Static Techniques to Improve Power Efficiency of Branch Predictors

Abstract

Current power-efficient designs focus on reducing the dynamic (activity-based) power consumption in a processor through different techniques. In this paper, we illustrate the application of two static techniques to reduce the activities of the branch predictor in a processor leading to its significant power reduction. We introduce the use of a static branch target buffer (BTB) that achieves the similar performance to the traditional branch target buffer but eliminates most of the state updates thus reducing the power consumption of the BTB significantly. We also introduce a correlation-based static prediction scheme into a dynamic branch predictor so that those branches that can be predicted statically or can be correlated to the previous ones will not go through normal prediction algorithm. This reduces the activities and conflicts in the branch history table (BHT). With these optimizations, the activities and conflicts of the BTB and BHT are reduced significantly and we are able to achieve a significant reduction (43.9% on average) in power consumption of the BPU without degradation in the performance.