static power reduction for asynchronous...
TRANSCRIPT
![Page 1: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/1.jpg)
Static Power Reduction for Asynchronous Circuits
Carlos Tadeo Ortega Otero
Asynchronous VLSI Group and Architecture
Computer Systems Laboratory
Cornell University
Ithaca, NY, 14853
May, 25th, 2012
![Page 2: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/2.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Low Duty-Cycle Power Constrained Designs
I Sensing every 5 minutes
I Battery life > 1 year
*Culler, D . Overview of Sensor Networks
Motivation:Applications 2/43
![Page 3: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/3.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Low Duty-Cycle Power Constrained Designs
I Biological implants
I Neurological stimulators
I Cardiac rhythm management
I RFID tags
I Remote sensing devices
Low Duty-Cycle and Power Constrained
Motivation:Applications 3/43
![Page 4: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/4.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Overview
1. Review leakage mechanisms
2. Review static power reduction techniques
3. Review Power Gating techniques
I Non-state preservingI State preserving
Our Contributions:
1. Power gating in the context of Async circuits
I Non-state preservingI State preserving
2. Evaluate Power Gating in Async circuits
3. ZZDRTO - Pipeline Power Gating
4. Evaluate ZZDRTO
Motivation:Applications 4/43
![Page 5: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/5.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Leakage Mechanisms
I Leakage: Current that �ows when the ideal current is 0A
I Source-Drain Leakage
I Subthreshold LeakageI Reverse biased diodeI Gate induced Drain leakage (GIDL)
I Gate Leakage
I Direct TunnelingI Hot carrier injection
Gate(G) Drain(D)Source(S)
Substrate
Iinv
ITUNNEL
IGIDL
ID,weak
IHC
Mechanisms:Overview 5/43
![Page 6: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/6.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Leakage Mechanisms: Source-Drain
I Leakage: Current that �ows when the ideal current is 0A
I Source-Drain Leakage
I Subthreshold LeakageI Reverse biased diodeI Gate induced drain leakage (GIDL)
I Gate Leakage
I Direct TunnelingI Hot carrier injection
Gate(G) Drain(D)Source(S)
Substrate
IinvIGIDL
ID,weak
Mechanisms:Overview 6/43
![Page 7: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/7.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Leakage Mechanisms: Gate Leakage
I Leakage: Current that �ows when the ideal current is 0A
I Source-Drain Leakage
I Subthreshold leakageI Reverse biased diode leakageI Gate induced Drain leakage (GIDL)
I Gate Leakage
I Direct TunnelingI Hot carrier injection
Gate(G) Drain(D)Source(S)
Substrate
ITUNNEL IHC
Mechanisms:Overview 7/43
![Page 8: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/8.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Source-Drain Leakage
Gate(G) Drain(D)Source(S)
Substrate
VdVg
ID,weak
>=0 >0.1
Subthreshold leakage (top)
Gate(G) Drain(D)Source(S)
Substrate
Iinv
VdVg
n-
p+
=0 >0
Reverse-biased diode leakage (bottom)
Mechanisms:Overview 8/43
![Page 9: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/9.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
The Impact of CMOS Scaling on Static Power
I Leakage expected to increase as devices shrink
I Major concerns:
I Gate Oxide Thickness (Tox) scalingI Channel MiniaturizationI Vdd ,Vth scalingI Source-Drain punchthroughI Doping concentration
I Scientists work hard to keep static power manageable
Mechanisms:Technology Scaling 9/43
![Page 10: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/10.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
General Static Power Reduction Techniques
