Empirical Study on the GPU-accelerated HPL Performance: Effects of PCIe Communication

Computing resources equipped with GPU devices are popularly adopted in various application areas, and the HPL benchmark is used to evaluate their standard performance. The performance of GPU-accelerated HPL however cannot be free from PCIe communication even in the most optimal case. Here, we investigate the PCIe communication overhead and its effect on the NVIDIA HPL performance using a simple model. This preliminary study intends to derive a way that can estimate the HPL performance in upcoming computing resources that support full cache coherence between CPU and GPU memory.