A Comparison of Performance Tuning Process for Different Generations of NVIDIA GPUs and an Example Scientific Computing Algorithm

We consider the performance of a selected computational kernel from a scientific code on different generations of NVIDIA GPUs. The code that we use for tests is an OpenCL implementation of finite element numerical integration algorithm. In the current contribution we describe the performance tuning for the code, done by searching a parameter space associated with the code. The results of tuning for different generations of NVIDIA GPUs serve as a basis for analyses and conclusions.