I Device Level: Active devices
I Doping, materials, Vdd ,Vth
I Circuit Level. Gates
I Natural stacks (Fig. A)I Forced stacks (Fig. B)
I System Level: macro blocks,datapaths and co-processors
I Power Gating
a
b
a
b
c
c
c
Vdd Vdd
z1 z2
Vm1
[B][A]
Vm2
M1
M2
M3 M4 M5
M6
M7
Leakage reduction: 10/43
![Page 11: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/11.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Power Gating
I Series Resistance
I Leakage ReductionI Performance Penalty
I Virtual Power Nets
I gvddvI gvssv
LOGIC
sleep
VDD
gvssv LOGIC
VDD
gvddv
sleep
Power Gating:Overview 11/43
![Page 12: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/12.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Power Gating Techniques
I Non-state preserving
I Value of dynamic nodes drift towards power rails
I State preserving
I Retain value of registersI Retain value of dynamic gates
Power Gating:Overview 12/43
![Page 13: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/13.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Non-State Preserving
I Cut-o� (CO)
I Internal Nodes �oat
I Foot transistor(Figure) N-typetransistor
I Head transistorsP-type transistor
I Transient behavior
I Long sleep settle timeI Long wake-up time
_sleep
VDD
gvssv
LOGIC LOGICLOGIC
Power Gating:Non-state preserving 13/43
![Page 14: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/14.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
State Preserving
I Zig-Zag Cut O� (ZZCO)
I Select Head or Foottransistor
I Head →0I Foot →1
I Better transient behaviorvs cut-o�
I Less e�ective for leakagereduction
VDD
gvssv
_sleep
VDD
VDD
sleep
0 1 0 1 0
Power Gating:State preserving 14/43
![Page 15: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/15.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
State Preserving - Sneaky paths
sleep
sleep
1 0 1
VDD
VDD
L2 L1
Power Gating:State preserving 15/43
![Page 16: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/16.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Overview
1. Review leakage mechanisms
2. Review static power reduction techniques
3. Review Power Gating techniques
I Non-state preservingI State preserving
Our Contributions:
1. Power gating in the context of Async circuits
I Non-state preservingI State preserving
2. Evaluate Power Gating in Async circuits
3. ZZDRTO - Pipeline Power Gating
4. Evaluate ZZDRTO
Async Power Gating: 16/43
![Page 17: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/17.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Asynchronous Power Gating Techniques
I Standard techniques work! ... mostly
I Conditions necessary for correct operation
_sleep
VDD
gvssv
LOGIC LOGICLOGIC
VDD
gvssv
_sleep
VDD
VDD
sleep
0 1 0 1 0
Async Power Gating:Overview 17/43
![Page 18: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/18.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Pseudo-Static Gates
PUN
PDN
VDD
z z
PUN
PDN
VDD
VDD
zz
M4
M3
M1
M2
Async Power Gating:Overview 18/43
![Page 19: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/19.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Async Non-State Preserving Power Gating
_sleep
gvssv
VDD
zz
PUN
PDN
VDD
VDD
I Cut-O� (CO)
I Similar approach
I Static CMOSgates
I Pseudo-staticgates
Async Power Gating:Non-state preserving 19/43
![Page 20: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/20.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Async Non-State Preserving Power Gating
I Isolation Circuits
I Wake-up sequence
I De-assert sleepI Exercise reset sequenceI Assert safe
Async Power Gating:Non-state preserving 20/43
![Page 21: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/21.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Async State Preserving Power Gating
