.
Figure 33.14 shows one possibility for averaging process gradients using a
common-centroid configuration (see Ch. 7) with two R-2R strings connected in series. In
this figure we are assuming that the process variations change linearly with position. For
example, the first resistor in the string may have an effective value of 1k, while the
second's value may be 1.01k, and the third's value is 1.02k, etc. The normalized change in
the resistance value is shown in the figure using numbers. However, we could show that
the process gradients average out no matter what numbers are used, when using this
layout topology, as long as the sheet resistance varies linearly with position. For example,
the MSB 2R in the top string of Fig. 33.14 (on the left) has a value of 14 (6 + 8). The
MSB 2R in the bottom string (on the right) has a value of 24. Adding the values of the
two resistors, by connecting them in series, results in a resistor value of 38 (2R = 38 while
R = 19). The middle resistor value in the top string has a value of 12, while the bottom
Chapter 33 Submicron CMOS Circuit Design 247
dummy
dummy
(a)
Simplified layout
view of a resistor.
Layout of actual resistor with large
width and length for better matching
and power dissipation.
(b)
Metal at a potential higher than
the resistor will attract electrons
here
Metal
Resistor
Figure 33.13 (a) Minimal layout of R-2R string, (b) actual layout of resistor, and (c)
conductivity modulation of the resistor value.
248 CMOS Mixed-Signal Circuit Design
resistor has a value of 7. Again, adding the two resistors results in a value of 19.
Fundamentally, the limiting factor in matching then becomes the voltage and temperature
(because of the different current densities through the resistors) coefficients of the
resistors and the finite resistance of the MOSFET switches used in the DAC (discussed in
more detail in the following chapter).
Dummy
6 7 8 9 10 11 12 13 14 15 16 17 18 +5
Process gradient
Dummy
Dummy
+0
1 2 3 4 5 6 7 8 9 10 11 12 13
Figure 33.14 Two-string layout for improving matching of R-2Rs. Assumes the
resistors are connected together on higher levels of metal to avoid
conductivity modulation.
ID vs VGS. While this is an important concern, and any SPICE model used for simulating
submicron circuits should show good agreement, it won't be what we focus on here. Here
we focus on the ability of the model to transition continuously from weak to strong
inversion. While we might think that looking at the DC curves of a device (e.g., ID vs. VGS )
would show the discontinuities (kinks) between weak and strong inversion, a much better
indication is to look at several devices operating under similar, related conditions.
VDD
W
L W 1W 1W 1W 1W
L 2L 4L 8L 16 L
Consider the binary weighted current mirror shown in Fig. 33.15. In the following
discussion we assume the lengths and widths of the devices are so large that oxide
encroachment and lateral diffusion are not an issue. Clearly, in Fig. 33.15, the MOSFETs
will all be operating in the same region, for example, strong inversion. What we are going
to do, with the binary weighted current mirror, is utilize the fact that MOSFETs in series
and parallel can be combined, as seen in Fig. 33.16, to implement a test circuit to evaluate
the performance of our submircon SPICE model. Note that our test circuit will have
nothing to do, directly, with short-channel behavior, but rather it will evaluate the
fundamental implementation of the model.
Drain
Drain
Drain Drain
W
L
Gate Gate 2W Gate Gate W
W
W L L 2L
W
L L
Source Source
Source
Equivalent Source Equivalent
VDD
This circuit is useful because it now relates the current flowing in a strongly
inverted device, say M1 or M2 in Fig. 33.17, to a weakly inverted device, say M3. If the
MOSFET model is operating with a truly continuous change from one region to another
the currents will be binary-related and sum to IREF. Figure 33.18 shows the simulation
results for an eight-stage W-2W current mirror modeled with the EKV model. The 1 to
2% error in the binary currents is related to the differing drain-to-source voltages, VDS , of
the devices. The minimum length of the device modeled by the EKV model in this
simulation is 0.15 µm while the actual length used in the simulation is 5 µm (33 times
minimum). The widths of the devices used in the simulation are 20 µm and 40 µm.
Figure 33.19 shows the simulation results using a MOSFET model that doesn't
model these transitions well. If one were to use this model where the W-2W section is
used in a DAC, the designer might think the DAC performance is circuit-limited and not
matching-limited (keeping in mind that all MOSFETs are perfectly matched in a SPICE
simulation). This also points out an important point: Simulations don't always tell the
truth! The good design engineer knows the limitations of the models and the simulator.
While there are other reasons for using the EKV model (simulation speed, scaling,
well-behaved, etc.), the topic of MOSFET modeling is outside the scope of this book [5].
Chapter 33 Submicron CMOS Circuit Design 251
Model Parameters
The EKV model parameters are listed below for both NMOS and PMOS devices. Notice
that the minimum length is 0.15 µm and the minimum width is 1.05 µm. Also note that the
level used depends on the simulator. For the simulations in this book, again, we will utilize
WinSPICE. While some of the model names are discussed in Ch. 5, information
concerning the remaining names can be found in [4]. Also note that these models are
located in the file models.txt in the zip file chap33_spice.zip located at cmosedu.com.
*** SPICE Models
An Important Note
We will be using the "scale" option available in WinSPICE with the EKV model. This
option is added to a netlist as follows
.option scale=0.15u
A MOSFET specified by
M1 1 2 3 4 NMOS L=1 W=10
would indicate that the MOSFET has a length of 0.15 µm and a width of 1.5 µm. A
possible mistake, when writing a netlist manually, would be to forget to add the scale
parameter.
Figures 33.20 and 33.21 show example IV plots for both NMOS and PMOS
devices. The NMOS device is sized 10/1 (actual width 1.5 µm and length 0.15 µm). The
PMOS device is sized twice as large as the NMOS, that is, 20/1 so that its current levels
are similar to the NMOS device. In our example process used in the book VDD is 1.5 V.
ID ID
V SB = 0 V GS = 1.5 V DS = 1.5 V
V GS = 1.25
V SB = 0
V SB = 1.25
V DS V GS
ID
10/1
D
G B
V DS
S
V GS V SB
ID ID
V BS = 0 V SG = 1.5 V SD = 1.5 V
V SG = 1.25
V BS = 0
V BS = 1.25
V SD V SG
ID
20/1
D
G B
V SD
S
V SG V BS
1.5
W/L
V DS
S D Rn =
ID
ID
V DS
W/L = 10/1
Rn
20/1
50/1
V DS
Figure 33.23 NMOS effective resistance from Fig. 33.22.
Chapter 33 Submicron CMOS Circuit Design 257
Example 33.3
Estimate, and verify with a SPICE simulation, the delay time in the following
circuit (Fig. 33.24).
Using Eq. (33.3) the effective digital resistance of the MOSFET is 1k. The
propagation delay (input going high and output going low) can then be estimated,
knowing the load capacitance is much larger than the MOSFET capacitances,
using
t PHL = R n ⋅ C L = 1k ⋅1 pF= 1 ns
The SPICE simulation results are shown in Fig. 33.25. T
V in
V out
t PHL
Figure 33.26 shows the test circuit used to determine the PMOS effective digital
switching resistance, Rp. From the simulation results shown in Fig. 33.27 we can write
1 + K p ⋅ (VDD − V THP ) L
R p = 20 kΩ ⋅ L ≈ ⋅ (33.4)
W KPp W
Again this is an average estimate for the resistance for all source-to-drain voltages. Also
note how it doesn't matter if we use actual device sizes or scaled sizes in this equation
because of the ratio.
258 CMOS Mixed-Signal Circuit Design
W/L
1.5 − VD
D S Rp =
ID
VD ID 1.5
Rp W/L = 20/1
40/1
100/1
VD
Figure 33.27 PMOS effective resistance from Fig. 33.26.
Bidirectional Switches
The NMOS and PMOS switches shown in Figs. 33.22 and 33.26 make use of the fact that
the NMOS source is connected to ground and the PMOS source is connected to VDD.
For complementary static CMOS logic design we can rely on Eqs. (33.3) and (33.4) to
estimate the MOSFET sizes for a particular drive strength. However, if the MOSFET is
used as a pass gate where current can flow bidirectionally, the effective switching
resistance can become very large. This is related to the fact that the PMOS switch can't
pass a logic low (0 V) well and an NMOS switch can't pass a logic high well (VDD). Of
course, combining the NMOS and PMOS into a transmission gate (TG) eliminates this
concern at the price of larger layout area and the need for two complementary clocks.
Chapter 33 Submicron CMOS Circuit Design 259
Figure 33.28 shows the NMOS device used as a switch with the input at VDD (=
1.5 V here). Figure 33.29 shows how the effective switching resistance of this device
changes with size under various output voltages. Notice that, as we would expect, Rn gets
very large as the output approaches VDD − VTHN (with body effect).
1.5
1.5 − V out
W/L Rn =
ID
In Out
ID
1.5 V out
W/L = 10/1
Rn
20/1 50/1
V out
Figure 33.29 NMOS effective resistance from Fig. 33.28.
Another application where a switch can be used bidirectionally was in the DAI
discussed in the last chapter. If the switches used in the DAI can be replaced with NMOS
devices the implementation is simpler. If we increase the voltage of the gate signal, it is
possible to turn the MOSFET all the way on and pass VDD from the switch input to its
output. Figure 33.30 shows the results if the gate signal in Fig. 33.28 is increased to a
value of 2.3 V.
260 CMOS Mixed-Signal Circuit Design
W/L = 10/1
Rn
20/1
50/1
V out
Figure 33.30 NMOS effective resistance from Fig. 33.28 with gate at 2.3 V.
Toward the goal of increasing the amplitude of the clock signal controlling the
switches consider the bootstrapped clock driver circuit shown in Fig. 33.31. This circuit is
a simple implementation of a voltage (charge) pump discussed in Ch. 18. The output
amplitude of the circuit approaches 2VDD so concern for oxide damage or long-term
failure is warranted as discussed in Ch. 6. The inverters are sized with 100/1 PMOS and
100f 1,000f
100/50
100/50 100/1
Clock in Clock out
100/1
50/1 NMOS devices. The other devices are sized to minimize power while supplying a
reasonable level of output drive. For example, the 100 fF capacitor only has to supply
charge to the gate of M2, but the 1,000 fF capacitor supplies charge to both the gate of
M1 and the load. Further scaling is possible to further reduce power and enhance output
drive. Note this circuit is noninverting and can be used in a nonoverlapping clock
generator, as shown in Fig. 33.32.
φ1
CLK
P P φ 1D
P P φ 2D
φ2
The clock generator of Fig. 33.32 provides the two phases of a clock signal as well
as slightly delayed clocks for use in a sample-and-hold, see Ch. 27. As the simulation
results show in Fig. 33.33, the output amplitude is approximately 2.5 V when a 100 fF
load is connected to each phase of the output. The time that all four clock signals are low,
the nonoverlap time (dead time), can be increased by increasing the delay in series with the
output of the NOR gates. Adding inverter pairs to the outputs of each NOR gate is a
common method of increasing the dead-time. Also note that the capacitors in Fig. 33.31
can be implemented using NMOS devices since they will always be in strong inversion.
A Clocked Comparator
One of the circuits that we used often in Ch. 32, in NS data converters, was a clocked
comparator. Let's develop a clocked comparator using inverters and switches. Examine
the evolution of circuits shown in Fig. 33.34. In parts (a) and (b) the basic inverter- based
latch is shown. In order to reduce power and make the comparator clocked, we add the
NMOS switch shown in part (c). Before each comparison, we want to ensure the
comparator is equilibrated. To erase the memory of the previous comparison, the PMOS
switch in part (d) is added to our circuit. When the clock, φ, goes high, the PMOS device
turns off and the NMOS device turns on allowing the inverters to latch in a stable
condition. The only thing we need to add to this circuit is circuitry to somehow create an
imbalance across the inverters when φ goes high. We can do this several ways. One
possibility is shown in Fig. 33.34e.
262 CMOS Mixed-Signal Circuit Design
φ
φ
out−
V out+
V out−
out+
V in+ V in−
V in− = 0.75 V
φ
out−
out+
V out−
V out+
One last comment before leaving this topic: Notice, after reviewing the simulation
netlist, how we used voltage sources for our comparator inputs. In any practical
simulation it is a better idea to add some series resistance between the comparator inputs
and the voltage sources to simulate finite source impedance. Resistors with 10k values are
typical for simulations of this nature. Using these resistors shows kickback noise on the
inputs to the comparator. This noise is often the fundamental limitation of the
comparator's performance in an actual circuit. Methods to reduce this kickback, such as
cascoding the input common-source amplifiers or using an additional amplification stage
such as a diff-amp, should be used in a general application. Note that kickback noise
shouldn't be a problem in the comparator used in a noise-shaping data converter due to the
large capacitance connected to the input of the comparator (the DAI's integrating
capacitance).
Common-Mode Noise Elimination
High-speed digital signals can often appear more like a sinewave than a square wave.
When these sinewave-like signals are applied to the inputs of a digital gate, the timing
delay between the gate's inputs and output can vary. This delay variation can be the result
of noise coupled onto the wires used to connect the circuits together or because of
power-supply fluctuations. To avoid these problems in analog circuits, as discussed on
page 191 and in Chs. 25 and 27, fully-differential outputs are used. Using fully differential
outputs in a digital system, however, can result in large layout area. The gates must have
both differential inputs and outputs. Because of this and the large noise margin of digital
signals, generally only signals that propagate over long distances will experience
Chapter 33 Submicron CMOS Circuit Design 265
detrimental noise corruption. However, any digital signal can be corrupted because of
power-supply fluctuations. In this section we discuss the idea of common-mode noise
elimination, CMNE. The CMNE circuit will, ideally, eliminate common-mode noise on a
wire pair while, at the same time, not affect the differential component. The technique can
be used to "square-up" digital signals. We apply this technique to the design of a
high-speed digital buffer.
Toward the design of a CMNE circuit, examine Fig. 33.36a. This circuit was the
heart, or decision circuit, of our comparator in Ch. 26. Assuming for the moment that all
transistors are sized equally we know that when io+ is greater than io−, the output vo+ is
greater than vo−. Any common signal to both io+ and io− will not be a factor in which signal,
vo+ or vo−, is larger. This circuit can only function if both io+ and io− are positive. Figure
33.36b shows the decision circuit of part (b) where we've added a PMOS decision circuit
so that the currents io+ and io− can be positive or negative. The schematic representation of
this circuit using inverters is shown in Fig. 33.36c.
VDD
i o+ i o−
i o+ i o−
v o+ v o−
v o+ v o−
i o+
v o+
v o−
i o−
To determine quantitatively how this circuit functions, consider the inverter output
vs. input plot shown in Fig. 33.37. The switching point, VSP , is defined as the point where
the inverter's input and the output voltages are equal. This is also the DC operating point
of the CMNE circuit in Fig. 33.36c with zero input current. The inverter's model is also
shown in this figure. When vIN increases, vOUT ( = g m ⋅ v IN ⋅ r o ) decreases. The actual
values of gm and ro are not needed to understand the operation of the CMNE circuit.
20/10
v IN v OUT v OUT
v OUT V SP
v IN g m ⋅ v IN ro
Inverter model v IN
V SP
Consider the CMNE circuit shown with the inverter currents in Fig. 33.38. In this
circuit the input current is made up of a common-mode component, ICM , and a difference-
mode component, idiff, so that
i o+ = i CM + i diff (33.5)
g m ⋅ v o+
v o+
g m ⋅ v o− ro
i CM i diff g m ⋅ v o+ 2
v o−
ro
i CM 2
i diff g m ⋅ v o−
and
i o− = i CM − i diff (33.6)
The resistors with values ro/2 connected to the two signal nodes model the parallel
combination of the two inverters' output resistance at each node. The fact that there are
only two signal nodes in this circuit is an important point since we don't want the CMNE
circuit to affect the desired, fully-differential signals. We want the desired signal to pass
through the CMNE circuit without any delay. If we sum the currents at each node, we get
v
i diff = g m v o− + g m v o+ + o+ + i CM (33.7)
r o /2
and
v o−
−i diff = g m v o− + g m v o+ + + i CM (33.8)
r o /2
Taking the difference in these two equations results in
v −v
i diff = o+ r o o− (33.9)
NMOS 10/1
PMOS 20/1
v in v out
V REF or v in
v out
v in
Figure 33.40 Simulating the circuit in Fig. 33.39 showing switching points
for references of 0.7 and 0.9 V.
One might wonder, after reviewing Fig. 33.40, if it's possible to increase the gain
of the buffer by adjusting the sizes of the inverters. We see that it takes over 300 mV input
signal to cause a full logic transition on the buffer's output. Also, it would be nice if the
buffer could be designed to have hysteresis, as discussed in Ch. 26. While we can increase
the length of the devices used in the inverters to both increase the gain and decrease the
power dissipation, the trade-offs are decreases in speed (the delay through the buffer) and
frequency response of the CMNE circuit. At very high frequencies the inverters aren't fast
enough to eliminate common-mode noise.
After reviewing how we introduced hysteresis into our decision circuit back in Ch.
26, we see that by increasing the length of the MOSFETs used in the shorted input/output
inverter we can achieve the same results in the buffer of Fig. 33.39. Consider Fig. 33.41, a
redrawn version of Fig. 33.38, with weak inverters. By weak we mean that the W/L ratio
Chapter 33 Submicron CMOS Circuit Design 269
v o+
g m ⋅ v o− ro
i CM i diff 2
g m ⋅ v o+
v o−
ro
i CM 2
i diff g mw ⋅ v o−
Figure 33.41 Figure 33.38 redrawn with weak inverters for hysteresis.
of the transistors used in this inverter is smaller than the other W/Ls used in the circuit.
We indicate that the modified inverter's gm is now weak, and thus less than the other
inverter's gm, by relabeling it gmw. We can rewrite Eq. (33.9) as
v −v v −v
i diff = o+ r o− + (g mw − g m ) ⋅ o+ o− (33.13)
o 2
or
2v diff
i diff = (v o+ − v o− ) ⋅ r1 + 1
(33.14)
o 2/(g mw − g m )
This equation can be written as
v diff = i diff ⋅ o g 1− g
r
(33.15)
2 mw m
Since gmw is less than gm, the effective resistor added in parallel with ro is negative having
the effect of increasing the resistive loading of the CMNE circuit. This increase has the
effect of boosting the differential gain of the buffer of Fig. 33.39 and introducing
hysteresis into the switching point. Note that if gmw is equal to gm, then Eq. (33.15) reduces
to Eq. (33.12).
The allowable range of VREF is set by the threshold voltages and the power
supplies. For example, VREF can be no larger than VDD − VTHP. In reality VREF is limited to
voltages less than this because the drive of the inverter connected to VREF decreases.
Notice, in Fig. 33.40, how the final output crossover point is creeping toward VDD.
The CMNE technique can be used in a variety of places including delay-locked
loops or input and output buffers or simply to balance two differential digital signals.
While we used CMOS inverters here the technique can be extended to use inverting
amplifiers. We will extend this technique to balancing the outputs of fully-differential
op-amps later in the chapter.
270 CMOS Mixed-Signal Circuit Design
Example 33.4
Resimulate the buffer shown in Fig. 33.39 if the length of the MOSFETs is
increased to 2. Rerun the simulation if weak inverters, Fig. 33.41, are used with
their Ls increased from 2 to 2.1.
Figure 33.42 shows the simulation results. In part (a) the increase in gain with the
increase in L from 1 to 2 is evident. The current drawn from VDD decreases from
1 mA to 0.5 mA. Note in this figure we show VREF = 0.5, 0.7, and 0.9 V. The
references at 0.5 and 0.9 are at the edge of the allowable reference voltages, VREF.
In part (b) the small increase in the weak inverter's length (only 5%) shows that the
gain is improved further. This small increase introduces little hysteresis into the
input buffer. Increasing the weak inverter's length further, say to 3, causes both the
gain and hysteresis to further increase. T
V REF = 0.9 V
v out and v out V REF = 0.7 V
V REF = 0.5 V
(a)
v in
V REF = 0.5 V
(b)
v in
Figure 33.42 Simulation results for the buffers described in Ex. 33.4.
Chapter 33 Submicron CMOS Circuit Design 271
Master Slave
In Out Out
In
clk clk
clk
Figure 33.43 Simple delay element using pass transistors and CMOS inverters.
Reset
Out
clk clk
In Out
In
clk
clk
clk
Master Slave
Both cells in Figs. 33.43 and 33.44 require the use of two clock signals. Figure
33.45 shows a delay element [7] that functions as an edge-triggered D flip-flop with a
single clock signal. The circuit technique used to implement this delay is termed true
single-phase clocking (TSPC). As seen in Fig. 33.45, both true and complement outputs
are available (the true output being buffered). The current pulled by this delay cell, when
clocked at 100 MHz, is below 10 µA. The layout area is comparable to the layout area
used by the clocked CMOS delay in Fig. 33.46.
Out Out
clk
Out
clk clk (Q) In
In clk
(D)
clk clk/2
clk/2
clk clk Q
clk
D
D Q b2 D Q
clk Q clk Q b2
D Q b1 D Q
clk Q clk Q b1
D Q bo D Q
clk Q clk Q bo
Ripple up counter Ripple down counter
Synchronous up counter
Example 33.5
Sketch the implementation of a synchronous up/down counter. Discuss its
operation.
Figure 33.48 shows the basic block diagram of the counter. An adder is used to
either add or subtract one from the contents of the register. Two's complement
numbers are used in the adder to avoid overflow problems (see Ex. 31.22). The
MUX selects either +1 (= 0000...1) if UP is high or −1 (= 1111...1) if UP is low to
add to the contents of the register. T
D Q bn
clk
select, UP
000...1 UP
MUX D Q b1
111...1 clk
UP
D Q bo
clk
clk
33.2.3 An Adder
The last digital building block we will look at in this chapter is the adder. An adder was
used in all of our digital filters discussed in Chs. 31 and 32. While there are many ways to
implement adders, here we discuss ripple adders using dynamic logic. Again these designs
result in low power and a small layout area. If the delay through the adder is too long, for
a specific application, we use pipelining, Fig. 33.49, to segment the adder's internal delays.
Note that using pipelining results in a delay in series with the adder. The adder in Fig.
33.49 would have a z−3 delay in series with the output signal. In a practical circuit we
would segment 4-bit, or more, adders instead of the single bit adders shown in the figure.
The output of a 1-bit adder can be written as
s out = a in ⋅ b in ⋅ c in + (a in + b in + c in ) ⋅ c out (33.16)
where ain and bin are the adder's inputs, while cin is the carry input. The carry output can be
written as
c out = a in ⋅ b in + c in ⋅ (a in + b in ) (33.17)
Chapter 33 Submicron CMOS Circuit Design 275
1-bit
b4 c out
Latch
Latch
Latch
a4 s4
c4
1-bit
b3
Latch
Latch
Latch
a3 s3
c3
1-bit
b2
Latch
Latch
Latch
a2 s2
c2
1-bit
b1 Carry out
Latch
Latch
Latch
a1 s1
Carry in c1
0
Figure 33.49 A 4-bit pipelined adder. The latches (clocked) behave as delay elements.
Figure 33.50 shows the implementation of a dynamic adder. The first stage generates the
carry out and is implemented in NMOS precharge-evaluate (PE) logic. The second stage is
implemented using PMOS PE logic. The overall gate can be thought of as a domino gate.
The output of the adder is valid when clk is high.
c out
a in
a in c in b in c in
b in a in
a in
b in b in
c in
clk clk
c out
s out
VDD VDD
V ot2 M7 M8 V ot1
VDD VDD
V biasp V biasp
VDD VDD
M1 M2
V inp V inm
M3
M4
V biasn V biasn
M5 M6
V ob2 V ob1
have the same threshold; 0.4 V at room temperature in a typical process run. We will
design our circuits so the PMOS and NMOS devices have the same gate-to-source
voltage. Writing KVL from VDD to ground through M8, M2, M3, and M5 results in
VDD = V SG8 + V DS2 + V SD3 + V GS5 (33.18)
or
VSG8 VDS2 VSD3 VGS5
10/1 40/4
30/3
20/2
20/2
40/4
10/1
(a) Drain current vs. drain-source voltage (b) Output resistance vs. drain-source voltage
V DS
0.5 V
Figure 33.52 NMOS curves for 40/4, 30/3, 20/2, and 10/1 devices.
20/1 80/4
60/3
40/2 40/2
80/4 20/1
(a) Drain current vs. drain voltage (b) Output resistance vs. drain voltage
1.5 V
V SG = 0.5 S
1.0
VD
Figure 33.53 PMOS curves for 80/4, 60/3, 40/2, and 20/1 devices.
Chapter 33 Submicron CMOS Circuit Design 279
region. The circuit-design techniques that we use for submicron circuit design should
allow reasonable gains even if our transistors move into the triode region with temperature
or process variations.
Small-Signal Transconductance, gm
Figures 33.54 and 33.55 show the test circuits and simulation results used to determine
each device's small-signal transconductance, gm. Note how we adjusted the DC drain-to-
source voltage so that the devices are operating in the triode region (see Figs. 33.52 and
33.53). Note also how we've tried to size the PMOS device to have similar characteristics
as the NMOS but the gm is still half that of the NMOS device.
i d = g m v gs g m , µA/V
20/2
0.1 V
v gs
0.5 V
g m , µA/V
1.5 V
40/2
v sg
i d = g m v sg
1.0 1.4
discussed back in Chs. 9 and 25. Figures 33.56 and 33.57 show how fT is determined from
simulation results. Large fT indicates high speed. From a circuit point of view
V GS
fT ∝ (33.20)
L
Using minimum-length devices with large gate-to-source voltages results in high-speed
operation with low gain (because of the small output resistance, as seen in Figs. 33.52 and
33.53) and reduced output swing (using a large VGS results in devices that enter the triode
region with relatively large drain-to-source voltages).
id
id
ig ig
20/2
1.5 V
v gs
0.5 V
fT
1.5 V
id
ig ig
40/2
v sg
id
1.0
fT
MOSFET's threshold voltage. The 80/1 device, M2, is large relative to M1 so that its VGS
is approximately 0.4 V (the threshold voltage). Because M2's VGS is 0.5 V, there is 100
mV dropped across the 10k resistor (and this sets the current). In a practical circuit,
subject to process variations, we can adjust the resistor value by adding series-shorted and
parallel resistors to the source of M2. This makes adjusting the current to a specific value
possible. Note, as discussed in the homework problems of Ch. 21, adding a capacitance to
ground at the source of M2 can result in an unstable circuit. The reference can actually
oscillate. This may be a problem if the resistor is bonded out to set the current value.
Adding capacitors to ground or VDD at Vbiasn and Vbiasp, however, can often be useful to
reduce coupled noise to the bias voltages.
VDD
M5
M3 M4 40/2
10/20 40/2 V biasp
10/1
M7
10/5 80/1
20/2 V biasn
M6
M1 M2
Start-up circuit 10k
T=0C
T = 25 C
Reference current
T = 100 C
T=0C
VDD
V bias1
Current through vtest.
40/2 T=0C
V bias2 T = 100 C
40/2
vtest
Figure 33.61 Using the general purpose biasing circuit of Fig. 33.60.
Figure 33.62 shows how the currents in each leg of the diff-amp of Fig. 33.51
change as we sweep Vinp with Vinm held at 0.75 V. Note how, with Vinp = Vinm = 0.75 V, the
current in each leg of the diff-amp is 6 µA or less than the current in the biasing circuits.
This is the result of the body effect experienced by the MOSFETs in the diff-amp.
T=0C
T = 25 C
T = 50 C
T = 75 C
T = 100 C
Figure 33.62 DC sweeps showing the currents in each leg of the diff-amp of Fig. 33.51.
284 CMOS Mixed-Signal Circuit Design
VDD = 1.5 V
400/2
vout
vin vout
200/2
vin
Looking at the transfer curve shown in Fig. 33.63, we see that the gain of the
inverter (the slope of the curve) is largest when the output falls between 0.25 and 1.25 V
(an output swing of 1 V). Throwing away 0.5 V, or 33%, of the power-supply voltage on
a MOSFET operating in the triode region is, of course, highly undesirable. Thinking about
this for a moment we may realize that if this (second) stage is preceded by a high-gain
(first) stage the op-amp may still function within specifications when the second stage gain
is dropping. (Though the distortion introduced by the output stage may be too high
because of this nonlinearity.) A more important concern then is the quiescent current
pulled through this stage. For the inverter in Fig. 33.63 this current is > 1 mA. For both
gain and power reasons we need to modify this basic output circuit.
Chapter 33 Submicron CMOS Circuit Design 285
To lower the quiescent current pulled from VDD and to increase the linear output
swing of the output stage consider adding batteries to the circuit as seen in Fig. 33.64.
We've selected the batteries so that when vin is 0.75 V the gate-source voltages of the
MOSFETs are 0.5 V. From Figs. 33.52 and 33.53 we can estimate the quiescent current in
the output stage as 150 µA (simulation results verify this estimate). Clearly, increasing the
battery voltages will make the output swing approach the power supply rails before the
MOSFETs enter the triode region. This increase in linear output swing comes at the cost
of speed, as indicated by Eq. (33.20).
VDD = 1.5 V
400/2
0.25
vout
vin vout
0.25
200/2
vin
The next question becomes, "How do we implement the batteries in Fig. 33.64?"
The batteries must track both process and temperature changes. What we want is
something like what is seen in Fig. 33.65. Using Vbiasn2 from the circuit of Fig. 33.60, we
can precisely set the current in the output stage. Looking at this figure for a moment, we
VDD = 1.5 V
400/2
V biasp2
VDD MC1
vout
V biasn2
MC2
200/2
Two ga
te-sourc
e voltag
es
see that we need to somehow couple our input signal to the gates of the output MOSFETs
and we need to provide a path for current flow in the added transistors. As drawn MC1
and MC2 are off.
Figure 33.66 shows the use of a floating current source (discussed in Ch. 25) to set
the bias current in the output stage. Note that since the current in the cascoded transistors
splits between MC1 and MC2, we have reduced their size by one-half in order to set the
current in the output MOSFETs to precisely ten times the current in the remaining
MOSFETs. By further reducing the size of MC1 and MC2 (reducing W or increasing L),
we can choke off the current flowing in the output transistors. While this will push the
frequency of the pole located on the output of the op-amp downwards, ultimately making
stability a concern, it may be used to reduce the quiescent power of the op-amp. The input
signals, Vot1 and Vob1, come from the diff-amp shown in Fig. 33.51. The cascode transistors
provide the "first stage gain" so that the circuit shown in Fig. 33.66 is an op-amp minus
the diff-amp. The structure is biased so that it can function with very low power supply
voltages. The limitation on how low the power-supply voltages can go is set by the
diff-amp of Fig. 33.51.
VDD
V ot1
VDD
V bias2
400/2
V biasp2
20/2
MC2 MC1
V biasn2 10/2
200/2
V bias3
M12
V ob1
M11
All unlabeled PMOS are 40/2
All unlabeled NMOS are 20/2
Figure 33.66 Biasing a push-pull output stage with floating current source.
Example 33.6
Consider the AC small-signal simplification of the floating current source shown in
Fig. 33.67. Assuming the NMOS cascode output resistance is labeled Rncas, what is
the small-signal resistance seen by the test voltage, vtest?
Chapter 33 Submicron CMOS Circuit Design 287
To PMOS cascode
i test
S
i dmc2
MC2 MC1
v test r on r op
i dmc1
S
R ncas i test
What we are going to show is that the floating current source will not load, or
decrease, the resistance seen by the cascode structures. Writing a KVL from vtest to
ground results in
v test = (i test − i dmc1 − i dmc2 ) ⋅ r on r op + i test ⋅ R ncas
where the drain currents of MC1 and MC2 are idmc1 and idmc2, respectively, and their
output resistances are ron and rop. The drain currents can be written as
i dmc1 = g mn ⋅ v gs = g mn ⋅ (−i test ⋅ R ncas )
and
i dmc2 = g mp ⋅ v sg = g mp ⋅ v test
Combining these equations
≈ g mp ⋅r on r op ≈ g mn ⋅r on r op ⋅R ncas
drain current. Now, however, the current through M12 is constant (its gate is held at Vbias3,
while its source is held at a fixed potential). What we need to add to this basic circuit is a
circuit that will hold the drain of M11 at a fixed potential while at the same time adjust the
gate voltage of M12 so that it can turn on.
Figure 33.68 shows adding N and P diff-amps to help with adjusting the biasing so
that slewing isn't a concern. Figure 33.69 shows the circuit implementation of the
amplifiers. The source followers were added in Fig. 33.69 so the inputs can go to the
power supply rails and to reduce the input capacitance. Adding the N and P amplifiers to
the amplifier of Fig. 33.66 (called gain-enhancement back in Ch. 25) also increases the
output resistance of the cascode stack. This is important because we want the pole
associated with this node (the output of the first stage or the input to the second stage) to
be dominant so that it compensates the op-amp. The voltages Vhigh and Vlow are generated
with the bias circuit of Fig. 33.60 for, once again, the widest possible operating range. As
discussed in Ch. 25, we can increase the gain of the op-amp by increasing the gains of
these added amplifiers.
