Transcript Presentation - Signal Processing Group

CMOS Switched-Capacitor Circuits for
Bio-Medical and RF Applications
David J. Allstot
Mackay Professor of EECS
University of California
Berkeley, CA 94720
Origin of Switched-Capacitors?
James C. Maxwell, A Treatise on Electricity and Magnetism
Oxford: Clarendon Press, 1873, vol. 2, pp. 374-375.
2
MOS Switched Capacitors - 1972
•
David L. Fried, “Analog Sample-Data Filters,” IEEE J. Solid-State Circuits,
pp. 302-304, Aug. 1972. – MOS SC “resistor” concept and SC n-path filter
Early MOS data converters and switched-capacitor filters for the perchannel voice-to-PCM interface of digital telephony – UC Berkeley
•
J.L. McCreary and P.R. Gray, “All-MOS charge redistribution analog-todigital conversion techniques: Part I,” IEEE JSSC, Dec. 1975.
•
R.E. Suarez, P.R. Gray and D.A. Hodges, “All-MOS charge redistribution
analog-to-digital conversion techniques: Part II,” IEEE JSSC, Dec. 1975.
•
Y.P. Tsividis and P.R. Gray, “An integrated NMOS operational amplifier with
internal compensation,” IEEE JSSC, Dec. 1976.
Paul R. Gray
David A. Hodges •
I.A. Young, D.A. Hodges and P.R. Gray, “Analog NMOS sampled-data
recursive filter,” IEEE ISSCC, Feb. 1977.
•
D.J. Allstot, R.W. Brodersen and P.R. Gray, “MOS switched-capacitor
ladder filters,” IEEE JSSC, Dec. 1978.
Key Paper on n-path filter analysis:
•
B.D. Smith, “Analysis of commutated networks,” IRE Trans. on Aerospace
and Navigational Electronics, pp. 21-26, 1953.
Robert W. Brodersen
3
Future Research Topics
N-Path Filters:
Blocker-tolerant
front ends
Switched
Capacitor:
High-efficiency,
high-power
transmitters;
Converters
Time-to-Digital
Converter:
Ring-oscillator
amplifiers;
Analog-to-digital
converters
Golden Age for RF-CMOS Design!
*Courtesy of Prof. James Buckwalter, UC Santa Barbara
4
Outline
 Challenges in CMOS Radio Design
 Switched-Capacitor N-path Filters
 Analog-domain Compressed Sensing for
Bio-signal Acquisition
5
Ubiquitous Wireless
Emerging IT platforms fundamentally change the way we
interact with and live in the information-rich world
Core
Projection: 1000 radios per
person on Earth by 2025
TRILLIONS OF
CONNECTED DEVICES
The Cloud
The Swarm
Mobile Access
Sensors
[J. Rabaey, ASPDAC’08]
J. M. Rabaey, "A Brand New Wireless Day: What
Does It Mean for Design Technology?," Asia and
South Pacific Design Automation Conf., 2008, p. 1.
Vision potentially doomed by
network deficiencies:
• lack of availability
• lack of reliability/robustness
• lack of security
6
RF Transceiver Coexistence
N-path
State-of-the-Art
• Without SAW filter:
IIP3 =
(PJAM + 2PTX - PXM - 5)
= 35dBm
2
• TX leakage needs at least 20dB of rejection to improve IIP3 so that
LNAs can handle input power
• Challenge: Reconfigurable, linear duplexer + SAW replacement
*Courtesy of Prof. James Buckwalter, UC Santa Barbara
7
“Brain Radio” Coexistence
Neural
Recording
LNA
Neural
Stimulation
PA
• Stimulator leakage needs
rejection to increase IIP3 so
LNAs can handle input power
8
Universal Receiver – Blocker Rejection
• Low Cost
- No Inductors 
- No Off-Chip Filters
•• Low
Low Noise
Noise Figure
Figure 
•• High
High Linearity
Linearity
•• Low
Low Power
Power Diss.
Diss. 
•• High
High Blocker
Blocker Tolerance
Tolerance
•• Wide
Wide Frequency
Frequency Range
Range 
LNA
GSM Example
*Courtesy of Prof. Behzad Razavi, UCLA, 2015 ISCAS
Keynote Presentation
9
Translational Filter à la Smith