I Zig-Zag Cut-O� (ZZCO)
I Same approach
I Static CMOS GatesI Pseudo-static gates
I Note forward inverter
sleep
_sleep
VDD VDD
gvssv
gvddv
PUN
PDN
VDD
1 0
M3
M4
Async Power Gating:State preserving 21/43
![Page 22: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/22.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Async State Preserving Power Gating
I Zig-Zag Cut-O�Weakened Statizicer(ZZCO-WS)
I Weakened Staticizer
I gvddv instead of VddI gvssv instead of GNDI Better power savingsI Better performance
I Low-cost
I No topology changeI Similar wiring costs
sleep
_sleep
VDD VDD
gvssv
gvddv
PUN
PDN
gvddv
gvssv
M3
M4
1 0
Async Power Gating:State preserving 22/43
![Page 23: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/23.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Power Gating Circuit Techniques: Evaluation
_sleep
gvssv
VDD
zz
PUN
PDN
VDD
VDD
CO
sleep
_sleep
VDD VDD
gvssv
gvddv
PUN
PDN
VDD
1 0
M3
M4
ZZCO
sleep
_sleep
VDD VDD
gvssv
gvddv
PUN
PDN
gvddv
gvssv
M3
M4
1 0
ZZCO-WS
Async Power Gating:Evaluation 23/43
![Page 24: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/24.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Power Gating Circuit Techniques: Steady State Evaluation
Pipeline Cluster Transistors FO4
Add Round Key (AK) 8400 2.4
Shift Rows (SR) 7567 2.6
Byte Substitute (BS) 84144 20.4
Mix Column (MC) 30000 16.8
Circuit Transistor Count
Control Circuitry 18000
Counter overhead 4300
Total Number of transistors: 153000
Async Power Gating:Evaluation 24/43
![Page 25: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/25.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Power Gating Circuit Techniques: Steady State Evaluation
I BSIM, T-T, conservative wire cap, 90nm, 298K
I AES pipeline functional blocks
Async Power Gating:Evaluation 25/43
![Page 26: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/26.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Power Gating Circuit Techniques: Evaluation
I CO o�ers the best steady state
I Power savingsI Performance
I ZZCO-WS is better in steady state than ZZCO
I Power savingsI PerformanceI Negligible implementation costs
Async Power Gating:Evaluation 26/43
![Page 27: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/27.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Pipeline Power Gating
I Async: Self throttling circuits
I Control circuitry
I Safe turn o� (Empty pipeline detection)I Correct dynamic operation (isolation circuits)I Quick wake-up
ZDRTO:Overview 27/43
![Page 28: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/28.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Zero-Delay Ripple Turn On (ZDRTO)
ZDRTO:ZDRTO 28/43
![Page 29: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/29.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Zero-Delay Ripple Turn On (ZDRTO)
I Pipeline Cluster = Power Gating Domain
C1 C2 C3 C4
Pipeline Clusters
ZDRTO:ZDRTO 29/43
![Page 30: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/30.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Zero-Delay Ripple Turn On (ZDRTO)
I Pipeline Cluster
I Power Gating Domain
I Choose
I Cluster size depends on the applicationI Power Gating technique for each cluster
C1 C2 C3 C4
Pipeline Clusters
ZDRTO:ZDRTO 30/43
![Page 31: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/31.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Zero-Delay Ripple Turn On (ZDRTO)
I Leverage pipeline stage computation latency
I Hide latency of powering up downstream stages
I Leverage asynchronous circuits robustness
I Do computation during power up
C1 C2 C3 C4
Pipeline Clusters
ZDRTO:ZDRTO 31/43
![Page 32: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/32.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Zero-Delay Ripple Turn On (ZDRTO)