VDD
V ot1
VDD
V high N
V biasp2
V out
V biasn2
V low
P
V ob1
Device sizes as seen in
Fig. 33.66
Figure 33.68 Adding amplifiers to our op-amp to boost gain and help slew-rate.
Reviewing Fig. 33.62 we see that the current sourced by our diff-amp is very
limited. Let's say that there is 5 µA of current available to charge the output transistors in
Fig. 33.68. We can estimate the input capacitance of these two transistors using
Chapter 33 Submicron CMOS Circuit Design 289
V biasp
out
P
Out
VDD
VDD VDD
Out
out
N
V biasn
Figure 33.69 Diff-amps with source-follower level shifters for use in Fig. 33.68.
C ox
Wp Wn
L
ε
C 1 = 400 + 200 ⋅ 2 ⋅(scale) 2 ⋅ ox = 1, 200 ⋅ (0.15 µ) ⋅
2 35.13 aF/µm
= 237 fF
t ox 0.004 µm
(33.21)
The rate we can charge this input capacitance is
5 µA dV
= = 21 mV/ns (33.22)
237 fF dt
At first glance this may appear to be a significant limitation if the output of our op-amp
has to change by 1 V or more during a 10 ns clock cycle. However, after reviewing Fig.
33.64 we see that well under 50 mV change on the input of these output transistors is
needed to cause the output to change from rail to rail. One final comment: To balance the
drive to the output transistors, a capacitor can be added in between each of the gates of
290 CMOS Mixed-Signal Circuit Design
the two transistors. This capacitor doesn't affect the compensation or the speed, it simply
makes the biasing appear more battery-like, as seen in Fig. 33.64. We also added
capacitors in the diff-amp of Fig. 33.51 so that the source-followers used for biasing
appear more battery-like (as discussed in Ch. 24).
Now that we've calculated the capacitance on the output of the first stage, C1, we
can estimate the location of the dominant pole. The voltage gain of the cascode section is
approximately (gmro)2. When the gain enhancement amplifiers (N and P in Fig. 33.69) are
added the voltage gain increases to (gmro)3 noting this is the gain to the output of the first
stage and not the final gain of the op-amp. If we pull the transconductance of the diff-amp
out of this equation, we estimate the cascode output resistance as g 2m r 3o (= R1). Using Figs.
33.52 - 33.55 we can estimate R1 as 150 MΩ. The location of the dominant pole is then
estimated as f 1 = 1/(2π ⋅ 150 MΩ ⋅ 237 fF) = 4.4 kHz . This pole can be pushed lower in
frequency, further compensating the op-amp, by adding capacitance from the gates of the
output MOSFETs to AC ground (ground or VDD).
The observant reader may ask, "If we can get away with only 5 µA of bias current
in our amplifier and avoid slew-rate limitations, why use the diff-amp of Fig. 33.51?" As
we discussed earlier, this diff-amp represents the weak link in the minimum power supply
voltage we can use with our op-amp and it dissipates more power than an equivalently
biased regular diff-amp. Also, the input common-mode range of this diff-amp is very
limited. We won't be able to use the op-amp as a simple voltage follower. Because of
these concerns/reasons we won't use this topology for our basic op-amp design. (The
noise performance, discussed in the next section, is also poorer for this diff-amp mainly
because the source followers used for biasing are in series with the input signal.)
Figure 33.70 shows one possible mixed-signal op-amp topology. We used diode
connected, cascoded MOSFETs to generate Vot1 and Vob1 in the diff-amp in an effort to
equalize all drain-source voltages. The minimum supply voltage, because of the diff-amp
used, can be significantly less than 1.5 V (approaching 1 V for a typical process run). One
potentially important concern is the negative common-mode range of the diff-amp. If the
inputs are held at 0.75 V, with the gate-source voltage of the diff-amp pair at 0.5 V, there
will be only 0.25 V left to drop across the diff-amp's current sink. The result is the
diff-amp's biasing current may decrease (and so will the common-mode rejection ratio).
Using a single MOSFET to bias the diff-amp or increasing the widths of the diff-pair can
improve this situation.
The gain of this topology is very high. This can lead to simulation problems (which
has led us to use HSPICE in the following simulation results). Figure 33.71 shows the
output of the op-amp as a function of the noninverting input voltage of the op-amp with
the inverting input held at 0.75 V. The systematic offset voltage is approximately 3 mV.
Not cascoding the diode-connected loads of the diff-amp can result in a significantly larger
offset voltage. Increasing the widths of the diff-amp pair reduces the systematic offset,
and, if laid out properly, reduces the random offsets.
Figure 33.72 shows the open-loop gain of the op-amp (approximately 110 dB at
DC). As is, the op-amp will be unstable if used in a unity gain configuration. The dominant
Chapter 33 Submicron CMOS Circuit Design 291
V ot1
V bias2 VDD
V high N
400/2
V biasp2
20/2
V out
V biasn2
10/2
V low 200/2
P
V bias3 40/2
V bias4 V ob1
40/2
pole, as discussed above, is at the output of the first stage (input to the second stage). We
might think that using different N and P diff-amps (or some other operational
transconductance amplifier) with a higher gain will help improve the stability. While
increasing the gain of these amplifiers will push the dominant pole to a lower frequency, it
will also increase the low-frequency gain having little effect on the stability.
Figure 33.73 shows how adding two 250 fF capacitors to the output stage pushes
the dominant pole downwards and makes the op-amp stable. The unity gain frequency of
the op-amp is approximately 70 MHz. Unfortunately, the settling time of the op-amp
increases. Figure 33.74 shows a test configuration where the op-amp, driving a relatively
large 5 pF capacitor, has a settling time of approximately 50 ns. (Figures 33.72 and 33.73
were generated without a load; so the unity gain frequency with a load will be less than
what is shown in these figures.)
While reducing the capacitive load decreases the settling time, we may not have
this option available. Let's discuss the design of an op-amp and the trade-offs if we use the
basic topology of Fig. 33.70 to implement a mixed-signal op-amp.
1. The load is purely capacitive. We may consider eliminating the floating current
source and the push-pull output stage (and note that the inverting and noninverting input
292 CMOS Mixed-Signal Circuit Design
1.5 V
VDD
V in
V out
V out 0.75
0V
V in
122 nV
600 MHz
V out
0.75
∠V out 1 100MEG
Unstable!
VDD
250 fF
Out
250 fF
Adding capacitors to
the output stage
In
10k
10k
C load
1.0 5 pF
In
0.5
Figure 33.74 Showing settling time of the op-amp in Fig. 33.70 with the compensation
capacitors shown in Fig. 33.73.
294 CMOS Mixed-Signal Circuit Design
terminals switch places). While the output range is reduced when not using the push-pull
output buffer, the stability is almost guaranteed with a reasonable-sized load capacitance.
The biasing current is increased so that it can drive the capacitive load in the time
required. A scaling increase in the biasing currents is followed by an increase in the size of
the devices. For example, if we increase our biasing current from nominally 15 µA (see
Figs. 33.52 and 33.53) to 150 µA, then our NMOS size is increased to 200/2 and the
PMOS size is increased to 400/2. Not scaling devices, as discussed earlier, can result in
too small of an fT or too large of a ∆V.
2. We use the circuit as seen in Fig. 33.70 but the settling time is too long. By
reducing the lengths of the push-pull output stage, the frequency of the second pole gets
pushed out (increases) and the gain of the output stage decreases (both helping with
stability). However, the current in the output stage increases, resulting in higher power
dissipation. This, in most situations, isn't enough alone to guarantee a stable op-amp.
3. Increasing the biasing current lowers the gain and improves the speed (and thus
decreases the settling time). The linear output range decreases (perhaps by too modest an
amount to be a concern because of the large first-stage gain) and the input diff-amp
minimum common-mode range may become too large, causing the diff-amp to shut off
when VDD/2 is applied to the op-amp input. This latter concern was discussed earlier.
This fix for the settling time is trivial to simulate by changing the resistor value in the
beta-multiplier biasing circuit used in the op-amp.
4. Increasing both the biasing current and the size of the devices in Fig. 33.70 so
that a given load capacitance becomes, effectively, less difficult to drive. This is the same
fix as 1), above, except that we still have the floating current source and output buffer
(and so the current needed to charge the load is supplied by the push-pull amp and is not
directly related to the biasing current in the diff-amp).
Before listing number 5, let's review from Ch. 27 how the op-amps unity gain
frequency, fu, is related to settling time. Assuming no slew-rate limitations, the op-amps
time constant can be written (assuming the op-amp has only a single dominant pole) as
τ= 1 (33.23)
2πf u ⋅ β
β is the feedback factor. Here we assume β = 1, where all of the output is fed back to the
input. For our op-amp response of Fig. 33.73 τ = 2.3 ns. The output signal for our
single-time constant dominant-pole op-amp circuit can be written as
V out = V outfinal (1 − e −t/2.3ns ) (33.24)
For 0.1% settling accuracy (Vout /Voutfinal = 99.9%), it takes
t = 15.9 ns (33.25)
While we used 70 MHz for fu, including the 5 pF load drops the unity gain frequency and
increases the bandwidth-limited settling time.
Chapter 33 Submicron CMOS Circuit Design 295
5. Use minimum channel lengths for any high-speed design. This increases the
device's fT while also increasing the drive current strength of the MOSFETs. Using
minimum channel lengths with a larger biasing current can push the settling time down to
under 10 ns. The decrease in the gain resulting from using minimum length devices and a
larger biasing current (larger ∆V) in the topology of Fig. 33.70 shouldn't be too much of a
concern since we are starting with 110 dB gain.
Differential Output Op-Amp
Let's build on our basic op-amp of Fig. 33.70 to implement a high-speed, low-power
fully-differential op-amp. Consider the gain and output stage shown in Fig. 33.75 and the
associated simplified schematic representation. Notice how the inputs labeled "Top" and
"Bottom" are low-impedance, cascode-connected current mirrors (with node voltage
labels, Vot1 and Vob1). Using the simplified model, Figure 33.76 shows the schematic of a
fully-differential op-amp (minus the common-mode feedback circuit).
VDD VDD
V ot1
V bias2 VDD
Top V high N
V biasp2
V out
V biasn2
Bottom V low
P
V bias3
V ob1
V ot1
Top
Out V out
V ob1 Bottom Simplified schematic symbol
VDD VDD
Top Top
Bottom Bottom
V bias3
V bias4
VDD VDD
V CM
V bias3
Circuit useful with Long L devices
DC feedback around V bias4
the op-amp Adjust current as needed
Figure 33.77 Adding auxilary input to the diff-amp to balance the op-amp's outputs.
v o+
R
v o+ + v o−
2
R
v o−
Figure 33.78 Using common-mode noise elimation to balance the outputs of an op-amp.
should have small shunt capacitors placed across their terminals to compensate for the
inverters' input capacitance.
We can analyze the operation of the circuit shown in Fig. 33.78 using the steps
shown in Eqs. (33.5)-(33.12). Assuming the output resistance of the inverter is large
compared to the resistors R, we get
v o+ + v o− v o+ − (v o+ − v o− )/2
i diff = g m ⋅ + + i CM (33.26)
2 R
and
v o+ + v o− v o− − (v o+ − v o− )/2
−i diff = g m ⋅ + + i CM (33.27)
2 R
Taking the difference in these equations shows, once again,
v −v
2i diff = o+ o− (33.28)
R
and that the common-mode component of the signal is removed (set to VSP).
298 CMOS Mixed-Signal Circuit Design
0.750003
VDD = 1.5 V
R
V out 0.75 V
σ
R 6σ ≈ 6 µV
2σ
0.749997
(a)
(b) Volts
RMS voltage = σ ≈ 1 µV
Figure 33.79 (a) A voltage divider and (b) the variation of the output voltage because of
thermal noise (an example PDF).
Chapter 33 Submicron CMOS Circuit Design 299
2. The mean-squared value (squaring the RMS value) indicates the average power
of the thermal noise. We could also say that the variance, σ2, indicates the average power
of the noise. Sometimes this is more correctly called the total average normalized power
because we assume a 1 Ω resistor when converting from RMS voltage or current to power
and we sum the total power in the spectrum, see Eq. (31.37).
3. The power (variance) in uncorrelated noise sources (random variables) can be
added directly, but the RMS voltage or current values (standard-deviation) cannot. For
example, if we have an RMS thermal noise voltage, σtherm, an RMS quantization noise
voltage, VQe,RMS, and a sampling error power due to jitter of PAVG,jitter (see Ex. 31.15)
corrupting a sinewave signal with a peak amplitude of Vp, then we would determine the
signal-to-noise ratio of the signal using
Vp / 2
SNR = 20 ⋅ log (33.29)
σ 2therm + V 2Qe,RMS + P AVG,jitter
The numerator is, of course, the RMS value of the desired signal while the denominator is
the square root of the total error power from each error source, i.e., thermal noise,
quantization noise, and clock jitter. It's interesting to note that averaging K samples of a
random variable, say thermal noise, results in a reduction of its RMS value
σ therm σ2
σ K,therm = or σ 2K,therm = therm (33.30)
K K
The shape of the Gaussian PDF gets taller and narrower as we average the random signal.
The area, however, remains constant and equal to one. Rewriting Eq. (33.29) and
including the effects of averaging the signal and noise results in
Vp / 2
SNR= 20 ⋅ log + 10 ⋅ log K (33.31)
σ 2therm + V 2Qe,RMS + P AVG,jitter
which shows, once again, that averaging can be employed to increase SNR.
The Spectral Characteristics of Thermal Noise
Examine Fig. 33.80 where we've eliminated the DC bias and have shown the random
current sources used to model each resistor's thermal noise contributions to Vout. The RMS
noise current has a value (as given in Ch. 7) of
4kT
i2 = ⋅B (33.32)
R
where k is Boltzmann's constant (1.38 × 10−23 Watt⋅sec / K ), T is the temperature in
Kelvin, and B is the bandwidth over which the noise is measured. For the circuit shown in
Fig. 33.80, the RMS output noise voltage is
4kT 4kT
σ therm = V out,RMS = (R 1 R 2 ) ⋅ ⋅B+ ⋅B (33.33)
R1 R2
300 CMOS Mixed-Signal Circuit Design
V out
4kT ⋅ B R1 4kT ⋅ B R2
R1 R2
As indicated in Ch. 7, to perform a noise analysis we (1) add RMS noise voltages to the
circuit, (2) determine the RMS output noise from each contribution using superposition
(i.e., with only one noise source in the circuit at a time), and (3) square each contribution
followed by summing and taking the square-root to get the output noise (as a function of
the bandwidth B).
Note that even with both sides of a resistor connected to ground (no DC current),
a noise current (electrons) will move back and forth from the ground to the resistive
material at each connection as long as T > 0 K . In other words, a noise current will still
be present in the circuit.
Reviewing Eq. (33.32) for a moment would reveal that if we reduce the bandwidth
of a measurement (pass the signal plus noise through a narrow band filter) we get a
corresponding reduction in the RMS noise present in the signal. Spectrum analyzers use
narrow band filtering for this reason: to get extremely large dynamic range. The PSD of
the thermal noise voltage specified by Eq. (33.33) is plotted in Fig. 33.81. The output
noise power is given by
fH
where
B = fH − fL (33.35)
We now need to discuss how to determine the bandwidth, B.
P therm ( f ), V 2 /Hz
4kT ⋅ (R 1 R 2 )
Figure 33.81 Thermal noise power spectral density for Eq. (33.33).
Chapter 33 Submicron CMOS Circuit Design 301
R
V in V out V out
C 4kT R C
R
Noise circuit
The output noise power for the circuit of Fig. 33.82 can be determined using
∞
V 2out,RMS = ∫ 4kTR ⋅ 1 ⋅ df (33.36)
0
1 + (2πf ⋅ RC) 2
Knowing
du 1 u
∫ a 2 + u 2 = a tan −1 a + C (33.37)
then
f 3dB
1 ∞
V 2out,RMS = 4kTR ⋅ ⋅[tan −1 2πf ⋅ RC] 0 (33.38)
2π ⋅ RC
or
V 2out,RMS = 4kTR ⋅ f 3dB ⋅ π (33.39)
2
The noise equivalent bandwitdth (NEB) is then
NEB = f 3dB ⋅ π (33.40)
2
Figure 33.83 shows the interpretation of this equation. The area from DC to the NEB is
equal to the area under the actual response. Note that Eq. (33.39) reduces to
V 2out,RMS = kT (33.41)
C
which is our familar result for the RMS thermal output noise power of an RC circuit (kay
tee over cee noise).
302 CMOS Mixed-Signal Circuit Design
V 2 /Hz
NEB = f 3dB ⋅ π
2
4kTR ⋅ 1
1 + (2πf ⋅ RC) 2 Same area
f
f 3dB
Figure 33.83 Showing the same area in the NEB and the actual spectrum.
Spectrum V( f )
analyzer
f
Figure 33.84 (a) Measuring op-amp output noise spectral density and, (b) noise
model where output noise is referred back to the input.
Chapter 33 Submicron CMOS Circuit Design 303
gain, AOL, using f 3dB = f u /A OL . Assuming the op-amp is the limiting bandwidth factor in
the circuit, the RMS input-referred noise is given by
fu π
V in,RMS = V( f ) ⋅ ⋅ (33.42)
A OL 2
NEB
This assumes V( f ) is a constant magnitude number (white noise) vs. frequency. (White
noise is analogous to white light where the spectrum of white light is occupying all
frequencies of interest with equal amplitude.)
In a CMOS op-amp we'll be able to use NEB to get an idea for the circuit noise,
especially if the circuit bandwidth is wide. However, because of flicker noise (1/f noise
discussed in Ch. 9) present in MOSFETs we will need to use Eq. (33.34) to get an exact
idea for the output RMS noise. Equation (33.34) is rewritten as
fH
V 2out,RMS ( f ) = ∫ V 2 ( f ) ⋅ df (33.43)
fL
where now the measured noise output spectrum, V( f ) , is not a constant but changes with
frequency. The RMS input-referred noise can be determined for a particular circuit
configuration using
fH
V2 ( f )
V 2in,RMS ( f ) = ∫ H( f ) 2
⋅ df (33.44)
fL
analyzer
se
dom
provides isolation
tes
power of both contributions from thermal noise and 1/f noise (flicker) in a MOSFET's
drain current as i 2noise or
i 2noise = i 2therm + i 21/f (33.45)
Because noise is always measured on the output of a circuit and referred back to
the input for comparison with the input signal we can use either of the circuits shown in
Fig. 33.86 for performing simple noise analyses in our CMOS circuits.
g m v gs = i d
i 2noise /g 2m
i 2noise
VDD VDD
In
M1 M1 i 2noise1
Out In
Out
Bias M2
Bias
i 2noise2
input-referred noise. Increasing the biasing current will also cause the noise currents to
increase (as discussed in Ch. 9). To increase gm1 without changing the noise currents we
must increase the width of M1. While increasing the width improves the noise
performance, it slows down the inherent speed of the circuit because of the drop in ∆V
(and thus fT ). However, speed is usually not a concern with a source-follower (under light
to reasonable load conditions) because its bandwidth approaches fT . The point here is that
increasing the width of the MOSFET whose gate is connected to an input node can be
used in any amplifier to reduce the input-referred noise assuming constant drain current.
An important use of the source-follower is in small-input capacitance amplifiers
(such as charge amplifiers used in charge-coupled devices [CCDs]). Because the AC input
voltage is ideally equal to the AC output voltage, the voltage change across the gate-
source capacitance is zero. This means that the input capacitance of the source-follower in
Fig. 33.87 is set by the gate-drain capacitance of M1 (and, compared to a common-
source amplifier, is considerably smaller). However, because the voltage gain of the
source-follower is one the input-referred noise contributions from the amplifier connected
to the output of the source-follower are referred directly back to the source-follower's
input without any amplitude reduction. To understand this statement in more detail, let's
consider the noise performance of a cascade of amplifiers.
Noise Performance of a Cascade of Amplifiers
Consider the cascade of amplifiers shown in Fig. 33.88. Here we are assuming the
amplifiers have infinite input resistance (the amplifier input is the gate of a MOSFET; the
amplifiers amplify an input voltage). If this isn't the case, then we need to model the
input-referred noise with both voltage and current generators to account for the loading
on the amplifier output. As seen in the figure, referring all three stage's noise contributions
back to the overall amplifier input results in
V22 ( f ) V 23 ( f )
V 2in ( f ) = V 21 ( f ) + + 2 2 (33.47)
A 21 A1A2
V 21 ( f ) V 22 ( f ) V 23 ( f )
In A1 A2 A3 Out
In A1 A2 A3 Out
V 21 ( f ) + V 22 ( f )/A 21 + V 23 ( f )/A 21 A 22
Stage two's input-referred noise, V 22 ( f ) , is referred back to the overall amplifier's input by
dividing by the first stage's gain, A 21 . If the first stage's gain is one then the second-stage's
input-referred noise appears directly on the overall amplifier's input (which is the case
when using a source-follower as the first amplifier). However, if the first stage has a large
gain, then the contributions from the following stages are negligible. The point here is that
to minimize the input-referred noise the gain of the first stage should be large.
For the basic op-amp shown in Fig. 33.70 the noise performance is dominated by
both the first-stage diff-amp and the second-stage amplifier made up of the cascoded
transistors and floating-current source. The voltage gain of the diff-amp will be close to
one making the second stage's input-referred noise reflect directly back to the input of the
op-amp. Again, for low-noise design, we want large first-stage gain.
To minimize a diff-amp's input-referred noise (see Ch. 24), we can increase the
widths of the diff-pair. As discussed earlier, this also minimizes both the systematic and
random offsets in an op-amp and increases the input common-mode range. The cost of
these benefits, again, is the lowering of the parasitic poles (ultimately affecting op-amp
stability) that the diff-pair introduces into the op-amp's overall transfer function. The fT of
the two MOSFETs used in the diff-pair decreases.
It's interesting to note that an amplifier's offset voltage can be thought of as a
special case of noise at DC. This means that Eq. (33.47) can be applied to determine the
importance of amplifier offset in a cascade of amplifiers. (Replace V[ f ] with VOS in Eq.
[33.47].) Note that because of the small voltage gain of the diff-pair used in our
mixed-signal op-amp of Fig. 33.70, the topology can have a relatively large systematic
offset voltage (e.g. 5 mV). The noise and offset performance of this op-amp, however, is
no worse than that of the folded-cascode amplifiers presented in Ch. 25. Folded cascode-
based op-amps also use a low-voltage gain in the input diff-pair. For circuit techniques to
reduce both noise and offsets (chopper stabilization, offset storage, and correlated double
sampling) the reader is referred to [9].
DAI Noise Performance
Figure 33.89 shows the DAI (see Fig. 31.78) with kT/C noise sources shown. The input-
referred noise is given by
V 2in,RMS = kT (33.48)
CI
in series with both v1 and v2. A total of 2kT/C I is sampled onto CI during each clock cycle.
If the input signal can swing from VDD to ground, then we can estimate the SNR using
VDD/ 2 2
SNR = 20 log (33.49)
2kT/C I
If VDD = 1.5 V, T = 300K, and CI = 100 fF then the maximum SNR is 68 dB (roughly
11-bits of resolution). Equation (33.49) is useful in determining the minimum value of
capacitors used in a DAI for a specific application.
Chapter 33 Submicron CMOS Circuit Design 307
φ1 φ2
CF
V CM CI
V out
v1
v2
kT/C I kT/C I
Sampled onto C I when φ 1 switches close.
REFERENCES
[1] R. J. Baker, H. W. Li, and D. E. Boyce, CMOS: Circuit Design, Layout, and
Simulation, Wiley-IEEE, 1998. ISBN 0-7803-3416-7
[2] Y. Taur and T. H. Ning, Fundamentals of Modern VLSI Devices, Cambridge
University Press, 1998. ISBN 0-521-55959-6
[3] C. Enz, F. Krummenacher, and E. Vittoz, "An analytical MOS transistor model
valid in all regions of operation and dedicated to low-voltage and low-current
applications," Journal on Analog Integrated Circuits and Signal Processsing,
Kluwer Academic Publishers, pp. 83-114, July 1995
[4] M. Bucher, C. Lallement, C. Enz, F. Théodolz, and F. Krummenacher, Electronics
Laboratories, Swiss Federal Institute of Technology (EPFL), Lausanne,
Switzerland. Available at http://legwww.epfl.ch/ekv/index.html
[5] D. P. Foty, MOSFET Modeling with SPICE: Principles and Practice,
Prentice-Hall, 1997. ISBN 0-13-227935-5
[6] D. I. Hariton, Floating MOS Capacitor, U.S. Patent 5,926,064, July 20, 1999.
[7] J. Yuan and C. Svenson, "High-Speed CMOS Circuit Technique," IEEE Journal
of Solid State Circuits, Vol. 24, No. 1, pp. 62-70, February 1989.
[8] M. J. M. Pelgrom, A. C. J. Duinmaijer, and A. P. G. Welbers, "Matching
Properties of MOS Transistors," IEEE Journal of Solid State Circuits, Vol. 24,
No. 5, pp. 1433-1440, October 1989.
[9] C. C. Enz and G. C. Temes, "Circuit Techniques for Reducing the Effects of
Op-Amp Inperfections: Autozeroing, Correlated Double Sampling, and Chopper
Stabilization," Proceedings of the IEEE, Vol. 84, No. 11, pp. 1584-1614,
November 1996. Available at http://cmosedu.com
308 CMOS Mixed-Signal Circuit Design
QUESTIONS
33.1 Regenerate the plots shown in Fig. 33.1 using an L of 5 µm and a W of 15 µm. Do
the curves look similar? Could we use the level 1 model with long L devices?
33.2 Can we use a native MOSFET in an application where we need a small threshold
voltage? What are the limitations?
33.3 Sketch the implementation of a floating capacitor that uses a MOSFET switch to
connect the source/drain/bulk to ground. Show the capacitor used in a DAI, Fig.
31.78. Assume the switch connected to the capacitor is clocked with the φ1 clock
(why?). Explain the operation of the circuit.
33.4 Sketch the practical layout of a 10k n+ poly resistor using a silicide block.
33.5 Verify the error in Fig. 33.18 is due to differing device drain-source voltages.
Show that using an even longer length device can result in less error.
33.6 If the drawn area of a source implant is 100 m2, what is the actual area if a scale
factor of 0.15 µm is used.
33.7 Using Eqs. (33.3) and (33.4), estimate the high-to-low and low-to-high delays in
the circuits shown in Fig. 33.90.
40/20 40/20
In Out In
1 pF
33.8 Using WinSPICE results, tabulate the output amplitude of the circuit in Fig. 33.32
against capacitive load.
33.9 Show that kickback noise on the inputs of the comparator shown in Fig. 33.34
using simulations. Show adding the circuit of Fig. 33.91 (a source-follower) to
each input of the comparator will drastically reduce kickback noise.
33.10 Modify the design of the digital input buffer shown in Fig. 33.39 using the
topology shown in Fig. 33.78. Show that the circuit still functions as expected
using simulations.
33.11 Show, using simulations, that the circuits in Fig. 33.47 do indeed behave as
counters.
33.12 Verify the operation of the element in Fig. 33.50 as a 1-bit adder.
Chapter 33 Submicron CMOS Circuit Design 309
VDD
10/50
Out
In
20/1
33.13 Do the capacitors in Fig. 33.51 slow down the operation of the diff-amp? Why?
33.14 Estimate the input common-mode range of the diff-amp shown in Fig. 33.51.
33.15 Regenerate the plots shown in Figs. 33.54 and 33.55 if the drain-source voltages
of the MOSFETs are increased to 1.5 V. Why did the transconductance increase?
33.16 Show, using simulations, that a MOSFET's transition frequency does increase with
increasing gate-source voltage and decreasing length.
33.17 Do any of the MOSFETs in Fig. 33.58 move into the triode region if the resistor's
value is decreased to 1k? Verify your answers with simulations.
33.18 Compare the cascode current mirror performance shown in Fig. 33.61 to a basic
current mirror. Show the increase in output resistance when using a cascode and
compare the minimum voltages across the mirrors.
33.19 What happens in Fig. 33.65 to the current flowing in the push-pull amplifier if we
increase the lengths of MC1 and MC2?
33.20 Estimate the minimum VDD allowed for proper operation of the diff-amps shown
in Fig. 33.69.
33.21 If, in Fig. 33.70, V GS = V SG = 0.5 V and ∆V = V DS,min = 0.1 V what is the minimum
allowable power supply voltage, VDD, for proper op-amp operation.
33.22 Sketch the implementation, based on the topology of Fig. 33.70, of a wide-swing
op-amp with rail-to-rail input common-mode range. Assume the tail currents used
in the diff-amps use 20/2 (NMOS) and 40/2 (PMOS) devices (half the current
flowing each MOSFET of the diff-pair) so that summing the currents in the circuit
of Fig. 33.75 at the top and bottom nodes doesn't cause any MOSFET to enter the
triode region.
33.23 Resketch Fig. 33.79 for the cases when the thermal noise is averaged. Show the
cases with K = 1, 2, 4, 8, and 16.
310 CMOS Mixed-Signal Circuit Design
34
Implementing Data Converters
Minimum gate lengths in CMOS technology are falling below 100 nm [1]. This feature
size reduction is driven mainly by the desire to implement digital systems of increased
complexity in a smaller area. This natural trend in feature size reduction, with
accompanying reduction in supply voltage, can present challenges for the mixed-signal
design engineer. The accompanying lower supply voltage results in an inherent reduction
in dynamic range, decrease in SNR, and increasing challenges when implementing analog
circuitry with little, ideally zero, voltage overhead. This chapter focuses on these issues,
and others, related to the implementation of data converters in a digital, submicron CMOS
technology.
Data converters are fundamental building blocks in mixed-signal systems. This
chapter presents and discusses methods and trade-offs for designing CMOS data
converters in submicron CMOS. For DAC design, we focus on converters implemented
with resistors using R-2R networks. The benefit of, and reason we are focusing on, using
R-2R networks over other methods for DAC implementation, such as charge
redistribution [2] or current steering topologies [3], is the absence of good poly-poly
capacitors in a digital CMOS process, the desire for small layout area, and/or the ability to
drive an arbitrary load resistance. R-2R-based DACs can be laid out in a small area while
achieving resolutions in excess of 12-bits. Charge-scaling DACs require linear capacitors.
The layout area needed for these capacitors can often be very large and practically limit
both the resolution and accuracy of the DAC. Similarly, the implementation of
current-steering topologies can result in a very large layout area with limited resolutions,
generally less than 8 bits if integral nonlinearity (INL) is a concern and, more importantly,
if limited output swing and the requirement of known, fixed-load resistances are a
concern.
The first section of this chapter reviews current- and voltage-mode R-2R-based
DACs. The second section of the chapter discusses the use of op-amps in data converters,
while the third section presents an overview of general ADC implementations in
312 CMOS Mixed-Signal Circuit Design
submicron CMOS process. While we briefly discuss the future direction of ADCs, we
concentrate our discussion on the implementation of pipeline data converters.
The goal in this chapter is not to provide an exhaustive overview of data converter
design but rather to provide discussions and practical insight. We assume that the reader is
familiar with data converter fundamentals (discussed in Ch. 28) and data converter
architectures (discussed in Ch. 29). For example, the reader knows the difference between
differential nonlinearity (DNL) and integral nonlinearity (INL) or the difference between a
two-step flash ADC and a pipeline ADC.
where bX is either a 1 or 0, or
N−1
V out = V REF− + R ⋅ Σ (bX ⋅ I TAPX ) for V REF+ < V REF−
X=0
(34.4)
Chapter 34 Implementing Data Converters 313
VDD
V REF+
V out
R 1 0
2R b N−1
TapN-1
R 1 0 V REF+ − V REF−
b N−2 1 LSB =
2R 2 N+1
TapN-2 (Output swing limited to VDD/2.)
R 1 0
b N−3
2R
1 0
Detail
b N−3 b N−3
R 1 0
b0
2R To resistor
Tap0
MOSFET channel resistance should be << 2R
under most circumstances only either NMOS
2R or PMOS devices need be used
V REF−
Using these equations, we can see the main problem with the basic current mode topology
of Fig. 34.1 in a submicron CMOS process using low-power supply voltages, namely,
limited output swing. If VREF− is set to 0 V, with VREF+ > 0, then the output of the DAC
must be negative, which, of course, can't happen when the only power supply voltage is
VDD. If VREF− is set to VDD, then we can see from Eq. (34.4) that this would require Vout
> VDD. After reviewing Eqs. (34.1)-(34.4), we see that the range of output voltages
associated with the current mode R-2R DAC is limited to VDD/2, e.g., 0 to VDD/2,
VDD/2 to VDD or 0.25VDD to 0.75VDD, etc. Giving up half of the power-supply range
in a DAC and correspondingly reducing the dynamic range, is usually not desirable.
