MITOPENCOURSEWARE

MASSACUSETTS INSTITUTE OF TECHNOLOGY

6.976
High-Speed Communication Circuits and Systems
Lecture 29
Lowpass and Bandpass Delta-Sigma Modulation

Richard Schreier

ANALOG
DEVICES
Copyright © 2003 Richard Schreier

1
 Outline
1 ∆Σ Basics
1st-Order Modulator

2 Advanced ∆Σ
High-Order ∆Σ Modulators
Multi-bit and Multi-Stage Modulation

3 Bandpass ∆Σ Modulation
4 Example Bandpass ADC

2
1. ∆Σ Basics

3
 st
CTMOD1: A 1 -Order
Continuous-Time ∆Σ Modulator
• The input signal, U, is converted into a sequence
of bits, V ∈ (0,1).

C
R
U
Y V

CK
0,I

4
 Properties of CTMOD1
DC Inputs
• Integrator ensures that input current is exactly
balanced by the (average) feedback current
“Infinite resolution”
• Signals which alias to DC are rejected
“Inherent anti-aliasing”

5
 Non-ideal Effects in CTMOD1
• Component shifts
R → R+∆R or I→ I+∆I merely changes full-scale.
C → C+∆C scales the output of the integrator,
but does not affect the comparator’s decisions.
• Op-amp offset, input bias current,
DAC imbalance
All translate into a DC offset, which is unimportant in
many communications applications.
• Comparator offset & hysteresis
Overcome by integrator.
• Finite op-amp gain
Creates “dead-bands.”

6
 Non-ideal Effects (cont’d)
• DAC jitter
Adds “noise.”
• Resistor nonlinearity (e.g. due to self-heating)
Introduces distortion.
• DAC nonlinearity
Introduces distortion and intermodulation of shaped
quantization noise.
• Capacitor nonlinearity
Irrelevant.
• Op-amp nonlinearity
Same effects as DAC nonlinearity, but less severe.

7
 CTMOD1 Model
• Normalize R=1Ω, C=1F, I=1A, Fs=1Hz
Full-scale range is [0,1]V.
• Assume comparator and DAC are delay-free

∫
uc yc y v

vc
DAC

8
 Waveforms/Timing
yc u = 0.2
1
Comparator
Threshold

-1 t
vc
v(4) 1
1 V = ---
5
v(0) v(1) v(2) v(3)
0 t
0 1 2 3 4 5 6 7 8 9 10
9
 CTMOD1 @ 5% 1’s density
l(r)
l(ck)
v(u)
1

-1
-v(yn)
1

-1

-2
l(v)
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1
time, x1e-6 Seconds

10
CTMOD1 @ 10% 1’s density
l(r)
l(ck)
v(u)
1

-1
-v(yn)
1

-1

-2
l(v)
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1
time, x1e-6 Seconds

11
CTMOD1 @ 51% 1’s density
l(r)
l(ck)
v(u)
1

-1
-v(yn)
1

-1
l(v)
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1
time, x1e-6 Seconds
12
CTMOD1 @1/π 1’s density
l(r)
l(ck)
v(u)
1

-1 1 7
-v(yn) --- ------
1 3 22
0

-1

-2
l(v)
0 .1 .2 .3 .4 .5 .6 .7 .8 .9 1
time, x1e-6 Seconds
13
Analysis of CTMOD1

∫
From the diagram: n
y c(n) = y c(n – 1) + ( u c(τ) – v c(τ) )dτ
n–1

1) Sample yc at integer time and identify y(n) = yc(n).

∫
n
2) Observe that v c(τ)dτ = v(n – 1)
n–1

∫
n
3) Define u(n) = u c(τ)dτ
n–1
THEN
y(n) = y(n – 1) + u(n) – v(n – 1)
Also, from the diagram
v(n) = Q(y(n))

14
 CTMOD1 Equivalent
MOD1
uc 1 u y v
∫0 dt
z-1

z-1

• CTMOD1 is the same as a discrete-time first-

order modulator (MOD1) preceded by a sinc
filter!