I Leverage pipeline stage computation latency
I Hide latency of powering up downstream stages
I Leverage asynchronous circuits robustness
I Do computation during power up - Domino e�ect turn on
C1 C2 C3 C4
Pipeline Clusters
ZDRTO:ZDRTO 32/43
![Page 33: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/33.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Zero-Delay Ripple Turn On (ZDRTO)
I Leverage pipeline stage computation latency
I Hide latency of powering up downstream stages
I Leverage asynchronous circuits robustness
I Do computation during power up - Domino e�ect turn on
C1 C2 C3 C4
Pipeline Clusters
ZDRTO:ZDRTO 33/43
![Page 34: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/34.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Zero-Delay Ripple Turn On (ZDRTO)
I Leverage pipeline stage computation latency
I Hide latency of powering up downstream stages
I Leverage asynchronous circuits robustness
I Do computation during power up - Domino e�ect turn on
C1 C2 C3 C4
Pipeline Clusters
ZDRTO:ZDRTO 34/43
![Page 35: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/35.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
ZDRTO Evaluation
I Result trade-o�s between
I Wake-up latencyI Power savingsI Operating frequency
I BSIM4, T-T conservative Wire cap, 90nm, 298K
I Example pipeline: 4-cluster AES pipeline
I Di�erent Power Gating techniques
Evaluation:Setup 35/43
![Page 36: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/36.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
ZDRTO Evaluation
Pipeline Cluster Transistors FO4
Add Round Key (AK) 8400 2.4
Shift Rows (SR) 7567 2.6
Byte Substitute (BS) 84144 20.4
Mix Column (MC) 30000 16.8
Circuit Transistor Count
Control Circuitry 18000
Counter overhead 4300
Total Number of transistors: 153000
Evaluation:Setup 36/43
![Page 37: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/37.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
ZDRTO Evaluation
AK SR BS MC
AK SR BS MC
Baseline
CO
ZZCO-WS
ZZDRTOZZCO-WS
Mixed-A
Mixed-B
Evaluation:Setup 37/43
![Page 38: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/38.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
ZDRTO Evaluation
Non-ZDRTO Wake-up(ns) Leakage(µW ) Frequency (Mhz)
Baseline 0 7.1 285
CO 32.9 1.5 262
ZZWS 5.9 6.34 180
ZDRTO Wake-up (ns) Leakage(µW ) Frequency(Mhz)
ZZWS 5.6 6.46 182
Mixed-A 18.4 6.05 226
Mixed-B 26.2 1.62 260
Evaluation:Results 38/43
![Page 39: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/39.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
ZDRTO Evaluation - Wake-up vs Leakage
0 5 10 15 20 25 300
1
2
3
4
5
6
7
Wake−up time (ns)
Leak
age
(µ W
)
Base
ZZWS(ZZDRTO) ZZWS
Mixed−A
Mixed−B CO
Wake−up time vs Leakage for multiple experiments
Evaluation:Results 39/43
![Page 40: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/40.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Conclusions
I Overview static power leakage mechanisms
I Source-Drain leakageI Gate leakageI Analyzed the impact of miniaturization on several currents
I Overview methods to reduce static power
I Device levelI Circuit levelI System level: Maximal gains with Power Gating
Conclusions:Conclusions 40/43
![Page 41: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/41.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Conclusions
I Asynchronous Power Gating techniques
I State preserving (Avg 25% savings)I Non-state preserving (Avg 80% savings)
I Pipeline Power Gating techniques
I Empty pipeline detectionI Zero-Delay Ripple Turn On (ZDRTO)
I Demonstrated trade-o�s
I Wake-up latencyI Power savingsI Operating frequency
Conclusions:Conclusions 41/43
![Page 42: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/42.jpg)