314 CMOS Mixed-Signal Circuit Design
By removing the requirement that the noninverting input of the op-amp be tied to
VREF− and that the feedback resistor be R (the same value used in the R-2R string), we can
increase the output range of the DAC. The output of the op-amp is level-shifted by the
voltage on the noninverting input of the op-amp and by increasing the closed-loop gain of
the op-amp. Similarly, we could add a gain stage to the output of the DAC (two op-amps
would then be used) to achieve wider DAC output swing. We don't cover these options
any further here because they either put more demand on the op-amp design, such as
increased op-amp open-loop gain and speed, or won't, in a practical implementation, result
in a rail-to-rail output swing.
34.1.2 The Voltage-Mode R-2R DAC
Figure 34.2 shows a schematic of a voltage-mode DAC. The voltage on the non-inverting
input of the op-amp can be written as
b N−1 ⋅ V REF+ + b N−1 ⋅ V REF− b N−2 ⋅ V REF+ + b N−2 ⋅ V REF−
V+ = + + ... + VREF− (34.5)
21 22
or, in general terms,
N
b N−k ⋅ V REF+ + b N−k ⋅ V REF−
V+ = Σ + V REF− (34.6)
k=1 2k
N b ⋅V +b ⋅V
V out = 1 + F ⋅ Σ N−k REF+ k N−k REF− + V REF−
R
(34.7)
R I k=1 2
If the input code is all zeroes, with VREF−= 0, VREF+ = VDD, and the op-amp in the follower
configuration, then Vout = VREF−. If the input code is all ones, then the output of the DAC is
VREF+ − 1 LSB.
By using the voltage-mode DAC, we would seem to have solved the problem of
the limited output swing associated with the current-mode DAC of Fig. 34.1. However,
consider how the finite common-mode rejection ratio (CMRR) of the op-amp in Fig. 34.2
can affect the linearity of the overall DAC design. We know the effects of finite CMRR
can be modeled as a variable offset voltage, ∆V OS (see Ch. 25), in series with the
noninverting input of the op-amp that is a function of the change in the op-amp common-
mode voltage, ∆V C , or
∆V C
∆V OS = (34.8)
CMRR
We should see the problem at this point: that is, ∆V OS is in series with the R-2R resistor
string and will ultimately limit the linearity of the DAC. To further illustrate the problem,
let's assume the CMRR of the op-amp in Fig. 34.2 is 20 dB at 1 MHz. Since the
common-mode voltage on the input of the op-amp, again assuming VREF+ = VDD, VREF− =
0, and the op-amp in the unity follower configuration can range from zero to
approximately VDD, the change in the offset voltage used to model finite CMRR when the
Chapter 34 Implementing Data Converters 315
RF
VDD
RI
V out
V REF+
1 2R
0
b N−1
R
2R
b N−2
V REF+
R
2R detail
b N−3 to resistor
b N−2
2R
V REF−
b1
R MOSFET channel resistance
should be << 2R
2R
b0
V REF+ − V REF−
2R 1 LSB =
2N
(Assuming op-amp is in the
follower configuration)
V REF−
DAC's inputs are changing at 1 MHz is 10% of VDD. At first glance we might simply
consider the resulting offset as a nonlinear gain error affecting only the large-signal
linearity (INL). However, it is unlikely in any practical op-amp design that the CMRR will
vary linearly with changes in the input common-mode voltage and so the small-signal
linearity (DNL) will be affected as well. Since, for this example, 1 LSB = VDD/2N, the
resolution of the DAC, because of the finite CMRR and assuming 1 LSB > ∆V OS , is
limited to 4 bits! Performing DC or audio-frequency tests on the voltage-mode DAC made
with an op-amp with a CMRR of, for example, 120 dB at DC results in no practical
316 CMOS Mixed-Signal Circuit Design
resolution limit (indicating that if DAC speed isn't a concern, the voltage-mode
configuration may still be used for high resolutions). Note how CMRR isn't a concern with
the current-mode R-2R DAC (assuming no secondary effects, such as common mode
substrate noise, are present on the input of the op-amp). For precision, high-speed data
converter design we must use an inverting op-amp topology where the inputs of the
op-amp remain at a fixed voltage.
34.1.3 A Wide-Swing Current-Mode R-2R DAC
We've shown that it is desirable to have a wide output swing, as is provided by the
voltage-mode R-2R DAC, while at the same time having a fixed input common mode
voltage, as is provided by the current-mode R-2R DAC. Figure 34.3 shows a wide-swing
current-mode R-2R DAC configuration that has a rail-to-rail output swing while keeping
the input common-mode voltage of the op-amp fixed at the common mode voltage, VCM ,
or (V REF+ + V REF− )/2 .
Like traditional current-mode R-2R DACs, the DAC scheme shown in Fig. 34.3
operates on currents. Using superposition and assuming VREF− is the reference for
calculations, we can show that the current flowing in the feedback resistor, RF , is given by
V REF+ − V REF− V REF+ − V REF−
IF = − + ⋅ 1 ⋅ b N−1 + 1 ⋅ b N−2 + ... + N−1
1 ⋅b
0 (34.9)
2R 2R 2 2
noting the inversion used in the control logic of Fig. 34.3. The output voltage of the DAC
is then given, assuming R = RF , by
V REF+ − V REF−
V out = V REF− + + IF ⋅ R (34.10)
2
or
R = RF
V REF+
2R VDD
0 MSB
1
b N−1 V out
R
2R
V REF+ + V REF−
b N−2 V CM =
2
R
0 2R
1 V REF+
b N−3
detail
0 2R to resistor
1 b N−2
b1
R
0 2R V REF−
LSB
1
b0
2R
V REF+ − V REF−
1 LSB =
2N
V REF−
DNL Analysis
It was shown back in Ch. 29 that for a binary-weighted DAC the worst case DNL
condition tends to occur at midscale when the code transitions from 01…11 to 10…00.
Let's assume in a worst-case scenario the 2R resistance of the MSB input in Fig. 34.3 has
a maximum positive mismatch of ∆R, and all other resistors have a maximum negative
mismatch of –∆R. In this case, the current provided by the MSB has to match the sum of
currents provided by all other lower input bits plus one LSB. Again using the
superposition principle we can verify that the step error of the current flowing through the
feedback resistor RF, caused by the resistor mismatch at the midscale transition, is
approximately equal to
318 CMOS Mixed-Signal Circuit Design
Switches
The switches (MOSFETs) used in the R-2R DAC should have an effective switching
resistance (see Eqs. [33.3] and [33.4]), much less than the resistors used in the R-2R
ladder. The inherent switching time of the switches is extremely fast (speeds comparable
to logic gate delays). Since the switches are in series with the branch resistances of the
R-2R ladder, the R-2R relationship is broken if the switch resistance is not negligible, and
this affects both the INL and the DNL. Also note that we can try to compensate for the
switch-effective resistance by making the length of the 2R resistor slightly shorter than the
length of the R resistor. However, if not careful, this may lead to problems over the
process corners and temperature.
Experimental Results
The wide-swing, current-mode R-2R DAC, based on the scheme in Fig. 34.3, was
fabricated in a 0.21 µm/1.8 V CMOS process (single poly, up to five layers of metal) for
resolutions of 8, 10, and 12 bits. The cell dimensions of the 12-bit DAC are 150 µm by
300 µm. The goal of the experimental results was to verify that the topology of Fig. 34.3
would indeed perform as predicted by Eqs. (34.14) and (34.17) and to generate a
low-power, small-area DAC cell for general-purpose, mixed-signal circuit designs.
Unsilicided n+ poly was used for the R-2R resistances as discussed earlier (see Table
33.1). The mismatch indicated in Table 33.1 for an unsilicided n+ poly resistor is 0.005
(= ∆R/R). Using Eqs. (34.14) and (34.17), we would estimate that our resolution is
limited to 8.6 bits if we want both INL and DNL less than 1 LSB. The results in Table
34.1, however, show that the resolution is better than estimated. This may be because of
our pessimistic assumption of how the resistor values change with position as discussed in
the derivation of these equations. The nominal resistor value used in these experimental
DACs is 10k. To enhance the resistance matching, dummy resistors are implemented at
both ends of the R-2R ladder (see Fig. 33.13). The output range of the DAC is
programmable by choosing the value of the feedback resistance or by the setting of the
reference voltages VREF+ and VREF−.
The measured INL and DNL profiles of the three DACs with resolutions of 8, 10,
and 12 bits are shown in Fig. 34.4. The outputs of the DACs are configured to swing to
both rails (VREF+ = VDD and VREF− = 0). The first several points, adjacent to the two rails,
are not shown in Fig. 34.4 due to the high nonlinearity of the op-amp in those regions.
Major performance results are maximum DNLs of 0.15 LSB, 0.45 LSB, and 2 LSB for
8-bit, 10-bit, and 12-bit resolutions, respectively, with no special circuit techniques (laid
out as shown in Fig. 33.13) or trimming (adjustments). The corresponding maximum DAC
INLs are 0.2 LSB, 1 LSB, and 3 LSB, respectively. Notice that the LSB of the 8-, 10-,
and 12-bit DACs are 7.03 mV, 1.75 mV, and 439 µV, respectively. The DNL/INL can be
written in terms of a voltage as 1.05 mV/1.4 mV for the 8-bit DAC, 0.788 mV/1.75 mV
for the 10-bit DAC, and 0.878 mV/1.31 mV for the 12-bit DAC. The measurements were
taken while the DAC was driving a 1k load. The power dissipated by the DAC, with 1.8 V
output, while driving a 1k resistor is 3.88 mW. The unloaded power dissipation of the
DAC is approximately 500 µW. The DACs were designed using op-amps with simulated
unity gain frequencies of 10 MHz. The measured DAC settling time was approximately
200 ns.
Figure 34.4 Experimental results for the wide-swing DAC of Fig. 34.3.
Chapter 34 Implementing Data Converters 321
V REF+
0 2048 µA IF
MSB
b 11 1
2R
R
1024 µA
b 10
2R
R
512 µA
b9 1 LSB = 1 µA
2R
0 2 µA
1
b1 2R
R
Contribution to the
0 1 µA
LSB feedback current
1 when bit is high
b0 2R
2R
V REF−
Figure 34.5 Showing how currents sum into the feedback current.
322 CMOS Mixed-Signal Circuit Design
makes attaining good DNL with less accurate components possible [5]. A segmented
wide-swing DAC is shown in Fig. 34.6. In this figure we've taken the upper four bits and
segmented their current contributions to the feedback resistor. If we use the numbers from
Fig. 34.5, then when the code 0000 1111 1111 (255 µA) transitions to 0001 0000 0000
(256 µA) the 1 output of the decoder goes high and the bottom resistor connected to the
output of the decoder contributes 256 µA to the feedback path. When the code changes
from 0001 1111 1111 (511 µA) to 0010 0000 0000 (512 µA), both lower outputs (1 and
2) of the thermometer decoder are high. Since the 1 decoder output continues to
contribute to the output current, the step height is set by the difference between the 2
decoder output and the contributions from the lower eight bits. This makes the accuracy
0 V REF+ R
3-bit thermometer decoder output
1
15 7 0 0 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0
b 11 6 0 0 1 1 1 1 0 0 0 0 0 1 0 0 0 0 0 0
Outputs
5 0 1 1 1 1 1 1 0 0 0 1 1 1 0 0 0 0 0
4 0 1 1 1 1 1 1 1 0 1 1 1 1 1 0 0 0 0
Thermometer
0 R
b 10 3 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0
decoder
1 2
14 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0
b9
0 R 011 101 011 101 000
(3) Inputs
b8 1
1 R F = R/16
2R VDD
0
1
b7 V out
0 2R
1 VREF+ + V REF−
b1 V CM =
2
R
0 2R
1
b0
2R
V REF−
requirements for 1/2 LSB DNL in a 12-bit converter set by 8-bit matching. Note that
while segmentation reduces DNL error, it does nothing for INL. Segmentation can also be
used to reduce the glitch area associated with the changing DAC output.
Trimming DAC Offset
Figure 34.7 shows how the op-amp's offset voltage shifts the DAC's output. It may be
desirable in some situations to trim or remove this offset. The offset may be the result of
an inherent systematic offset in the op-amp or the result of random variations in the
characteristics of the MOSFETs used in the op-amp. An offset may also result because of
the voltage dependence of the resistors used to generate the common-mode voltage, VCM.
V out
V out
± V OS
+V OS
V CM −V OS Digital input code
(a) Showing offset voltage in an op-amp.
(b) DAC transfer curves showing offset.
Figure 34.7 Showing how an op-amp offset affects the DACs transfer curves.
Figure 34.8 shows one possible method to generate a common-mode voltage that
is adjustable with a digital code. Here again we are assuming that VCM is ideally 0.75 V.
We should recognize the R-2R ladder from Fig. 34.2. The output voltage of this ladder, as
seen in Eq. (34.6), is an analog voltage related to the digital input word (assuming the
voltage divider made up of Rbig and the two R resistors connected to the output in Fig.
34.8 doesn't load the circuit). Figure 34.9 shows the output of this circuit for all possible
digital input words when R is 10k and Rbig is 100k. The inset in Fig. 34.9 shows that the
adjustability of the output is approximately 1 mV. To decrease this value, we can either
increase Rbig (resulting in a decrease in the output swing) or increase the number of bits in
the R-2R DAC. The value, R, of the resistors on the output can be decreased, but this can
result in an increase in power dissipation.
Note that the accuracy required of the 5-bit DAC can be very loose. n-Well
resistors can be used to implement the offset trimming circuit to reduce area and power.
The main concerns are considering the possibility of substrate noise injection and making
sure that the same resistive material is used for the entire circuit. We wouldn't want the
temperature behavior of an n-well resistor used in a circuit with a poly resistor because the
temperature dependencies are different (the offset trimming would only be effective at the
temperature it was performed). Finally note that in a practical circuit it is a good idea to
324 CMOS Mixed-Signal Circuit Design
VDD = 1.5 V
VDD = V REF+
R
2R R big
b4 To op-amp + input
R
2R R
b3 Adjustable voltage
R
2R
b2
R Selected large so it doesn't load the R-2R
2R
b1 ladder and so we get a large attenuation
to the output.
R
2R
b0
1 LSB ≈ VDD ⋅ R/2
2R 2 N R/2 + R big
centered around V CM
Ideal output
Figure 34.9 Output of the circuit in Fig. 34.8 for all possible digital codes.
Chapter 34 Implementing Data Converters 325
add capacitors from the output of the circuit to both VDD and ground to ensure that the +
op-amp input is connected to a good AC ground.
Trimming or calibrating out the offset can be performed at a time prior to
packaging the chip, or it can be performed with some autocalibration sequence after the
chip has been fabricated where the output of the DAC is compared to a known voltage
reference. The concern, as with any calibration, is to adjust only one known error at a time
(known as orthogonal tuning in filter design). For example, the DAC may not have any
offset but may have an INL error for a given input code, Fig. 34.10a. If we were only to
look at this one input code, say 10000... (VCM in binary offset), we wouldn't know if the
error is an INL error or an offset error. After the offset is calibrated out, Fig. 34.10b, we
would then perform an INL calibration to pull the end-points of the transfer curve back to
the ideal straight line transfer curve.
V out V out
INL error
Ideal curve
V out V out
Ideal gain (slope)
Actual gain
INL
(a) DAC transfer curves with gain error. (b) DAC transfer curves after offset
calibration with gain error.
Figure 34.11 Showing gain error and how it can cause problems in an offset calibration.
R-2R
ladder
Fig. 34.3 V CM
Variable current
(a)
R-2R
ladder
Fig. 34.3
Variable voltage
(b)
Figure 34.12 Trimming the output of the DAC using (a) current and (b) voltage.
Digital 12 R-2R
input (12) ladder
Fig. 34.3 V out
5
Upper five bits
Address R-2R
5
5-bit Register 31 circuit
5-bit Register 30 MUX Fig. 34.8
Out
5-bit Register 2
5-bit Register 1 00010
5-bit Register 0
00000 (upper five bits all zeroes)
VDD
2R V out
b4
R CL RL
2R
b3
R
2R
b2
2R
R 1 LSB = VDD
2N
b1
R
2R
b0
2R
Example 34.1
Suppose a 10-bit, voltage-mode DAC with the topology given in Fig. 34.14 is
implemented where R = 10k and CL = 10 pF. Estimate the maximum clocking
Chapter 34 Implementing Data Converters 329
frequency we can use to clock the register supplying the input words to the DAC.
Verify your answer using SPICE.
For complete settling we require that the DAC be 10-bit accurate to within 0.5
LSBs over its full-scale range
N+1
Accuracy = 0.5 LSB = VDD/2 = 111 = 0.04883%
Full scale range (VDD) VDD 2
The time constant associated with the DAC and capacitive load is
RC L = 10k ⋅ 10p = 100 ns
We can use this time constant to relate the final ideal output voltage, Voutfinal, to the
actual output voltage, Vout, using
V out = V outfinal (1 − e −t/RC L )
or, relating this to the required accuracy,
1 = 1 − V out = e −t settling /RC L
2 N+1 V outfinal
The required settling time is then
t settling = RC L ⋅ ln 2 N+1 (34.18)
Using the numbers from this example results in tsettling = 762 ns. The SPICE
simulation results are shown in Fig. 34.15. The maximum clock frequency is then
estimated as
f clk,max = 1 = 1 (34.19)
t settling RC L ⋅ ln 2 N+1
For this example, fclk,max= 1.3 MHz. Note that the fundamental way to decrease the
settling time is to decrease the resistance in the R-2R ladder (assuming we have no
control over the load capacitance). The practical problem then becomes
implementing the switches (MOSFETs) with a resistance small compared to R. T
Figure 34.15 Example output for the 10-bit DAC in Ex. 34.1 showing settling time limitations.
330 CMOS Mixed-Signal Circuit Design
Example 34.2
Suppose that the 2R MSB resistor in the DAC described in Ex. 34.1 experiences a
0.5% mismatch. Estimate the resulting DACs INL and DNL. Use SPICE to verify
your answer.
The 0.5% mismatch (∆R/R or 1 σ [standard deviation]) is the mismatch specified
for the unsilicided n+ polysilicon resistors specified in Table 33.1. Again it is
desirable to use poly resistors because they sit above the substrate on the field
oxide and are more immune to substrate noise. Note that the voltage coefficient
can (will) also cause nonlinearities. However, instead of the worst-case situation of
an abrupt mismatch between the lower resistors and the MSB 2R resistor, as used
in this example, a first-order voltage coefficient error will cause a linear variation
of the resistor values from the LSB resistor up to the MSB resistor (and so the
effects of the voltage coefficient, for reasonably small values, are generally not
significant compared to the random mismatch effects).
Rewriting Eqs. (34.14) and (34.15) to estimate the maximum number of bits
possible with 1 LSB INL or DNL results in
Figure 34.16 Output if MSB resistor in Fig. 34.14 experiences a 0.5% mismatch.
Chapter 34 Implementing Data Converters 331
that the DAC is nonmonotonic (DNL < −1 LSB). An increase in the digital input
code results in a decrease in the output voltage. Nonmonotonic DACs can result in
circuits that don't function properly (an example being a successive approximation
ADC). A DNL of −1 LSB would indicate the output voltage of the DAC doesn't
change when the input code changes.
To improve the DNL, the upper bits of the DAC must be segmented as seen in
Fig. 34.6. Improving the INL relies on calibrating out the mismatch errors (see
Figs. 34.12 and 34.13). Also note, again, that mismatch can be improved by layout
techniqes (e.g., common-centroid) and by averaging the outputs of multiple
resistor strings. T
We can characterize the effects of a DC load resistance, RL , as seen in Fig. 34.14,
by noticing that RL forms a divider with the R-2R ladder. The LSB with a load can be
written as
RL
1 LSB = VDD ⋅ (34.21)
2N R + RL
Notice that if R L → ∞ , this equation reduces to the LSB value given in Fig. 34.14. The
time constant associated with driving an output capacitance can now be written as
τ = R R L ⋅ CL (34.22)
Two Important Notes Concerning Glitches
Note that we have assumed that the RC delay through the resistors used in the R-2R
ladder is negligible. This may not be the case in many practical situations (especially if
diffused or implanted resistors are used), resulting in a DAC output glitch. Also, we have
been simulating with perfectly aligned digital signals, that is, signals that change at the
exact same moment. When the digital signals are slightly misaligned, a significant glitch
can occur in the DAC's output. This means that the inputs to the DAC should be provided
by the same digital hold register. Using segmentation with the required thermometer
decoder can result in the digital signals driving the R-2R ladder seeing differing delays.
Care must be exercised when designing the DAC input clocking circuit (e.g., add small
dummy delays).
Example 34.3
Repeat Ex. 34.2 if a 200 ps skew is experienced by the lower nine bits in the digital
inputs with relation to the MSB.
The simulation results are shown in Fig. 34.17. When comparing this result to Fig.
34.16, the magnitude of the glitch in relation to the much smaller final step in the
output voltage should be obvious. The small 200 ps skew in the digital inputs
causes a code of 00 0000 0000 to be applied to the DAC for 200 ps. For this very
short period of time the output begins to discharge from 750 mV down to ground.
Note in this figure and in Fig. 34.16 the load capacitance was reduced to 0.1 pF to
decrease the settling time. T
332 CMOS Mixed-Signal Circuit Design
Figure 34.17 Showing glitch if the lower 9-bits are skewed by 200 ps in Ex. 34.2.
b b
b b
R LL R RL R LL R RL
R LL R RL R LL R RL
b7 6 I REF 6
Segments
Decoder
5
b6 4 R LL R RL
3
b5 2
1 I REF 1
2 −1 I REF
b4
W-2W current mirror (see Fig. 33.17)
2 −2 I REF b
3
2 −4 I REF b 2
2 −8 I REF b 1
2 −16 I REF b 0
can be laid out adjacent to the W-2W and can also benefit from averaging the outputs of
several DAC layouts. In some cases the number of bits used in the W-2W ladder equals the
number of bits used for the upper segments. For example, a 10-bit DAC would use a 5-bit
W-2W ladder (whose MSB is IREF /2) and 31 segments with values of IREF (a total of 36
outputs as seen in Fig. 34.20a). To improve INL, the layouts are connected (see the
example common-centroid layout in Fig. 34.20b) together to, hopefully, average the
random mismatch effects and increase the linearity of the DAC. Of course, the current
output in the circuit of 34.20b is four times the current in 34.20a for the same reference
biasing levels.
Segments
Outputs
31
Segments
and W-2W mirror.
3
2
Segments
1
and W-2W mirror
Segments
Segments
20 log [A OL ( f )]
A OLDC
A OL ( f ) = f
A OLDC 1+j⋅ f
3dB
mixed-signal op-amp. The phase margin of this op-amp was 70 degrees. The step response
shows a moderate amount of ringing. Decreasing the phase margin increases the peak
amplitude of the ringing and can lengthen the settling time (the time it takes the op-amp's
output to settle to within 1/2 LSB of the ideal final value). Note that the settling time, in
Fig. 33.74, was measured using an inverting op-amp topology. While we concluded, in
Sec. 34.1.2, that we must use an inverting op-amp (or a fully-differential topology where
the input common-mode voltage remains constant), it is still useful to look at the basic
speed (bandwidth) differences between noninverting and inverting topologies.
Gain Bandwidth Product of the Noninverting Op-Amp Topology
Figure 34.22 shows the basic topology of a noninverting op-amp amplifier. The voltage on
the inverting op-amp input can be written as
β
R1
v − = V out ⋅ (34.23)
R1 + R2
where β is the feedback factor for this series-shunt feedback amplifier (the ideal closed-
loop gain, ACL , is 1/β or 1 + R2/R1). The output of the amplifier is
V out = (V in − v − ) ⋅ A OL ( f ) (34.24)
Solving these equations for the closed-loop bandwidth of the amplifier, fCL,3dB, gives
f CL,3dB ≈ β ⋅ A OLDC ⋅ f 3dB = β ⋅ f u (34.25)
336 CMOS Mixed-Signal Circuit Design
R2
R1
v−
V out = A OL ( f ) ⋅ (v + − v − )
v+
V in
R2
R1
V in v−
V out = A OL ( f ) ⋅ (v + − v − )
v+
Example 34.4
Compare the bandwidth of a +1 gain amplifier implemented using a noninverting
op-amp topology (Fig. 34.22) to the bandwidth of a −1 gain amplifier using the
inverting op-amp topology (Fig. 34.23).
Using Eq. (34.26), the bandwidth of the +1 gain amplifier is fu. This amplifier is
commonly known as a unity voltage follower and has R 1 = ∞ (an open) and R 2 = 0
(a short). The bandwidth of the inverting, −1, gain amplifier can be determined
using Eq. (34.32) with R1 = R2 and is 0.5fu. This result is important because it
shows that for the fastest speed the noninverting op-amp topology offers the best
choice. Practically, however, the nonlinearities related to the finite CMRR (see Eq.
[34.8]) force the use of inverting op-amp topologies. As discussed earlier, the
input common-mode voltage must remain constant in any precision application.
Note that a fully-differential op-amp topology is also a shunt-shunt amplifier with a
gain bandwidth product given by Eq. (34.32). T
Example 34.5
Comment on the derivations of Eqs. (33.23)-(33.25) in the last chapter.
The equations are still valid; however, it would be more correct to use the value of
β given by the resistive divider in Eq. (34.23) instead of β = 1. For the test setup
shown in Fig. 33.74 β = 0.5 and so the settling time would more accurately be
estimated as 31.8 ns. T
34.2.1 Op-Amp Gain
In this section we answer the question of how large the DC open-loop gain of the op-amp,
AOLDC , must be in a data converter with a resolution of N bits. We know that the op-amp
must amplify signals to within 1/2 LSB of the ideal value. Further we know that the
closed-loop gain of an amplifier can be written as
A OL ( f )
A CL = (34.33)
1 + β ⋅ A OL ( f )
The feedback factor can be written, after reviewing Fig. 34.23 or Fig. 34.24, as
338 CMOS Mixed-Signal Circuit Design
R1 CF 1/jωC I
β= or = (34.34)
R1 + R2 C F + C I 1/jωC I + 1/jωC F
where the capacitive dependence, or second term in this equation, is used when estimating
the feedback factor present in a DAI.
CF
CI
V in
V out
As we discussed in Ch. 29 the output of the amplifier will be equal to its ideal
value minus some maximum deviation, ∆A. We can write the gain of the DAI over one
clock cycle (treating the integration, z −1 /[1 − z −1 ] , as an initial DC condition on the
feedback capacitor) as
CI
A CL = (34.35)
CF
Then we can write
CI A OLDC
A CL = − ∆A = CF
(34.36)
CF 1 + A OLDC ⋅ C F+C I
If the maximum value of ∆A is at most 1/2 LSB of the ideal gain, or,
CI 1/2 LSB C 1/2 ⋅ (V REF+ − V REF− )/2 N C I
∆A = ⋅ = I ⋅ = ⋅ 1 (34.37)
C F Full scale output C F (V REF+ − V REF− ) C F 2 N+1
then we can estimate the minimum required DC open-loop gain as
τ= 1 (34.41)
2π ⋅ β ⋅ f u
The minimum required op-amp unity gain frequency is then given by
f clk ⋅ ln 2 N+1
fu ≥ (34.42)
2π ⋅ β
or, again assuming β =1/2,
f u ≥ 0.22 ⋅ (N + 1) ⋅ f clk (34.43)
If we design a 12-bit ADC that is clocked at 100 MHz, we need to use op-amps with unity
gain frequencies, fu , of 286 MHz (and a DC gain of at least 16k). Again, this estimate for
the unity gain frequency is optimistic. A good design would use a larger fu than what is
specified by Eq. (34.43).
34.2.3 Op-Amp Offset
A critical characteristic of any op-amp used in a data converter is its offset voltage. We
introduced the concept of reducing the offset voltage of an op-amp back in Ch. 27. Here
we provide additional comments and possibilities for offset reduction.
Adding an Auxiliary Input Port
A simple method of nulling the offset voltage of an op-amp is shown in Fig. 34.25 [6]. In
this figure the added MOSFETs, M1 and M2 (which operate in the triode region) are
essentially used to balance the current flowing in the current mirror load. We can think of
the added MOSFETs as providing an auxiliary input port for offset calibration.
VDD VDD
To diff-amp
Figure 34.25 Trimming offset using an auxiliary input port.
340 CMOS Mixed-Signal Circuit Design
Figure 34.26 shows how we would use the auxiliary input port to remove (lower)
the offset. When zeroing out the offset, the op-amp is removed from the circuit by opening
S1 (and possibly a switch [not shown] in series with the op-amp's output). This is followed
by closing S2 and S3 so that a control voltage is stored on C. Note that we have assumed
that the op-amp is used in an inverting configuration (that is, the noninverting input of the
op-amp, +, is tied to VCM). The offset removal is dynamic and will have to be performed
periodically. We could also use a simple R-2R DAC with a topology similar to what is
seen in Fig. 34.8 to calibrate out the offset (eliminating the dynamic nature of the method).
The output of the DAC would be connected to the auxiliary input port. Note that an
increase in voltage on the auxiliary input port must result in a decrease in output voltage.
(There must be negative feedback when connecting the output of the op-amp to the input
port.)
S1 S2
V CM V out
S3
C
Auxilary input port
The practical problem with the topology of Fig. 34.26 is the charge injection and
capacitive feedthrough resulting from shutting off (opening) S3. This "glitch" of charge
causes a change in the auxiliary port's input voltage and can place a significant limitation
on the minimum possible offset voltage attainable after calibration. The amplitude of the
glitch can be reduced by increasing C or by increasing the length of the MOSFET used in
the op-amp (M2 in Fig. 34.25). Increasing the length results in a decrease in the
MOSFET's transconductance (keeping in mind that the MOSFET is operating in the triode
region) making the amplitude of the glitch less harmful. The drawback of increasing the
MOSFET's length is that the range of offset voltages we can remove is reduced.
Example 34.6
Suppose perfect switches are available for the circuit of Fig. 34.26. Estimate the
residual offset voltage in terms of the op-amp's gain, AG, from the auxiliary port to
the op-amp output.
If the offset voltage before reduction is VOS, then the offset voltage after reduction
is VOS/AG. For reasonable values of AG the final inherent offset voltage is negligible.
The point of this example is that the charge injection and capacitive feedthrough
from the switches is the dominant source of offset error using this technique. T
Chapter 34 Implementing Data Converters 341
We've seen the problem of charge injection and capacitive feedthrough before. The
most common technique for reducing its effect is to use a fully-differential topology.
Figure 34.27 shows a modification of Figs. 34.25 and 34.26 to compensate for charge
injection. The idea is that when S4 and S3 turn off (open) the variation in voltages on the
gates of M1 and M2 are equal resulting in a common change in each MOSFET's
resistance. Ideally then the current will remain balanced in the diff-amp. Note that while
we've shown the use of triode-operating MOSFETs M1/M2 in Figs. 34.25 and 34.27 in
series with the load of a diff-amp on the input of an op-amp, we could also use this
concept in later stages of the op-amp.
VDD VDD S2
S1
V CM V out
M1 M2
S3
S4 C C
To diff-amp
Figure 34.27 Using an auxiliary input port to lower offset (two terminals).
Figure 34.28 shows another possible topology for offset removal using an auxiliary
input [7]. An additional diff-amp is added in parallel to the main input diff-amp stage of an
op-amp to balance the currents and zero out the offset voltage. Again, long length
MOSFETs are used in the added input so that the glitches resulting from the imperfections
in the MOSFET switches (S4 and S3 in Fig. 34.27) have the least effect on the operation
of the circuit.
Long L devices
Figure 34.28 Using an auxiliary diff-amp for balancing current in an op-amp's input.
342 CMOS Mixed-Signal Circuit Design
We can estimate the maximum offset voltage we can zero out using the technique
of Fig. 34.28 by writing the imbalance in the main diff-amp's currents because of its offset
voltage as
g m ⋅ V OS,max = i d (34.44)
The auxiliary input must sum the opposite of this current in the main diff-amp's load to
balance the currents in the main diff-amp (and hence eliminate the offset voltage). If we
label the transconductance of the diff-amp used in the auxiliary input gmaux and the
maximum allowable differential voltage on the auxiliary input for linear operation Vaux,max
then we can write
g m ⋅ V OS,max = g maux ⋅ V aux,max (34.45)
Because we are using long length devices in the auxiliary input g maux << g m for the same
biasing current levels. If V aux,max = 200 mV (a differential voltage of ± 200 mV will cause
all of the diff-amp tail current to flow through one side of the diff-amp) and
g m /g maux = 10, then we can zero out at most 20 mV of op-amp offset.