15
 CTMOD1 NTF and STF
• The NTF is the same as MOD1:
z-plane:

NTF(z) = 1 – z –1

• MOD1’s STF is 1, so the overall STF is just the TF

of the prefilter:
∞
∫0
s-plane: 4π
STF(s) = e –st g p(t) dt
2π
1 (1 – e –s )
= ∫0 e –st dt = ----------------------
s
( 1 – z –1 )
= ---------------------- , where z = e s
s

16
 Frequency Responses
10
NTF
0

-10
STF
dB

-20

-30 Q. Noise Inherent

Notch Anti-Aliasing
-40

-50
0 1 2 3
Frequency (Hz)
17
 CTMOD1 Spectra
spec1
0
u = 1/32

-50
dB

-100
spec2
0
-6dB peak

-50
dB

u = FS sine-wave
-100
0 5 10 15 20 25 30 35 40 45 50
freq, x1e6 Hertz
18
 Properties of MOD1
• Single-bit quantization yields “inherent linearity.”
The DAC defines two points and two points can
always be joined with a line. (Not so simple in
continuous-time.)
• 0≤u≤1⇒ y ≤1
MOD1 is stable for inputs all the way up to full-scale.
The quantizer in MOD1 does not “overload.”
• Assuming the quantization error is white with
power σ e2 , the in-band noise power is
π 2 σ e2
N 02 ≅ ------------------------3- . ~12-bit performance at OSR=256.
3 ( OSR )
Doubling OSR reduces noise power by a factor of 8.
“1.5 bits increase in SNR per octave increase in
OSR”
19
 MOD1 Properties (cont’d)
• DC input u = --a- results in period-b behavior.
b
The spectrum of the error is not white! Spectrum
consists of a finite set of harmonics of f s ⁄ b .
• Irrational DC inputs result in aperiodic behavior.
Nonetheless, the spectrum of the error is still
discrete!
Spectrum consists of an infinite number of tones
with frequencies that are irrational fractions of f s .
• Finite op-amp gain shifts NTF zero inside the
unit circle and allows a range of u values to
produce the same limit cycle.
1
Worst case is around u = 0, 1, --- etc. ;
yields “dead bands.” 2
• The behavior of MOD1 is erratic.

20
2. Advanced ∆Σ

21
 A Single-Loop ∆Σ Modulator
E
U L0
Y
V
L1

Y = L0 U + L1 V V = GU + HE , where
1
V = Y+E H = --------------- & G = L 0 H
1 – L1
Inverse Relations:
L1 = 1 – 1/H, L0 = G/H

• The zeros in H come from the poles in L1

22
 th
A 5 -Order Lowpass NTF
Zeros optimized for OSR=32
• Pole/Zero diagram:

optimization
1 flag

OSR = 32;
H = synthesizeNTF(5,OSR,1);
...
0

Zeros spread across

the band-of-interest to
minimize the rms value
of the NTF.
-1 Poles such that ||H|| = 1.5.
-1 0 1

23
 th
Example: 5 -Order Modulator
1

-1

0 50 100
Time (sample number)

0
-20
-40
dBFS

-60
-80
-100 NBW = 1.8x10–4 fs
(8K-Point FFT)
-120
0 Normalized Frequency (1 → fs ) 0.5

24
 SQNR Limits for Binary
Modulators

N=8
N=7
N=6
N=5
140 N=4 N=3
N=2
120
Peak SQNR (dB)

100
N=1
80

60

40

20

0
4 8 16 32 64 128 256 512 1024 OSR
25
 Multi-Bit Quantization
Toolbox Conventions
• Single-bit quantizer output interpreted as ±1
instead of 0,1.
Quantizer step size, ∆, is 2; input range is [-1,+1].
• Convention for multi-bit quantization is:
M=1 M=2 M=3
v
3
2 v
1 v 1
y y y
-1 -1 e
-2 e
e -3
mid-tread quantizer; mid-rise quantizer;
v: even integers v: odd integers

∆ = 2; # of Q. levels is nlev = M+1, from –M to +M;

no-overload range (|e| ≤ 1) is –nlev to +nlev.

26
SQNR Limits for 3-bit Modulators

N=8
N=7
N=6
N=5
N=4 N=3 N=2
140

120
N=1
Peak SQNR (dB)

100

80

60

40

20

0
4 8 16 32 64 128 256 512 1024

OSR
27
 Theoretical SNR Limits
for Multi-Bit Modulators
H = 32 N=8
120 ∞
with 1 LSB (peak) input

H ∞ = 16 N=7
SNR (dB) @ OSR = 8

100 H ∞ = 8 N=6

N=5
80 H ∞ = 4
N=4

60 N=3
H ∞
= 2 N=2
40
H ∞ is the max. gain of the NTF
over all frequencies.
20
100 101

Total RMS Noise Power (LSBs)

28
 Multi-Bit Quantization
Pros and Cons
• Multi-bit quantization overcomes stability-
induced restrictions on the NTF
Dramatic improvements are possible!
• Multi-bit quantization loses the inherent linearity
property of a binary DAC
DAC levels are not evenly spaced and so cannot be
joined with a straight line.
DAC errors are effectively added to the input, and
thus are not shaped.
Can be overcome with calibration, digital correction
or mismatch-shaping.

