islped 2003 power efficient comparators for long arguments in superscalar processors *supported in...
TRANSCRIPT
ISLPED 2003
Power Efficient Comparators for Long Arguments in Superscalar Processors
*supported in part by DARPA through the PAC-C program and NSF
Dmitry Ponomarev, Gurhan Kucuk, Oguz Ergin, Kanad Ghose Department of Computer Science
State University of New YorkBinghamton, NY 13902-6000
http://www.cs.binghamton.edu/~lowpower
International Symposium on Low Power Electronics and Design (ISLPED’03), August 27th 2003
ISLPED 2003
Outline
Motivations8-bit comparator designs
Traditional ComparatorDissipate on Match Comparator (ICCD’02)
32-bit Comparator DesignsResults : Application to the Load-Store QueueConclusions
ISLPED 2003
Motivation
Equality comparators are very pervasive in today’s superscalar datapaths.
Wake-up logic of the Issue Queues
Dependency checking logic
Load-Store queues
Translation Lookaside Buffers (TLB)
Caches
Branch Target Buffers (BTB)
ISLPED 2003
Motivation (continued)
Traditional comparators dissipate energy on mismatches in any bit position of the arguments
In many cases, mismatches are much more frequent than matches
Issue queue : Only 3% of all comparisons result in a match (Ergin et.al. , ICCD’02)
For energy efficiency, dissipate-on-match designs can be considered
ISLPED 2003
Use of the Long Comparators
Load-Store queues – to allow loads to bypass earlier stores
TLBs – for associative lookup
Caches – for associative lookup
BTBs – for associative lookup
ISLPED 2003
Comparison of larger operands : Some Alternatives
~270 ps ~270 ps
~270 ps >400 ps
>400 ps
ISLPED 2003
Choosing the Right Alternative
Need to consider impact on :Delay
Energy
Have to look at distributions of bit values that are compared
More than one alternative may be acceptable
ISLPED 2003
Experimental Setup (AccuPower, DATE’02)
CompiledSPEC
benchmarks
Datapathspecs
Performance stats
VLSI layoutdata
SPICEdeck
SPICE
MicroarchitecturalSimulator
Energy/PowerEstimator
Power/energystats
SPICE measures ofEnergy per transition
Transition counts,Context information
ISLPED 2003
0.4%
32.5%
22.4%23.7% 21.0%
8.2%
60.5%
8.7%
14.0%
8.6%
68.1%
2.8%1.9%4.7%
22.6%
70.1%
4.1%1.7%0.7%
23.5%
0%
10%
20%
30%
40%
50%
60%
70%
80%
0 1 2 3 4
Number of matching bit pairs in each of the 4 8-bit groups
group of bits 0-7
group of bits 8-15
group of bits 16-23
group of bits 24-31
Matching Statistics of 32-bit Addresses in the LSQ
ISLPED 2003
Matching Statistics of 32-bit Addresses in the LSQ
0.4%
32.5%
22.4%23.7% 21.0%
8.2%
60.5%
8.7%
14.0%
8.6%
68.1%
2.8%1.9%4.7%
22.6%
70.1%
4.1%1.7%0.7%
23.5%
0%
10%
20%
30%
40%
50%
60%
70%
80%
0 1 2 3 4
Number of matching bit pairs in each of the 4 8-bit groups
group of bits 0-7
group of bits 8-15
group of bits 16-23
group of bits 24-31
ISLPED 2003
Main Results
We discussed some energy-efficient 32-bit
wide comparator designs
19% comparator-related energy reduction in
LSQ is achieved by using a hybrid design (3
TRADs + 1 DMC) compared to the use of 4
TRAD comparators
Results can be extended to TLBs and BTBs
ISLPED 2003
THANK YOU !
*supported in part by DARPA through the PAC-C program and NSF
LOW POWER RESEARCH GROUP Department of Computer Science
State University of New YorkBinghamton, NY 13902-6000
http://www.cs.binghamton.edu/~lowpower
International Symposium on Low Power Electronics and Design (ISLPED’03), August 27th 2003
ISLPED 2003
32-bit Matching Statistics: LSQ
0%
10%
20%
30%
40%
50%
60%
70%
80%
0 1 2 3 4 5 6 7 8
group of least significant 8 bits
group of bits 8-15
group of bits 16-23
group of most significant 8 bits