Branch Prediction using Advanced Neural Methods

Among the hardware techniques, two-level adaptive branch predictors with two-bit saturating counters are acknowledged as best branch predictors. They accomplish very competitive performance at low hardware cost. However, with the rapid of evolution of superscalar processors, the more accurate predictors are desired for more correct branch prediction as one of speculation method. They will lead to higher performance of processors with no doubt. This article presents alternative new and potential neural net methods for branch prediction. The advanced applications of the neural networks more than perceptron or backpropagation are examined as alternative methods. They are radial basis networks, Elman networks and Learning vector quantization (LVQ) networks. I demonstrate that these neural methods can achieve misprediction rate comparable to the conventional two-level adaptive predictors, representatively a Gshare method, without consideration of hardware and prediction latency. I also present the effects of the history length of the global history shift register (HR) and the size of the pattern history table (PHT) on the misprediction rate for each neural method.