Bias-Free Neural Predictor

Prior research in neurally-inspired perceptron predictors have shown significant improvements in branch prediction accuracy by exploiting correlations in long branch histories. However, systems with moderate hardware budgets typically restrict such perceptron predictors from correlating beyond 64 to 128 past branches and limit their capability to learn distant branch correlations, such as on the order of 1024 to 2048 branches deep. In this work, we propose Bias-Free Neural predictor that is structured to learn correlations only with prior non-biased conditional branches, aka. branches whose dynamic behavior varies during a program’s execution. This, combined with a recency-stack-like management policy for the global history register, opens up the opportunity for a modest history length to include much older and much richer context to predict future branches more accurately. Bias-Free Neural predictor achieves 2.73 MPKI for a 32KB storage budget and 2.1 MPKI for an unlimited budget.