Predicting GPU Performance and System Parameter Configuration Using Machine Learning

Zhuren Liu, Trevor Exley, Austin Meek, Rachel Yang, Hui Zhao, Mark V. Albert

Abstract: GPUs are widely used in accelerating computation-intensive applications. Performance models are important for designing high-performance and cost-efficient GPUs. In this work, we developed machine learning models that can accurately predict GPU system performance. Our model can identify important features that can provide insights to designers on the most important hardware system parameters when executing their applications. We also developed a model for predicting minimum system configuration parameters based on performance. Our model can provide system configuration recommendations for users to meet their performance requirements.