A conditional branch predictor based on weightless neural networks

Abstract

Conditional branch prediction allows the speculative fetching and execution of instructions before knowing the direction of conditional statements. As in other areas, machine learning techniques are a promising approach to building branch predictors, e.g., the Perceptron predictor. However, those traditional solutions demand large input sizes, which impose a considerable area overhead. We propose a conditional branch predictor based on the WiSARD (Wilkie, Stoneham, and Aleksander’s Recognition Device) weightless neural network model. The WiSARD-based predictor implements one-shot online training designed to address branch prediction as a binary classification problem. We compare the WiSARD-based predictor with two state-of-the-art predictors: TAGE-SC-L (TAgged GEometric-Statistical Corrector-Loop) and the Multiperspective Perceptron. Our experimental evaluation shows that our proposed predictor, with a smaller input size, outperforms the perceptron-based predictor by about 0.09% and achieves similar accuracy to that of TAGE-SC-L. In addition, we perform an experimental sensitivity analysis to find the best predictor for each dataset, and based on these results, we designed new specialized predictors using a particular parameter composition for each dataset. The results show that the specialized WiSARD-based predictor outperforms the state-of-the-art by more than 2.3% in the best case. Furthermore, through the implementation of specialized predictor classifiers, we discovered that utilizing 90% of the specialized predictor for a specific dataset yielded comparable performance to the corresponding specialized predictor.