A performance estimation model for GPU-based systems

Abstract

GPUs - Graphics Processing Units are now used in a wide variety of computing systems, to solve a wide variety of computational problems. Even though they were initially developed to accelerate graphics processing, their parallel architecture has demonstrated to be extremely useful for other applications, including high-performance computing. Due to their widespread use, it is important to understand and estimate its performance, which depends on several architecture parameters, particularly core frequency, memory frequency, and number of cores. In this paper, we present an analytical model to estimate GPUs performance, and we demonstrate its accuracy using a set of benchmarks: 3D games, namely Crysis and Company of Heroes, a 3D Wave benchmark, namely DirectCompute. We also apply the model to a High Performance Computation benchmark, SGEMM, which is based on floating-point single precision matrix multiplications. Comparison between the output of the estimation model and measured data for different benchmarks and GPU architectures is less than 10% for all tested cases.