Transcript Slide 0

DYNAMIC SYNCHRONIZER FLIP-FLOP PERFORMANCE IN
FINFET TECHNOLOGIES
NOCS 2014
Mark Buckler1, Arpan Vaidya2, Xiaobin Liu2, Wayne Burleson2,3
Cornell University1, University of Mass. Amherst2, AMD Research3
September 18, 2014
WHY ARE WE LOOKING AT SYNCHRONIZERS?
 Low power NoCs utilize
fine grain DVFS
 Leads to Inter ClockDomain Communication
• Requires synchronizers!
• Error rates similar to
that of transient faults
Core 0
Core 1
Core 2
Cache Router
Cache Router
Cache Router
Core 3
Core 4
Core 5
Cache Router
Cache Router
Cache Router
Core 6
Core 7
Core 8
Cache Router
Cache Router
Cache Router
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 2 of 35
PROBLEM: METASTABILITY
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 3 of 35
THE BRUTE-FORCE SYNCHRONIZER
 Synchronizer reliability is quantified as the
Mean Time Between Failures (MTBF)
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 4 of 35
SYNCHRONIZER FLIP-FLOP EVALUATION
 τ is the strongest circuit level term in
calculating MTBF
 Goal 1: Provide NoC designers with τ
simulations over design parameters
 Goal 2: Find the best Flip-Flops, and ensure
that they can be used by NoC designers
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 5 of 35
SYNCHRONIZER FLIP-FLOP EVALUATION
 Flip-Flops:
• PowerPC Flip-Flop
• Dynamic Latch Flip-Flop
• Psuedo-NMOS Flip-Flop
 Technology Nodes
• Planar (50nm and 22nm)
• FinFET (20nm, 10nm and 7nm)
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 6 of 35
SIMULATION TECHNIQUE
Zhou and Ashouei et al, Microelectronics Journal, 2011
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 7 of 35
EFFECT OF TECHNOLOGY CHANGE
DLFF
Power PC
PseudoNMOS
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 8 of 35
VOLTAGE AND TEMPERATURE SENSITIVITY
PseudoNMOS in 22nm
DLFF in 22nm
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 9 of 35
VOLTAGE AND TEMPERATURE SENSITIVITY
PseudoNMOS in 20nm
DLFF in 20nm
 FinFETs are more sensitive to changes in
threshold voltage from lower temperatures
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 10 of 35
EFFECT OF FORWARD-BODY BIASING
22nm (Planar)
20nm (FinFET)
 DLFF is best for high voltage, but we may
have a problem...
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 11 of 35
THE DYNAMIC LATCH FLIP-FLOP
 A reset signal is necessary…
Dike and Burton, IEEE SSC 1999
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 12 of 35
THE DYNAMIC SYNCHRONIZER
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 13 of 35
THE DYNAMIC SYNCHRONIZER
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 14 of 35
THE DYNAMIC SYNCHRONIZER
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 15 of 35
BENEFITS OF THE DYNAMIC SYNCHRONIZER
 The Dynamic Synchronizer allows DLFFs to
be used in typical synchronous systems
 Improves the latency-reliability tradeoff!
• MTBF of 5,000 yrs over 5 yrs (same # of stages)
or
• Latency of 2 cycles over 3 cycles (same MTBF)
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 16 of 35
CONCLUSIONS
 Evaluated synchronizer τ values
• τ continues to track with FO4 in FinFET
• FinFET increases sensitivity to temperature
• Body biasing still works for high voltage
 The Dynamic Synchronizer
• We improved the latency-reliability tradeoff by
enabling the use of low τ flip-flops
Motivation • Flip-Flop Evaluation • The Dynamic Synchronizer • Conclusions
Page 17 of 35