Fast Path-Based Neural Branch Prediction Daniel A. Jimenez Presented by: Ioana Burcea
Outline Research motivation Neuran prediction: perceptron prediction Staggered algorithm Experiments and results
Research motivation Branch prediction –Accuracy –Latency Neural learning predictors –Most accurate –High latency => can we do any better?
Branch Prediction with Perceptrons Global history shift register that stores outcomes of branches –History length h Perceptron predictor –Weight matrix: n x (h + 1) weights Every row stores a vector of h + 1 weights W 0 is often called the bias weight
Prediction
Update
Staggered Algorithm
Prediction
Update
Experiments 17 integer benchmarks (SPEC2000 & SPEC95)
Simulated Predictors 2Bc-gskew with two level overriding –Hybrid predictor One bimodal and 2 gshare predictors Perceptron predictor Gshare.fast Fixed length path predictor Path-based neural predictor
Tuned history lenghts
Estimated latencies
Average misprediction rates
Average IPC
Misprediction rates at 8KB hw