Advanced branch predictors for soft processors

Abstract

This work studies implementations of the Perceptron [1] and TAGE [2] branch predictors for general purpose, in-order pipelined single core soft processors. It proposes FPGA-friendly optimizations whose goal is to achieve high operating frequency. This work discusses the design tradeoffs and proposes a highly accurate and fast branch predictor variant based on TAGE, O-TAGE-SC. It operates at 270MHz, the maximum frequency of Altera's highest performing soft-processor Nios II-f. Using a representative subset of the SPECCPU2006 benchmarks, this work shows that O-TAGE-SC delivers 5.2% better instruction throughput versus the previously proposed gRselect predictor [3].