The offset storage technique shown in Fig. 34.27 relies on the removal of the
op-amp from the circuit while autozeroing the offset. The scheme in Fig. 34.29 shows
how the technique can be extended to remove the offset while leaving the main op-amp,
O1 in the circuit at all times. When S2, S3, and S4 are closed, the offset of O2 is zeroed
out. At this time switches S1 and S5 are open. After O2's offset is stored, S2, S3, and S4
are then opened. Next S1 and S5 close. O2 is used to precisely set the inverting input of
O1 to VCM through the feedback around O1 (not shown). When O2 goes back to zeroing
out its own offset (S2-S4 close) the capacitor connected to the auxiliary port of O1 retains
the charge, and thus voltage, needed to keep O1's offset nulled out to zero. Again this
capacitor should be large to avoid problems from the imperfections of S5.
O1
V CM V out
V OS
S2 O2
S1 V CM
S5
S3
S4
φ2 φ3
CF φ1
φ3 φ3
CL
V in+ CI
V out+
V CM
V in− CI V out−
CL
CF
φ1
to t1 t2 t3
φ3
φ1
φ2
φ3
φ2 φ1
CF φ3
V in+ V out+
V in− V out−
φ2 CF φ3
to t1 t2 t3 φ1
φ1
φ2
φ3
φ3
φ φ
The difference between Q I 1 and Q I 3 is transferred to CF when φ3 goes high. The output
voltage is then determined knowing charge must be conserved
C F ⋅ (V out − V CM ± V OS )
φ φ φ
Q F1 QI 1 QI 3
= C F ⋅ (V in − V CM ± V OS ) + C I ⋅ (V in − VCM ± V OS ) − C I ⋅ (V CM − V CM ± V OS )
(34.48)
or when φ3 goes high
V out = 1 + I ⋅ V in − I ⋅ V CM
C C
(34.49)
CF CF
Notice how the op-amp offset is autozeroed out. The ideal residual offset is VOS/AOL.
Practically the residual offset is limited by the imperfections in the switches (which, once
again, forces us to use fully-differential topologies). Also note, in Fig. 34.30, that we have
drawn the input capacitance of the next stage as a load, CL. This was so that the output of
the S/H would appear to change only on the rising edge of φ3 (plus the output settling
time). Finally, if the S/H is clocked at fclk = 1/Tclk, the output of the S/H must settle to less
than 1/2 LSB, worst-case, in a time of Tclk/2. This means that the value of op-amp unity
gain frequency given by Eq. (34.42) should be doubled.
Chapter 34 Implementing Data Converters 345
Equation (34.49) can be used to determine the relationship between Vin and Vout for
fully-differential signals
V in S/H V out
CI
1+
CF
Figure 34.32 Block diagram for the S/H of Fig. 34.30.
Example 34.7
Simulate the operation of the data converter S/H building block shown in Fig.
34.30. Assume CI = CF = 1 pF and fs = 100 MHz.
The simulation results are shown in Fig. 34.33. In part (a) the clock signals are
shown. Unlike the clock signals shown in Fig. 34.30 where the falling edge of φ2 is
delayed from φ1, the simulation sets the signals so they go low at the same time.
This was to avoid the outputs of the op-amp changing to very large values for the
small amount of time the op-amp operates open-loop with an input signal applied.
In part (b) we show the op-amp outputs. Note how, when φ1 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
op-amp to VCM when φ1 is high if offset isn't important). When φ3 goes high, the
circuit behaves as an S/H with a gain of two. Part (c) of the figure shows the
outputs connected through φ3 switches, as seen in Fig. 34.30, driving 10 pF load
capacitances. T
A Single-Ended to Differential Output S/H
Note how we have assumed in the S/H of Fig. 34.30 that the input voltage was fully-
differential. In most practical situations at the input of an ADC, this isn't the case. While
we can connect Vin− to VCM in an attempt to change the single-ended input into a
fully-differential sampled output, the practical problem is the variation of the op-amp's
input common-mode voltage. As we've already discussed, precision data converters must
use op-amp configurations where the input common-mode voltage is constant. Also, and
346 CMOS Mixed-Signal Circuit Design
φ1
(a)
φ2
Clock signals
φ3
(b)
Op-amp outputs
Inputs
(c)
Outputs across C L
perhaps more practically, the range of allowable common-mode voltages can be very
restricted when designing low-voltage circuits. As we saw with our basic mixed-signal
op-amp in the previous chapter, Fig. 33.70, the minimum input common-mode voltage can
be very close to VCM when only an NMOS diff-pair is used. A technique to force the
op-amp's input common-mode voltage to VCM when a single-ended input is applied to the
S/H is seen in Fig. 34.34 [8]. The error amplifier senses the op-amp's input common-mode
Chapter 34 Implementing Data Converters 347
φ2 φ3
CF φ1
φ3 φ3
CL
V in + V CM CI v o+
V out+
CI V out−
V CM v o−
CL
CF
φ1
φ3
V CM
Error amplifier
voltage. It adjusts the value of the voltage applied to the bottom plates of the input
capacitors when the φ3 switches are closed until the common-mode voltage is
approximately VCM. We say "approximately" to indicate that we don't want the gain of the
error amplifier to be so large that stability is a concern. The settling time of this circuit is
not too important because any variation in its output simply represents a deviation from
VCM on the op-amp inputs. As long as the deviation is small and falls within the
common-mode range of the op-amp, the single-ended to differential S/H functions
properly. The addition of this error amplifier will increase the CMRR and thus reduce
op-amp distortion (see Eq. [34.8]). Of course, when the φ1 and φ2 switches are closed, the
op-amp's input common-mode voltage is V CM ± V OS . Figure 34.35 shows a possible design
for the error amplifier. The error amplifier is simply an operational transconductance
amplifier.
A common-mode feedback (CMFB) circuit is still required to precisely balance the
outputs of the op-amp. Figure 34.36 shows one possible design. In this figure we assume
the reader is familiar with the designs and notation used in the last chapter for op-amp
design in a submicron process. When the φ1 switches are closed, the outputs are connected
to the CMFB amplifier (see Fig. 34.36a). Also when the φ1 switches are closed, from Figs.
34.30 or 34.34, the op-amp is placed in the unity feedback configuration. Because the gain
of the op-amp is large, the inputs must be at the same voltage. The CMFB circuit is used
to ensure that this voltage is VCM (±V OS ) . A typical mixed-signal op-amp is seen in Fig.
34.36b. This op-amp is derived from the topology given in the last chapter, without the
output buffers. The benefit of removing the output buffers when driving purely capacitive
loads is the ease of attaining a 90-degree phase margin (and so clean settling behavior).
348 CMOS Mixed-Signal Circuit Design
VDD
W/2
Out
W/2 W
Bias
The drawbacks of not using an output buffer are the reduced gain and the need to increase
the biasing currents and device sizes to drive a given load capacitance. Again, the
gain-boosting amplifiers labeled N and P in part (b) can be compensated, if needed, using
capacitors at their outputs to ground (or VDD). If a basic diff-amp, as seen in Fig. 33.69,
is used then in most situations no additional capacitance is needed. The CMFB amplifier is
seen in Fig. 34.36c. It is simply a PMOS diff-amp with diode loads. The gain of this
amplifier should be similar to the gain of the diff-amp used in the main op-amp so that the
same load capacitances can be used for compensating both the op-amp and the CMFB
loop. Note that when the φ1 switches are open in Fig. 34.36a, the capacitors essentially
average the outputs maintaining a balanced condition.
Example 34.8
Simulate the operation of the S/H shown in Fig. 34.31. Assume the S/H is clocked
at 100 MHz, Vin+ is a sinewave that swings from ground to VDD, and Vin− is
connected to VCM (the input signal is single-ended and covers the entire supply
range). Show how the op-amp's input common-mode voltage range changes (that
is, doesn't remain at VCM as it would if the input signals were fully-differential).
Show how a 10% mismatch in the two capacitors affects the output of the S/H.
Figure 34.37 shows the simulation results. In part (a) the input common-mode
voltage of the op-amp is shown. When the φ1 switches are closed, the voltage
returns to VCM (the op-amp is placed in a follower configuration where the input
signal charges the two capacitors). In part (b) the rail-to-rail input signal is
converted into a differential S/H output signal. Note that the gain of the S/H is
one. Note also how, unlike the op-amp outputs in Fig. 34.33, the outputs are
limited to V CM + V p /2 where Vp is the peak amplitude of the input sinewave (= 0.75
V here). When the input sinewave has an amplitude of 1.5 V (0.75 V above the
Chapter 34 Implementing Data Converters 349
φ1
v o+
V CM
v o−
CMFB amp
detail shown
op-amp in (c) on the
detail V CMFB next page
(a)
Common-mode feedback circuit
V bias2
N V high V high N
v o− v o+
V low V low
P V bias3 P
2W
V bias4
2W
V CMFB
(b)
Figure 34.36 Mixed-signal op-amp for use in a S/H with CMFB.
VCM ), the positive output is 1.125 and the negative op-amp output is 0.375.
Subtracting the op-amp outputs results in 0.75 V (the same voltage as the input
signal referenced to VCM). It's important to understand how going from a
single-ended signal to a fully- differential signal results in a reduction in the op-amp
output swing.
Finally, the simulation results were generated using 0.9 pF and 1.0 pF sampling
capacitors (labeled CF in Fig. 34.31). Because the feedback factor is unity, these
capacitors will not affect the gain of the S/H. The point here is that the matching of
the capacitors isn't important for a precise gain of one when using this (Fig. 34.31)
S/H topology. (The op-amp open-loop gain, however, is still important.) T
350 CMOS Mixed-Signal Circuit Design
VDD
V bias1
2W
V bias2
2W
W/2
W/2
Out
V bias3
(c)
Figure 34.36 (cont'd) CMFB amplifier circuit.
Input
Clocked comparator
2
V CM Select MUX
1 0
V CM 0
f clk
In
Clk N-bit shift register
Clock pulses
÷N Clk/N
Hold register
LSB MSB
(Assuming N = 8)
Example 34.9
Determine the output of the ADC in Fig. 34.38 if the input voltage is 1.5 V.
We begin by sampling the input voltage of 1.5 V. The output of the comparator is
a logic 1 (MSB). Next, VCM (= 0.75 V) is subtracted from the S/H output resulting
352 CMOS Mixed-Signal Circuit Design
Example 34.10
Repeat Ex. 34.9 if the Cyclic ADC input is 1.1 V.
1. Sample the 1.1 V input voltage. The comparator output goes high (MSB, b7 , =
1). The output of the multiply by 2, after subtracting VCM (= 0.75) from the S/H
output, is 0.7 V.
2. Sample the 0.7 V fed back voltage. The comparator output goes low (b6 = 0).
The output of the multiplier is 1.4 V.
3. Sample 1.4 V. The comparator output goes high (b5 = 1). The output of the
multiplier is 1.3 V.
4. Sample 1.3 V. The comparator output goes high (b4 = 1). The output of the
multiplier is 1.1 V.
5. Sample 1.1 V (b3 = 1) output of the multiplier is 0.7 V.
6. Sample 0.7 V (b2 = 0) output of the multiplier is 1.4 V.
7. Sample 1.4 V (b1 = 1) output of the multiplier is 1.3 V.
8. Sample 1.3 V (b0 = 1) output of the multiplier is 1.1 V.
9. Sample the new input voltage and begin conversion again. The output word in
the hold register is 1011 1011 (binary offset). T
Comparator Placement
We showed the inverting input of the comparator in Fig. 34.38 connected to the common-
mode voltage. In practice, however, we know that the comparator will have an offset or
that the fed back signal may have a common-mode voltage slightly different than the ideal
value. If the common-mode voltage of the fed back signals was, for example, 10 mV
different than the ideal value, the comparator can make a wrong decision. Further, if the
common-mode voltage is varying because of power supply, noise, or temperature changes,
we can make a wrong decision even if some calibration scheme is employed. To avoid a
wrong decision, the comparator is most often used in a fully-differential configuration, as
seen in Fig. 34.39, with offset storage.
In Fig. 34.39 the clocked comparator shown in Fig. 33.34 is used without the
NAND gates on the output. This comparator can have significant kickback noise. By
Chapter 34 Implementing Data Converters 353
adding the φ2 switches in series with the comparator input, we ensure the kickback noise
doesn't corrupt the S/H input voltage. Note that since φ3 and φ2 are nonoverlapping, we
guarantee that the comparator and S/H are disconnected when the comparator is clocked
(and the kickback noise is generated). When the φ1 switches are closed, the offset voltage
of the comparator or the op-amp is zeroed out. The performance requirements of the
comparator (gain and offset) can be greatly reduced (offset storage is not required) if we
use 1.5 bits per clock cycle instead of the 1 bit per cycle used here [9, 10]. We discuss this
further in the next section.
φ2 φ3
CF φ1
φ3 φ3
CL
V in+ CI v o+
V outa+
V CM CI V outa−
v o−
V in−
CL
CF
φ1
φ3
φ1
V outd+
φ3 Digital
V outd−
φ1
Clocked comparator,
φ3 see Fig. 33.34 where
the NAND gates are
not used
Figure 34.39 Implementation of the comparator with an S/H for use in a cyclic ADC.
Example 34.11
Estimate the gain required of a comparator used in a 10-bit cyclic ADC if VREF+ =
1.5 V and VREF− = 0.
The LSB of this converter is 1.5/2 10 = 1.465 mV . This means that the comparator
gain must be large enough so that the output can fully transition to either VDD or
ground with less than 1.465 mV difference on its inputs (assuming the offset
354 CMOS Mixed-Signal Circuit Design
voltage is zeroed out). In other words, the gain must be well above 1.5/1.465 mV
or 210 or 1,024. Clearly this presents a real design concern. The gain of the positive
feedback comparator of Fig. 33.34 may be very large because of the positive
feedback used. However, the delay time of the comparator (time delay between the
clock going high and the outputs transitioning all the way to VDD and ground)
may be too long with such a little input voltage difference. Increasing the gain of
the comparator, without care, can result in the comparator being unstable when
placed in the unity feedback condition (the φ1 switches closed in Fig. 34.39). To
increase the comparator gain and avoid instability, a diff-amp (or two) can be
added as a pre-amp in front of the basic comparator of Fig. 33.34. The offset of
the diff-amp can then be zeroed out by placing the φ1 switches between the
diff-amp's output and its input. T
Implementing Subtraction in the S/H
Notice in Fig. 34.38 how we can implement the S/H and then multiply by two by simply
setting C F = C I in Fig. 34.30. Reviewing Fig. 34.38 we see that it would also be nice to
implement the subtraction in the S/H. In this figure we see that if the output of the MUX
is 0 V, nothing needs to be changed in Fig. 34.30. However, if the MUX output is VCM ,
then the S/H output must be reduced by VCM. Consider what happens if, when φ3 goes
high, instead of connecting the bottom plate of CI to VCM in Fig. 34.30 we connect it to a
voltage VCI+ (see Fig. 34.40). Doing this, after reviewing Eqs. (34.46) to (34.49), results in
φ
Q I 3 = C I ⋅ (V CI+ − V CM ± V OS ) (34.51)
or
φ2
CF
φ3
CI V in
V in+ S/H V out
V CI+
V CI− CI CI
V in− 1+ ⋅ (V CI+ − V CI− )
CI CF CF
(Partial view) CF
V in S/H V out
V CI+ − V CI− CI
CF
1+
+1 CF
CI
Figure 34.41 Block diagram of Fig. 34.30 with bottom plates of C I tied to V CI.
Example 34.12
Simulate the operation of the S/H shown in Fig. 34.42 if fs = 100 MHz, CF = CI = 1
pF, VCI+ = 1.5VCM , and VCI− is 0.5VCM. Comment on the resulting output.
The simulation results are shown in Fig. 34.43. We only show the situation when
we would want to subtract VCM /2 from the differential input signal of the cyclic
ADC, that is, when V in+ > V in− . When the inputs are approximately the same
voltage, the + input is approximately the same voltage as the − input and Vin is 0.
φ2 φ3
CF φ1
φ3
φ3
CL
V in+ CI
V CI+ V out+
V CI− CI V out−
V in−
CL
CF
φ1
φ3
+ input signal
+ output signal
We subtract 0.375 V (VCM /2) from the input and multiply by 2. The result, when
the input is 0, is − 0.75 V (−VCM ). When this happens, Vout+ is 375 mV and Vout− is
1.125 V. Taking the difference in these signals results in − 0.75 V. At 100 ns in
Fig. 34.43, for example, Vin is + VCM. After we subtract VCM /2 and multiply by 2,
we get VCM again (as indicated in the figure). T
Example 34.13
Repeat Ex. 34.12 if we want to add VCM /2 to the input signal.
We only want to add VCM /2 to the input signal when V in+ < V in− . In this situation
we set VCI+ = 0.5VCM and VCI− to 1.5VCM. The simulation results are shown in Fig.
34.44. T
input signal
output signal
V out
V in
V REF− = 0 V CM = VDD/2 V REF+ = VDD
= 2V CM
Figure 34.45 Transfer curve for the cyclic ADC (single-ended case).
Figure 34.46 shows the transfer curve of Fig. 34.45 redrawn to indicate that the signals,
both input and output, are fully-differential. Note, as seen in Fig. 34.39, that the
comparator output transitions from a 0 to a 1 when V in+ > V in− (Vin > 0).
V out
VDD/2 V CM
b=0 b=1
V in
−VDD/2 −V CM
−VDD/2 0 VDD/2
Figure 34.46 Transfer curve for the cyclic ADC when using fully-differential signals.
We can rewrite Eq. (34.54) for fully-differential signals by looking at Fig. 34.46
and noticing that now because the inputs and outputs are fully-differential and referenced
to VCM instead of 0 so must the voltages we add and subtract from the input. Equation
(34.54) can be written as
or
Example 34.14
Using the fully-differential S/H stage of Fig. 34.42 simulate the transfer curve
shown in Fig. 34.46.
The results of the simulation are shown in Fig. 34.47. To simulate Vin, the Vin+
input is a ramp that varies from 0.375 V to 1.125 V while, at the same time, the
Vin− is a ramp that varies from 1.125 V to 0.375 V. This results in Vin changing
linearly from −VCM to +VCM (knowing VCM = VDD/2). When the two ramps are
equal, that is, when they are both VCM (Vin = 0), VCI+ changes from 0.5VCM to
Chapter 34 Implementing Data Converters 359
1.5VCM, and VCI− changes from 1.5VCM to 0.5VCM so that we go from adding VCM /2
when Vin < 0 to subtracting VCM /2 when Vin > 0. T
be an analog voltage within 2 LSBs of its ideal value. The third stage output, V N−3 , must
be an analog voltage within 4 LSBs of its ideal value and so forth. The important point
here is that because the required accuracy of each stage decreases as we move down the
line, the settling time, gain accuracy, and offsets all become less important. Smaller (and
thus lower power) stages can be used for the later stages having, possibly, a dramatic
effect on both layout size and power dissipation. While we're showing 1 bit/stage in Fig.
34.49, most commercially available pipeline ADCs use the digital error correction present
in the 1.5 bit/stage topology discussed later.
Total N stages
V in V N−1 V N−2 V0
1-bit/stage 1-bit/stage 1-bit/stage 1-bit/stage
V N−3
digital digital digital digital
b0
N latches
latch latch
latch
latch
latch b N−3
latch b N−2
t
word ou
b N−1 Digital
Figure 34.49 Pipelined ADC based on the cyclic stages discussed in the last section.
Chapter 34 Implementing Data Converters 361
We know from our analysis of pipeline ADC errors in Ch. 29 that they can be the
result of comparator offsets, gain or linearity errors, and amplifier offsets. In the remainder
of this section we discuss how to reduce the effects of these errors. We begin with a
discussion of using the 1.5 bit/stage topology to make the comparator offsets (and the
reference voltages used with the comparators) a "don't care." While we might think that
using six stages with 1.5 bits/stage would result in an ADC with 9-bit resolution, we find
that the extra 0.5 bit/stage is used to correct for errors introduced by the comparators.
Using six stages will still result in a 6-bit ADC. We then cover the use of capacitor error
averaging [12] to set the gains of the amplifiers to precisely two. The cost of this
technique is the increase in the number of clock cycles required for each stage's operation
and a slightly more complex switching scheme. Finally, we cover some other topologies
useful in amplifier offset removal and discuss offsets in general.
Using 1.5 Bits/Stage
As we saw in Ex. 34.11, the gain of the comparator (and the offset) can present a practical
limitation to the operation of the ADC at high resolutions. The transfer curve of Fig. 34.45
relates the analog input of the cyclic ADC to its analog output voltage. The important
point in this figure, besides the gain and linearity, is where the output of the comparator,
b, transitions from a 0 to a 1. One-bit corresponds to two levels, i.e., a 0 or a 1. Two bits
corresponds to four levels, i.e., 00, 01, 10, and 11. If we were to use three levels then we
would get 1.5 bits of resolution. Using a thermometer code or decimal numbers, the three
levels would then be
Thermometer, ab Decimal
11 3
01 1
00 0
Next consider the transfer curves shown in Fig. 34.50 where three levels are used
(1.5 bits). We can rewrite Eq. (34.54) for the 1.5 bit case as
V out = 2 ⋅ (V in − ab ⋅ 0 − ab ⋅ V CM − ab ⋅ 2V CM ) + V CM (34.59)
or
V CM V 3V
V out = 2 ⋅ (V in + ab ⋅ − ab ⋅ CM − ab ⋅ CM ) (34.60)
2 2 2
or if ab = 00 (V in < V CM /2) , then V out = 2 ⋅ (V in + VCM /2) , if ab = 01 (V CM < V in < 3V CM /2)
then V out = 2 ⋅ (V in − 0.5V CM ) , and if ab = 11 then V out = 2 ⋅ (V in − 1.5VCM ) . For the
fully-differential situation Eq. (34.59) can be rewritten as
V out+ − V out− = 2 ⋅ (V in+ − Vin− + ab ⋅ V CM − ab ⋅ 0 − ab ⋅ V CM ) (34.61)
where all we did was reference, to VCM, the voltages added/subtracted to Vin as seen in Eq.
(34.58).
362 CMOS Mixed-Signal Circuit Design
V CM VDD/2 V in
V in −VDD/2
V CM V REF+ = VDD = 2V CM −V CM V CM
0
V REF− = 0 V CM V CM
−
V CM 3V CM 2 2
2 2
Single-ended input and output Double-ended input and output
Figure 34.50 Transfer curves for using 1.5-bits per clock cycle.
Figure 34.51a shows how the comparators would be set up to determine ab and
how we would provide the outputs. The outputs of the comparators are used as address
inputs to a MUX. The MUX is used to set the bottom plate voltage of the CI capacitor in
Fig. 34.42. Figure 34.51b shows how we would implement the comparators if the input
and output signals are double-ended.
V in + V CM
(a) Single-ended input To bottom plate ofC I (V CI )
3V CM 2 addr out
ab
2 MUX
00 01 11
0 2V CM = VDD
V CM
V CM
2
V in− V CI+
ab MUX
3V CM 2 00 01 11
(b) Double-ended input
4 0 V CM 2V CM
ab 11 01 00
MUX
V in+
V CI−
Figure 34.51 Implementing comparators and MUX for 1.5 bits.
Chapter 34 Implementing Data Converters 363
Example 34.15
Repeat Ex. 34.14 if 1.5 bits/stage are used.
The simulation results are shown in Fig. 34.52. The same signals were used for the
inputs (two ramps) here as used in Ex. 34.14. The voltage VCI+ is 0 V when Vin is
less than − 375 mV and is 0.75 when Vin is between − 375 mV and + 375 mV,
while it is 1.5 V (= VDD = 2VCM ) when Vin is greater than + 375 mV. Notice how
we only need three precision voltages, unlike the 1-bit/stage case, that is, VDD,
VCM , and 0. T
VDD = 1.5 V
Before going any further, let's answer, "How can using 1.5 bits/stage eliminate the
need for precision comparators?" After reviewing Fig. 34.50, we see that an output of 11
cannot be followed by another output of 11 because we subtract VCM from the input prior
to multiplication by two. Even if the comparator has a reasonable offset (say less than 100
mV) or low gain (say less than 50) resulting in a wrong decision, it's impossible for a 11 to
be followed by another 11 or for a 00 to be followed by another 00. It is possible,
however, to get a continuous string of 01 outputs (a simple example is when VCM is
applied to the ADC).
Let's now discuss how the outputs of the 1.5-bit stage are combined into the final
ADC output word. Let's assume, again, that VREF+ = VDD = 1.5 V, VREF− = 0, and VCM =
0.75 V. We know that the final N-bit output word can be converted back into an output
voltage (with the unwanted quantization noise) using a DAC with the following weighting,
Eq. (30.27),
1 LSB
V CM V V CM
V out (if a DAC or V in if an ADC) = b N−1 ⋅ V CM + b N−2 ⋅ + b N−3 ⋅ CM + ... + b 0 ⋅ N−1
2 4 2
(34.62)
364 CMOS Mixed-Signal Circuit Design
Reviewing Eq. (34.54) note how if, on the first cycle, b = 1, we subtract VCM from the
input and then multiply the result by two. After a little thought, we should be able to see
how we derive Eq. (34.62) from Eq. (34.54) after N clock cycles.
If the first thermometer code output of the 1.5-bit stage is labeled a 1.5N−1 b 1.5N−1 and
the second output is a 1.5N−2 b 1.5N−2 , etc., then we can write, with the help of Eq. (34.59),
the relation between the ADC input (analog) and the ADC outputs (digital) as
V CM
V in = a 1.5N−1 b 1.5N−1 ⋅ V CM + a 1.5N−1 b 1.5N−1 ⋅ 2V CM −
2
V 2V V
+ a 1.5N−2 b 1.5N−2 ⋅ CM + a 1.5N−2 b 1.5N−2 ⋅ CM − CM
2 2 4
V 2V V
+ a 1.5N−3 b 1.5N−3 ⋅ CM + a 1.5N−3 b 1.5N−3 ⋅ CM − CM + ... (34.63)
4 4 8
noting that we can group
= 0 here
−
V CM V CM V CM
− − − ... = −V CM ⋅ − 1 = −V + 0.5 LSB + V REF−
2N (34.64)
CM
2 4 8 2N 2N
which is nothing more than a level shift. Clearly we can combine outputs in the following
manner to arrive at an equation similar to Eq. (34.62)
V CM
V in = (a 1.5N−1 b 1.5N−1 + a 1.5N−2 b 1.5N−2 ) ⋅ V CM + (a 1.5N−2 b 1.5N−2 + a 1.5N−3 b 1.5N−3 ) ⋅
2
V CM
+ (a 1.5N−3 b 1.5N−3 + a 1.5N−4 b 1.5N−4 ) ⋅ + ...
4
a 1.5N−1 b 1.5N−1 ⋅ 2V CM − V CM + 0.5 LSB (34.65)
Next notice that, when using a thermometer code, the only time a 1.5N−X b 1.5N−X (the logical
AND of ab) can be high is when both are high. The term a 1.5N−X cannot be high while
b 1.5N−X is low (there is no such output code as 10, see Fig. 34.50). This means that we can
replace a 1.5N−X b 1.5N−X with simply a 1.5N−X . We can rewrite Eq. (34.65) as
V CM
V in = a 1.5N−1 ⋅ 2V CM + (a 1.5N−1 b 1.5N−1 + a 1.5N−2 ) ⋅ V CM + (a 1.5N−2 b 1.5N−2 + a 1.5N−3 ) ⋅
2
V CM V CM
+ (a 1.5N−3 b 1.5N−3 + a 1.5N−4 ) ⋅ + ... + a 1.50 b 1.50 ⋅ N−1 − (V CM − 0.5 LSB) (34.66)
4 2
The + symbol in Eq. (34.66) indicates addition rather than a logical OR. If
a 1.5N−1 b 1.5N−1 = 1 and a 1.5N−2 = 1, then the second term in this equation is 2VCM and the first
term must be 0. When this occurs, the addition of the two terms is 0 and a carry is
generated. We can now use this information to write the relationship between Eq. (34.62)
and Eq. (34.63) knowing ⊕ indicates an exclusive OR
b 0 = a 1.50 b 1.50 with carry = c 0 = a 1.50 (34.67)
b 1 = a 1.51 b 1.51 ⊕ c 0 with c 1 = a 1.51 b 1.51 c 0 (34.68)
Chapter 34 Implementing Data Converters 365
b 2 = a 1.52 b 1.52 ⊕ a 1.51 ⊕ c 1 with c 2 = a 1.52 b 1.52 a 1.51 + c 1 (a 1.52 b 1.52 + a 1.51 ) (34.69)
noting each bit and carry are the outputs of a full adder. To simplify the carry equation,
and the subsequent equations, we can substitute c 1 to get
c 2 = a 1.52 b 1.52 a 1.51 + c 1 a 1.52 b 1.52 (34.70)
We can write the general form of b 2 through b N−1 , using full adders, as
b N−1 = a 1.5N−1 b 1.5N−1 ⊕ a 1.5N−2 ⊕ c N−2 and
c N−1 = a 1.5N−1 b 1.5N−1 a 1.5N−2 + c N−2 (a 1.5N−1 b 1.5N−1 + a 1.5N−2 ) (34.71)
The bit bN has a weighting of 2VCM and thus the final output word size is N + 1
b N = a 1.5N−1 ⊕ c N−1 (34.72)
Before we sketch the implementation of this digital circuit, let's make a few
comments. To begin, notice that the word size is one larger than N (the resolution). In the
1 bit/stage circuit our maximum output is (assuming N = 8)
1111 1111 = VDD − 1 LSB = 2VCM − 1 LSB (1 bit/stage)
Now our maximum output is
1 0000 0000 = VDD = 2VCM (1.5 bit/stage)
The resolution is still essentially eight bits; we just have a slightly larger (1 LSB) output
range. To make the words exactly match, we can throw out the MSB and assume 1 0000
0000 is an out-of-range condition. Note that an input of VDD − 1 LSB using 1.5 bits/stage
gives an output code of 0 1111 1111.
One more comment: if we go through the logic in Eq. (34.66), we don't get the
correct outputs unless we subtract VCM − 0.5 LSB. The binary offset representation can be
written as
V CM − 0.5 LSB = 0011111111... (34.73)
For example, applying VCM (single-ended, see Fig. 34.51) to the ADC input results in a
continuous output (ab) of 01. The output prior to subtraction, from Eqs. (34.67) to
(34.72), is then 01111111... (bN = 0, bN−1 = 1, bN−2 = 1, etc.). After subtracting
0011111111..., we get 01000000 or VCM knowing the weighting of the second bit is VCM.
Figure 34.53 shows one possible implementation of Eqs. (34.67)-(34.72) for a
cyclic ADC. The state shown, i.e., the ab values, is valid at the end of the conversion
(after N clock cycles). When starting the algorithm (on the first rising edge of clock), all
latches are reset to zeroes and the first comparator outputs, ab, are applied. On the second
rising edge of clock the output b2 corresponds to the final bN . After N clock cycles, the
hold register is clocked (and the latches are reset). The final ADC output is the contents of
the hold register after subtracting 00111111.... Changing the numbers to two's
complement may be useful when implementing this stage (assuming the MSB has a
weighting of VCM , i.e., throw out bN ). Note the subtraction can precede the hold register.
366 CMOS Mixed-Signal Circuit Design
a 1.50 a 1.51
D Q
clk
c1 Full c 2
c0 Adder
a 1.50 b 1.50
b 0 LSB b1 b2 b3 b4 bN
f clk /N
clk Hold register
LSB MSB
001111...
Subtraction
ADC out
Figure 34.53 Combining the outputs of the cyclic ADC when 1.5 bits/stage is used.
Example 34.16
Repeat Ex. 34.10 if 1.5 bits/stage are used. Assume the converter is ideal and the
comparators switch precisely at VCM /2 (= 0.375 V here) and 3VCM /2 (= 1.125 V
here). Assume all latches initially contain zeroes.
Vin Comparator outputs, ab Vout Digital out
1.1 (N − 1 = 7) 01 1.45 b7 = 0 c7 = 1 b8 = 1
1.45 (N − 2 = 6) 11 0.65 b6 = 0 c6 = 0
0.65 (N − 3 = 5) 01 0.55 b5 = 1 c5 = 0
0.55 (N − 4 = 4) 01 0.35 b4 = 1 c4 = 0
0.35 (N − 5 = 3) 00 1.45 b3 = 1 c3 = 0
1.45 (2) 11 0.65 b2 = 0 c2 = 0
0.65 (1) 01 0.55 b1 = 1 c1 = 0
0.55 (0) 01 0.35 b0 = 1 c0 = 0
Chapter 34 Implementing Data Converters 367
We can reorder the bits so the MSB is on the left, the LSB is on the right, and
subtract 0 0111 1111 yielding
1 0011 1011 (315 decimal)
- 0 0111 1111 (127)
0 1011 1100 (188)
This is the result given in Ex. 34.10 (0 1011 1011) plus 1 LSB. The 1 LSB
discrepancy can be traced to Eq. (34.66) where we used 0.5 LSBs. Because our
resolution is at best 1 LSB, sometimes the result will experience a round-off error.
