Perceptron learning for predicting the behavior of conditional branches

Branch prediction, i.e., predicting the outcome of a conditional branch instruction, is essential to the performance of current and future microprocessors. We show how perceptrons can be used to improve the state of the art in branch prediction. We explore the unusual challenges this domain presents for neural systems, and we show why other neural methods, such as backpropagation, provide no additional accuracy in this context. Finally, we identify other areas where neural systems can be applied to microprocessor implementation.