Low-Power Perceptron Branch Predictor

Branch predictors relying on neural networks have received increasing attention in recent years. Unfortunately, such designs are often impractical as they come with high latency and power dissipation. In this work we introduce a low-power perceptron predictor which utilizes as much resources as needed according to the branch behavior, effectively reducing overall number of computations. We reduce predictor energy consumption by not assigning computation resources to unnecessary computations. While reducing predictor energy consumption, we also improve overall performance as we reduce prediction latency. We reduce the predictor computational power dissipation up to 34% while improving the processor performance by up to 19%.