CMA-BP: A Clustered Multi-Task Learning and Branch Attention Based Branch Predictor

Branch prediction stands as a key bottleneck in enhancing CPU performance, particularly evidenced by an average of around 10 mispredicted hard-to-predict(H2P) branches per benchmark in SPEC 2017 by current neural network methods. To improve, this paper proposed a Clustered Multi-task Learning and Branch Attention Mechanism-Based Branch Predictor (CMA-BP). Clustered multi-task learning enhances model generalization, and branch attention extracts preferences of different branches for global history. Thus, CMA-BP efficiently aggregates branches with similar features, reducing training complexity. Experimental results show that CMA-BP outperforms existing predictors in accuracy significantly and in the number of parameters required. By advancing the state-of-the-art in branch prediction, this work has important implications for future high-performance computer architecture design