Rollback-Free Value Prediction
with Approximate Loads
Bradley Thwaites
Gennady Pekhimenko
Amir Yazdanbakhsh
Jongse Park
Girish Mururu
Hadi Esmaeilzadeh
Onur Mutlu
Georgia Institute of Technology
Carnegie Mellon University
Todd Mowry
Mitigating Memory Wall with
Approximation
Rollback-Free Value Prediction
– Microarchitecturally-triggered approximation
– Predict the value of an approximate load when it
misses in the cache
– Do not check for mispredictions
– Do not rollback from mispredictions
Mitigate long latency memory accesses
Rollback Free Value Prediction
Design Principles
Maximize opportunities for performance
and energy benefits
Minimize the adverse effects of
approximation on quality degradation
Design Challenges and Solutions
Target Performance-Critical Safe
Loads
– Profile-directed compilation
– Usually, < 32 loads cause 80% of
cache misses
Utilize Fast-Learning Predictors
– Two-delta stride predictor
– Prediction: table lookup plus an
addition
Integrate RFVP with existing
architecture
Experimental Results with a Modern
OoO Processor
2MB LLC, 4-Wide, Performance Results
Two-Delta Value Prediction - Quality Loss
1.25
100%
1.20
80%
Quality Loss
Performance Benefit
90%
1.15
1.10
70%
60%
50%
40%
30%
1.05
20%
10%
1.00
0%
8%
Average
19%
Maximum
0.8%
Average
1.8%
Maximum
More CPU configurations and value predictors are in the paper
Ongoing Work
Mitigate both Memory Wall and Bandwidth Wall
• Extend rollback-free value prediction to GPUs
• Drop a fraction of the missed requests
• Preliminary results: Up to 2x improvement in
energy and performance with only 10%
quality degradation