• Scaled transistors are good
switches with low Ron on Coff
N-path filter basics
• Each “path” behaves as a passive
mixer that translates the
baseband impedance to an RF
impedance
S21 S11 (dB)
Shunt RLC filter
that is tuned with
local oscillator
0.0
-5.0
-10.0
-15.0
-20.0
-25.0
S21
S11
500.0M
1.0G
1.5G
Frequency (Hz)
2.0G
• Large switches reduce insertion
loss but limit tunability
* Luo and Buckwalter, MWCL 2014
10
Shunt vs. Series N-path Filters
• Shunt filter: Bandpass response
• Series filter: Bandreject response
• compatible with digital CMOS
• Benefits from faster switches (e.g., CMOS SOI process)
* Luo and Buckwalter, MWCL 2014
11
How Many Paths?
• Number depends on the tunability of the filter
• Aliasing is prevented to
the N-1 LO harmonic.
• Low OOB rejection is a
problem in spite of high
linearity.
* Luo and Buckwalter, MWCL 2014
Harmonic aliasing (dBc)
• Require each path to be switched with 1/N duty cycle
-10
-20 Luo and Buckwalter, MWCL 2014 simulation
-30
measurement
-40
-50
-60
-70
-80
-90
-100
2 3 4 5 6 7 8 9 10 11 12 13 14 15
Harmonic
12
Can We Filter at the Antenna?

N-path filter basics
• For
BW = 200 kHz: Ctot = 28 nF
• For 20-dB rejection: Rsw = 5 W
• Switch linearity with 0-dBm blocker?
*Courtesy of Prof. Behzad Razavi, UCLA, 2015 ISCAS Keynote Presentation
13
Miller Resistance
*Courtesy of Prof. Behzad Razavi, UCLA, 2015 ISCAS Keynote Presentation
14
Miller Bandpass Filter
Ctot=2 nF
NF ~ 1.6 dB
*Courtesy of Prof. Behzad Razavi, UCLA, 2015
ISCAS Keynote Presentation
• Low Cost
- No Inductors 
- No Off-Chip Filters 
• Low Noise Figure 
• High Linearity?
• Low Power Diss. 
• High Blocker Tolerance?
• Wide Frequency Range 15
Miller Multiplication / Harmonic Rejection
100 pF
50 W
Fundamental
*Razavi, 2014 CICC; Weldon, et al., Dec. 2001 JSSC
Third Harmonic
16
Outline for Compressed Sensing
 Motivation for Compressive Sampling
 Intuition and Key Ideas
 Reconstruction
 Experimental Results
17
Motivation for Compressive Sampling

(Medical) Body Area Networks
 Many wireless sensors linked to Smartphone, nearby IPAD, etc.
 Personal mobile units linked to Dr. via internet/cellular network
 Dr. feedback for real-time control of detail vs. energy efficiency

Reduce data rates to increase sensor lifetime and energy efficiency
18
CS Sensor System
Compressed Sampling
Bio-Signal Acquisition System
x(t)
LNA
Electrode
Sensor
Feedback
CS
AFE
[Y]
ADC
Antenna
Power
Amplifier
Compressed Data Rate

Ultra-low-power CS Analog Front-end

RF PA is Dominant Energy Consumer; ADC Next

CS Compresses Data Rate and PA/ADC Duty Cycles

Compressed Data [Y] is Digitized and Transmitted
19
Conventional Sampling
1
2
3
4
5
6
7
8
9
10
11
12