Motivation Mechanisms Leakage reduction Power Gating Async Power Gating ZDRTO Evaluation Conclusions
Acknowledgements
I Prof Manohar, Prof Myers and Prof Suh
I Jonathan Tse
I λ-team: Rob Karmazin and Benjamin Hill
I Async Members: Nabil, Fang, Filipp, Basit, Ilya, Stephen
I CSL faculty and students
I Funding agencies:
Conclusions:Conclusions 42/43
![Page 43: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/43.jpg)
Static Power Reduction for Asynchronous Circuits
Carlos Tadeo Ortega Otero
Asynchronous VLSI Group and Architecture
Computer Systems Laboratory
Cornell University
Ithaca, NY, 14853
May, 25th, 2012
![Page 44: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/44.jpg)
Backup Cut-O� Backup slides Optimizations
Reverse-Biased Diode Leakage
IINV = Is × (e(Vd/Ut) − 1) (1)
I Reverse biased diode current
I Is =Reverse Saturation Current
I Ut= Temperature voltage
IINV = Ad × JINV (2)
Gate(G) Drain(D)Source(S)
Substrate
Iinv
VdVg
n-
p+
=0 >0
Backup:Source-Drain leakage 44/43
![Page 45: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/45.jpg)
Backup Cut-O� Backup slides Optimizations
Source-Drain Leakage
Id ,weak =W
L× Io × e(Vgs−Vth) × (1− eVds(mUT )−1) (3)
I Condition Vg < Vth,|Vd | ≥ 0.1 and Vs = Vb = 0
I Vgs =Voltage Gate-Source - exponential dependence
I Vds =Voltage Drain-Source - linear dependence
I UT =Thermal Voltage
Gate(G) Drain(D)Source(S)
Substrate
VdVg
ID,weak
>=0 >0.1
Backup:Source-Drain leakage 45/43
![Page 46: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/46.jpg)
Backup Cut-O� Backup slides Optimizations
Source-Drain Leakage: Subthreshold Slope-RVT - 90nm
Vg
Ioff
Vd
W=7*Lmin
Vgs
(V)
log(
I d)
Subthreshold Slope n−fet 90nm Rvt
, W=7×Lmin
Id0, V
d=0.2 (Reference)
Vd=0.5
Vd=0.8
Vd=1.2
Backup:Source-Drain leakage 46/43
![Page 47: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/47.jpg)
Backup Cut-O� Backup slides Optimizations
Source-Drain Leakage: Subthreshold Slope-HVT - 90nm
Vg
Ioff
Vd
W=7*Lmin
Vgs
(V)
log(
I d)
Subthreshold Slope n−fet 90nm Hvt
, W=7×Lmin
Id0 (Reference)
Vd=0.2
Vd=0.5
Vd=0.8
Vd=1.2
Backup:Source-Drain leakage 47/43
![Page 48: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/48.jpg)
Backup Cut-O� Backup slides Optimizations
Source-Drain Leakage: Subthreshold Slope-HVT - 65nm
Vg
Ioff
Vd
W=7*Lmin
Vgs
(V)
log(
I d)
Subthreshold Slope n−fet 65nm Hvt
, W=7×Lmin
Id0 (Reference)
Vd=0.2
Vd=0.5
Vd=0.8
Vd=1.2
Backup:Source-Drain leakage 48/43
![Page 49: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/49.jpg)
Backup Cut-O� Backup slides Optimizations
Direct Tunneling
I Dependent on large voltages VgdI Gate to Source-Drain-channel-body
I Electron tunnelingI Hole tunneling
Gate(G) Drain(D)Source(S)
Substrate
Iinv
ID,weak
IGIDL
Backup:Gate Leakage 49/43
![Page 50: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/50.jpg)
Backup Cut-O� Backup slides Optimizations
Hot Carrier Injection
I Dependent on large voltages VgdI Gate-to-Drain tunneling
Gate(G) Drain(D)Source(S)
Substrate
Iinv
ID,weak
IGIDL
Backup:Gate Leakage 50/43
![Page 51: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/51.jpg)
Backup Cut-O� Backup slides Optimizations
Gate-Induced Drain current
I Dependent on large voltages VgdI Gate-to-Drain tunneling
Gate(G) Drain(D)Source(S)
Substrate
V>>0g
IGIDL
Backup:Gate Leakage 51/43
![Page 52: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/52.jpg)
Backup Cut-O� Backup slides Optimizations
CMOS Scaling: Gate Oxide Thickness Scaling
I Le� = 45× Tox(Intel)
I What happens at 100nm?
I TOX ∼ 12Amstrongs to16Amstrongs
I GIDL. Imposes a limit on TOX as Electric �eld increasessigni�cantly
I GIDL. Less relevant as voltages reduce below the energy bandgap of the Silicon
I Gate Direct Tunneling.