To understand this in the subtraction above, the more correct decimal
representation of VCM − 0.5 LSBs is 127.5 and the more correct decimal output is
187.5 (the answer given in Ex. 34.10 was decimal 187). When we discussed the
output of the ADC with an input voltage of VCM, on the bottom of page 365, our
output prior to subtraction was 011111... For eight bits this is decimal 255 (nine
bits out). When we subtract 127, we get the correct output of 128 for the final
output. In this situation the round-off worked in our favor (we round 127.5 up to
128). Because of this 0.5-bit offset, one might simply subtract VCM (= 0 1000
000...) to simplify the subtraction circuitry. While this appears complicated, the
significant benefit is relaxed comparator and reference voltage (used with the
comparators) requirements. T
Example 34.17
Repeat Ex. 34.16 if the comparators switch at 0.4 V (a 25 mV offset) and 1.05 V
(a 75 mV offset).
Vin Comparator outputs, ab Vout Digital out
1.1 (N − 1 = 7) 11 − 0.05 b7 = 0 c7 = 0 b8 = 1
− 0.05 (N − 2 = 6) 00 0.65 b6 = 0 c6 = 0
0.65 (N − 3 = 5) 01 0.55 b5 = 1 c5 = 0
0.55 (N − 4 = 4) 01 0.35 b4 = 1 c4 = 0
0.35 (N − 5 = 3) 00 1.45 b3 = 1 c3 = 0
1.45 (2) 11 0.65 b2 = 0 c2 = 0
0.65 (1) 01 0.55 b1 = 1 c1 = 0
0.55 (0) 01 0.35 b0 = 1 c0 = 0
Which is the exact same result! While the comparator performance can be
extremely poor, the circuit of Fig. 34.42 must subtract and amplify to an accuracy
set by the least significant bit of the converter. When we calculated values in Ex.
34.16 and here we assumed subtractions of exactly 0, VCM , and 2VCM followed by a
multiplication of exactly two. Finally, notice that the negative output of −0.05 V
(single-ended) is easily accommodated when using fully-differential op-amps. T
368 CMOS Mixed-Signal Circuit Design
φs φh
C+
φh φs
C + + ∆C +
V in+
V CI+ V out+
V CI− V out−
V in−
C − + ∆C −
φs
C−
one cycle φh
φs Sample Sample
φh Hold Hold
V out valid
Next consider what happens if, instead of sampling the input voltage again, we
simply switch the positions of the C and C + ∆C capacitors, that is, we connect C to VCI
and C + ∆C to Vout. Figure 34.55 shows this switch and the modified S/H using an extra
half-clock cycle. The sample and amplify phases of the clock are exactly the same as
before, and so Eqs. (34.74) and (34.75) are still valid. We denote the outputs at the end
(falling edge) of the amplify phase as V outa+ and V outa−
∆C + (V
V outa+ = 2 ⋅ V in+ − V CI+ + ⋅ in+ − V CI+ ) (34.76)
C+
and
∆C − (V
V outa− = 2 ⋅ V in− − V CI− + ⋅ in− − V CI− ) (34.77)
C−
where the ideal situation is ∆C = 0 .
φh
φa φh φs φa
C+
φs
C + + ∆C +
V in+
V CI+ V out+
V CI− V out−
V in−
C − + ∆C −
φs
C−
φa
φh
V outa valid
φs Sample Sample
φa Amplify
φh Hold (average)
At the end of the amplify phase the charge on the capacitors is (assuming the
op-amp input voltages are 0 and only looking at the + path to simplify the equations)
Q a+ = (C + + ∆C + ) ⋅ (V CI+ ) + C + ⋅ (V outa+ ) (34.78)
and at the end of the hold phase
Q h+ = (C + + ∆C + ) ⋅ (V outh+ ) + C + ⋅ (V CI+ ) (34.79)
Because charge must be conserved, Q a+ = Q h+ , and therefore
C+ C+
V outh+ = V CI+ + ⋅ V outa+ − ⋅ V CI+ (34.80)
C + + ∆C + C + + ∆C +
It will be useful to use
C+ 1 ∆C +
= ≈1− (34.81)
C + + ∆C + 1 + ∆C + /C + C+
Rewriting Eq. (34.80) gives
∆C + ∆C +
V outh+ = 1 − ⋅ V outa+ + ⋅ V CI+ (34.82)
C+ C+
Substituting Eq. (34.76) for V outa+ results in
∆C + (V
V outh+ = 2V in+ − V CI+ − ⋅ in+ − V CI+ ) (34.83)
C+
where the terms containing (∆C + /C + ) 2 are assumed negligible. If the matching is 1%, then
(∆C + /C + ) 2 = 10 −4 .
Clearly averaging Vouta+ (Eq. [34.76]) and Vouth+ (Eq. [34.83]) results in the precise
desired gain. The question now becomes: "How do we perform the averaging without
introducing more error?" Ideally we want
V outa+ + V outh+
V avg+ =
2
∆C + (V ∆C + (V
= 1 ⋅ 2V in+ − V CI+ + ⋅ in+ − V CI+ ) + 2V in+ − V CI+ − ⋅ in+ − V CI+ )
2 C+ C+
= 2V in+ − V CI+ (34.84)
or more generally
V avg = V avg+ − V avg− = 2(V in+ − V in− ) − (V CI+ − V CI− ) (34.85)
This equation should be compared to Eq. (34.61).
Consider the averaging amplifier shown in Fig. 34.56. We can write the charge on
the four capacitors when the φ a switches are closed (actually at the falling edge of the φ a
clock, assuming complete settling) as
φa
Q + = (V outa+ − V CM ± V OS ) ⋅ C F + (Vouta− − V CM ± V OS ) ⋅ C I (34.86)
and
Chapter 34 Implementing Data Converters 371
φh
φa CF
φa
CI
V outa+ ⋅ φ a + V outh+ ⋅ φ h
V avg+
V outa− ⋅ φ a + V outh− ⋅ φ h V avg−
CI
φa
φa CF
φh
Figure 34.56 Averaging amplifier for use on the output of Fig. 34.55.
φa
Q − = (V outa− − V CM ± V OS ) ⋅ C F + (Vouta+ − V CM ± V OS ) ⋅ C I (34.87)
Note that the common-mode voltage and op-amp offset subtract out when we take the
difference between the balanced signals and so we do not include VCM or VOS in the
remaining analysis. (The offset from the common-mode feedback circuit results in an
unharmful variation in VCM and is discussed later in this section.) On the falling edge of φ h ,
and following the same procedure as used in Eq. (34.48), we can write
C F ⋅ V avg+ = C F ⋅ V outa+ + C I ⋅ V outa− − C I ⋅ V outh− (34.88)
and
C F ⋅ V avg− = C F ⋅ V outa− + C I ⋅ V outa+ − C I ⋅ V outh+ (34.89)
We can then write
CI
V avg+ = V outa+ + (Vouta− − V outh− ) ⋅ (34.90)
CF
CI
V avg− = V outa− + (Vouta+ − V outh+ ) ⋅ (34.91)
CF
noting that if V outa− = V outh− and V outa+ = V outh+ , the matching of the capacitors in Fig.
34.55 is perfect and V avg+ − V avg− = V outa+ − Vouta− (the output perfectly follows Eq.
[34.61]). Taking the difference of these two equations
Error adjustment term
∆C + ∆C −
V avg+ − V avg− = 2(V in+ − Vin− ) − (V CI+ − V CI− ) + (V in+ − V CI+ ) ⋅ − (V in− − V CI− ) ⋅ −
C+ C−
2(V in+ − V CI+ ) ∆C + 2(V in− − V CI− ) ∆C −
⋅ + ⋅ (34.93)
C F /C I C+ C F /C I C−
which reduces to Eq. (34.85) or (34.61) assuming CI /CF is 1/2 (CF is twice as large as CI ).
Note how the selection of (and matching) of the capacitor ratios, C F /C I , is not that
important. The capacitors do not have to match to the final ADC resolution. A matching
of 1% will result in an error term that is one-hundreth of the error that would be present if
the error averaging were not used. Also note, as indicated at the beginning of the section,
that only the first stages (say the first five in a 14-bit ADC) need use capacitor error
averaging because of the reduced accuracy requirements placed on the later stages as we
move through the converter.
The penalty for this precision technique is the increase in conversion time used in
each stage (and increased noise because two op-amps are used in each 1.5-bit section). As
seen in Fig. 34.55, an extra half-clock cycle is required. For 20 Msamples/s a 30 MHz
clock would be required. Again while one stage is in the hold (average) state, the next
stage in the pipeline can be in the sample state so that the conversion time is shared (but
still requiring 1.5 clock cycles). Using the capacitor error averaging technique for
precision gain (and subtraction) and using 1.5-bit/stage sections a 14-bit ADC, using
fully-differential input signals, has been attained at 20 Msamples/s without trimming or
calibration [13].
Example 34.18
Simulate the operation of the circuit shown in Fig. 34.57. Comment on the ideal
outputs and the simulation results.
φs φh
1.00p
φh φs
1.02p
V in+ 1.00 V
V out+
0.75 V
V out−
V in− 0.50 V
1.04p
φs
1.06p
φh
The capacitor values were chosen arbitrarily. The input voltage, Vin, is 1 − 0.5 or
0.5 V. The ideal output voltage, Vout, is then 1 V. Vout+ should ideally be 1.25 V,
and Vout− should ideally be 0.25 V. The simulation results are shown in Fig. 34.58.
The error plotted in this figure is the result of taking the difference between the
ideal output and the actual output. The error for the capacitor values shown in Fig.
34.57 is approximately 5 mV. Clearly, at the risk of stating the obvious, capacitor
mismatch is a fundamental limitation to ADC accuracy. T
Error, millivolts
Figure 34.58 Simulation results from Ex. 34.18 showing amplification error.
Example 34.19
Simulate the operation of the circuit shown in Fig. 34.59 (a cascade of Figs. 34.55
and 34.56). Comment on the ideal outputs and the simulation results.
This circuit shows the same mismatched capacitors in the multiply-by-two stage as
we saw in Ex. 34.18. The averaging stage also shows mismatched capacitors with
1.00p 2.05p
1.02p 0.91p
1.00 V V out+
0.75
0.50 V 1.04p
V out−
1.10p
1.06p 1.94p
arbitrary values. Again, as in Ex. 34.18, the ideal output voltage is 1.0 V (Vout+ =
1.25 V and Vout− = 0.25 V). The simulation results are shown in Fig. 34.60. The
error has dropped from 5 mV to −70 µV. It may be instructive to resimulate the
capacitor error averaging topology of Fig. 34.59 using different values of
capacitors in order to get a feeling for just how forgiving the topology is to
mismatches.
Note that in the simulation netlist, where we are using near ideal op-amps with
open loop gains of 100 million, we added switches in series with the CI capacitors
in the averaging circuit to avoid the situation of the op-amp operating open-loop
with its outputs going to millions of volts when both φa and φh are low. (The
op-amp operates open loop when φs is high, which is usually not a problem in a
practical circuit). T
Error, microvolts
Figure 34.60 Simulation results from Ex. 34.19 showing amplification error.
Example 34.20
Repeat Ex. 34.19 if the op-amps used have open-loop gains of 1,000.
The simulation results are shown in Fig. 34.61. The error has increased by a factor
greater than 10. As indicated earlier in Eq. (34.39), the minimum op-amp open-
loop gain is set by the resolution of the data converter. If we were designing a
14-bit ADC, the op-amp used would require open-loop gains greater than 216 or
64k (96 dB). T
Comparator Placement
The implementation of a pipeline ADC showing how the clock signals are used in each
error averaging stage is shown in Fig. 34.62. The important thing to note is the use of only
three phases of an input clock signal. The connection of the clock phases to the switches
changes in each stage allowing a stage to sample an input signal while the previous stage is
in the hold mode. We'll discuss the generation of these clock signals in the next section;
here we discuss comparator placement and performance.
Chapter 34 Implementing Data Converters 375
Error, microvolts
Figure 34.61 Regeneration of Fig. 34.60 using op-amps with open-loop gains of 1,000.
Clock (3 phases)
One Two Three One
V in+ Error avg. Error avg. Error avg. Error avg.
V in− Fig. 34.59 Fig. 34.59 Fig. 34.59 Fig. 34.59
Amplify, one
Hold, one
Sample, two
Amplify, three
Figure 34.62 Implementation of a pipeline ADC using error averaging (Fig. 34.59).
376 CMOS Mixed-Signal Circuit Design
As seen in Fig. 34.55 the voltages VCI+ and VCI− must be valid and stable when the
amplify-and-hold clock phases are high. Reviewing Fig. 34.51 we can implement the two
comparators using a clock signal followed by a latch, Fig. 34.63. The comparators can be
clocked on the rising edge of the amplifying clock (which is a different clock phase in each
of the three possible clocking schemes). It's important to place a separate clocked latch on
the outputs of each comparator (clocked with a slightly delayed clock signal which isn't
shown in the figure) to ensure that comparator metastability isn't a factor when generating
the MUX addressing (select inputs). We can get away with this type of clocking scheme
because we are using three levels/stage (1.5 bits/stage). The comparators don't have to be
N-bit accurate, as discussed earlier, but can withstand offsets/errors approaching VCM /4.
The signals VCI+ and VCI− do have to be N-bit accurate, though, and must settle and stay
settled to this accuracy by the end of the amplify phase.
Clock (3 phases)
One Two Three
Error avg. Error avg. Error avg.
Fig. 34.59 Fig. 34.59 Fig. 34.59
Amplify, one
Amplify, two
Amplify, three
D Q
D Q
D Q
D Q
D Q
clk
clk
clk
clk
clk
clk
The two bits coming out of the comparators (actually the latches connected to the
comparator outputs) can be thought of as the first delay element shown in Fig. 34.49
(except now it is a 2-bit delay element). Because each of these first delays are clocked on
the rising edge of one phase of the clock signal, we have a problem with synchronizing the
bits together prior to application to the digital correction logic (similar to Fig. 34.53).
Reviewing the clock signals in Fig. 34.62, we see that if we clock all other delay elements
in the pipeline of Fig. 34.49 on the rising edge of "amplify, one" we can synchronize all of
the digital data together.
Chapter 34 Implementing Data Converters 377
Clock Generation
Generating the nonoverlapping clock signals used in, for example, Fig. 34.55 can be
accomplished by using the basic circuit shown in Fig. 34.64. It might be helpful to review
Fig. 33.32 before we discuss the operation of this circuit. To understand the operation of
this circuit (Fig. 34.64), note that when one of the input signals goes low the
corresponding output phase goes high. The feedback ensures that the output doesn't go
high until the previous phase goes back low (both inputs of the NOR gate must be low
before its output goes high). The amount of nonoverlap time, again, is set by the delay in
series with the output of the NOR gates.
In1
Phase1
Phase3
In3
Figure 34.64 Circuit used for generating three phases of a nonoverlapping clock.
The input signals (overlapping clock signals) are generated with the circuit shown
in Fig. 34.65. The outputs of this circuit change states on both the rising and the falling
edges of the input clock signal. The latches are clocked on both the rising and falling
edges of the input clock. They can be implemented with a parallel connection of the
dynamic latches given in the last chapter (one latch clocked on the rising edge and one
clocked on the falling edge with both latches sharing the common output MOSFETs). If
the input clock signal is 25 MHz with a period of 40 ns, then the width of each of the
output nonoverlapping clock phases in Fig. 34.64 is less than (but approximately) 20 ns.
Note that there are other ways to generate the input clock signals for the circuit of Fig.
34.64. (A shift register with one bit set low is one example.) The method shown starts
with both latches' outputs set low, 00 (= Q 1 Q 2 ). On the rising edge of the clock signal the
output changes to 01. On the following falling edge the output changes to 11. On the next
rising edge the output changes to 01 and the sequence repeats itself.
378 CMOS Mixed-Signal Circuit Design
Q1
D Q
In1
Clkin
clk Q
Q2 In2
D Q
clk Q
In3
Figure 34.65 Generating the overlapping input clock signals for Fig. 34.64.
Example 34.21
Simulate the operation of the circuit shown in Fig. 34.66. Show the input and
output signals as the difference between the two differential input signals. Assume
the common-mode voltage coming out of the op-amp is precisely 0.75 V.
V CI+
0.75 V V out+
V out−
V CI−
The simulation results are shown in Fig. 34.67. Note how, as we would expect, the
gain of the circuit is two. The signals shown are ideal. T
V out+ − V out−
V in+ − V in−
Figure 34.67 Input and output for the circuit of Fig. 34.66.
Example 34.22
Repeat Ex. 34.21 if the common-mode voltage coming out of the op-amp sees an
offset of 100 mV, i.e., it is 850 mV.
The differential input and output signals with this output common-mode offset
look exactly like what is seen in Fig. 34.67. The single-ended output signals, Fig.
34.68, show the offset (the signals swing around 850 mV in Fig. 34.68). Clearly a
common-mode offset in the op-amp output signals isn't a concern (unless it's so
large that it limits the op-amp output swing range). Likewise an offset in the
common-mode voltage of the input signals isn't a concern (this comment is easy to
verify with simulations). What is a concern, though, is the value of the voltages
used for VCI+ and VCI−. T
V out+
V out−
Figure 34.68 Output signals for the circuit of Fig. 34.66 if common-mode voltage is 0.85 V.
380 CMOS Mixed-Signal Circuit Design
In Fig. 34.66 VCI+ and VCI− are shorted together and connected, through a switch,
to 0.75 V (VCM). It can be shown that when this occurs, the absolute value of the voltage is
irrelevant because it is common to both input signal paths, see Eq. (34.53). This means
that if we used ground instead of VCM in Fig. 34.66, we would get the same outputs. A
problem does occur if the difference in the voltages (when adding or subtracting a voltage
from the input signal) isn't exactly what is desired, see Eq. (34.61).
While an offset in the common-mode voltage isn't important, the op-amp's offset is
a concern. The op-amp offset voltage is zeroed out when using the topology shown in Fig.
34.66. Equation (34.50) was derived assuming the op-amp had a nonzero offset voltage
and shows the offset will not (ideally) affect the building block's output signals. To show
the removal of the offset using simulations, consider the following example.
Example 34.23
Simulate the operation of the circuit shown in Fig. 34.69. This figure is Fig. 34.66
redrawn with a 100 mV op-amp offset.
The simulation results are shown in Fig. 34.70. Figure 34.70 should be compared
to Fig. 34.67. Note how the unrealistically large offset has no effect on the building
block's output signals. T
0.75 V V out+
100 mV
V out−
If the op-amp offset is zeroed out and doesn't affect the circuit's outputs when
using the basic topology shown in Fig. 34.30, why would we want to consider some other
topology? The answer to this question comes from the realization that the op-amp's
outputs must settle to the final accuracy of the ADC, referring to Fig. 34.55, by the falling
edge of all three phases of the clock. It would be nice to make the settling during one of
these three (or two) phases irrelevant. Also, and perhaps more importantly, it would be
nice to use other types of CMFB (discussed in the next section).
Chapter 34 Implementing Data Converters 381
V out+ − V out−
V in+ − V in−
Figure 34.70 Input and output for the circuit of Fig. 34.69 with op-amp offset of 100 mV.
Consider the building-block topology shown in Fig. 34.71. During the sampling
phase, the inputs of the op-amp are shorted to the common-mode voltage. During this
time the op-amp operates open loop, and settling time is meaningless. The status of the
op-amp outputs during this time is discussed in the next section. The charge stored on the
capacitors during the storage phase is
φs
Q I,F = C I,F ⋅ (V in − V CM ) (34.94)
while during the hold phase
φ φ
Q I h = C I ⋅ (V CI − V CM ± V OS ) and Q Fh = C F ⋅ (V out − V CM ± V OS ) (34.95)
V CM
φs
φs
φh
CF
φh
φh CI
V in+
V CI+ V out+
φs
V CI− V out−
V in− CI
CF φh
φs
V CM
Once again knowing charge must be conserved, the output voltage can be written as
unwanted term
Example 34.24
Repeat Ex. 34.23 using the topology shown in Fig. 34.71.
The simulation results are shown in Fig. 34.72. The ratio of CI to CF is one and so,
ideally, the output is just twice as large as the input. The output should swing from
2 V down to −2 V. However, because of the 100 mV offset, which is also
multiplied by a factor of 2, the output is shifted downwards so that it swings from
1.8 V to −2.2 V. T
V out+ − V out−
V in+ − V in−
Figure 34.72 Input and output for the circuit of Fig. 34.71 with op-amp offset of 100 mV.
Dynamic CMFB
The CMFB circuits discussed earlier employ an amplifier to sense the average of the
outputs and feedback a correction to center the signals around VCM. In Ch. 27 we
discussed a CMFB technique used in an op-amp that was dynamic and doesn't employ an
amplifier. A simplification of this dynamic CMFB can be realized by noting that during the
sampling phase in Fig. 34.71 the op-amp inputs are forced to VCM. The scheme we are
about to present won't force the inputs to V CM ± V OS during the sample phase as in the
other topologies based on Fig. 34.30.
Chapter 34 Implementing Data Converters 383
The basic dynamic CMFB circuit is shown in Fig. 34.73. During the sample phase
of the clock the inputs and outputs of the op-amp are shorted to the common-mode
voltage. Also during this time the common-mode feedback voltage, VCMFB, is set to a bias
voltage (Vbias4 if the op-amp of Fig. 34.36 is used [see Fig. 33.60]). During the hold phase,
the CMFB capacitors on the output of the circuit are disconnected from both VCM and
Vbias4 and are used to sense the average value of the outputs. If the outputs move in a
balanced fashion, then VCMFB remains equal to Vbias4. If the average of the outputs moves
upwards above VCM, then VCMFB increases, pulling the output common-mode voltage
downwards. Again, because the CMFB loop utilizes negative feedback, an increase in
VCMFB must result in a decrease in (v o+ + v o− )/2 .
V CM
φs φh
v o+
v o−
V CMFB
φs V CM
V CM
V bias4
CMFB capacitors φs
Looking at Fig. 34.73 we see that because of the op-amp's offset voltage the
outputs of the op-amp will source/sink a current into VCM during the sample phase of the
circuit's operation. By adding an extra pair of switches to disconnect the CMFB capacitors
from the op-amp outputs we can avoid this situation. Adding the switches causes the
op-amp outputs to approach the power supply rails during the sample phase (because of
the offset voltage). This output railing isn't a problem if an OTA (single-stage) topology
like the one in Fig. 34.36 is used. The outputs have no capacity to hold charge and so
when the CMFB capacitors are reconnected to the op-amp outputs, the outputs
immediately go to VCM (neglecting the connection of the feedback capacitors which are not
shown in Fig. 34.73). If a two-stage op-amp is used then care must be taken to ensure that
the compensation capacitor (which, of course, does have the capacity to hold charge)
doesn't charge to the rail resulting in a large recovery time when the op-amp circuit is put
in the hold mode.
384 CMOS Mixed-Signal Circuit Design
One solution to the problem of offset (when using the topology of Fig. 34.71) is to
AC couple the output of the first op-amp stage to the input of the second op-amp stage, as
in Fig. 34.74 (add a capacitor between the two stages to remove the DC component). The
first stage's offset is stored on the capacitors during the sample phase, while the second
stage's inputs are shorted to Vbias1. When the op-amp is used in the hold phase of
operation, the second stage's offset is referred back to the input of the op-amp after
dividing by the first stage's gain. For example, if the first stage of the op-amp has a gain of
50 and the second stage's offset is 10 mV, then the input-referred offset of the op-amp in
the hold mode is only 0.2 mV. Note that the gain of the first stage can't be too large. If,
for example, the first stage's gain were 1,000 and its offset were 10 mV, the outputs
would saturate and the offset would not be stored on the capacitors. If the gain of the first
stage were 50 with a 10 mV offset, then the (differential) outputs of the first stage would
change by 500 mV. As long as the MOSFETs remain in the saturation region, the offset
storage works as desired.
N V CM V CM N
φs φs
v o− v o+
P 2W P
2W
V CMFB
φs
V bias4
φs φs
V CM
Depending on the design parameters the op-amp shown in Fig. 34.74 would be
compensated using either a load capacitance (as if the op-amp were a single stage) or
using Miller compensation by adding a capacitor from the output of the op-amp to the
output of the diff-amp (first stage) with a series resistance to cancel the right-hand plane
zero (as discussed in Ch. 25).
Layout of Pipelined ADCs
Before leaving this chapter, let's comment on the layout of pipeline ADCs. Using a fixed-
height layout structure, with the automatically routed power and ground busses (when the
cells are laid end-to-end), is a good place to start when laying out the op-amp. As seen in
Fig. 34.75, a height (with example values shown) is selected with the width variable. It's
important, as discussed in Ch. 28, to keep the analog signals separate, both physically and
by distance, from the digital signals.
Analog VDD 10 um
Analog ground 10 um
A possible block diagram of the placement of the fixed-height cells together with
the capacitors and switches used to implement a stage of the ADC is seen in Fig. 34.76.
Looking at the input signal, we notice that they are laid out next to each other and routed
as close as possible to the input of the first stage of the ADC. All of the differential analog
signals in the ADC should follow this practice to help make any coupled noise truly
common-mode. Notice how we have placed ground pads adjacent to the input signals.
These pads are not used for ground connections on-chip. The pads are used to help reduce
noise coupling into the input signals. Ideally, these ground pins provide a termination for
the noise keeping the input signals "clean." This is especially important when bonding
wires connect the integrated circuit to a padframe in the final packaged part. The bonding
wires used for digital signals tend to radiate more than enough signal to corrupt the input
signal and ruin the ADC's SNR when placed close to an analog low-level signal.
Next notice that we must have a clock signal in our analog domain. This signal, as
we have seen, is used for clocking the switches in the S/H stage. Although in the figure we
show the placement of the clock adjacent to the input signals, it may be better to move the
pad and, if possible, the routing of the clock signals away from the inputs. The main goal
386 CMOS Mixed-Signal Circuit Design
VDD Reference
VDD ground VDD ground VDD ground
GND Reference
Op-amp Op-amp Op-amp
VDD Power with bias with bias with bias
GND Power capacitors capacitors capacitors
clock clock
drivers switches switches switches
GND
V in+
V in−
signals
signals
signals
clk
clk
clk
GND
Analog domain
Digital domain
Digital power
when routing the clock signals is to keep the layout regular. Routing clock signals all over
the layout is asking for problems.
Two sets of power and ground pads are used (more if possible) in the analog
domain for the implementation of the ADC. One set of pads is used for supplying power
to the op-amps while the other set is used for supplying the reference voltages to each
stage. The power and ground supplies are common to both digital and analog sections
off-chip. Off-chip the supplies are connected together and decoupled (bypassed) using
large capacitors (actually a wide range of capacitor values are connected in parallel
between the VDD and ground to avoid the increase in a single large capacitor's effective
series resistance with frequency). On-chip decoupling capacitors can be used as well. The
analog and digital power and ground connections are not shared, and so care must be
taken not to decouple the analog VDD to digital ground or digital VDD to analog ground.
Figure 34.77 shows one example of how the decoupling capacitors can be connected.
Chapter 34 Implementing Data Converters 387
VDD
Off-chip decoupling capacitor.
Digital circuitry
Pin Pad Pad Pin
Figure 34.77 Showing how decoupling capacitors are used in a mixed-signal chip.
In general we don't want any DC current flowing on our reference voltages (those
voltages used for VCI in our op-amp) because of the possible voltage drop along the
supplying line. There may also be a voltage drop along the metal lines supplying power to
the op-amps. However, small (DC) changes in these voltages are usually not a significant
factor in the precision of the ADC. These changes could be a factor if the output signal
approaches the power-supply voltages where the op-amp runs out of head room. AC
changes in the power-supply voltages can be a significant factor limiting the ADC's
performance.
Ground planes and wide conductors should be used where possible. Using power
and ground planes not only provides good distribution of power and ground but also
increases the capacitance between the supplies. Areas, labeled "gutter" in Fig. 34.76
should be provided for low-level analog signals. These areas are free to allow the quiet
routing of the switch inputs and outputs to the op-amp outputs.
It's also a good idea to use guard rings around the sensitive analog circuits (e.g.,
the switched capacitors) to avoid coupled substrate noise. Figure 34.78 shows the basic
idea. In this figure the capacitors are laid out over an n-well. The n-well is tied to analog
VDD through an n+ implant in the n-well and metal. Surrounding the n-well is a ring of
p+. This ring is tied to analog ground. The idea is that the p+ will provide a sink for any
current injection from the surrounding circuitry. Because ground is the lowest potential in
the circuit, the noise will terminate on this p+ and not penetrate the area under the
capacitors (and then hopefully not couple into the capacitors). While this works well by
itself, it may not be enough. Noise currents may still move deep in the substrate and work
their way up under the capacitors. Because the n-well under the capacitor is held at the
most postive potential in the circuit, any noise that does get into the n-well will hopefully
be swept out through VDD and not couple into the capacitors.
388 CMOS Mixed-Signal Circuit Design
p+ tied to ground
n-Well
n+ tied to VDD
Capacitor layout area
Figure 34.78 Using guard rings for protection in sensitive analog blocks.
Finally, if a sensitive analog signal does need to cross a digital signal (an example
being the potential need to feed the switch inputs and outputs across the three phases of
the switch clock signals in Fig. 34.76) a shield should be used, Fig. 34.79. In this figure
the sensitive analog signal is assumed to exist on Metal1, while Metal2 is used for an
analog ground shield from the digital signal on Metal3. This shield is used for isolation
providing a terminating plane for the electric fields resulting from the voltages on both the
digital and analog signals.
Metal2 shield
(tied to analog ground)
Layout view Digital signal, Metal3
Analog signal
REFERENCES
[1] Y. Taur and T. H. Ning, Fundamentals of Modern VLSI Devices, Cambridge
University Press, 1998.
[2] R. E. Suarez, P. R. Gray, and D. A. Hodges, “All-MOS Charge Redistribution
Analog-to-Digital Conversion Techniques - Part II,” IEEE Journal of Solid-State
Circuits, Vol. 10, No. 6, pp. 379-385, December 1975.
[3] R. J. Baker, H. W. Li, and D. E. Boyce, CMOS Circuit Design, Layout, and
Simulation, Wiley-IEEE, 1998.
[4] S. Franco, Design with Operational Amplifiers and Analog Integrated Circuits,
Second Edition, McGraw-Hill, 1998.
[5] J. A. Schoeff, An Inherently Monotonic 12-bit DAC, IEEE Journal of Solid-State
Circuits, Vol. SC-14, No. 6 December 1979, pp. 904-911.
[6] C. G. Yu and R. L. Geiger, "An Automatic Offset Compensation Scheme with
Ping-Pong Control for CMOS Operational Amplifiers," IEEE Journal of
Solid-State Circuits, Vol. 29, No. 5 May 1994, pp. 601-610.
[7] C. C. Enz and G. C. Temes, "Circuit Techniques for Reducing the Effects of
Op-Amp Inperfections: Autozeroing, Correlated Double Sampling, and Chopper
Stabilization," Proceedings of the IEEE, Vol. 84, No. 11, pp. 1584-1614,
November 1996.
[8] I. E. Opris, L. D. Lewicki, and B. C. Wong, "A Single-Ended 12-bit 20 Msample/s
Self-Calibrating Pipeline A/D Converter," IEEE Journal of Solid-State Circuits,
Vol. 33, No. 12, December 1998.
[9] E. Fong, private communication, May 2001.
[10] B. Ginetti, P. G. A. Jespers, and A. Vandemeulebroecke, "A CMOS 13-b Cyclic
RSD A/D Converter," IEEE Journal of Solid-State Circuits, Vol. 27, No. 7, July
1992.
[11] W. C. Black and D. A. Hodges, "Time Interleaved Converter Arrays," IEEE
Journal of Solid-State Circuits, Vol. SC-15, No. 6, December 1980.
[12] B. S. Song, M. F. Tompsett, and K. R. Lakshmikumar, "A 12-bit, 1-MSample/s
Capacitor Error-Averaging Pipelined A/D Converter," IEEE Journal of Solid State
Circuits, Vol. 23, No. 6, pp. 1324-1333, December 1988.
[13] H. S. Chen, B. S. Song, and K. Bacrania, "A 14-b 20-Msample/s CMOS Pipelined
ADC," IEEE Journal of Solid State Circuits, Vol. 36, No. 6, pp. 997-1001, June
2001.
QUESTIONS
34.1 Assuming the DAC shown in Fig. 34.1 is 8 bits and VREF+ = 1.5 V and VREF− = 0,
what are the voltages on each of the R-2R taps?
390 CMOS Mixed-Signal Circuit Design
34.2 Give an example of how the traditional current-mode DAC will have limited
output swing.
