TinyBERT for branch prediction in modern microprocessors

Recent progress has highlighted the crucial importance of branch prediction (BP) in optimizing computer performance, especially in reducing computational delays by preventing stalls in modern microprocessors (also known as CPUs). In this paper, we investigate the use of machine learning (ML) models to improve BP accuracy, focusing on the capabilities of transformer models based on their exceptional predictive and classification performance. Although existing studies have employed various ML methods for BP, their selected models are computationally expensive and impractical for such task. Hence, we present an advanced ML-based dynamic BP technique utilizing tiny bidirectional encoder representations from transformers (TinyBERT), notable for its efficiency, simplicity, and low resource utilization. This method not only streamlines the BP process but also offers a more effective alternative to conventional strategies. A key aspect of our approach is the application of local post hoc explanations, which provide deep insights into the model’s predictive actions. Our empirical findings reveal that this methodology secures a substantial 13% reduction in the rate of mispredictions compared to top predictors like TAGE-SC-L, across various multimedia and integer application benchmarks. These results underscore the potential of using compact transformers in establishing significant criteria for efficient and effective BP.