A Hybrid Neural Based Dynamic Branch Prediction Unit


Modern high performance processor architectures have come to depend uponhighly pipelined operation in order to achieve improvements in operating speed. As a result, the cost associated with flushing the pipeline and refilling it when a branch instruction is mis-predicted can significantly impact processor performance. Many schemes, from the extremely simple to the highly complex, have been proposed toimprove branch prediction accuracy. Conventional two-level branch predictors predict the outcome of a branch either based on the( local branch history) information, comprising the previous outcomes of a single branch (intra-branch correlation), or based on the (global branch history) information, comprising the previous outcomes of all branches (inter-branch correlation). The mispredictionrates for these predictors are very high when they predict branch instructions with hybrid correlations. In this paper we suggest a hybrid perceptron based predictor which employs up to 31-bits of both local and global branch history information to minimize the misprediction rates. The software written for simulation and testingshows that the suggested hybrid predictor achieves a high accuracy. Our results shows that the best response of the predictor is obtained on history length of 16- bits.