29
 Digital Correction
Look-up vdig
∫
u v
Table

vdac
DAC

• Lookup table contains the digital equivalent of

each DAC level
In practice, the look-up table only needs to store the
differences between the actual and ideal DAC levels.
• Thus vdig = vdac, so DAC errors are now shaped
by the loop!

30
 Mismatch-Shaping
• Shapes mismatch-induced noise by ensuring
that each element in a unit-element DAC is
driven by a shaped sequence
Two popular forms of mismatch-shaping are
element-rotation and element-swapping.
Thermometer Rotation Swapping
16 16 16

8 8 8

1 1 1
Time Time Time

31
 Multi-Stage Modulation
u 2nd-order v1
z -1(2 - z -1)
modulator mod = mod2;
ABCD = mod.ABCD;
x2 [v1 x] = simulateDSM(input,ABCD);
0.25 v v2 = simulateDSM(x(2,:)/4,ABCD);
v = filter([0 2 -1],1,v1) + ...
4*filter([1 -2 1],1,v2);

2nd-order v2
4(1 - z -1)2
modulator OSR=32
0
-20
G = z -1, -40
v1
SQNR = 56 dB
dBFS

H = (1 - z -1)2 -60
-80 v
-100 SQNR = 81 dB

-120 NBW=1.8E-4
Composite NTF is 4H 2 10
-3
10
-2
10
-1

32
 ∆Σ Toolbox Summary
http://www.mathworks.com/matlabcentral/fileexchange/
Click on Control Systems, then delsig
Specify OSR,
lowpass/bandpass,
no. of Q. levels.

synthesizeNTF scaleABCD
calculateTF
stuffABCD
NTF (and STF) avail- realizeNTF Parameters for a spe- ABCD: state-space
able. cific topology. mapABCD description of the
modulator.
predictSNR, findPIS,
simulateDSM, find2dPIS
simulateSNR
Time-domain simula- Also designLCBP Convex positive
tion and SNR mea- invariant set.
surements. designHBF
simulateESL…

33
3. Bandpass ∆Σ

34
 A Bandpass ∆Σ ADC
• Like a lowpass ∆Σ ADC, a bandpass ∆Σ ADC
converts its analog input into a bit-stream
The output bit-stream is essentially equal to the
input in the band of interest.
• A digital filter removes out-of-band noise and
mixes the signal to baseband

u Bandpass v Bandpass complex

∆Σ Filter/ w digital
analog Modulator Decimator output
input 1 bit @fs
2xn @fB
baseband
signal shaped signal
noise

–f s ⁄ 2 fs ⁄ 2 –f s ⁄ 2 fs ⁄ 2 fB

35
 BP∆Σ Perspective #1
It is just Filtering and Feedback
• Putting the poles of the loop filter at ω0
forces H to have zeros at ω0
• Example system diagram:

Lots of gain at ω0
ω 02
------------------
-
s + ω0
2 2

DAC

36
 BP∆Σ Perspective #2
It is just the result of an
“N-Path Transformation”
• z → – z 2 (a “pseudo 2-path transformation”)
applied to H(z) = 1 – z –1 yields H'(z) = 1 + z –2
• This transformation can be applied to any
system that processes DT signals, including SC
filters, digital filters, ∆Σ modulators and even
mismatch-shaping logic
By replacing the state storage elements (registers),
or by interleaving two copies of the original system
and negating alternate inputs and outputs

37
 BP∆Σ Perspective #3
It is something New and Valuable
• BP∆Σ offers a way to make a “tuned” ADC
Possibly the only way.
Ideally-suited to narrowband systems, i.e. radios.
• BP∆Σ keeps the signal away from 1/f noise as
well as low-frequency distortion products
Like regular narrowband bandpass systems,
second-harmonic distortion is not problematic.

1,1,-1,-1 1,1,-1,-1
H(z)
z-1 z-1 z-1 -1 OR
H(z)

38
 th
A 6 -Order Bandpass NTF
• Pole/Zero diagram:
OSR = 64;
f0 = 1/6;
H=synthesizeNTF(6,OSR,1,[],f0);...