12 Ball Problem: 11 Light Balls (1 g); 1 Heavy Ball (100g)

Goal: Identify Heavy Ball in Fewest Measurements

Conventional Sampling requires 12 measurements
20
Intuition for CS
1
7
2
8
3
9
4
10
5
11
6
12
1g
1g
1g
1g
1g
1g
1g
1g
1g
100g
1g
1g
Y
=
100000000000
010000000000
001000000000
000100000000
000010000000
000001000000
000000100000
000000010000
000000001000
000000000100
000000000010
000000000001
1g
1g
1g
1g
1g
1g
1g
1g
1g
100g
1g
1g
=
F
X
(Measurement
matrix)
(Signal
Vector)
(Measurement
Vector)

Key Idea: Extend Group Sampling Fewer Measurements
•
•
R. Dorfman, “The detection of defective members of large populations,” The Annals of
Mathematical Statistics, vol. 14, pp. 436-440, Dec. 1943.
M. Sobel and P.A. Groll, “Group testing to eliminate efficiently all defectives in a binomial sample,”
Bell System Technical Journal, vol. 38, pp. 1179-1252, Sept. 1959.
21
Random Sampling – 1
102g
1
1
8
10 11
6
5
7 11
3
=
4
9 10 12
000000010110
2
8
1g
1g
1g
1g
1g
1g
1g
1g
1g
100g
1g
1g

Random Sample to Find Y11

Use 1-b Random Numbers (e.g., Bernoulli, Toeplitz,
Circulant, etc.) Incoherent Between Rows
22
Random Sampling – 2
102g
5g
1
1
8
10 11
6
5
7 11
3
4
9 10 12
2
8
2
8
=
000000010110
100011100010
1
1
8
10 11
6
5
7 11
3
4
9 10 12
1g
1g
1g
1g
1g
1g
1g
1g
1g
100g
1g
1g

Random Sample to Find Y21

Use 1-b Random Numbers (e.g., Bernoulli, Toeplitz,
Circulant, etc.) Incoherent Between Rows
23
Random Sampling – 3
1
1
8
10 11
6
10 11
7 11
1
8
6
3
5
9 10 12
Random Sample to Find Y31

Reconstruction: Two Heavy
1
Measurements—Only #10 Common
3
4
12
2 8
9 10Fewer
Measurements
(e.g., 3)

CS Works for Sparse Signals
1
10 11
6
5
7 11
3
1g
1g
1g
1g
1g
1g
1g
1g
1g
100g
1g
1g
 Other (unlikely) Possibilities:
1
8
=
000000010110
100011100010
101100001101

5
7 11
4
102g
5g
105g
2 8
4
9 10 12
 Solution in 1 Measurement
2
8
 No Solution in M Measurements
24
Sparsity vs. Compressibility
22
Compression Factor, C = N/M
18
14
10
8-bit ECG
6
2 50
60
70
80
90
100
Sparsity (%)

Limit: M > K log(N/K); K Nonzero Samples; Heuristic: M > 2K

Error Bounds: E. Candès, “An introduction to compressive sampling,” IEEE
Signal Processing Magazine, vol. 25, pp. 21-30, Mar. 2008.

E. Candès and T. Tao, “Near optimal signal recovery from random
projections: Universal encoding strategies,” IEEE Trans. Info. Theory, vol. 52,
pp. 5406-5425, Dec. 2006.
25
Compressed Sampling - I
…
[F]MXN = [F11, …, F1N ]
]
[
[
]
[FM1, …, FMN ]
[Y]MX1 =
[Y11, …, YM1]
[Y] = [Φ][X]
[X]NX1 =
[X11, …, XN1]
K=3

[X]16X1; [F]8X16; [Y]8X1; C = 2

[F] is Gaussian, Uniform, Bernoulli, Toeplitz, etc.

