A Study on Deep Belief Net for Branch Prediction

Abstract

Since 2006, there have been significant advances in deep learning algorithms, and they have shown superior performance in audio and image processing. In this paper, we explore the feasibility of applying deep learning algorithms to branch prediction. We treat branch prediction as a classification problem and compare the effectiveness of deep learning with existing branch predictors. We make several interesting observations from our study. The first is that for branch prediction, the deep learning algorithm based on deep belief networks outperforms the prior work, but only outperforms state-of-the-art branch predictors, such as the TAgged GEometric length (TAGE) predictors, for several benchmarks. Compared with the much simpler perceptron branch classifier, the deep learning classifier reduces the average misprediction rate by 3%-4% for the benchmarks in this paper. Second, we analyze the impact of the length of hashed program counter, local history register, global history register, and branch global addresses of deep learning classifiers on the misprediction rate. Our results show that an adaptive length of the history information is a better choice than the longest history. Third, compared with TAGE, the hardware budget of our model is less than 1% of the TAGE predictor.