-1
-1 0 1

39
 Example Waveform
th
8 -Order fs/8 Bandpass Modulator
v
1
u

–1

0 50 100
Time (sample number)

40
 Example Spectrum
th
8 -Order fs/8 Bandpass Modulator
0
SQNR = 100 dB
-20 @ OSR = 64

-40
dBFS/NBW

-60

-80

-100

-120
NBW = 1.8x10–4 fs
-140
0 0.1 0.2 0.3 0.4 0.5
Normalized Frequency (1→fs)
41
 Bandpass Modulator Structure
u
Q
Resonator Resonator v
Quantizer

Loop Filter

• The loop filter consists of a cascade of

resonators
The resonance frequencies determine the poles of
the loop filter and hence the zeros of the NTF.
• Multibit and multistage variants are also
possible

42
4. Design Example

43
 A Dual-Conversion
Superheterodyne Receiver
LNA AGC ADC DSP

LO1 LO2 CK

RF 1st IF 2nd IF Sample Rate

50-2000 MHz 10-300 MHz 2-4 MHz 10-750 ksps

• A bandpass ADC fits naturally into this

narrowband system
Perfect I/Q, high dynamic range. Low power?
44
 System Partitioning

Front End Back End Custom

DSP
N

LO1 LO2 CK

• Goal: a general-purpose, high-performance,

low-power back-end

45
 Traditional Implementation
MIXER
IF1 IF2
LNA AAF VGA BP ADC

LO2

Numerous high-dynamic range blocks

Noise and power budgets are very tight

Large VGA range needed

46
 Eliminating the AAF with a
Continuous-Time BP Σ∆ ADC
IF2 gm gm FLASH
IDAC

IDAC

Anti-alias filtering is inherent

But still need a low-noise, linear V-I converter

47
 Eliminating the Input gm
Mixer
IF1 Remainder
LNA gm
of ADC

LO2 IDAC

The output of the mixer is available in

current form, so …

48
 Merge ADC with Mixer!
MIXER ADC
ADC input is No DC drop!
a current!

TO/FROM

IDAC
LO2 ADC
BACKEND

IF1
LNA V-I CONVERTER

• Eliminates redundant I-V & V-I conversion

• Gives mixer and IDAC more headroom
49
 Merge ADC with Mixer!
No noise!
No distortion!
No power!

IDAC
LO2

IF1
LNA V-I CONVERTER

LC tank effectively adds gain, without adding

noise, adding distortion or consuming power
50
 Noise Analysis
vn
vn
Z
g mZ vn
–+ –+
Vin gmVin
Z

LNA/Mixer Tank
• Noise in the ADC backend is attenuated by gm
times the tank impedance
In this work, gm ≈ 10 mA/V

51
 Zeff
• Near resonance, |ZL| = |ZC|
|ZL,C| ≈ 300Ω in this design
• At resonance, |Z| ≈ Q • |ZL,C|
About 6kΩ for Q = 20 ⇒ gmZ ≈ 60.
• More generally, the effective tank impedance is
found by integrating the input-referred noise
over the band of interest:

∫ ∫
v 2 v 2
v
 -------------------
2
  2 dω =  ------------------
- dω =  -------
n n n
g m Y (ω )  g m Z eff
 g m Z(ω)

⇒ Z eff = ( Y rms ) –1

52
 Z vs. Frequency
10K
Ideal
Zeff = 7K Q = 20
5K Zeff = 4.4K
Z
2K

1K
0.07 f0

0.8 0.9 1 1.1 1.2

f/f0
• Q = 20 reduces Zeff by about 4 dB
53
 Tuning the LC Tank
• ∆f0/f0 = 2% ⇒ 3 dB reduction in Zeff
Inductor accuracy is 10%, so tuning is required

Make an oscillator:

Off-chip
tank

Cap. array

Negative-gm

54
 Remainder of ADC?
ADC Flash
Backend
IDAC

• Add resonator stages until the quantization

noise of the flash is low enough

55
 Second Resonator?
ADC Flash
Backend
IDAC

• LC: Needs more external components plus

associated pins
• Active-RC: 2 mA for 50 nV/ Hz input-referred noise
• Switched-Cap: est. >10 mA for same noise
56
 Third Resonator?

–
+ ? Flash

–
+

• Active-RC: Q ≈ 10 ⇒ Need 4th resonator

• Switched-Cap: Q is high & drift is low;
<1 mA for 300 nV/ Hz i.r.n.

