Evaluation of Subthreshold Design Techniques

Download Report

Transcript Evaluation of Subthreshold Design Techniques

Circuit Design Techniques for Low
Power DSPs
Simone Gambini
Marghoob Mohiyuddin
Melinda Ler
Motivation
 Energy Per Operation (EOP) important
 For energy-constrained systems, e.g., battery-powered
devices
 Supply voltage scaling can be used to reduce
energy consumption
 Leakage limits scaling to above a certain supply voltage
 Conventional techniques for low power/energy design
may not be beneficial
 Low power designs also have performance
constraints (apart from power)
 Design should meet throughput constraints while
minimizing energy consumption
 Want to explore tradeoffs in design given these
constraints
7/17/2015
Free Template from www.brainybetty.com
2
Problem Statement
 Study the impact of low power design
techniques for different circuits with
performance constraints




Effects of process variations and temperature
Supply voltage scaling
Using parallelism to reduce power
Architectural approaches
 Multiplier designs used as case studies
 Fundamental block in many DSP systems
7/17/2015
Free Template from www.brainybetty.com
3
Prior Work
 If energy per operation (EOP) used as an
optimization metric then an optimal choice
of Vdd exists [1]
 Technology, micro-architecture and architecture
affect the EOP
 Minimum EOP point at EOP optimal Vdd
shown to shift with different microarchitectures [2]
 Performance constraint not taken into account
7/17/2015
Free Template from www.brainybetty.com
4
Project Outline
 Technology characterization
 Simulated leakage and delay for 90nm technology
node across process corners and temperature
 Modeling EOP
 Using simulated delay and leakage data for FO1 and
FO4 ring oscillators
 Extrapolated to get predictions for EOP behavior for
different micro-architectures
 Characterizing test circuits to validate our
predictions
 Delay validation
 Power validation
7/17/2015
Free Template from www.brainybetty.com
5
Technology Characterization: Process
and Temperature Variation Effects
 EOP Model
EOP
   V dd1  V th




n

U
T
2
C  K  L  e
 C tot    V dd
L
fit
d


 Variations in EOP behavior for SVT and LVT across corners
 Variations of minimal EOP with temperature
7/17/2015
Free Template from www.brainybetty.com
6
Technology Characterization:
Process Options
 Tradeoff exists between LVT and HVT process options for
different operating frequencies with intermediate switching
activities
7/17/2015
Free Template from www.brainybetty.com
7
Parallelism and Pipelining
 Pipelining
 For constant throughput,
pipelining allows for lower
energy per operation at
lower supply voltages
 Parallelism
 Due to overhead in
hardware and increase in
logic delay, parallel
structures increases
minimal energy per
operation
7/17/2015
Free Template from www.brainybetty.com
8
Case Study: Multipliers
 Investigated multiple architectures
 Wallace tree multiplier
 Array multiplier
 Serial multiplier
 Impact of parallelism on EOP behavior
 Technology characterization predicts that parallelism
increases EOP for low supply voltages
• Leakage becomes the dominating factor at low supply
 Higher leakage factor means optimal Vdd should
increase with parallelism
7/17/2015
Free Template from www.brainybetty.com
9
Flow for EOP estimation
 Circuit synthesis (Module Compiler)
 Extraction of activity factor (ModelSim)
 Correlated input vectors generated with Matlab
 Gates annotated with activity factors and
capacitances
 Delay simulation for critical paths (Spectre)
 Over multiple Vdds
 Power estimation (PowerPrime)
 Dynamic and leakage power for single supply voltage
 Power scaling with Vdd extrapolated using scaling
factors from FO4 inverter chain
7/17/2015
Free Template from www.brainybetty.com
10
Simulated EOPs for Multipliers
 Parallelism yields
energy benefits only
above a certain Vdd
 Tradeoff between
different
architectures at
different bitwidths
 Wallace tree is
better than carry
save for higher
bitwidths
 Architectural
decisions affect EOP
strongly
7/17/2015
Free Template from www.brainybetty.com
11
Delay Validation
 Simulated critical path delays vs.
extrapolated delays from inverter chain
 Delay scales proportionally to inverter delay
7/17/2015
Free Template from www.brainybetty.com
12
Leakage Validation
10
 A large variation in the
leakage power over
input vectors
 Total leakage power
depends on the
statistical distribution
of input vectors over
time
 Leakage does not scale
the same way as
inverter leakage
7/17/2015
Normalized Leakage
16b Adder Critical Path
64b Adder Critical Path
Inverter
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Vdd
Normalized Leakage for Different Input Vectors
1
0.9
0.8
0.7
Normalized Leakage
 Leakage power for the
critical paths for a
specific input vector
 Leakage power for a
NAND4 gate with
different input vectors
0
0.6
0.5
0.4
0.3
0000
0001
0011
0111
1111
Inverter
0.2
0.1
0
0.1
0.2
0.3
Free Template from www.brainybetty.com
0.4
0.5
0.6
Vdd
0.7
0.8
0.9
1
13
Leakage Validation
Leakage Current
10
10
10
10
-7
0000
0001
0011
0111
1111
-8
-9
-10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Vdd
-7
10
Weighted Leakage Current
 Top figure shows
leakage current
for different
input vectors
 Bottom figure
shows leakage
current for
different input
vectors weighted
by the
probability of
the input vectors
1111
0111
0011
0001
0000
-8
10
-9
10
-10
10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Vdd
7/17/2015
Free Template from www.brainybetty.com
14
Leakage Validation
 Sensitivity of the optimal Vdd point with
respect to the leakage energy
 For sub-threshold operation, the optimal Vdd is not
very sensitive to leakage energy [2]
7/17/2015
Free Template from www.brainybetty.com
15
Conclusions & Future Work
 For low power systems operating below a
certain operating frequency, parallelism
would not be the ideal option due to leakage
 This trend is expected to be reinforced in the future
technology nodes
 Fast and accurate leakage estimation tools
needed
 Should take into account the state-dependent
behavior of leakage
7/17/2015
Free Template from www.brainybetty.com
16
References
[1] D. Markovic, V. Stojanovic, B. Nikolic, M. A. Horowitz, and R. W.
Brodersen, “Methods for True Energy-Performance Optimization,”
IEEE Journal of Solid-State Circuits, vol. 39, pp. 1282–1293, Aug. 2004
[2] B. H. Calhoun and A. Chandrakasan, “Characterization and Modeling
of Minimum Energy Per Operation Point,” in Proc. IEEE International
Symposium on Low Power Electronics and Design, Newport Beach,
California, Aug. 2004, pp. 90–95.
[3] A. Wang and A. Chandrakasan, “A 180mV FFT Processor Using
Subthreshold Circuit Techniques,” in Proc. IEEE International SolidState Circuits Conference, Feb. 2004.
[4] Y. Taur and T. H. Ning, Fundamentals of Modern VLSI Devices.
Cambridge University Press, 1999.
7/17/2015
Free Template from www.brainybetty.com
17