A neural network-based approach for the performance evaluation of branch prediction in instruction-level parallelism processors

Branch prediction is essential for improving the performance of pipeline processors. As the number of pipeline stages in modern processors increases, an accurate branch prediction is important. Traditional branch predictor uses the concept of counter and history for the prediction of conditional branch instructions. Furthermore, this concept is replaced with the number of perceptrons using neural networks. In this paper, neural network-based approaches like perceptron neural branch predictor, global perceptron neural branch predictor, and a learning vector quantization neural branch predictor are applied to the different trace files to predict the conditional branch instructions. Furthermore, a backpropagation neural branch predictor scheme is proposed, providing more accuracy than other neural network techniques. The statistics results are obtained regarding accuracy, misprediction rate, precision rate, recall rate, and F1-score rate. The average results suggest that the proposed backpropagation neural branch predictor improves the accuracy of perceptron branch predictor, global perceptron branch predictor, and learning vector quantization neural branch predictor by 13.82%, 5.85%, and 1.11%, respectively.