GPU design space exploration: NN-based models

Abstract

Different applications have different memory and computational demands. Therefore, obtainable performance and energy efficiency on a GPU depends on how well the GPU resources and application demands are balanced. In this study, we are presenting a Neural Network based predictor to model power and performance of GPGPU applications. The proposed model accurately predicts power and performance for most of the configurations in the design space with average prediction error of less than 6.5%. For configurations with high prediction errors, we have developed an outlier detection method to filter them out from the output of the model. The proposed filter captures most of the extreme outliers and improves the accuracy of the model.