Application Performance on the Newest Processors and GPUs
TimeTuesday, July 243:45pm - 4pm
DescriptionThis paper discusses the capabilities of the newest processors and GPUs to run a mixture of the most common scientific applications. The baseline system for these comparisons is the 32-core Intel Broadwell processor which has been around for two years. Comparisons are made to the newer Intel Skylake and the AMD EPYC processors. The EPYC architecture has typically twice as many cores so one point of comparison is whether each code can effectively make use of the higher core count. These codes can be accelerated using GPUs with some taking advantage of 32-bit acceleration while others need good 64-bit performance. The consumer grade NVIDIA GeForce GTX 1080Ti cards are used as the baseline for the GPU comparisons. Higher level NVIDIA Quadro GP100 and Titan V cards are evaluated using each code. All applications use CUDA to enable GPU acceleration. AMD provides tools in its HIP package that allow translation of C and C++ CUDA code into source code that can be compiled with either NVIDIA’s NVCC or AMD’s HCC compilers. This project also involves investigating the performance and ease of converting CUDA code to run on the AMD Radeon Vega Frontier Edition GPU card.