34.3 Repeat question 34.1 for the DAC shown in Fig. 34.2.
34.4 For the wide-swing current mode DAC shown in Fig. 34.3, what are the voltages
at the taps along the R-2R string assuming 8 bits, VREF+ = 1.5 V, VREF− = 0, and a
digital input code of 0000 0000?
34.5 Can the op-amp shown in Fig. 34.36 be used in fully-differential implementations
of the DACs shown in Figs. 34.1 - 34.3? Why or why not?
34.6 Show the detailed derivation of Eqs. (34.12)-(34.14).
34.7 Why would we want to use both current segments and binary-weighted currents to
implement a current-mode DAC? (Why use segmentation?)
34.8 Why do we subtract ∆A in Eq. (34.36)? Why not add the gain variation?
34.9 Does the matching of the capacitors matter in the S/H of Fig. 34.31? Why or why
not?
34.10 Neglecting offsets and assuming the φ1 switches are closed in Fig. 34.34, can the
two inverting terminals of the error amplifier (the input terminals of the fully-
differential op-amp) be at different potentials?
34.11 When the φ3 switches are closed in Fig. 34.34, is it possible for the inverting inputs
of the error amplifier to be at different potentials? Again, neglect offsets.
34.12 Repeat Ex. 34.10 if the cyclic ADC's input is 0.41 V.
34.13 Is kick-back noise from the comparator a concern for the circuit of Fig. 34.39?
34.14 Derive the transfer function for the circuit shown in Fig. 34.80.
φs φh
CF φs
V CF+ φh
φh
CL
CI
V in+ V out+
V CI+
V CI− CI V out−
V in−
CL
V CF− CF
φs
φh
In1
Phase1
Phase3
In3
Phase4
In4
34.22 What is the main advantage of using dynamic CMFB over other CMFB circuits?
What is the main disadvantage?
392 CMOS Mixed-Signal Circuit Design
34.23 Can MOSFETs be used to implement the on-chip decoupling capacitors in Fig.
34.77?
34.24 Sketch the cross-sectional view of the layout in Fig. 34.78.
34.25 Figure 34.82 shows the implementation of a pipeline DAC. How would we
implement this DAC using a topology similar to Fig. 34.42? Sketch the DAC's
implementation and the timing signals (clock phases) used.
35
Integrator-Based CMOS Filters
We've covered the detailed design of analog interfaces used for analog-to-digital and
digital-to-analog conversion in the past five chapters. While we can perform signal
processing and filtering in the digital domain, as seen in Fig. 30.2, analog antialiasing and
reconstruction filters are still required in our system. Analog continuous-time filters (a
simple example being the RC circuit shown in Fig. 30.9) can be faster (have wider
bandwidths) and take up less area than their discrete-time counterparts. However, unlike
discrete-time filters, continuous-time filters cannot be fabricated with precise transfer
functions and must be tuned. This is especially true if passive resistors and capacitors are
used where each can have a variation of ± 20% . By using active CMOS integrators in the
filter implementations instead of passive elements, we can electrically tune the filters. Also,
we can more easily implement higher order filters while minimizing the effects of loading.
In this chapter we discuss analog filters (both continuous- and discrete-time filters
where the input and output are analog) made using continuous-time analog integrators
(CAIs or active-RC integrators), MOSFET-C integrators, transconductor-capacitor
(gm-C) integrators, and discrete-time analog integrators (DAIs). We also discuss digital
filter design (which, of course, is also a discrete-time filter) based on the topologies
developed using the digital integrator (e.g., see Fig. 31.34) where the input and output are
digital.
f 3dB = 1
2πRC
v in (t) R v out (t) 0 dB f (Hz)
V out ( f ) 20 dB/decade
C
Vin ( f )
V out ( f )
∠
s=− 1 Im Vin ( f ) 0
RC
45
Re 90
where ω = 2π ⋅ f and f is the frequency of the input (and thus the output). Next consider
the block diagram in Fig. 35.2. This figure shows an integrator and a summing block
(which by now we know can be implemented with a single op-amp). The output of the
block diagram can be determined by solving
V out ( f ) = G
s ⋅ (Vin ( f ) − V out ( f )) (35.2)
or
V out ( f )
= 1 (35.3)
V in ( f ) 1 + s/G
where for a sinewave input s = jω . Comparing this equation to Eq. (35.1), we see that if
we set the integrator's gain, G, using
In G Out
V in ( f ) s V out ( f )
Integrator
R
V in+ V out−
V in− V out+
R
Switched so gain is
positive
C
Reviewing Fig. 35.2, we see that the CAI of Fig. 35.3 alone will implement the
needed integration but not the summing (difference) block. By adding an additional
feedback path, as we did in Fig. 32.24, the entire block diagram of Fig. 35.2 can be
implemented. Figure 35.4 shows the integrator-based implementation of the circuits in
R C
V out−
R
V in+ V out−
R
V in− V out+
R
Figs. 35.1 and 35.2 (noting the op-amp must be able to drive a resistive load). This filter is
called an active RC filter 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.
Example 35.1
Simulate the operation of the filter in Fig. 35.4 from DC to 100 MHz if R = 10k
and C = 10 pF. Show both the magnitude and phase responses of the filter.
Assume the op-amp is ideal.
From Fig. 35.1 we know the 3 dB frequency of the filter is 1.59 MHz. The
simulation results are shown in Fig. 35.5. The magnitude and phase response
follow, as expected, the responses for the simple RC filter shown in Fig. 35.1. T
V out ( f )
V in ( f )
V out ( f )
∠
V in ( f )
Figure 35.5 Magnitude and phase responses for the first-order filter in Fig. 35.4
if R = 10k and C = 10 pF.
What would happen if we switched what we define as V out+ and V out− in the filter
described in Ex. 35.1 without changing any other connections? Perhaps it is trivial, but the
answer is that the output will be inverted. We can modify the block diagram of Fig. 35.2
by simply multiplying the output by −1, as seen in Fig. 35.6. The phase shift in Fig. 35.5
would shift up or down by 180 degrees. It would vary from 180 to 90 degrees, or from
Chapter 35 Integrator-Based CMOS Filters 397
V in ( f )
In G1 Out
s 1 V out ( f )
Integrator
G2
Switch for inversion
see Fig. 35.4
RF C
G1 = 1
RI RIC
V in+ V out+ R
G2 = I
RI RF
V in− V out−
G1G2
RF f 3dB =
2π
C
−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.
35.6, 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. 35.4), we can write
V in V out V out
− = (35.6)
RI R F 1/sC
It's important to note (for use later) that in order to subtract the output from the input, the
voltages are first changed to currents. This equation can be rewritten as
RF
V out RI
= (35.7)
V in 1 + sR F C
Using the block diagram in Fig. 35.6, we can write
1
V out G2 G G
= and f 3dB = 1 2 (35.8)
V in 1 + G sG 2π
1 2
398 CMOS Mixed-Signal Circuit Design
Example 35.2
Modify the filter in Ex. 35.1 so that the low-frequency gain is 20 dB.
Using Eq. (35.7) or Eq. (35.8), we leave C = 10 pF and RF = 10k. To get the gain
of 10, we make RI = 1k. The simulation results are shown in Fig. 35.7. T
V out ( f )
V in ( f )
where v+ and v− are the voltages on the noninverting and inverting op-amp input terminals,
respectively. (Note how we are using f3dB in both Figs. 35.6 and 34.21 to indicate the 3 dB
frequency of a frequency response.) When a practical op-amp is operating at 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
A OLDC ⋅ f 3dB f u 2πf u ω u
A OL ( f ) ≈ = = s = s (35.11)
j⋅f j⋅f
Chapter 35 Integrator-Based CMOS Filters 399
where, again from Fig. 34.21, fu is the frequency where the op-amp's open loop gain is
unity (0 dB). Rewriting Eq. (35.6) to include the op-amp's finite gain bandwidth product
(that is, fu ) 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
V in − v − V out − v −
− = sC ⋅ (V out − v − ) (35.12)
RI RF
After some algebraic manipulation with v − = −V out /A OL ( f ) , we get
RF
V out RI
= (35.13)
1 + sCR F + A OL ( Ff ) + A OL1( f ) 1 − RFI
V in sCR R
Desired response
or, with A OL ( f ) = ω u /s
RF
V out RI
= (35.14)
+ s ⋅ CR F + ω1u 1 − RFI + 1
V in CR F R
s2 ωu
This equation is very revealing and shows just how significant a limitation the op-amp can
be in a filter. For the moment, to simplify things, let's assume ω 2 << ω u /CR F so that the s2
term in Eq. (35.14) is negligible. We can then write the magnitude and phase responses as
RF
= −tan−1 ωCR F + ωω 1 − F
V out RI V out R
= and ∠
V in 2 Vin u RI
1 + ωCR F + ωωu 1 − RFI
R
(35.15)
Example 35.3
Suppose a first-order filter is designed based on the topology seen in Fig. 35.6,
where ω u = 10/R F C and R F /R I = 10. Assuming ω 2 << ω u /CR F , comment on how
the magnitude and phase responses of the filter will be affected by the finite
op-amp, fu.
The op-amp's unity gain frequency is only 10 times larger than the bandwidth of
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 10 op-amp
circuit is f u /10, which here is equal to the ideal filter 3 dB frequency of 1/2πR F C .
The magnitude of the filter's response can be written as
RF
= −tan−1 ω ⋅
V out RI V out CR F
= and ∠
V in CR F 2 Vin 10
1 + ω ⋅
10
The point of the preceding example is, in general lowpass filter design, to minimize
the effects of the op-amp's finite fu a low value of closed-loop DC gain should be used. In
the remaining discussion let's assume R F /R I = 1 , so Eq. (34.14) simplifies to
V out 1
= CR F
(35.16)
V in ⋅ s + s ⋅ CR F + 1
2
ωu
noting that if ω u → ∞ , then s p1 = ∞ , and s p2 = −1/CR F (the ideal position of the pole, see
Fig. 35.1).
To get some idea of the required op-amp fu (ω u = 2πf u ) , let's assume that we want
the pole to vary no more than 1% from the ideal location due to finite op-amp bandwidth
CR
−1 = 99 ⋅ −CR F − (CR F ) − 4 ω u
2 F
CR
(35.18)
CR F 100 2 ⋅ ω uF
This can be rewritten as
ω u ⋅ CR F 1
1.01 = − (ω u ⋅ CR F ) 2 − 4(ω u ⋅ CR F ) (35.19)
2 2
If we let x = ω u ⋅ CR F , then we need to solve
x − x 2 − 4x = 2.02 (35.20)
knowing x is positive and much larger than one (ω u >> 1/CR F ). Solving Eq. (35.20) for x,
results in x = 100. This means the op-amp's unity gain frequency must be 100 times larger
than the filter's f3dB in order for the variation of this frequency (the pole) to deviate less
than 1% from the ideal. If we can withstand a 10% decrease in the filter's cutoff frequency,
then fu need only be 10 times larger than the filter's f3dB. Clearly, from Eq. (35.14), the
first-order filter's frequency response is actually second-order when the op-amp's gain
bandwidth product fu is a factor. Therefore, the shapes of the magnitude and phase
responses will deviate from the ideal first-order shapes seen in Fig. 35.1 so we can draw
two very practical conclusions. First, even if it were possible to fabricate precise resistor
and capacitor values, the limitations of the op-amp's finite bandwidth may still require the
use of tuning when filtering with active-RC integrator-based filters. Tuning would consist
of adding or removing resistors and capacitors to adjust the precise filter cutoff frequency
(adding/removing the elements using either fuses or, if possible, MOSFET switches).
Second, the op-amp's fu should be at least 10 times larger than the cutoff frequency ( f3dB )
of the filter (again assuming a closed-loop DC gain of unity, i.e., RF /RI = 1). This is a
general "rule-of-thumb." Precision filters (well-defined magnitude and phase responses)
would require wider bandwidth op-amps. Consider the following example.
Chapter 35 Integrator-Based CMOS Filters 401
Example 35.4
Repeat Ex. 35.1 if an op-amp is used with a DC gain of 10,000 and an fu of 10
MHz.
Because the op-amp's AOLDC is 10,000 and fu = 10 MHz, the op-amp's open loop
f3dB is 1 kHz (see Fig. 34.21). We can use the circuit shown in Fig. 35.8 (see also
Fig. 32.45) in our SPICE simulation to model an op-amp with finite fu. The RC in
Fig. 35.8 is selected to give an op-amp open loop f3dB of 1 kHz.
v o+
1k
10k 1
v−
E1
V CM 159n
op-amp outputs
v+ 159n
E2
1k
v o−
10k
1
Figure 35.8 SPICE modeling a differential input/output op-amp with finite bandwidth.
The 3-dB frequency of the filter described in Ex. 35.1 is, under ideal
conditions, 1.59 MHz. Because our op-amp's unity gain frequency is only 10 MHz,
we would expect, from Eq. (35.16), the op-amp to affect the frequency response
of the filter. Figure 35.9 shows the simulation results using the op-amp model of
Fig. 35.8. Comparing Fig. 35.9 to Fig. 35.5, we see differences in both the
magnitude and phase responses of the filters. The magnitude response of Fig. 35.9
initially rolls off at 20 dB/decade below the ideal 1.59 MHz. Around 10 MHz the
response transitions to 40 dB/decade. Clearly this faster roll off is the result of the
op-amp's closed-loop pole coming into play. The limiting behavior of the op-amp,
when looking only at the magnitude response, may be welcome (the filter's
response rolls off faster) in a lowpass filter. However, it is not welcome in other
filters (a highpass filter, for example). Figure 35.10 shows what happens if we
decrease the filter's 3 dB frequency to 159 kHz by increasing the resistors used to
100k. What we are doing here is showing how making the op-amp's bandwidth
much larger than the filter's affects the frequency response of the circuit. The
magnitude response starts to fall off at −40 dB/decade at the op-amp's unity gain
frequency, fu, of 10 MHz. Also seen in Fig. 35.10 is the phase response of the
filter. The op-amp's (closed-loop) phase response, which starts rolling off one
decade below fu, results in the final phase shift of the filter approaching −180
degrees. T
402 CMOS Mixed-Signal Circuit Design
V out ( f )
V in ( f )
V out ( f )
∠
V in ( f )
Figure 35.9 Magnitude and phase responses for the first-order filter in Fig. 35.4
if R = 10k and C = 10 pF using an op-amp with a 10 MHz unity-gain
frequency.
−20 dB/decade
−40 dB/decade
Figure 35.10 Increasing the resistance to 100k and replotting Fig. 35.9.
Chapter 35 Integrator-Based CMOS Filters 403
V out ( f )
= 1 ⋅ 1 (35.21)
V in ( f ) 1 + s/G 1 + j f
fu
This result could have been used in the previous example to predict how the op-amp
affects the filter's behavior. If the filter has gain (> 1) in the passband, see Eq. (35.8), we
can modify this equation to read
Undesired
1
V out ( f ) G2 1
= ⋅ (35.22)
V in ( f ) 1 + G 1sG 2 1 + j f
G 2 ⋅f u
For a higher order filter we would multiply the desired frequency response by the
undesired term's (the op-amp's) response for each op-amp used in the circuit. Clearly, this
limits the order of the filter (limits the number of op-amps used in a circuit; a first-order
filter uses one op-amp, a second-order filter uses two op-amps, etc.). This is especially
true if the filter has a passband approaching the fu of the op-amps used.
Active-RC SNR
Consider the single-ended active-RC filter shown in Fig. 35.11. Let's assume an ideal
op-amp with a maximum RMS output voltage of VDD/(2 2 ) . The RMS input-referred
noise of the filter, assuming thermal noise dominates over the bandwidth of interest, is
simply kT/C . The filter's SNR can then be written as
VDD/ 2 2 2
SNR = 20 ⋅ log = 10 ⋅ log VDD /8 (35.23)
kT/C kT/C
This result shouldn't be too interesting at this point. As we've already seen, the size of the
integrating capacitor fundamentally sets the SNR in integrator-based data converters or
C V out
VDD
VDD
V in R
VDD /2 V CM
V out
V CM 0
time
modulators. But consider the following: the maximum electrical energy stored in the
capacitor used in an integrator is
2
Maximum electrical energy = 1 C ⋅ VDD (35.24)
2 2
Equation (35.23) can then be rewritten as
2
1 VDD
2
C
2 2 Electrical energy
SNR = 10 ⋅ log VDD /8 = 10 ⋅ log = 10 ⋅ log (35.25)
kT/C kT Thermal energy
This equation can also be used to estimate the fundamental dynamic range, DR, of a filter.
That is, we can estimate DR by assuming DR = SNR. Of course, as discussed in Ch. 31, a
more accurate estimate of DR is the measured signal-to-noise plus distortion ratio, SNDR.
Practically, DRs approaching 90 dB (15 bits) can be attained using active-RC filters with
good polysilicon resistors (to avoid the large nonlinear voltage coefficient associated with
diffused or implanted resistors) and linear capacitors. Bandwidths approaching 50 MHz,
assuming 500 MHz fu op-amps are used, can be attained (at, of course, lower DRs).
35.1.3 MOSFET-C Integrators
Let's now look at a variation of the active-RC filter where the resistors are replaced with
MOSFETs. Figure 35.12 shows a MOSFET-C filter. In order for the MOSFETs to behave
as resistors they must remain in the triode region. Using long length devices helps ensure
triode operation. Because the MOSFETs are operating as resistors, their speed is not
governed by their gate-source voltage or channel length, as indicated in Eq. (33.20).
However, the linearity of the MOSFET resistors is still very important as is the possiblity
that the MOSFETs will introduce a parasitic pole into the filter's frequency response
because of the distributed resistance/capacitance of the channel (Fig. 35.12). Figures 33.23
and 33.29 show how the channel resistance of an NMOS device changes with VDS. For
large input signals, the active MOSFET resistors become nonlinear, resulting in filters with
SNDRs of around only 40 dB. The bandwidth of the MOSFET-C filters parallels that of
the active-RC filters.
We might be questioning the usefulness of the MOSFET-C filter with an SNDR of
only 40 dB. Clearly this filter will only find use in data conversion systems using six bits of
resolution or less (36 dB DR) or in systems that process continuous-time signals. The big
benefit of this filter over the active-RC filter is its ability to be tuned. Tuning the active-RC
filter required adding or removing, via switches or fuses, resistors or capacitors in parallel
or series with the existing resistors and capacitors. Tuning the MOSFET-C filter shown in
Fig. 35.12 can be accomplished by adjusting Vtune upwards or downwards. If we assume
long-channel behavior, we can write the resistance (see Ch. 9) of the MOSFETs in terms
of Vtune (assuming the input common-mode voltage of the op-amp is 0) as
Rn = 1 (35.26)
KP ⋅ WL ⋅ V tune − V THN − V DS
=Vin
Chapter 35 Integrator-Based CMOS Filters 405
V tune
MOSFET-C filter
C
M1
V in+ V out−
V in− V out+
G
Drain current
S D
One over the slope of n+ n+
this line is the MOSFETs
resistance
V tune+ V tune−
V in+
M1A
V in−
M1B
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 35.14 shows the
implementation of a first-order MOSFET-C filter using linearized MOSFETs.
V tune+ V tune−
M1A M1B
V in+
V out−
V in− V out+
i out+ − i out−
i out+
v in+ v out+
Slope = g m
C
v in−
v out− v in+ − v in−
i out−
(a) Schematic symbol of an OTA
or transconductor. (b) Transfer curves for an OTA.
VDD
i out− i out+
Tuning current
v out− v out+
V CM
v in+ v in−
V bias3
Before discussing these issues in more detail let's look at an important limitation of
gm-C filters; namely, the fact that the transconductor's input voltage must vary. As we
discussed in Sec. 34.1.2, 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 SNDR of gm-C
filters to around 50 dB. Again, not too useful if used as an antialiasing or reconstruction
filter unless the system's resolution is less than eight bits (48 dB SNR). The gm-C filter
finds extensive use in continuous-time signal processing.
We can relate the input voltage difference to the output voltage difference for the
circuit in Fig. 35.15a using
I out+ I g m (Vin+ − Vin− )
V out+ − V out− = = − out− = (35.30)
jωC jωC jωC
where, for example, V out+ = V out+ ( f ) , is the frequency domain representation of the output
voltage v out+ (t) . Comparing this result to Eq. (35.5), we can use the same design
techniques if we require
gm
G= 1 = or f 3dB = G (35.31)
RC C 2π
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.
The circuit of Fig. 35.15a implements an integrator, as does the active-RC circuit
of Fig. 35.3. However, as seen in Fig. 35.2, we also need to implement a summing block
in a first-order filter. Toward the goal of implementing the summing block, consider the
transconductor circuit shown in Fig. 35.16a. The output current of a single transconductor
g m (v in+ − v in− )
v in+
v in−
(a)
Switched for subtraction
v in+ v out−
C
v in− v out+
(b)
is, as shown in the figure, g m (V in+ − V in− ) . We can sum this current with the output current
from the second transconductor to implement the summing block in Fig. 35.2. As seen in
Fig. 35.16b, the outputs of the two OTAs are combined, so the currents output from each
transconductor subtract. Assuming each transconductor has the same transconductance,
we can write
g m (V in+ − V in− ) − g m (V out+ − V out− ) − jωC(V out+ − V out− ) = 0 (35.32)
or
V out+ − V out− 1
= (35.33)
V in+ − V in− 1 + jωC ⋅ g1m
If the transconductors have different gms, we can design a filter with a DC gain (see
Problem 35.10), which can be characterized using Eq. (35.8).
Example 35.5
Repeat Ex. 35.1 using a gm-C filter with a gm of 100 µA/V.
To simulate a transconductor using SPICE, a voltage-controlled current source
can be used as seen in Fig. 35.17. 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.
Not using these resistors after reviewing Fig. 35.16 would result in an unknown
common-mode voltage on the second transconductor input.
In order to have the same time constant, and thus pole location, as in Ex. 35.1,
let's set the capacitor value, in the schematic of Fig. 35.16 to 10 pF. The value of
the transconuctance, 1/gm, is 10k. The simulation results are shown in Fig. 35.18.
As we would expect, the shape follows, exactly, that of the active-RC filter in Fig.
35.5. Also, although not shown, the phase response matches as well. T
i out+ V CM
v in+ v out+
big
v in− i out+
v out− v out+
i out− voutp
v in+ vinp
v in− 100u
vinm voutm
v out−
SPICE i out−
big
G1 voutm voutp vinp vinm 100u
V CM
Notice how SPICE defines positive current flow as current
flowing from the + terminal to the terminal
i out+ 2C
i out+
v in+ v out+ v in+ v out+
C
v in− v in− v out−
v out−
i out− i out− 2C
Figure 35.19 Showing how we break the capacitor up to provide a load for
the CMFB circuit.
Chapter 35 Integrator-Based CMOS Filters 411
A High-Frequency Transconductor
A transconductor, Fig. 35.20, can be implemented using the common-mode noise
elimination discussed in Ch. 33 [6]. To increase the input common-mode range, the
lengths of INV1 and INV2 can be increased as in the diff-amp of Fig. 35.15. This, again,
lowers the location of the parasitic poles introduced into the transconductor's response.
Note, in this circuit, that other than the power supplies and the tuning voltage, there are
only two sets of nodes: the input and the output nodes. This allows the transconductor's
capacitances to sum with the load capacitances and be tuned out.
V tune
Inverter power supply
INV1
v in+ v out+
INV2
v in− v out−
V out (z) C −1 C fs 1
= I ⋅ z −1 ≈ I ⋅ =G⋅ s (35.40)
V in (z) − V out (z) C F 1 − z C F j2πf
where
CI 1 and
G= ⋅ fs = f 3dB = G (35.41)
CF C F R sc 2π
placing the gain of the integrator in the same form as Eqs. (35.4). From Ch. 27 we should
recognize Rsc as a switched capacitor resistor
1 T
R sc = = s (35.42)
CI ⋅ f s CI
noting that in Ch. 27 we used fclk to indicate the frequency of the clock waveform used
with the filter (not fs as we are here). The equivalent block diagrams for first-order filters
using CAIs and using the DAIs (again assuming f << fs) are compared in Fig. 35.21.
V in ( f ) In G Out
V out ( f )
s
Integrator
(a) CI
G= ⋅ fs
CF
CI −1
V in ( f ) In ⋅ z
Out
V out ( f )
C F 1 − z −1
Integrator
(b)
Example 35.6
Sketch the implementation of a DAI-based (switched-capacitor), first-order filter
with characteristics like the one in Ex. 35.1. Using SPICE simulate the design.
The schematic of the filter is shown in Fig. 35.22. 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, CI, is then, from Eq. (35.42) and knowing Rsc is
10k, 1 pF. The 3 dB frequency of the filter is, once again,1/2πR sc C F = 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
peak-to-peak). The results are seen in Fig. 35.23.
Chapter 35 Integrator-Based CMOS Filters 413
Indicates plate
10 p closest to the
φ1 φ2 substrate
1p V CM v out
v in
f s = 100 MHz
Figure 35.22 A switched capacitor, first-order filter similar to the one described in Ex. 35.1.
See Fig. 31.80a for additional information concerning the DAI topology used in
this filter.
Let's comment on the exact transfer function of the DAI in Fig. 35.22. It might
be helpful, at this point, to review Fig. 31.80a. We see in Fig. 35.22 that the output
is indeed fed back to the input through a φ1 controlled switch. The result is that the
output, through the feedback loop, sees one clock cycle delay, z−1. The output is
assumed settled on the falling edge of φ2 during each clock cycle. Because of this,
the input, which is settled on the falling edge of φ1, sees only a half clock cycle
delay, z−1/2. This means that the DAI in Fig. 35.22 really has a transfer function of
CI −1
V out (z) = ⋅ z ⋅ [V in (z) ⋅ z 1/2 − V out (z)] (35.43)
C F 1 − z −1
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 f << fs , is a small phase difference. We can delay the input
signal a full clock cycle by adding a φ1 controlled switch on the output of the filter.
However, because this switch may be part of the next filter section, we don't
discuss this option further. T
Filter input
Filter output
Example 35.7
Sketch the switched-capacitor implementation of the discrete-time lowpass (first-
order) filter shown in Fig. 35.24.
In G1 Out
V in ( f ) s V out ( f )
1
V out ( f ) G2
=
G2 V in ( f ) 1 + G 1sG2
The implementation is seen in Fig. 35.25. The coefficients, see question 35.12, are
C I1 C C
G1 = ⋅ f s and G 2 = I2 ⋅ f s ⋅ 1 = I2 (35.44)
CF CF G 1 C I1
The DC gain, as seen in Eq. (35.8) is set by the ratio of C I2 to C I1 (1/G2 ), and the
filter's 3 dB frequency is
G1G2
f 3dB = (35.45)
2π
Note that the DAI used in this filter has a transfer function of
V CM
C I1 V CM v out
v in
C I2
The big benefit of switched-capacitor-based filters is the fact that the filters' poles
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. SNDRs in the 60-70 dB range
have been achieved when the filter is operating in the MHz range.
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. Back in Secs. 34.2.1 and 34.2.2 we derived the requirements placed on
the op-amp's open-loop gain and unity-gain frequency for a given data converter
resolution (which can easily be converted to SNR for the ideal situation using Eq. [31.4]).
An analog antialiasing filter, AAF, (that is, not discrete-time filter) must be used to remove
the potential aliased signal prior to sampling. As with the noise-shaping modulator-based
data converters the fact that the sampling frequency, fs, is much larger than the input
frequencies of interest allows a relaxed AAF design. Indeed, the resistance of the
MOSFET switches used combined with the switched input capacitance (CI ) form a
lowpass filter. This filtering may serve as the switched-capacitor filter's AAF.
An Important Note
If we pause for a moment and think about the filters we have covered in this chapter (and
the data converter topologies in Ch. 34), we come to the realization that all require precise
analog components. High-speed, wide-bandwidth op-amps and/or components with
precise matching or absolute values are needed. We might argue that this would be a
reason to focus our discussion on digital filtering (filters using only multipliers, delays, and
adders) instead of filters using analog components. However, digital filters can't alone
filter an analog waveform without first running the signal through an ADC. Further,
traditional digital filters that use general-purpose multipliers at reasonable speeds can be
very large (take up a significant layout area). So large in fact as to not be practical in a
general purpose filtering application. Special-purpose chips have been fabricated
specifically for digital filtering (called digital signal processors, DSPs).
At this point we should remember that the reason we spent a whole chapter, Ch.
32, on noise-shaping topologies (see Sec. 31.3.2) was that they reduce the requirements
placed on the analog components in the circuit. Isn't it logical then to attempt to combine
noise-shaping with purely digital filtering for the design of future analog interfaces? The
answer is obviously, "yes"; however, as mentioned above, we have some caveats. While
the resulting interface will place lower demands on the precision of the analog circuitry,
we'll need to (1) develop digital filters that don't rely on complex multipliers. The
multiplications we do use should be simple, perhaps requiring an additional adder, or
trivial (shift) multiplication. Also we'll need to use the digital filters to not only filter the
input signal but to (2) filter out the modulation noise present in the output of the NS
modulator. We used a single-bit quantizer in the majority of our examples in Ch. 32. For a
general analog interface (where we are using the term "analog interface" to indicate both
analog-to-digital conversion and filtering), with up to 60 dB DR a better choice is to use a
416 CMOS Mixed-Signal Circuit Design
multibit quantizer. We say "better" to indicate that we can achieve a higher DR at a lower
oversampling ratio when using a multibit modulator. We'll briefly cover this type of
interface at the end of this chapter. For now let's show how the filter of Fig. 35.21b can be
implemented as a purely digital filter. This will also allow us to derive the exact transfer
function of the filter of Fig. 35.22 when the input frequency gets large (where the DAI
doesn't behave like a CAI, as indicated by Eq. [35.40]).
Exact Frequency Response of a First-Order Discrete-Time Digital (or Ideal SC) Filter
Figure 35.26 shows the digital only equivalent of Figs. 35.21a and 35.21b. We have
replaced the ratio of capacitors, C I /C F , with the variable A in the figure. To determine the
transfer function we can write
z −1
1 − z −1 Integrator
V in ( f ) V out ( f )
A z −1
In Out
Figure 35.27 shows the z-plane plot and magnitude response for this first-order
filter. Note the similarity to Fig. 31.62. Remembering all digital filters have a periodic
frequency response (a period of fs or one complete revolution around the unit circle), we
can compare this filter's response to the filters in the previous examples. To do so let's
assume f s = 100 MHz and write, from Eq. (35.41),
H(z) = A
z − (1 − A) z-plane
H( f )
1
A
A
2−A
f s /2 fs 3f s /2 f
Figure 35.27 The z-plane representation and magnitude response of a first-order digital filter.
Figure 35.28 shows the responses of the first-order discrete-time filter (or the SC filter
using an ideal, that is, no dominate pole, op-amp). The maximum attenuation of the filter
occurs at fs/2 or 50 MHz here and is, from Fig. 35.27 with A = 0.1, 0.0526 or − 25.6 dB.
f s /2 fs 3f s /2 Continuous-time f s /2 fs 3f s /2
filter response
(c) dB vs. log of frequency
fs
Discrete-
Continuous-time time filter
filter response response
f s /2 3f s /2
1 GHz
100 MHz
Figure 35.28 The magnitude response of the discrete-time first-order filter
of Fig. 35.26 with an A of 0.1.
418 CMOS Mixed-Signal Circuit Design
or
V out ( f ) 1 1 + 1/Gs 3
= ⋅ (35.52)
V in ( f ) G 2 1 + G 1sG 2
In G1 Out
V in ( f ) s V out ( f )
sG 3
G2
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
shifting), and highpass filters (keeping in mind that the highpass filter will ultimately
change into a bandpass filter response because of the op-amp's or transconductor's high
frequency rolloff). The location of the filter's pole is given by
G1G2
f 3dB, pole = (35.53)
2π
while the filter's zero is located at
A DC = 1 (35.55)
G2
Chapter 35 Integrator-Based CMOS Filters 419
Active-RC Implementation
The active-RC implementation of the bilinear transfer function is seen in Fig. 35.30. 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
the summation is implemented by changing the input/output voltages to currents. The
currents are summed at the inputs of the op-amp (which remain, ideally, at the
common-mode voltage, VCM). This is important to note in both the active-RC and
switched-capacitor implementations.
We won't discuss the implementation of the MOSFET-C-based bilinear transfer
function. It should be obvious that replacing the resistors in Fig. 35.30 with MOSFETs or
linearized MOSFETs (see Figs. 35.12-35.14) provides a MOSFET-C implementation.
