Towards the Specification of the GPU using Performance Parameters

Abstract

The characteristics of graphics processing units (GPUs), especially their parallel execution capabilities and fast memory access, render them attractive in many application areas. They promise more than an order of magnitude speedup over conventional processors for some non-graphics computations. The use of GPUs in general-purpose computing is becoming a very accepted alternative. In addition, the CUDA programming model gains acceptance. Each of these arguments make necessary count with tools that allow to evaluate GPUs. The performance parameters allow to model an architecture to predict the execution time of any application with any parallelism level. Furthermore, they are a useful tool to compare different architectures and determine its advantages and troubles. This work presents suitable benchmarks to evaluate different performance parameters of GPUs. The presented measurements focus on two issues of GPU performance: computing power and the global memory bandwidth. Their estimation will allow us determine technical characteristics of GPU and, in consequence, the analysis and optimization of the performance of applications that could run on actual or future GPUs.