Improved composite confidence mechanisms for a perceptron branch predictor

In 2001, Jiménez and Lin [Dynamic branch prediction with perceptrons, Proceedings of the 7th International Symposium on High Performance Computer Architecture, 2001, pp. 197-206] introduced the perceptron branch predictor, the first dynamic branch predictor to successfully use neural networks. This simple neural network achieves higher accuracies (95% at a 4 KiB hardware budget) compared to other existing branch predictors and provides a free confidence level. In this paper, we first gain insight into this inherent confidence mechanism of the perceptron predictor and explain why (additional) counter based confidence strategies can complement it. Second, we evaluate several composite confidence estimation strategies and compare them to the described technique by Jiménez and Lin [Composite confidence estimators for enhanced speculation control, Tech. rep., Department of Computer Sciences, The University of Texas at Austin, 2002]. We conclude that our overruling AND-combination of perceptron confidence and resetting counter mechanism outperforms the previously proposed confidence scheme.