Optimized Design of Branch Predictors Based on Hummingbird E203

Abstract

As the data processed by modern processors becomes increasingly complex, enhancing the computational capabilities of processors has become an unavoidable challenge. In this context, the optimization of branch predictors becomes important. As an open-source and low-power RISC-V processor core, The Hummingbird E203 has demonstrated great potential in the fields of education, academic research, and IoT. However, limited by the lack of prediction accuracy of the static branch predictor, the performance of the Hummingbird E203 also suffers a bit. In order to improve the branch prediction accuracy, this paper optimizes the original branch predictor under the premise of controlling the hardware overhead, and implements four dynamic sub-predictors to improve the performance of Hummingbird E203. The experimental results show that the highest prediction accuracy among the four predictors is the hybrid predictor consisting of the Gshare branch predictor and the local history-based branch predictor. In the test results of Dhrystone and CoreMark, the average prediction accuracy of the hybrid branch predictor has increased by 15.5%, and the score has increased by 1.76%.