Improving the GPU performance prediction models to design space exploration

Abstract

Recently, GPUs have also been used plenty in the scientific calculations for high-performance in parallel computing power and low energy consumption. Offering GPU performance prediction models base on Micro-architecture parameters to optimal design in the hardware process of the GPU, has been the subject of prior works. In this article, we defined design space NVIDIA Fermi GPU bigger than the prior similar work and with a 264 minimum size point, and we made performance prediction models by sampling only 45,000 design points from that, and then we offer a efficient search algorithm that carefully explore the design space by helping the performance models. Finally, we compare the our models with models of previous work by the help of this algorithm, and find out, that ,making models by more parameters and levels of values, help to more carefully explore in design space. Also, by analysis the models and results, we analyze the program estimator behavior with respect to micro-architecture parameters