Transcript Slides in ppt

Optimizing Static Power Dissipation By
Functional Units in Superscalar Processors
Siddharth Rele, Santosh Pande, Soner Onder, and Rajiv Gupta
Motivation and Research Goals
●
●
●
ILP distributed non-uniformly throughout a
program.
Functional Units idle for prolonged periods of
time during program execution.
Dynamic power vs static power
Solving The Static Power Problem
●
Dual Threshold Voltage
●
Power Gating
-Turn Off Devices by Cutting Off the
Supply Voltage
●
Turning On Latency
Power Gating
●
Using hardware to identify idle regions
-Additional Power Requirements
-Warmup Time
●
The Compiler Approach
-Identify Suitable Regions
-Determine the number and type of the
Functional Units to be Turned Off
Issues
●
●
●
Tolerating the Latency of Turning a Functional
Unit Off
Tolerating the Latency of Turning a Functional
Unit On
Dealing with Variable Length Idle Regions
Architectural Support
●
Power Aware Instruction Set
-on and off directives as instruction suffixes
●
On and Off Semantics for an OOO Superscalar
Processor
-processor-deadlock and pending-off stage
●
Nullifying Spurious Off-On Pairs
Example
mul.off
if(x>10){
wait=0;
while(1){
wait++;
if(wait==1000) break;
}}
mul.on
Compiler Support
●
Identify Hot Blocks and corresponding
Functional Unit Requirements
-Access to ILP required when multiple
Functional Units of the same type
●
Computing the PAFG
-identifying longer periods of time over
which a functional unit can be turned off
●
Place the off and on directives at right places
Experimental Methodology
●
Compiler-lcc
●
Code Generator-lburg
●
Cycle-level Simulator-FAST system
●
Benchmarks:
-MediaBench(rawcaudio.c ,rawdaudio.c)
-DSPstones(fir2dim.c,n-real-updates.c, fir.c)
-SPEC95(compress.c)
Experimental Results
●
Utlization improved upto 90%
●
Performance Degradation less than 1%
●
●
Average Duration for which Units turned off quite
long-several hundred to thousands cycles
Nullification Strategy quite effective for certain
benchmarks.