The Combined Perceptron Branch Predictor

Previous works have shown that neural branch prediction techniques achieve far lower misprediction rate than traditional approaches. We propose a neural predictor based on two perceptron networks: the Combined Perceptron Branch Predictor. The predictor consists of two concurrent perceptron-like neural networks, one using as inputs branch history information, the other one using program counter bits. We carried out experiments proving that this approach provides lower misprediction rate than state-of-the-art conventional and neural predictors. In particular, when compared with an advanced path-based perceptron predictor, it features 12% improvement of the prediction accuracy.