Multiply and sum for each Yij is a Random Linear Projection

[Y] is compressed analog signal with global information

K < M < N for sparse signal such as ECG, EMG, etc.
26
Compressed Sampling - II
[X]
[Y]
 [X]1024 X 1: Analog ECG samples
 [Y]256 X 1: Compressed analog output
 [F]256 X 1024: Measurement Matrix
 C = 4X
27
CS Reconstruction
Compressed Sensing BioSignal Reconstruction System
Antenna
y(t)
LNA
Baseband DSP
CS Optimization/
Reconstruction
DAC
Original Nyquist Data Rate

Reconstruction of Compressed Signal (e.g., Smartphone)

[Φ] is Non-square; Under-determined System with Many Solutions

Optimize; e.g., Convex Optimization with L1-Norm Minimization

“Feature Extraction” in DECODER Using [Y]—Sparsifying Matrix; e.g.,
Mexican Hat Wavelet to extract QRS Complex of ECG Waveform
A.M.R. Dixon, E.G. Allstot, D. Gangopadhyay, and D.J. Allstot, “Compressed sensing system considerations for ECG and
EMG wireless bio-sensors,” IEEE Trans. on Biomedical Circuits and Systems, vol. 6, pp. 156-166, April 2012.
28
CS Reconstruction - II
[X]
[Y]

Accuracy depends on:
 Compression Factor, C = N/M
 PDF of random coefficients and # bits
 Algorithm—Convex Optimization with L1 Norm
29
Switched-capacitor CS CODER

Electr
ode
CSADC
Structure Matrix
operations so that
input is pipelined.
Eliminates many
explicit S/H circuits
CSAD
C
30
Switched-capacitor CS CODER
Compressed Sensing BioSignal Acquisition System
Antenna
Ultra-low Power Analog Circuits
LNA
Electrode
Sensor
SC Multiplying
Digital-Analog
Converter
CS
AFE
ADC
[Y]Power
= [Φ][X]
Amplifier

64 Rows Implemented:

C-2C 6-b MDAC/ADC

C-2C 10-b SAR ADC
31
Switched-capacitor CS CODER
Fig. 3. CSADC circuits. Counterclockwise from top: Opamp, C-2C MDAC/integrator, C-2C SAR ADC (with
pre-amp offset cancellation), and comparator. Device stacking to reduce W/L and dual-gate switches
and logic gates are used to minimize leakage.

64 Rows digitally selectable

N is programmable
32
CSADC Measured Results (ECG)
Raw ECG
Compressed Y values
2X (32 rows; 0.9 uW)
4X (16 rows; 0.4 uW)
6X (10 rows; 250 nW)
Measured reconstruction of an ECG signal sparse in Daubechies-4 wavelet
domain using 8 frames each of N=128 samples. (Not thresholded at input.)
33
CSADC Results (ECG Bio-signals)
Raw ECG
Amplitude (mV)
Compressed Y values
2X (64 rows; 0.9 uW)
4X (32 rows; 0.45 uW)
8X (16 rows; 225 nW)
16X (8 rows; 112 nW)
time (s)
Measured reconstruction of an ECG signal sparse in the time domain using
8 frames each of N=128 samples. (thresholded at input.)
34
Switched-capacitor CSADC
IBM8RF
64 6-b C-2C MDAC
64 10-b C-2C SAR ADC
8 pad drivers
64 Comparators
64 SAR logic blocks
64 10-b C-2C SAR Cap-DAC
64 Op Amps
64 6-b Word Fibonacci / Galois LFSR
64 6-b C -2C MIDACs
3 mm
IBias,
Timing
0.13 µm CMOS
2 mm x 3 mm
M = 1 … 64 (selectable)
N = 128, 256, 512, 1024
C = N / M (Comp. Ratio)
Test Structures : MIDAC and SAR
2 mm
28 nW/row
D. Gangopadhyay, E.G. Allstot, A.M.R. Dixon, S. Gupta, K. Natarajan and D.J. Allstot, “Compressed sensing analog frontend for wireless bio-sensors,” IEEE JSSC, vol. 49, pp. 426-438, Feb. 2014.
35
Future Research Topics
N-Path Filters:
Blockertolerant front
ends
Switched
Capacitor:
High-efficiency,
high-power
transmitters;
Converters
Time-to-Digital
Time-to-Digital
Converters;
Converter:
Ring-oscillator
amplifiers;
Analog-to-Digital
Analog-to-digital
Converters
converters
Open Area of Research for Wireless and Biomedical!
*Courtesy of Prof. James Buckwalter, UC Santa Barbara
36
Mulţumesc
37