Investigating hierarchical temporal memory networks applied to dynamic branch prediction

Abstract

As a consequence of the high application of instruction-level parallelism techniques in modern processors, the branch prediction are a of study remains relevant after 40 years of research. This work applies neural networks based on the Hierarchical Temporal Memory (HTM) theory to the branch prediction task and explores their adequacy to the problem’s characteristics. More specifically, the problem is faced asa sequence prediction task and tackled by the HTM sequence memory. Four traditional branch prediction schemes adapted to operate with an HTM system and two variations of the previous designs were evaluated on a slice of the traces provided by the 4th Championship Branch Prediction contest. The leading result was achieved by the HTM predictor based on the g share branch predictor, that for 8 million instructions was able to improve them is prediction rate by 14.3% incomparison to it straditiona l2-bitcounters version when both used a 13-bithi storyl ength. However, high level so faliasing were found to prevent the HTM system to scale and compete again stlarger conventional branch predictors.