Width Prediction for Reducing Value Predictor Size and Power

Value prediction has been proposed for breaking datadependencies and increasing instruction level parallelism. One of the drawbacks of many of the proposed techniques is that the value predictors require very large hardware structures which use up many transistors and can consume a large amount of energy. In this study, we use data-widths to partition the value prediction table into several smaller structures, and then use datawidth prediction to select from these multiple tables. This width-partitioned value predictor requires less space because small bit-width values get allocated to smaller storage locations instead of using a full 64 bits. The total energy consumption of the predictor is also reduced because only a single smaller predictor table needs to be accessed. An 8KB width-partitioned last value predictor achieves nearly the same load-value prediction rates as a conventional 16KB last value predictor, while simultaneously consuming 41.1% less energy. A width partitioned finite context matching (FCM) predictor achieves the same prediction accuracy as a regular FCM predictor with twoto four-times the number of second-level table