SVMs for Improved Branch Prediction

This technical report is a preliminary investigation into the use of Support Vector Machines (SVMs) as a method of branch prediction. We present a new dynamic branch predictor based on SVMs. The SVM predictor, at the cost of a much larger hardware budget, can return a greater accuracy than current state-of-the-art predictors by exploiting its ability to learn linearly inseparable boolean functions, a limitation of many well-known dynamic branch predictors. Our untuned SVM predictor yields a 24% improvement over the best available dynamic neuralmethod branch predictor and a 16% improvement over gshare on the SPEC95 go benchmark at a cost of 10 MB. Tuning the SVM parameters would likely result in further performance gains. Branch prediction with SVMs is largely unexplored in the literature. These favorable results suggest it is worthy of further investigation.