RF
CI CF
RI
V in+
V out−
V in− V out+
RI 1
G1 =
R ICF
R
CI CF G2 = I
RF
RF G3 = R ICI
Transconductor-C Implementation
Figure 35.31 shows the implementation of the bilinear transfer function using
transconductor stages. Again, as with the active-RC filter, signals are summed using
currents. Summing the currents at the output nodes results in
V in+ − V out+ V
g m1 (V in+ − V in− ) − g m2 (V out+ − V out− ) + − out+ = 0 (35.56)
1/s2C 1 1/s2C 2
where we know V out+ = −V out− and V in+ = −V in− . It will be helpful to write
420 CMOS Mixed-Signal Circuit Design
V out+ − V out− g m1 1 + g m1 /C 1 s
=g ⋅ (35.59)
V in+ − V in− m2 1 + s
g /(C +C m2 1 2)
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.
2C 1
G 1 = g m1 /(C 1 + C 2 )
2C 2 g m2
G2 = g
m1
v in+ v out−
C1
g m1 g m2 G3 = g
v in− v out+ m1
2C 2
2C 1
Switched-Capacitor Implementation
The switched-capacitor, SC, implementation of the bilinear filter is seen in Fig. 35.32. This
filter is directly derived from the active-RC filter of Fig. 35.30. From Eq. (35.42) we can
write
RI = 1 and R = 1 (35.60)
F
C I1 ⋅ f s C I2 ⋅ f s
and so
C I1 C C I3
G1 = ⋅ f s , G 2 = I2 , and G 3 = (35.61)
CF C I1 C I1 ⋅ f s
Note how, in this discrete-time filter, the passband gain is C I3 /C 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 C I1 /C I2 .
Chapter 35 Integrator-Based CMOS Filters 421
φ1 φ2
C I11
G1 = ⋅ fs
CF
C I21
v in+ CF C I21
G2 =
C I11 C I11
C I1 C I1
v out− G3 =
C I11 ⋅ f s
V CM V CM
v out+
C I1
C I11 1 + 1/Gs 3
V out ( f ) 1
v in− CF = ⋅
C I21 V in ( f ) G 2 1 + G 1sG 2
G 3D
G 3D (1 − z −1 ) ≈ ⋅s (35.63)
fs
1 − z −1
In Out In Out
G 3D G3s
Digital implementation
Figure 35.34 shows the digital implementation of the bilinear filter of Fig. 35.29.
It's important, at this point, to see how the continuous-time implementation in Fig. 35.29
is directly implemented in Fig. 35.34. In particular we note that
G 3D
G 1 = G 1D ⋅ f s and G 3 = (35.64)
fs
and so
f
V out ( f ) 1 + 1/Gs 3 1 + j ⋅ f s /(2πG )
= 1 ⋅ = 1 ⋅ 3D
(35.65)
V out ( f ) G 2 1 + G 1G 2 G 2 1 + j ⋅
s f
f s G 1D G 2 /2π
In z −1
V in ( f ) 1 − z −1 Integrator
1 − z −1
Out
G 3D G 1D z −1
V out ( f )
z −1
Differentiator
G2
Let's attempt to simplify this filter. If we look at the transfer function, Eq. (35.65),
we see that the feedback gain, G2, simply scales the amplitude of the transfer function and
can be used to further adjust the pole of the filter. 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 G1D, we can, without loss of functionality, set
G2 to 1. We can then rearrange the summing, delaying, and multiplying blocks, as seen in
Fig. 35.35.
Chapter 35 Integrator-Based CMOS Filters 423
In
Out
G 3D G 1D z −1
z −1
In
Out
G 3D G 1D z −1
z −1
1 − G 1D
In Out
1 + G 3D G 1D z −1
G 3D z −1
1 − G 1D
In 1 + G 3D Out
G 3D
G 1D G 3D z −1
1 − G 1D
z −1
In 1 + G 3D Out
z −1 G 1D G 3D
G 3D
1 − G 1D
z −1
In 1 + G 3D Out
z −1 G 1D G 3D
G 3D
1 − G 1D z −1
G 1D z − G 3D /(1 + G 3D )
z − (1 − G 1D ) (1 + G 3D ) ⋅ z
Using the results of Fig. 35.35 we can write the z-domain respresentation of the
transfer function, Eq. (35.65), as
V out (z) G 1D (1 + G 3D ) z − G 3D /(1 + G 3D )
= ⋅ (35.68)
V in (z) z − (1 − G 1D ) z
Note how, for a lowpass filter with G3D set to 0, this equation reduces to Eq. (35.48) with
A = G1D. Before attempting to simplify the filter implementation seen in Fig. 35.35 further,
let's show that, indeed, Eq. (35.68) is equivalent to Eq. (35.65) when f << fs. It will be
helpful to remember that
2
z ≈ 1 + s and z −1 ≈ 1 − s if f << f s or z ≈ 1 + s ≈ 1 s ≈ 1−1 if s 2 ≈ 0 (35.69)
fs fs f s 1 − fs z fs
where s = jω = j2πf . Rewriting Eq. (35.68) in the frequency-domain gives
V out ( f ) G 1D (1 + G 3D )
= ⋅ 1 − 1 − s G 3D /(1 + G 3D ) (35.70)
V in ( f ) 1 + fss − 1 + G 1D fs
f
1 + f s /Gs 1 + j ⋅ f s /(2πG
3D )
= 3D
= (35.71)
1 + G s ⋅f s 1+j⋅ G
f
1D ⋅f s /2π
1D
Example 35.8
Sketch the digital filter equivalent of the following RC circuit. Assume the digital
filter is clocked at 100 MHz.
10k
10p
V out ( f ) 1 1 + jω ⋅ 100 ns
v in v out = ⋅
V in ( f ) 2 1 + jω ⋅ 50 ns
10k
In 1 + G 3D Out
z −1 G 3D
G 1D G 3D
1 − G 1D
z −1
In
z −1 1.1 Out
0.8
z −1
H(z) = 1.1 ⋅ z −
10/11 1.1
= ⋅ z −1 ⋅ (1 − z −1 ⋅ 10/11)
z(z − 0.8) 1 − 0.8z −1
1 + G 3D Out
G 1D G 3D
G 3D V out (z)
In
z −1
V in (z)
1 − G 1D
We can, again, derive the transfer function for the first-order bilinear digital filter.
This time, however, let's use the variables in Fig. 35.39. Again, assuming f << f s , and
using Eq. (35.69) results in
1+ s
f s 1+ B1
B
B0 + B1
H( f ) = ⋅ 0
(35.74)
1 − A 1 1 + f (1−A
s
)
s 1
where
f s (1 − A 1 )
f 3dB, pole = (35.75)
2π
and
fs
1+ 1
B
f 3dB, zero = (35.76)
2π B0
and the gain at DC is
B 0 + B1
A DC = (35.77)
1 − A1
Chapter 35 Integrator-Based CMOS Filters 427
Example 35.9
Using the canonic form of the first-order digital filter, repeat Ex. 35.8.
Comparing Eq. (35.74) with the transfer function in Fig. 35.36, we can write
and thus B 1 = −0.9 . The filter's sketch is seen in Fig. 35.40. We will discuss how to
implement the multipliers in the final section of the chapter.
Let's do a quick check to see if the filter functions as desired at DC. If we
apply 0.1 to the input of the filter then, according to the transfer function in Fig.
35.36, the output of the filter should be 0.05 or one-half the input. Because the
input to the filter is a DC signal, both sides of the delay will have the same value.
According to Fig. 31.62, this value will be 0.5 (the output of the weighted
integrator is 1/[1 − 0.8] times the input signal, here 0.1, at DC). The output will
then be 0.5 − 0.45 or 0.05 (as we would expect). T
Out
In
z −1 0.9
0.8
Figure 35.40 Canonic form of the first-order digital filter in Ex. 35.8.
Example 35.10
Sketch the digital filter implementation of the lowpass filter in Ex. 35.1 that has a
DC gain of one and a 3 dB frequency of 1.59 MHz. Assume the filter is clocked at
100 MHz.
The filter's continuous-time frequency response is given by
H( f ) = 1
f
1 + j ⋅ 1.59 MHz
428 CMOS Mixed-Signal Circuit Design
In Out
z −1 0.1
0.9
Before leaving this section, let's show in Fig. 35.42 the general form of an nth-order
canonic digital filter (where n indicates the number of poles in the transfer function). The
z-domain transfer function of the filter is given by
m m
Σ
i=0
B i z −i Σ
i=0
B i z n−i
H(z) = n = n (35.78)
1 − Σ A i z −i z n − Σ A i z n−i
i=1 i=1
If we want to write the frequency domain transfer function, we write, again assuming
f << f s and using Eq. (35.69),
n−1 i n−1 n−i
Σ B i 1 − j f Σ B i 1 + j f
2πf 2πf
s s
H( f ) ≈ i=0
n i
≈ i=0
n n n−i
(35.79)
1 − Σ A i 1 − j f s 1 + j 2πf − A i 1 + j 2πf
Σ
2πf
i=1
fs
i=1
fs
While we can design higher order digital filters using the topology of Fig. 35.42, we will
restrict our analysis in the remainder of the book to first- and second-order filters where
hand calculations are relatively easy to do. We can increase the attenuation of a filter using
several of these sections, either cascading the stages in series or taking the outputs of
several first- or second-order sections and adding them together.
Chapter 35 Integrator-Based CMOS Filters 429
B0
B1
Out
Bm
In
z −1 z −1 z −1 z −1
A1
A2
A n−1
An
Number of poles n ≥ number of zeroes.
2
πf 0 1 2πf 0 2
p1, p2 = ± − 4(2πf 0 ) (35.81)
Q 2 Q
or
2
πf 0
p1, p2 = ± j ⋅ 2πf 0 1 − 1 (35.82)
Q 2Q
V out1
In G1 G4 V out ( f )
V in ( f ) s s
sG 3 G2 sG 6
G5
Toward the goal of implementing a biquad, consider the block diagram in Fig.
35.43. This block diagram is essentially the cascade of two first-order blocks (as seen in
Fig. 35.29) except that instead of feeding the output of the second stage back to its input,
we feed it back to the input of the first stage. We can determine the transfer function of
this filter by writing
G
V out1 = (V in + sG 3 V in − G 5 Vout − G 2 V out1 ) ⋅ s1 (35.83)
or
s + G1G2 (1 + sG 3 )G 1
V out1
G G
s = V in s − V out 1s 5 (35.84)
R F1 C F1
R F2
C I1
C I2
C F2
V in+ R I1 V out1− R I2
V out+
V in− V out−
R I1 V out1+
R I2
C I2 C F2
C I1
R F2
R F1 C F1
1 R
G1 = G 2 = I1 G 3 = R I1 C I1 G 4 = 1 R
G 5 = I1 G 6 = R I2 C I2
R I1 C F1 R F1 R I2 C F2 R F2
C I1 C I2 C I1 C I2 a0 = 1
a2 = a1 = + R I1 C F1 R I2 C F2
C F1 C F2 R I2 C F1 C F2 R I1 C F1 C F2
ω 0 2πf 0 1 C I2
= = + f0 = 1 ⋅ 1
Q Q R F1 C F1 C F1 R F2 C F2 2π C F1 R I2 C F2 R F2
Figure 35.45 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
2
DC gain of the biquad filter can be set to a 0 /(2πf 0 ) . Note that if the pole quality factor,
Q, is greater than 1/ 2 the response will show peaking. Setting Q to 0.707 results in the
Butterworth or maximally flat response.
L 1
f0 = 1
v in (t) R v out (t) V out
= LC
V in s 2 + s R + 1 2π LC
L LC
C Q= 1 L
R C
Q
V out ( f ) max =
V in ( f ) 1 + 4Q1 2
Im, jω
0 dB f (Hz)
ω0 zeroes at infinity
- 40 dB/decade
Re, σ
f0
ω0
s plane
2Q
s = σ + jω
f max = f 0 ⋅ 1 − 1 2
2Q
Example 35.11
Design an LRC circuit with a Q of 0.707 and a cutoff frequency ( f0) of 1.59 MHz.
From Fig. 35.45 we have two equations we need to solve
3 dB down at f 0
- 40 dB/decade
Figure 35.46 Second-order magnitude response for the circuit described in Ex. 35.11.
Q=5
Q=1
Q = 0.2
Q = 0.5 Q = 0.707
Figure 35.47 The effect of Q on the frequency response of a second-order lowpass filter.
Example 35.12
Simulate the design of an active-RC filter that has frequency characteristics similar
to Fig. 35.46.
Using the basic topology of Fig. 35.44, we see that for a lowpass filter,
C I1 = C I2 = 0 and therefore G3 = G6 = 0. Further
f 0 = 1.59 MHz = 1 1
2π C F1 R I2 C F2 R F2
which we shall use to set R I2 = R F2 = 10k and C F1 = C F2 = 10 pF . The Q of the
filter is given by
Q = 1 = 2πf 0 ⋅ R F1 C F1 → R F1 = 7.07 kΩ
2
2
Knowing a2 = a1 = 0, the gain at DC is a 0 /(2πf 0 ) or R F2 /R I1 (1/G5), which is 1
here. The simulation results are shown in Fig. 35.48. T
434 CMOS Mixed-Signal Circuit Design
magnitude phase
Figure 35.48 Magnitude and phase responses for the active-RC filter of Ex. 35.12.
Notice that at DC, when used in the lowpass configuration, the outputs of the first
integrator in Fig. 35.44, Vout1+ and Vout1− , must be equal. If not, the difference is integrated
by the second section. As the frequency increases, so does the difference in these voltage
levels.
Figure 35.49 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
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).
1
V out s RC 1
v in (t) R v out (t) = f0 =
V in s 2 + s 1 + 1 2π LC
RC LC
L C
V out ( f ) Q=R C
L
V in ( f ) f 0 /Q
0 dB f (Hz)
Im, jω - 3 dB
Example 35.13
Repeat Ex. 35.11 for the bandpass LRC circuit.
Again, we can set C = 100 pF and L = 100 µF. Solving for Q using the equation in
Fig. 35.49 results in
100p
Q = R C = 0.707 = R → R = 707
L 100µ
The simulation results are seen in Fig. 35.50. T
Example 35.14
Repeat Ex. 35.13 if the Q is increased to 20.
Figure 35.51 shows the simulation results. To attain a Q of 20, we use a resistor of
20k with the inductor and capacitor values remaining unchanged. T
Example 35.15
Use an active-RC filter to implement a filter with the response shown in Fig. 35.51.
Let's begin by writing the filter's transfer function
V out a1s a1s
= =
V in 2πf 0
s + Q s + (2πf 0 )
2 2 10×10 6
s + 20 s + (10 × 10 6 )
2 2
Looking at this equation, Eq. (35.80), and Fig. 35.44 we see that a2 = a0 = 0 and
so CI2 = 0 and RI1 = ∞ . Further then
C I1
a1 = f0 = 1 1 = 1.59 MHz
R I2 C F1 C F2 2π C F1 R I2 C F2 R F2
2πf 0 2π ⋅ 1.59 MHz 1
= =
Q 20 R F1 C F1
The passband gain (the maximum gain) occurs at f0 and is calculated by replacing s
in the transfer function above with j2πf 0 . It is given by
a1 Q
A passband =
2πf 0
If, again, we set CF1 = CF2 = 10 pF and RI2 = RF2 = 10k, we get an f0 of 1.59 MHz.
Further then, with a Q of 20, we can set RF1 to 200k. Finally, setting the passband
gain to unity results in
2πf 0 C I1
a1 = = → C I1 = 0.5 pF
Q R I2 C F1 C F2
While these values do result in a biquad with the shape seen in Fig. 35.51, the
values are not practical. Redoing the calculations while trying to minimize the
component spread gives another possible solution: RI2 = 100k, CF1 = 20p, RF1 =
100k, CF2 = 5p, CI1 = 5p, and RF2 = 1k. The simulation results are seen in Fig.
35.52. T
magnitude phase
Figure 35.52 Outputs of the biquad of Ex. 35.15 using active-RC elements.
Chapter 35 Integrator-Based CMOS Filters 437
Switched-Capacitor Implementation
The switched-capacitor implementation of the biquad circuit is shown in Fig. 35.53. Note
how this circuit is a simple translation of the active-RC circuit of Fig. 35.44. Again, if the
filter designed using this section has a lowpass or bandpass response, it can be simplified.
For example, from Figs. 35.45 and 35.49 (the implementation of lowpass and bandpass
filters), we see that a2 is zero. This indicates that G6 can be set to zero (removing CI2 in
Figs. 35.44 or 35.53). The resulting second-order filter response can be written as
V out a1s + a0 sG 1 G 3 G 4 + G 1 G 4
= = (35.94)
V in 2πf 0 s + (2πf ) s + sG 1 G 2 + G 1 G 4 G 5
2
s2 + Q 0
φ1 φ2 φ1 φ2
C I21 C I22
v in+ C F2
C I11
C I2 C I12
C F1
C I1 v out1− v out+
V CM V CM
C I1 v out1+ v out−
C I11 C F1
C I2 C I12
C F2
v in−
C I21 C I22
High Q
We have a major concern alluded to in Ex. 35.15 when using either of the topologies of
Fig. 35.44 or 35.53 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
magnitude different (100k and 1k). This large difference can be traced to, again assuming
G6 = 0,
2πf 0 G4G5 C F1
= G 1 G 2 and 2πf 0 = G 1 G 4 G 5 or Q = = ⋅ R F1 (35.95)
Q G1 G2 R I2 C F2 R F2
This equation shows that RF1 has the largest direct dependence on Q. Using a large value
of RF1 results in a smaller feedback signal (a smaller amount of current is fed back to the
input of the first op-amp). In other words G2 in Fig. 35.43 is small.
In order to minimize the amount of signal, Vout1, 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. 35.43, biquad block diagram shown in Fig. 35.54. All we
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
(35.88) can be rewritten, assuming G6 is zero (a bandpass or lowpass response), as
V out s(G 1 G 3 G 4 ) + G 1 G 4
= (35.96)
V in s 2 + sG 1 (G 2 − G 2Q ) + G 1 G 4 G 5
or, equating the coefficient of s in the denominator of this equation with the coefficient of
s in the denominator of Eq. (35.80), results in
2πf 0
= G 1 (G 2 − G 2Q ) (35.97)
Q
The implementation of the "high-Q" biquad is seen in Fig. 35.55 (with G6 included). The
additional gain from the figure is
R I1 Q−1
G 2Q = Q
= G2 ⋅ (35.98)
R F1 Q−1 Q
V out1
In G1 G4 V out ( f )
V in ( f ) s s
sG 3 G2 sG 6
G 2Q
G5
R F1 C F1
R F2
C I1 Q
R F1 Q−1 = R F1Q
C I2
C F2
V in+ R I1 V out1− R I2
V out+
V in− V out−
R I1 V out1+
R I2
C F2
Q
C I2
R F1 Q−1 = R F1Q
C I1
R F2
R F1 C F1
R I1 Q−1
G 2Q = Q
= G2 ⋅
R F1 Q−1 Q
Figure 35.55 Implementation of the "high-Q" active-RC biquadratic transfer function filter.
The bold lines indicate the added components.
Example 35.16
Repeat Ex. 35.15 using the high-Q circuit of Fig. 35.55.
The passband gain is one so we know that
a1
2πf 0 G G
a1 = = G 1 G 3 G 4 = G 1 (G 2 − G 2Q ) = 1 2
Q Q
or 2πf 0 = 10 × 10 6 = G 1 G 2 = 1/R F1 C F1 . We also know that
C I1 1 1
G1G3G4 = = G 1 (G 2 − G 2Q ) = ⋅
C F1 R I2 C F2 C F1 R F1 Q
and
2πf 0 = 10 × 10 6 = 1
C F1 R I2 C F2 R F2
440 CMOS Mixed-Signal Circuit Design
In an attempt to minimize component spread let's set RF1 to 5k, RI2 to 20k, CI1 to
4p, CF2 to 20p, CF1 = 20p, RF2 = 1.25k, and finally RF1Q = 5.25k (roughly). The
simulation results and schematic are shown in Fig. 35.56. T
5k
20p
1.25k
5.25k
20p
4p
V in+ 20k
V out+
V in− V out−
4p 20k
20p
5.25k
1.25k
20p
5k
magnitude phase
Example 35.17
Repeat Ex. 35.16 using a switched-capacitor implementation. Assume that the
filter is clocked at 100 MHz.
To implement the filter, we need to replace the resistors in Fig. 35.56 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. 35.56. To do this let's divide all capacitors by 10 and
multiply all resistors by 10. Therefore, we can write CI1 = 0.4p, CF1 = 2p, and CF2 =
2p. The resistors can be calculated using
Chapter 35 Integrator-Based CMOS Filters 441
φ1 φ2 φ1 φ2
(0.16p)
0.2p 0.19p
0.8p
(0.8p) 2p 5p
0.4p
V in+ 0.125p
V out+
V CM
V CM
V out−
V in− 0.125p
0.4p
(0.8p) 2p 5p
0.8p
0.2p 0.19p
(0.16p)
the correct output. Looking at Fig. 35.56, we see that if we apply a 1 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. 35.58, the filter is unstable and oscillates. Indeed, using simulations it's
easy to show that even if we ground the inputs of the filter, the outputs will
oscillate at f0 (1.59 MHz). To understand why, remember in Fig. 35.49 that as the
Q of the bandpass filter is increased, the poles move closer to the right-hand plane.
If, for some reason, the poles move into the right-hand plane, the filter will become
an oscillator (unstable). It's important to remember that when we designed the
2πf
filter we approximated our discrete-time variable z as 1 + fss = 1 + j f s (when f << fs
2πf 2πf 2πf
cos fs
≈ 1 and sin fs
≈ fs
). We could be more exact and write
2πf
jf 2πf 2πf
z=e s = cos + j sin (35.100)
fs fs
2πf
which clearly will not follow 1 + f s for frequencies f approaching the sampling
frequency fs. As we discussed in Ch. 30, 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. In
Eqs. (35.68) - (35.71) we wrote, for a pole,
2πf f
z − (1 − G 1D ) ≈ 1 + j ⋅ − 1 + G 1D = G 1D + j ⋅ (35.101)
fs f s /2π
The pole is ideally located at (G 1D f s )/2π (the frequency where the imaginary part
is equal to the real part). Using Eq. 35.100, we can write, more exactly,
Real Imaginary
2πf 2πf
z − (1 − G 1D ) = cos − 1 + G 1D + j sin (35.102)
fs fs
It should be clear from this equation that 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 Fig. 35.49, causes the pole to move
closer to the right-hand plane (causing the Q to increase).
Output
Input
Input
Output
Figure 35.59 Output of the filter in Fig. 35.57 after lowering the Q to maintain stability.
Using this result, we can rewrite the pole locations of Eq. (35.80) (see Eqs. [35.81] and
[35.82]) as
1 ⋅ 1 (35.104)
s 1 + s + p s 1 + s + p
2πf u 1 2πf u 2
This subtraction results in a shift in the pole toward the right-hand plane, increasing the Q
of the circuit; Fig. 35.60. Reviewing Eq. (35.82), we can subtract the unwanted term in
Eq. (35.105) to estimate the shift in the Q or
2
πf 0 (2πf ) 2f 0
− or at f = f0 we can write πf 0 1 − (35.106)
Q 2πf u Q fu
The shifted Q is then
1 2f 0 Q
= 1− → Q shift = Q2f
(35.107)
Q shift Q fu 1 − fu 0
Im,
Poles move towards right-hand plane
due to finite op-amp gain-bandwidth.
The result is an increase in the filter's Q.
Re
s plane
ω0
Ideal distance is
2Q
Figure 35.60 Showing Q peaking resulting from the op-amp finite gain bandwidth product.
Example 35.18
Resimulate the filter in Ex. 35.16 using op-amps that have an fu of 100 MHz.
The center frequency, f0, of this filter is 1.59 MHz and the Q is 20. Using Eqs.
(35.107) and (35.108), we can estimate the increase in Q due to op-amp finite
gain-bandwidth product as
Q2f 0 20 ⋅ 2 ⋅ 1.59
= = 0.636 → Q shift = 55
fu 100
Practically, this is too high of a Q (the poles are too close to the right-hand plane),
and the filter will be unstable (noise in the circuit, or simulation noise in the
simulation, will push the poles into the right-hand plane). Figure 35.61 shows the
simulation results (see also Ex. 35.4). The inputs to the filter are grounded. The
unstable oscillation frequency is close to the ideal, f0, but is shifted by a small
amount. T
Chapter 35 Integrator-Based CMOS Filters 445
Figure 35.61 Showing how the filter of Ex. 35.16 becomes unstable due to
finite op-amp bandwidth.
Transconductor-C Implementation
Let's redraw the bilinear filter in Fig. 35.31 as seen in Fig. 35.62. We redraw it like this to
show how the feedback gain, G2, is implemented. In the block diagram of the biquad filter
shown in Fig. 35.43, we used a similar scheme to implement the feedback gain, G5. Figure
35.63 shows the implementation of a biquad filter using transconductors where we have
drawn it so that the transconductors appear to be connected in series. This topology can
be redrawn so that it looks similar to Fig. 35.62 (showing a direct correspondence
between it and Fig. 35.43). Note how 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 output on the top of the output instead of the plus output).
2C 1
2C 2
v in+ v out+
g m1
v in− v out−
2C 2
2C 1
G 1 = g m1 /(C 1 + C 2 ) g m2
g m2
G2 = g
m1
C1
G3 = g
m1
2C 1 2C 3 2C 4
v out−
2C 2
v in+
v out1− g m4
g m1 g m2 g m3
v in− v out1+
2C 2
v out+
2C 1 2C 3 2C 4
g m2 C g C
G 1 = g m1 /(C 1 + C 2 ) G 2 = g G 3 = g 1 G 4 = g m3 /(C 3 + C 4 ) G 5 = g m4 G6 = g 3
m1 m1 m1 m3
Out
B0
B1 B2
In
z −1 z −1
A1
A2
To translate this transfer function into the frequency domain, we use Eq. (35.69) and
assume our frequencies of interest are much less than the sampling frequency
2
B 0 1 + fss + B 1 1 + fss + B 2
H( f ) = 2
(35.110)
1 + s − A 1 + s − A
fs 1 fs 2
After some algebraic manipulation we can put this equation in the form seen in Eq. (35.80)
B 0 ⋅ s 2 + f s (2B 0 + B 1 ) ⋅ s + f s2 (B 0 + B 1 + B 2 )
H( f ) = (35.111)
s 2 + f s (2 − A 1 ) ⋅ s + f s2 (1 − A 1 − A 2 )
where
a2 = B0 (35.112)
a 1 = f s (2B 0 + B 1 ) (35.113)
a 0 = f s2 (B 0 + B 1 + B 2 ) (35.114)
2πf 0
= f s (2 − A 1 ) (35.115)
Q
fs
f0 = ⋅ (1 − A 1 − A 2 ) (35.116)
2π
Example 35.19
Repeat Ex. 35.11 using the digital biquad clocked at 100 MHz.
In this example a lowpass filter is designed with f0 = 1.59 MHz and Q = 0.707.
Reviewing Fig. 35.45, we see that for a lowpass filter a2 and a1 are zero. This
means, in Eq. (35.111), B0 and B1 are zero. Further,
2π ⋅ 1.59 × 10 6
Q= = 0.707 → A 1 = 1.859
100 × 10 6 (2 − A 1 )
and
2
2πf 0
1 − A 1 − A2 = → A 2 = −0.869
fs
and finally, because the gain at DC is 1,
B 2 = 1 − A 1 − A 2 = 0.01
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 shown in Fig. 35.65. To implement this simulation in
WinSPICE, we used transmission lines for the delay elements and voltage-
controlled voltage sources for both the multiplications and the adders. Note how
the frequency response is periodic with the filter's clocking frequency. T
448 CMOS Mixed-Signal Circuit Design
magnitude
Analog Noise-shaping
modulator Digital filter Digital output
input
Assuming we don't design the filter to decimate (reduce the digital word output rate) the
modulator's output, this filter will require more than 3K registers. As K gets large, this can
result in a large layout area. To reduce this layout area, as discussed in Sec. 31.2.2, we can
decimate the modulator's output. However, the big concern when using decimation to
reduce the filter's size and complexity is aliasing. It would be nice to have a small-area
filter to remove modulation noise.
Chapter 35 Integrator-Based CMOS Filters 449
Consider the frequency response of the Sinc decimating filter shown in Fig. 35.67.
This filter uses a clocking frequency of 100 MHz and assumes K is 16. Note how, at the
bandwidth of the desired signal, B, there is significant droop in the filter's response. It will
be highly desirable to design a filter that doesn't have this droop or, even more desirable,
contains a small amount of peaking at B to compensate for the eventual needed decimation
filter response (a Sinc shape with 3.9 dB droop at B). Using Eqs. (35.114) - (35.116) (and
some simulations), we can set, for a digital biquad, B2 = 0.03125, A1 = 1.75, A2 =
−0.78125 (there are several other solutions, depending on the desired complexity or
performance of the filter). Knowing that the sinc filter was almost ideal for removing
modulation noise and that a second-order (biquad) filter rolls-off at −40 dB/decade, we
can estimate, with the help of Fig. 35.67, the number of biquads required to remove the
modulation noise. One decade above 3.125 MHz (i.e., at 31.25 MHz) the attenuation of
the sinc filter is approximately −70 dB. This might make us think that two biquads in
cascade would be enough to remove the modulation noise since one decade above f0 we
would have −80 dB attenuation (and two may suffice in many applications). However, as
seen in Fig. 35.67, the sinc filter response is not monotonic. We will use three biquads to
remove the modulation noise (or again the number of filters cascaded is one more than the
order of the modulator). The simulation results comparing the Sinc and biquad filters are
seen in Fig. 35.68. The transfer function of the biquad filter is
−16 3
H(z) = 1 ⋅ 1 − z −1
f s /2 = 50 MHz 16 1 − z
In z −1 z −1 0.03125 Out
Example 35.20
Redesign the filter in Fig. 35.69 so that the response has a small amount of peaking
at 3.125 MHz. Compare, with simulations, the new response to the sinc filter seen
in Fig. 35.67.
Reviewing Eq. (35.115), we see that to increase the Q we need to increase A1.
Keeping in mind that we want to have simple multiplications relying heavily on
shifts, let's try increasing A1 to 1.78125 and A2 to 0.8125. The simulation results
are seen in Fig. 35.70. In this figure we compare the modified filter response to the
response of the Sinc filter. Note how we have a couple of dB peaking in the
cascaded biquad filter's output. T
Chapter 35 Integrator-Based CMOS Filters 451
Example 35.21
Suppose a noise-shaping modulator uses a 4-bit quantizer where 1111 is the most
positive output code and 0000 is the most negative output code. How would these
codes be converted into two's complement format?
Inherent in the binary offset format shown back in Fig. 31.36 is an LSB offset. This
resulted in our LSB being defined by, see Eq. (30.23),
V REF+ − V REF−
1 LSB= (35.127)
2N
In a noise-shaping modulator (see question 30.14) we may use a quantizer that
doesn't have this offset. When using a 1-bit quantizer, an output of 1 corresponds
to VREF+ and an output of 0 corresponds to VREF−. We can write 1 LSB for this
situation as
Chapter 35 Integrator-Based CMOS Filters 453
V REF+ − V REF−
1 LSB = (35.128)
2 N−1
To convert the 4-bit modulator outputs ranging from 0 to 15 into two's
complement, we simply complement the MSB. This would mean the most positive
output in two's complement is 0111, while the most negative output is 1000. In
order to eliminate the offset (asymmetry in the digital output code) we may use an
additional output bit in our quantizer so that the maximum modulator output is 1
0000 corresponding to VREF+. If this is the case, then the modulator outputs can be
converted to two's complement as seen below.
Maximum modulator output 1 0000 → 0 1000 (+8 in two's complement)
Middle modulator output 0 1000 → 0 0000 (0 in two's complement)
Minimum modulator output 0 0000 → 1 1000 (−8 in two's complement)
T
We can extend our discussion of multiplying by 0.5 to multiplying by 0.25, 0.125,
0.0625, 0.03125, etc., by simply extending the sign-bit of the input word. To multiply a
word by 0.03125, we simply shift the word to the right five times. For example,
multiplying a two's complement code of 01 by 0.03125 would give
+1 or +0.5 +1 or 0.015625
B= Multiply by
1 0
In Out 0.5 0.5
0.25 0.75
0.125 0.875
B 0.0625 0.9375
0.03125 0.96875
REFERENCES
[1] R. E. Bogner and A. G. Constantinides (eds.), Introduction to Digital Filtering,
John Wiley and Sons, 1975. ISBN 0-471-08590-1
[2] P. R. Gray, D. A. Hodges, and R. W. Brodersen (eds.), Analog MOS Integrated
Circuits, Wiley-IEEE, 1980. ISBN 0-471-08964-8
[3] P. R. Gray, B. A. Wooley, and R. W. Brodersen (eds.), Analog MOS Integrated
Circuits II, Wiley-IEEE, 1989. ISBN 0-87942-246-7
[4] P. A. Lynn, An Introduction to the Analysis and Processing of Signals,
Hemisphere Publishing Corporation, 1989. ISBN 0-89116-981-4
[5] Y. P. Tsividis and J. O. Voorman (eds.), Integrated Continuous-Time Filters:
Principles, Design, and Applications, Wiley-IEEE, 1993. ISBN 0-7803-0425-X
[6] B. Nauta, Analog CMOS Filters for Very High Frequencies, Kluwer Academic
Publishers, 1993. ISBN 0-7923-9272-8
[7] A. S. Sedra and K. C. Smith, Microelectronic Circuits, Oxford University Press,
1998. ISBN 0-1951-1663-1
[8] D. A. Johns and K. Martin, Analog Integrated Circuit Design, John Wiley and
Sons, 1997. ISBN 0-471-14448-7
[9] E. P. Cunningham, Digital Filtering: An Introduction, John Wiley and Sons, 1995.
