A Study On The Neural-Based Percetron Branch Predictor and Its Behavior

Abstract

Branch predictors are very critical in modern superscalar processors and are responsible for achieving high performance. As the depth of pipeline and instruction issue rate of high-performance superscalar processors increase, a branch predictor with high accuracy becomes indispensable

In recent times, neural based branch predictors, like perceptron predictor, are found to have higher accuracy than other popular two-level branch predictors. One major advantage of perceptron predictors over the two-level schemes is that we can have longer global or local history length, and consequently the perceptron predictor is robust to aliasing, resulting in better prediction accuracy.

In this thesis, the behavior and the intricacies of the perceptron predictor are extensively studied. The perceptron predictor has outperformed the classic Gshare predictor with lesser hardware resource. For a memory size of 64KB, the perceptron branch predictor has prediction accuracy about 2-10% higher than that of Gshare. The advantage of having longer history lengths was exploited to determine the performance and the IPC values for the perceptron predictor and showed commendable results. Also, varying the training parameter and the number of perceptrons for prediction helped in analyzing the behavior of the perceptron predictor under different environments.