LAFI: Look-Ahead Mechanism for Energy-Efficient Branch Prediction

To feed the execution core with uninterrupted instruction streams, modern processors access complex branch predictors every clock cycle, incurring high energy consumption. In this paper, we conduct a systematic study on the energy consumption of representative branch predictors, and identify several sources of energy inefficiency. Based on this study, we introduce an improved branch predictor design, called Looking Ahead and FIltering mechanism (LAFI), to reduce branch predictor energy consumption. LAFI employs two simple look-ahead structures to capture most recently executed branches, identify well-behaved branches and detect BTB misses one cycle prior to branch prediction; therefore, it avoids most unnecessary accesses to branch predictors. Experimental results show that LAFI reduces energy consumption in the branch predictor by 48.8% on average. Compared to previous techniques, LAFI saves the predictor energy more effectively and incurs minimal performance overhead.