Rollback-Free Value Prediction with Approximate LoadsBradley Thwaites, Gennady Pekhimenko, Amir Yazdanbakhsh, Jongse Park, Girish Mururu

Hadi Esmaeilzadeh, Onur Mutlu, Todd C. MowryRFVP Overview

Design Principles Design Challenges

Key Experimental Results Ongoing Work

Motivation

Mitigate long latency memory accesses Microarchitecturally-triggered approximationPredict the value of an approximate load when it misses in the cacheDo not check for mispredictionsDo not rollback from mispredictions

swim

fma3d

bwavesmcf

cactu

sADM

soplex

GemsFDTD

geomean1.0

2.0

3.0

4.0

5.0

6.0

Perf

orm

ance

Ben

e-fit

Perfect Prediction

Maximize opportunities for performance and energy benefits

Minimize the adverse effects of approximation on quality degradation

(0%, 0%, 0%) (1.06x, 1.03x, 2%)

(1.13x, 1.06x, 6%) (1.27x, 1.10x, 11%)

Extend rollback-free value prediction to GPUsDrop a fraction of the missed requestsPreliminary results: Up to 2x improvement in energy and performancewith only 10% quality degradation

Mitigate both Memory Wall and Bandwidth Wall

swim

fma3d

bwaves

mcf

cactu

sADM

soplex

GemsFDTD

geomean

1.00

1.05

1.10

1.15

1.20

1.25

Perf

orm

ance

Ben

efit

2MB LLC, 4-Wide, Performance Results

Target Performance-Critical Safe LoadsProfile-directed compilationUsually, < 32 loads cause 80% of cache misses

Utilize Fast-Learning PredictorsTwo-delta stride predictorPrediction: table lookup plus an addition

Integrate RFVP with existing architecture

swim fma3d bwaves mcf cactusADM soplex GemsFDTD0%

20%

40%

60%

80%

100%

17.7% 15.4%

2.8% 0.0%

18.3% 19.5%

2.1%0.2% 1.6% 1.8% 0.0% 0.3% 1.8% 0.0%

Value Prediction - Quality Loss

Stride

TwoDelta

Qua

lity

Loss

(Performance Gain, Energy Gain, Quality Loss)