Comprehensive regression-based model to predict performance of general-purpose graphics processing unit

Abstract

Recently, the use of graphics processors has been significantly increased in fast and accurate scientific calculations. These processors provide a heterogeneous design space, and make designers capable of performing more accurate designs with higher efficiency. In this paper, a regression model is proposed to predict the performance of various applications on general-purpose graphics processors units. We present the main challenges for predicting the efficiency of graphics processing units (GPUs) based on simulation experiments. Also, we build the regression statistical inference from the result of the simulation, which predicts the efficiency of GPUs in various performances with approximately 7% of error of measurement. We have used AMD Southern Island and SDK 2.5, OpenCL which are both based on OpenCL. The first version of the design is built by very large design space, approximately about 17 billion points, from which 8000 points were randomly chosen, and the performance of graphic processors was calculated based on the results of the simulation. The model of non-linear regression is capable of predicting the performance of graphics process with the average error rate of 7%.