ISBN 0-471-12475-3
[10] R. Schaumann and M. E. Van Valkenburg, Design of Analog Filters, Oxford
University Press, 2001. ISBN 0-19-511877-4
QUESTIONS
35.1 Resketch Fig. 35.2 for the following circuit.
L
v in (t) v out (t)
35.2 Show that Eq. (35.6) is still valid if the circuit's inputs and outputs are referenced
to the common-mode voltage, VCM. (The op-amp inputs should also be at VCM.)
Chapter 35 Integrator-Based CMOS Filters 455
35.3 Sketch the implementation of a first-order lowpass filter using a CAI with a 3 dB
frequency of 10 MHz and a DC gain of 6 dB. Simulate your design to verify if it
works as expected.
35.4 Plot, in the complex plane, the ideal pole location and the actual pole locations due
to finite op-amp unity gain frequency for the filter described in Ex. 35.4.
35.5 Regenerate Fig. 34.21 of the last chapter using SPICE and the op-amp model
shown in Fig. 35.8.
35.6 Suppose an antialiasing filter was required for a 12-bit data converter. Further
assume the filter is to be implemented using an active-RC topology. If VDD = 1.5
V, estimate the minimum value of the integration capacitor used, assuming the
filter's noise performance is dominated by thermal noise. Is it wise, for 12-bit
system performance, to design the filter so that its SNR is equal to the SNR of the
data converter?
35.7 Repeat question 35.6 if the op-amp used in the filter has a linear output swing of
80% of the power supply voltage.
35.8 Show, using the topology shown in Fig. 33.22 (and the same SPICE models), how
using the two MOSFETs linearizes the change in resistance with VDS.
35.9 Repeat Ex. 35.4 using a MOSFET-C filter. Use the MOSFET SPICE model given
in Ch. 33. After performing the AC simulations, try a transient simulation with an
input sinusoid at 1 MHz. Show how the output of the filter becomes distorted as
the amplitude of the input signal increases. Determine the filter's SNDR when the
input signal has a frequency of 1 MHz and an amplitude of VDD peak-to-peak.
35.10 Derive the transfer function for the filter shown in Fig. 35.16 if the transconductors
have different gms. Sketch the block diagram, similiar to the one seen in Fig. 35.6,
for the filter.
35.11 Derive the transfer function for the following first-order transconductor filter.
v in1+
v in1−
v out+
C
v in2+ v out−
v in2−
35.12 Show the derivation details that result in Eqs. (35.44) and (35.46).
456 CMOS Mixed-Signal Circuit Design
35.13 From the derivation of Eq. (35.48), what would happen if we repeated Ex. 35.6
with an input frequency of 101.59 MHz?
35.14 Show that if the values of A and B are restricted to 1, 0.5, 0.25, 0.125, etc. that the
circuit of Fig. 35.75 can be used to implement multiplication by coefficents that
aren't directly powers of two. How would a multiply by 0.75 be implemented? a
multiply-by-0.9375? a multiply-by-0.5625?
In A Out
Figure 35.75 A simple multiplier where A and B simply shift the data.
35.15 From the results of the preceding question, sketch the implementation of one
possible design of a multiplier topology that multiplies an input word by 0.8789
and 0.3164.
35.16 Show the details of how the gains, G, are derived in Fig. 35.30.
35.17 In Fig. 35.35 a filter section has a transfer function that can be written
z − A/(1 + A)
H(z) = (1 + A) ⋅ z
For this transfer function generate a z-plane plot and a magnitude plot similar to
what is seen in Fig. 35.27.
35.18 Plot the time-domain output of the filter in Fig. 35.37 when the input is a zero to
one step function.
35.19 Design a first-order canonic digital filter that is clocked at 100 MHz and has a
transfer function, in the frequency domain, given by
H( f ) = 1
f
1 + j 4 MHz
35.20 Redraw Fig. 35.42 using only two input adders.
35.21 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. 35.48.
35.22 What happens to the poles in the biquadratic equation, Eq. (35.80), if the Q is less
than 0.5? (Hint: The filter behaves like the cascade of two first-order filters.) Is the
fmax equation in Fig. 34.45 valid?
Chapter 35 Integrator-Based CMOS Filters 457
35.23 Compare the size of the elements used in Exs. 35.11 and 35.12. Is there a benefit
to using an active element for monlithic implementation?
35.24 Show, using the simulations from Ex. 35.17, that increasing the switch resistance,
and thus the spectral content present in a switched capacitor circuit, can help to
stabilize high-Q switched-capacitor bandpass filters.
35.25 Redesign and simulate the operation of the filter discussed in Ex. 35.17, with a Q
of 5, while trying to minimize the difference between CI1 and CF2. Suggest a
possible modification to the filter topology (similar to how we add G2Q in Fig.
35.54) to reduce this component spread.
35.26 Derive the transfer function of the transconductor-C biquad shown in Fig. 35.63.
Can this filter be orthogonally tuned? If so how?
35.26 Repeat Ex. 35.12 using the transconductor-based biquad.
35.27 How would a "high-Q" biquad be implemented using transconductors? Repeat Ex.
35.15 using the transconductor-based biquad.
35.28 Repeat Ex. 35.13 using a digital filter.
35.29 Repeat Ex. 35.14 using a digital filter.
35.30 Show, using biquad sections, how the following lowpass ladder filter would be
implemented.
v in v out
Figure 35.76 Implementing a ladder filter using biquads, see problem 35.30.
In Out
z −1 A
1−A A<1
36
At the Bench
In this chapter we present some practical prototyping techniques to illustrate a few of the
concepts discussed in this book. The goal of the chapter is to simply provoke thought and
show alternative possibilities (other than hand calculations and simulations) for looking at
the performance of a mixed-signal circuit or system.
6
0.01u 0.01u
13
In Out In Out
8
6
VSS -9V
Figure 36.1 (a) Push-pull amplifier and (b) prototyping the amplifier.
460 CMOS Mixed-Signal Circuit Design
14 2 11
6 13 3 1 10
12
8 5
7 4 9
Figure 36.2 Pin diagram for the 4007. Note how the bodies of all PMOS
devices are tied to pin 14, while the bodies of the NMOS devices
are tied to pin 7. This means that pin 14 must be tied to the highest
potential in the circuit (if the PMOS devices are used), and pin 7
must be tied to the lowest potential.
While the detailed datasheet for the 4007 can be found at the book's website
(http://cmosedu.com), we will comment that the threshold voltage for these transistors is
approximately 2 V and that they can drive around 1 mA or so into a load (this is a serious
limitation and will limit the size of the load we can drive). The large threshold voltage and
limited drive capability are the reasons we used ± 9 V power supplies.
Deadbug Prototyping
Figure 36.3 shows a chip flipped upside down and placed on a copper conductor (a glass
epoxy, FR-4 material coated with copper used in printed circuit board manufacuturing).
The reason this technique is called "deadbug prototyping" should be obvious (unless the
reader is so lucky they've never seen a dead cockroach). We use this approach instead of
the common white protoboards found in most undergraduate electronics laboratories for
prototyping because of the ability to use a good ground plane (the copper conductor). The
big drawback is the need to solder all of the components together. A good (meaning
equipotential) ground plane is essential to any low-noise, wide-dynamic range,
measurement. We'll also use BNC connectors to pipe our signals on-to and off-of the
board to avoid long wires, which tend to pick up coupled noise.
Figure 36.4 shows the prototype of the amplifier in Fig. 36.1b implemented using
the deadbug technique. The black and red wires coming into the circuit provide power and
ground. At the connections of power and ground on the board we add a decoupling
capacitor (a capacitor soldered from the power-supply connection to the ground plane).
This capacitor provides charge for any fast transients that may occur in the circuit. The
capacitor leads can be twisted into small loops and soldered to the ground plane or to a
pin on a chip and used as a contact point for the power supply clips. Before looking at
some measurement results, we need to discuss probe loading and measurement techniques.
Chapter 36 At the Bench 461
Probing
If we're not careful, we can load the circuit we are testing with our measuring system.
Consider the connection of a piece of coaxial cable to the oscilloscope shown in Fig. 36.5.
The scope has an input resistance of 1 MΩ and an input capacitance of 15 pF. When a
coax cable is driven by a large impedance and is terminated with a large impedance, we
think of it as a capacitor. If a 5-foot piece of coax is used to connect the push-pull
amplifier to the oscilloscope, we would get the circuit shown in Fig. 36.6. Clearly, the
measuring system will load the amplifier and keep us from accurately measuring the
response of the circuit.
15 pF
1 MEG
Shield of coax connected to ground
Figure 36.5 The loading when probing with a piece of coaxial cable.
To reduce the loading by the required coaxial interconnect cable (150 pF in Fig.
36.6), a scope probe trades off sensitivity for lighter loading. Figure 36.7 shows a 10:1
compensated scope probe. The term 10:1 represents the attenuation factor from the probe
tip to the input of the scope. One-tenth of the voltage on the tip of the probe actually
makes it to the input of the scope. The term "compensated" indicates that the probe is
designed to compensate for the large loading of the coaxial cable. If the probe is
compensated correctly, the impedance in the probe's tip (9Z) is exactly, independent of
+9V
9V
16.5 pF
5 feet Scope input
9MEG 30 pF/ft 15 pF
Probe tip 1 MEG
Probe cable
16.5 pF
Scope
9MEG 150 pF
15 pF
1 MEG
Probe tip
9Z Z
Approximation to calculate
10 MEG 15 pF effects of loading
Probe tip
frequency, nine times larger than the impedance of the combination of the coaxial cable
and the scope (Z) . Note how the loading of the probe at DC is 10 MEG while at high
frequencies the loading is roughly 15 pF.
Testing the Circuit
To test this circuit, we'll use a vector signal analyzer, VSA, (an instrument similar to a
spectrum analyzer with the capability to perform an inverse Fourier transform for viewing
a signal in the time domain). A test setup is seen in Fig. 36.8. We'll use an input resistance
of 50 Ω to avoid the need for a compensated probe. Because of the limited drive capability
of the amplifier, we'll add a 5k resistor in series with the output. This results in a 100:1
attenuation (− 40 dB) from the amplifier's output to the input of the VSA. The schematic
of the amplifier is seen in Fig. 36.9. Note that we also added a resistor to ground on the
input of the amplifier to avoid a floating node.
Figure 36.10 shows the input signal to the amplifier in the time domain. It is a 100
mV sinewave with a frequency of 100 kHz. The spectrum of this signal is seen in Fig.
36.11. Note the units on the y-axis are dBm or decibels with respect to 1 mW of power.
464 CMOS Mixed-Signal Circuit Design
Figure 36.8 A test setup showing a VSA, spectrum analyzer (not in use), and power
supply.
+ 9 V 0.01u
0.01u 0.01u
5k
In to 50 ohm input VSA
0.01u
9V
Figure 36.9 Final schematic of the push-pull amplifier shown in Fig. 36.3.
Chapter 36 At the Bench 465
Because this is a 50 Ω system, we can verify the power in the input sinewave is, as seen in
Fig. 36.11, − 10 dBm by writing
2
RMS voltage of the input sinewave
(Peak voltage amplitude)/ 2 /50 Ω
dBm = 10 ⋅ log = 10 ⋅ log 0.1 mW = −10 dBm
1 mW 1 mW
(36.1)
−10 dBm
The spectrum of the amplifier's output is seen in Fig. 36.12. Keeping in mind that we have
an attenuation of −40 dB between the amplifier's output and the VSA's input, we can
estimate the amplitude of the output as the − 37 dBm + 40 dB or 3 dBm. The gain is then
13 dB. This can be converted into a voltage amplitude (of the output sinewave) using
V 2outpeak /(2 ⋅ 50)
3 dBm = 10 ⋅ log → V outpeak = 447 mV (36.2)
1 mW
Figure 36.12 Spectrum of the amplifier's output with the input seen in Fig. 36.11.
Figure 36.13 shows the output spectrum if the amplitude of the input sinewave is
increased to 1 V. Note the additional tones at multiples of the input frequency. The ideal
resulting sinewave peak output amplitude is 4.47 V. Clearly this amplitude is well within
the bounds of the power supply voltages. However, knowing the MOSFETs in the 4007
can supply only 1 to 2 mA and that our load is nominally 5k (because of the added
attenuating resistor seen in Fig. 36.9), we may run into some loading problems (resulting
in the output becoming distorted). Further, we might expect some distortion simply
because the amplifier is operating open-loop and, as indicated back in Ch. 22, the large
signal gain varies with the input amplitude. Toward characterizing this distortion, we can
specify the total harmonic distortion (THD), as
a 22 + a 23 + a 24 + ... + a 2n
THD = (36.3)
a 21
where a1 is the amplitude of the fundamental, a2 is the amplitude of the second harmonic
(or the tone at twice the desired frequency), a3 is the amplitude of the third unwanted
tone, etc. The THD is usually specified as a percentage, e.g., 0.01%. We can determine
the amplitude of the tones from the plot, neglecting the division by 1 mW (making the
Chapter 36 At the Bench 467
Figure 36.13 Output spectrum showing distortion when the input amplitude is
increased to 1 V.
actual units dBm) because of the ratio in Eq. (36.3), as −18 dB = 10 ⋅ log a 21 or
a 21 = 0.0159 , assuming the second harmonic's amplitude is −40 dB a 22 = 0.0001 , assuming
the third harmonic's amplitude is −45 dB, then a 23 = 0.0000316 , and finally the fourth
harmonic's amplitude is approximately 2 × 10 −6 . The THD can then be calculated as
E(s)
V in /R in
V in V out
R in C
−V out /R f Rf
The desired signal is lowpass filtered, while the quantization noise is, again, highpass
filtered (resulting in the modulation noise). Again, assuming the comparator gain is
infinite, passing the output of the moduator through a digital filter with a bandwidth less
than 1/2πR f C results in a digital replica of the analog input signal. The practical problems
with this topology are the importance of the comparator's gain and the kickback noise
injected into the input signal when the comparator switches states.
Prototyping the Modulator
Figure 36.16 shows the schematic of the prototype modulator. The D flip-flop was added
to make the LM339 comparator appear as though it were a clocked comparator. Also, the
74HC74 is implemented using CMOS and so its outputs swing all the way down to
ground and up to +5 V. This is important when we use its output as the fed-back signal in
our modulator. The resistors and capacitor on the input of the modulator form a lowpass
filter (as seen in Eq. [36.7]) with a time constant of 100 µs. The 3 dB frequency
associated with this circuit is then 1.59 kHz. Input signal frequencies above this value will
experience an attenuation. Figure 36.17 shows the deadbug prototype of the modulator.
Analog out
5V 100k
In 5V 5V 1,000p
100k 1.2k 1,4 14
5 3
2
1,000p LM339 D Q
4 2 5
12 74HC74
100k 3 clk Q Digital out
5V 1k 1k clk 6
(1 MHz)
7
The input to the modulator used to generate some test results, see Fig. 36.18, is a
4 V peak-to-peak sinewave at a frequency of 500 Hz centered around 2.5 V. The digital
modulator output is shown in this figure as well. Looking at this digital data alone is
somewhat meaningless using the oscilloscope (and so we'll look at the spectrum of this
data). Figure 36.19 shows the spectrum of the digital data. A 3-foot coaxial cable is
connected between the digital output in Fig. 36.16 and the VSA (with a 1 MEG input
resistance so the loading will affect signal frequencies greater than roughly 100 kHz).
Figure 36.20 shows the spectrum of the modulator's output up to 200 kHz. Note how, as
seen back in Fig. 32.15, the modulation noise increases with increasing frequency. The
resolution of the measurement, in Fig. 36.20, is 25 kHz. This causes the desired signal at
500 Hz to appear as though it were occurring at DC. Finally, the bottom trace in Fig.
36.18 shows what happens if we pass the digital data output from the modulator through
an RC lowpass filter with a 3-dB frequency of 1.59 kHz. As expected, the resulting analog
output is a very close replica of the input signal.
470 CMOS Mixed-Signal Circuit Design
RC filtered output
DC 5 kHz 10 kHz
Figure 36.20 The spectrum of the modulator's output data up to 200 kHz.
Resolution bandwidth is 2.5 kHz (and so our input signal is smearing
with DC).
472 CMOS Mixed-Signal Circuit Design
MOSFETs connected), we simply remove C2 and connect the output to the spectrum
analyzer (dynamic signal analyzer). The resulting spectrum is seen in Fig. 36.23. Note that
the output spectral noise density of the LNA is roughly −80 dBV/ Hz (100 µV/ Hz ) at
1 Hz (dividing this by the op-amp's gain of 100 results in 1 µV/ Hz at the MOSFET's
drain). In the following, we ignore the LNA's contribution (to simplify the discussion).
-80 dB
10
MOSFET Noise
Figure 36.24 shows the noise spectrum when we put C2 back in the circuit, bias the
MOSFET at a specific operating point, and connect C1 to ground. The spectral density at
1 kHz is roughly −70 dBV/ Hz or
−70 dBV/ Hz = 316 µV/ Hz at 1 kHz (36.8)
or, calculating the 1/f spectral density (see Ch. 9)
FNN
2
316 µV 100 × 10 −9 V 2 = Flicker noise numerator (FNN) = 100 pV
2
v 21/f,out = =
Hz Hz 1 kHz f
(36.9)
As a quick check, the spectral density of the noise at 10 kHz can be calculated as
100 pV 2 10 nV 2 100 µV
v 21/f,out = = → v 21/f,out = = −80 dBV/ Hz (36.10)
10 kHz Hz Hz
which is what we see at 10 kHz in Fig. 36.24. (To determine the MOSFET's output noise
spectral density alone we divide the spectral density in Fig. 36.24 by the LNA's gain of
100.) To determine the RMS output noise, we can integrate the 1/f noise spectral density
1/2
1/2
fH
v 2on = ∫ v 21/f,out ⋅ df
fH
= FNN ⋅ ln (36.11)
fL fL
-80 dB
1 kHz 10 kHz
Figure 36.24 Measured Flicker noise from the MOSFET/LNA in Fig. 36.22.
Chapter 36 At the Bench 475
Chopper Stabilization
Consider the OTA shown in Fig. 36.25a and the associated noise spectral density shown in
Fig. 36.25b. This circuit is essentially an integrator. In an ideal integrator, connecting the
inputs together and to the common mode voltage would result in the outputs remaining
unchanged. However, in a real integrator, the OTA's offset and the 1/f noise results in the
outputs of the integrator eventually reaching the supply rails. It would be nice if we didn't
have to worry about either the offset or the 1/f noise. What we are going to do in the CHS
scheme is modulate the offset and noise to a place in the frequency spectrum where it
won't interfere with our desired signal.
Output noise, V/ Hz
10 dB/decade
In Out 20 dB/decade
1/f noise
(a) dominates
cout
In
cout
clk
offsets contributed by the OTA will directly affect the input sensitivity. Consider what
would happen if we chopped (or switched back-and-forth) the input/output terminals of
the OTA as seen in Fig. 36.27. When clk is high, the OTA is connected through switches
so that it behaves as seen in Fig. 36.26. The OTA's offset, for example, causes a current to
charge/discharge the capacitors. When clk is low, both the input and output terminals are
switched so that the gain of the amplifier remains the same polarity. The offset now causes
a current to flow in the capacitors in the opposite direction from the flow when clk was
high. This effectively, if the rate at which we switch back-and-forth is fast, results in net
zero current flow into the capacitors. A similar argument can be made for the low-
frequency 1/f noise. The chopping, or switching, reduces both the offset and the Flicker
noise on the output of the integrator (and so the input-referred noise is decreased as well).
Figure 36.27 Switching (chopping) the inputs and outputs of the OTA integrator.
clk clk
Consider the block diagram of the chopper and OTA of Fig. 36.27 shown in Fig.
36.29. Here we set the chopping frequency to one-half of the clock frequency (the clock
used to strobe the comparator in the noise-shaping modulator of Fig. 36.26). We do this
so that no aliasing occurs in our signal of interest from sampling the D signal at fclk (the
OTA noise doesn't fold into the signal of interest after sampling). If the settling time, when
A B C D
OTA
Note integration
is not occurring in
cos 2πf chop cos 2πf chop the traces below
frequency
OTA noise
not to scale
frequency
f chop f clk
Desired signal
Signal at D after demodulation
OTA noise
Figure 36.29 How chopping affects the noise and signal in an OTA.
Chapter 36 At the Bench 479
chopping at fchop, of the amplifier is longer than 2/fchop, then a lower chopping frequency
can be used (ultimately set by the integrator's bandwidth). For example, if we are clocking
the NS modulator of Fig. 36.26 at 100 MHz, then we might chop the OTA's input/output
at a rate of 12.5 MHz (divide the NS modulator's clock by eight using a cascade of three
of the circuits in Fig. 33.46).
In a second-order noise-shaping modulator the input-referred noise is mainly due
to the first integrator as discussed in Ch. 32. The input-referred 1/f noise is passed directly
to the output of the modulator. Modulators that use an autozeroed integrator don't have
this problem because the autozeroing operation removes both the offset at DC and
attenuates the 1/f noise spectral density. Looking at the power spectral density of 1/f noise
on the output of an integrator, when not used in a modulator with feedback, results in a
1/f 3 spectral shape. Averaging this noise results in linear growth with averaging time.
CF
φ1 φ2
CI V CM v out
v in
Ts
φ1
φ2
n−1 n t
n − 1/2
noting the z1/2 term in the numerator is simply a phase shift (a time delay), which will be
neglected as long as our input frequencies, f, are much less than the filter's clocking
frequency, fs. Remembering from Eq. (35.69) that
z ≈ 1 + s when f << fs (36.20)
fs
we can rewrite Eq. (36.19) as
V out (s) 1
= (36.21)
V in (s) 1 + sR sc C F
where
1
R sc = (36.22)
f sCI
The filter's 3-dB frequency is located at
f 3dB = 1 (36.23)
2πR sc C F
Clock Generation
The first thing we need to build is the clock generation circuit. Figure 36.31 shows the
basic schematic of a nonoverlapping clock generator circuit. We use ± 9 V supplies. The
two phases of the clock should transition between these voltages. We, again, use the 4007
Clk
φ1
φ2
CMOS transistors shown in Fig. 36.2 to implement the generator. Further, since this is a
purely digital circuit, we breadboard the design (see Fig. 36.32). Figure 36.33 shows the
outputs of this generator.
φ1 φ2
1000p
5 10
4 3 9 12 2 +9V
8
5 10
LT1365 Output
4 9 12 3 1
Input 4
3
130p
-9V
Using Eq. (36.23), we calculate the filter's 3-dB frequency as 2 kHz when the filter
is clocked at 100 kHz. Figure 36.36 shows the filter's input and output at this frequency. If
Chapter 36 At the Bench 483
one looks closely at the output signal, the discrete nature is obvious (see the steps in the
output waveform shown in the simulation in Fig. 35.23). Figure 36.37 shows how the 130
pF capacitor (the node at the bottom of the schematic) charges to the input signal and then
discharges back to ground (making the parasitic capacitance on this node unimportant).
Input Output
Figure 36.36 First-order filter's input and output at the 3-dB frequency of 2 kHz.
Before leaving this section, let's show example input/output spectrums for this
first-order switched-capacitor filter; Fig. 36.38. The desired signal is at 2 kHz and its peak
amplitude has been decreased to 500 mV (to avoid overloading the VSA). The input
signal amplitude is then 10 ⋅ log([(0.5/ 2 ) 2 /1 MΩ]/1 mW) or −39 dBm. Because we are
applying the 3 dB frequency, we expect our output amplitude to be −42 dBm (and it is).
Finally, to show that the filter is indeed a sampled circuit, we increase the input frequency
to 10 kHz and show the output images around the 100 kHz sampling frequency, Fig.
36.39.
Filter output
spectrum
Input spectrum
Figure 36.38 Input and output spectrums for the filter of Fig. 36.34.
Filter output
spectrum
Input spectrum
V Qe,RMS = ∫ V out ( f ) ⋅ df
2
(36.24)
0 (DC)
The signal V out ( f ) represents the data converter's output spectrum (after removing the
desired signal and any distortion spikes) and has units of V/ Hz . The maximum
frequency we integrate to, fmax, is generally the Nyquist frequency, f s /2 . We assume that,
when actually using the data converter, a reconstruction filter removes spectral content in
the output signal above the Nyquist frequency. In a noise-shaping modulator, a digital
filter sets fmax. Note that we can't accurately calculate the quantization noise added to an
input signal unless the input to the data converter is busy. A sinewave of sufficiently large
amplitude can be used to exercise the data converter and "whiten" the quantization noise.
Input signal
Quantization noise
Distortion tones
V out ( f )
After removing the desired signal and distortion tones.
Before going any further, let's show some example spectrums and the
corresponding y-axis units. The top waveform in Fig. 36.41 shows the resulting spectrum
when the input to the VSA is a 0.5 V (peak) waveform at 2 kHz. The RMS value of this
waveform is 354 mV. In dBm (using a 1 MΩ VSA input resistance) this is
10 ⋅ log (0.354) 2 /1 MΩ /1 mW or −39 dBm; see the top trace in Fig. 36.42. The units
for the top trace in Fig. 36.41 are volts, root-mean-square. Many of the spectrums we
used in earlier chapters used peak voltages for the y-axis units.
The bottom trace in Fig. 36.41 shows the voltage spectral density of our 0.5 V
peak sinewave at 2 kHz. The units of a voltage spectral density are V/ Hz (or more
precisely volts RMS per root Hz). Because the same exact waveform is input to the VSA
for each spectrum, we can relate the top and bottom waveforms in Fig. 36.41 simply by
knowing the resolution bandwidth of the measurement
V RMS V RMS
= (36.25)
Hz Resolution bandwidth
The resolution bandwidth used by the VSA, for the waveforms of Fig. 36.41, was 100 Hz.
This means the amplitude of the sinewave is now 354 mV/ 100 Hz or 35.4 mV/ Hz . In
dB this would be 20 ⋅ log 0.0354 or −29 dB. For a power spectral density, see the bottom
trace in Fig. 36.42, we can use
V 2 (or Watts) V 2 (or Watts)
= (36.26)
Hz Resolution bandwidth
Because our resolution bandwidth is 100 Hz, we expect the power spectral density to be
20 dB less than the power spectrum (top trace in Fig. 36.42) or −59 dBm/Hz. It should be
obvious how to change from dBm/Hz to V2/Hz.
Chapter 36 At the Bench 487
The RMS value of the desired signal in Fig. 36.19 is 1.41 V. The SNR, for this
modulator's output spectrum, is
0.01 +5 V
11,13,14,15,17,18
20k Reconstructed
b 10
5
Output
10k
20k
Analog input 19 In 9
b3
20k 10k
8 20k
ADC b2
10k
20k
Clock 12 Clk b 11 7
10k
20k 20k
6
b0
1,2,20,21,23,24 20k
The ADC we selected is the TLC5540. It is an 8-bit ADC. However, we will only
use the upper five bits of the ADC to illustrate the quantization process. The maximum
reference voltage, V REF+ , is 5 V, while the minimum reference voltage, V REF− , is ground.
The clock pulse we'll use for our measurements will oscillate between ground and 5 V at 1
Chapter 36 At the Bench 489
MHz. Because we are using five bits we can estimate the weighting of the LSB using Eq.
(30.23) as 156 mV. Further, from Eq. (30.30), we can estimate the RMS value of the
quantization noise as 45 mV. A picture of the prototyped ADC/DAC is seen in Fig. 36.44.
The TLC5540 comes in a plastic small outline package (SOP). This SOP package is
difficult to solder by hand in our deadbug prototyping scheme so we soldered it into a
dual-in-line package (DIP) carrier. This makes prototyping the circuit much easier.
Figure 36.45 shows the output spectrum for the ADC and resistive DAC seen in
Fig. 36.43. Again, the clock frequency is 1 MHz. The input signal is a sinewave at 50 kHz
with 0.5 V peak and centered around 2 V. The reason we don't see a DC signal in the
spectrum is that the VSA's input was AC coupled. Again, the VSA's input resistance is 1
MΩ. Note how the spectrum rolls off with increasing frequency. We connected the output
in Fig. 36.43 to the VSA's input through a piece of coax cable. The coax was 3 feet long
and resulted in a capacitance of approximately 100 pF shunting the VSA's input, Fig.
36.46. If we model the ADC/DAC as a voltage source with 10k output resistance, then
the frequency response of the measuring circuit is lowpass with a corner frequency of
Figure 36.45 Output spectrum showing quantization noise for the 5-bit ADC in Fig. 36.43.
10k
100p 15p
1MEG
Looking at Fig. 36.45 and knowing that the spectrum rolls off because of the
measuring system, we can get an estimate for the quantization noise power at low
frequencies as −65 dBm. Knowing the resolution bandwidth of the measurement was
10-kHz, we can estimate the power spectral density using
V 2RMS /1 MΩ
−65 dBm = 10 ⋅ log → V 2RMS = 316 × 10 −6 V 2
1 mW
or
−6 2
PSD = V 2out ( f ) = 316 × 10 V = 31.6 × 10 −9 V 2 /Hz
10 kHz
The Nyquist frequency is 500 kHz. If we again assume a filter is used to bandlimit the
ADC output to the Nyquist frequency, we can calculate the RMS quantization noise as
Chapter 36 At the Bench 491
1/2
500k
V Qe,RMS = ∫ 31.6 × 10 −9 ⋅ df = 125 mV
0
This RMS noise is three times larger than what we calculated earlier (45 mV). We might
speculate that the difference is due to not adequately randomizing the noise by using too
high of an input frequency, relative to the sampling frequency, or too small an input
amplitude (all of which are easy to verify at the bench).
Finally, let's show some time-domain waveforms showing quantization effects.
Figure 36.47 shows the input and output waveforms when the input frequency is 5 kHz.
Note how, as we calculated earlier, 1 LSB is 156 mV. Figure 36.48 shows the output
when the input frequency is increased to 50 kHz, while Fig. 36.49 shows the circuit's
inputs and outputs when the Nyquist frequency is applied to the circuit. Note how, as we
would expect, the DAC output is simply a square wave at a frequency of 500 kHz. After
this output is passed through a reconstruction filter with a frequency of just over 500 kHz
we get our exact replica of the input signal simply shifted in time. While the signal in Fig.
36.45 was measured by providing a connection between the circuit and the VSA using a
piece of co-ax cable, Fig. 36.46, the signals in Figs. 36.47 - 36.49 were measured using a
compensated scope probe, Fig. 36.7. The significantly reduced loading resulting from
using the compensated scope probe eliminates the spectrum roll off that was present in
Fig. 36.45.
156 mV
Figure 36.47 ADC input frequency of 5 kHz and the DAC output.
492 CMOS Mixed-Signal Circuit Design
Figure 36.48 ADC input frequency of 50 kHz and the DAC output.
Figure 36.49 ADC input frequency of 500 kHz (the Nyquist frequency) and the DAC output.
While the circuits built in this chapter represent a small number of examples,
relative to the material covered in the book, it is hoped that they are representative enough
to make the engineer/student want to spend some time at the bench.
bandpass filter, 132
circuits, 54, 89-91
clock jitter, 93-94
W
W-2W current steering DAC (mirror), 250-251,
333-334
White noise, 97, 303, 485
Windowing, 13-14, 46-48
WinSPICE, See also SPICE.
loading external data to perform a DFT, 67
Wireless, 230
I/Q extraction, 230-231
Z
z
defined, 17
domain, 16-17, 141
to s-transform (practical), 424, 480
z-plane, 99-102
Zeroes padding, 131
About the Author
Dr. Baker holds over 20 granted or pending patents in integrated circuit design and is a
member of the electrical engineering honor society Eta Kappa Nu. He is a co-author of
CMOS: Circuit Design, Layout, and Simulation, Wiley-IEEE, 1998, and DRAM
Circuit Design: A Tutorial, Wiley-IEEE, 2001. His research interests are in the areas of
CMOS mixed-signal integrated circuit design and the design of memory in new and
emerging fabrication technologies. Dr. Baker was a corecipient of the 2000 Prize Paper
Award of the IEEE Power Electronics Society.