Backup:Technology Scaling 52/43
![Page 53: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/53.jpg)
Backup Cut-O� Backup slides Optimizations
Channel Miniaturization
I Add here
Backup:Technology Scaling 53/43
![Page 54: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/54.jpg)
Backup Cut-O� Backup slides Optimizations
Source-Drain punchthourgh
I Not a concern f
Backup:Technology Scaling 54/43
![Page 55: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/55.jpg)
Backup Cut-O� Backup slides Optimizations
Empty Pipeline DetectionI Detect when it is safe to power gate
I Constant reponse time counter
Backup:Control 55/43
![Page 56: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/56.jpg)
Backup Cut-O� Backup slides Optimizations
Empty Pipeline DetectionI Interleaved counter allows full throughput operation
I Minimum overhead
Backup:Control 56/43
![Page 57: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/57.jpg)
Backup Cut-O� Backup slides Optimizations
Cut-o� techniques
I Cut-o�. Both logic and sleep use RVT devices
I MTCMOS: Logic implemented using low/regular VT whilesleep transistors -> high-VT devices
I BGCMOS: Boosted Gate: high-Vt thick oxide sleep transistors-> hurt performance -> overdrive Vdd during active mode
I Super Cut-O�: Gate of sleep transistors driven past supplyvoltages when idle
I Problems: Foundry support and biasing
Cut-O� Backup slides: 57/43
![Page 58: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/58.jpg)
Backup Cut-O� Backup slides Optimizations
Cut-o� transient behaviour
0 50 100 150 20010
−2
10−1
100
101
102
103
Curr
ent (µ
A)
(a) Supply Current
0 50 100 150 2000
0.2
0.4
0.6
0.8
1
1.2
1.4
Voltage (
V)
Time (µs)
(b) Virtual Ground and Enable Signal
vssv in.e
Cut-O� Backup slides: 58/43
![Page 59: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/59.jpg)
Backup Cut-O� Backup slides Optimizations
Self-throttling Asynchronous Circuits
I Stability: G->t is stable when G can change from true to falseonly in states where R(t) holds.
I State wont change unless it is acknowledged
I Problems: Noise margins are reduced.
I You should wait to assert �safe� until enough noise margin exist
I Assumption: A monotonic change on input will create amonotonic change on output
I Noise margins again. Charge Sharing and Capacitive Couplingcan break this assumption
Cut-O� Backup slides: 59/43
![Page 60: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/60.jpg)
Backup Cut-O� Backup slides Optimizations
Problems and Future Research
I Stability: G->t is stable when G can change from true to falseonly in states where R(t) holds.
I State wont change unless it is acknowledged
I Problems: Noise margins get diminished
I If there is not enough �sleep� time, then you can end-up usingmore power
I We should compute break-even point
I We need some tools to power gate in a systematic andprovably correct way
Cut-O� Backup slides: 60/43
![Page 61: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/61.jpg)
Backup Cut-O� Backup slides Optimizations
Gate Leakage
I Current that dribbles through the gate of the transistors
I Direct TunnelingI Hot carrier injection
I Increasing component of the absolute leakage currents
Cut-O� Backup slides:Gate Leakage 61/43
![Page 62: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/62.jpg)
Backup Cut-O� Backup slides Optimizations
Device Level
I Device Engineering. Transistors and devices
I Le� ,Tox , Substrate depth, nominal values of Vdd and Vth
I Choice of materials (semiconductor, metal, dielectric)
I Doping pro�le and doping halo
I Requires expertise on device physics
I Choice of fabrication process
Optimizations:Device Level 62/43
![Page 63: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/63.jpg)
Backup Cut-O� Backup slides Optimizations
Circuit Level
I Natural transistorstacking
I Forced stacking
a
b
a
b
c
c
c
Vdd Vdd
z1 z2
Vm1
[B][A]
Vm2
M1
M2
M3 M4 M5
M6
M7
Optimizations:Circuit level 63/43
![Page 64: Static Power Reduction for Asynchronous Circuitsvlsi.cornell.edu/~cto3/pdf/ortega_ms_presentation.pdf · Asynchronous VLSI Group and Architecture Computer Systems Laboratory Cornell](https://reader034.vdocuments.mx/reader034/viewer/2022050418/5f8e1f56de637040920c442c/html5/thumbnails/64.jpg)
Backup Cut-O� Backup slides Optimizations
System Level
I Design speci�c
I SRAM topologyI Amount of dark silicon
I General techniques
I Power gating
Optimizations:System level 64/43