Parallel Path Based Local Perceptron Branch Predictor

As the number of pipeline stages increases we become hostage to the penalty imposed by misprediction of branches. Previous works have shown that a neural branch predictor remains victorious among its peers by achieving much lower misprediction rates on similar sized hardware budget than traditional approaches. In this paper we implement 3 variations of perceptron based predictor and give each its verdict. We look at the modified base-line perceptron predictor, then at the history and address based path perceptron, and finally we look at a history based path guided perceptron combined with a local predictor. We discover that on average an improvement of 5% in misprediction rates can be seen by moving towards the latter kind of predictor.