Multi-GPU Server Deign Parameters Selection based on Empirical Observation of HPL Behavior

Design parameters of a multi-GPU based server that affect the performance of HPL are evaluated and suggested for optimized GPU server to achieve optimal HPL performance. GPUs are widely used in High-performance Computing and AI Computing. To obtain more performance, multi-GPU servers are generally used in various configurations. The configurations in a Multi-GPU server affect the overall performance of a system. The major parameters for a multi-GPU server configurations are the number of cores per GPU and number of PCI Express per GPU in HPL running. In this paper, the two parameters are tested, measured, and analyzed to figure out its impact on HPL performance. The experimental results show that the optimal number of cores has to be more than three per GPU, and the optimal number of GPUs per PCIe is three in a multi-GPU server. The empirical observation reveals that there exists the optimal number of GPUs in a multi-server for best HPL performance.