57
 Complete ADC
L C 9-Level
RC SC Flash Data
out

IF1 LNA/
LO2 Mixer IDAC DEM

9 mA 2 mA 3 mA 1 mA 1 mA

• Eliminates high-power VGA & AAF

2/3 of the total power used by LNA, Mixer & IDAC
• Uses cts-time and discrete-time elements, plus
multi-bit quantization and mismatch-shaping
58
 ADC in More Detail
External Tunable
Elements fCLK = 9-36 MHz
LC Tank
OSR = 48 to 960

RC SC Flash

IF1
LNA Mixer IDAC DEM
10-300
MHz
LO2
Full-Scale Adjust

• Can save power under small-signal conditions

by reducing IDAC’s full-scale
By a factor of 4, in this ADC

59
 Noise vs. Full-Scale
15
Input-Referred Noise
(dB relative to a 300Ω resistor)

10 Total
LNA/Mixer

0 RC a t ch
i s m AC T h e r m al o n
M ID t i
IDAC
n tiza
u a
Q
SC
-30 dBm -18 dBm
50 mVpp Full-Scale 200 mVpp
60
 Measured STF & NTF
0
STF
Measured
–20 STF
dBFS/NBW

–40 f0 = fCLK/8
NTF (scaled)
–60
Measured
PSD
NBW = 5.9 kHz
–80

fCLK = 32 MHz
–100
0 4 8 12 16
Frequency (MHz)
61
 In-Band Spectrum (OSR=48)
0
SNR = 81 dB fIF1 = 103 MHz
-20 fCLK = 26 MHz
dBFS/NBW

-40
SFDR = 103dB
-60

-80

-100

-120
BW = 270 kHz
NBW = 200Hz
-140
-150 -100 -50 0 50 100 150
Frequency Offset (kHz)
62
 In-Band Spectrum (OSR=900)
0
SNR = 92 dB fIF1 = 73 MHz
-20 fCLK = 36 MHz
dBFS/NBW

-40
SFDR = 106dB
-60

-80 -125 dBc/Hz

@ 1 kHz offset

-100

-120 NBW = 15Hz

-10 -5 0 5 10
Frequency Offset (kHz)
63
 SNR vs. Input Power
100
OSR =900 fIF1 = 273 MHz
fCLK = 32 MHz
80
SNR (dB)

60
OSR = 48

40

20
FS range: –30 to –18 dBm
DR = 90 dB
0
-110 -90 -70 -50 -30 -10 Pin (dBm)
64
 Architectural Highlights

Merging a mixer with a continuous-time

bandpass ADC containing an LC tank yields a
flexible, high-performance, low-power receiver
backend.

Performing a component-count/performance/
power trade-off in each resonator section results
in a multi-bit, hybrid continuous-time/discrete-
time architecture.

A variable full-scale saves power and reduces

noise.
65
 Performance Summary
Bandwidth 5 - 375 kHz
Input Frequency 10-300 MHz
Clock Frequency 9 - 36 MHz
Full-Scale Range 12 dB
Die Area 5 mm2
@ fIF = 73MHz, BW = 333kHz, fCLK = 32MHz, VDD = 3V:
Dynamic Range 90 dB
Current Consumption 16.5 mA
Noise Figure @ min FS 9 dB
IIP3 0 dBm
66
 Bandpass ∆Σ ADCs
120
Dynamic Range (dB)

Sch
100 reie
r 20
02
Jantzi/Ferguson Vel
dho
1994 ven
200
80 Salo 2002 3
Ueno 2002 Tabatabaei 2000
Hairapetian 1996
Norman
1996
60
Henkel
2002
Bulzachelli
2003 Raghavan
40 Salo 2002 1997
10 kHz 1 MHz 100 MHz

Bandwidth
67
 Summary
• ∆Σ is fun
All kinds of exotic behavior: limit-cycles, dead-bands
sub-harmonic locking and even chaotic dynamics!
• ∆Σ is a rich field
ADCs and DACs; Single-bit and Multi-bit; Single-
stage and Multi-stage; Lowpass and Bandpass;
Discrete-time and Continuous-time…
• A bandpass ∆Σ ADC converts an IF signal into
digital form and can do so with high dynamic
range and low power consumption
With wideband or tunable modulators, conversion of
RF to digital may soon be feasible.
ADC = “Antenna to Digital Converter”

68