Overcoming Challenges in GPU Computing for Scientific Applications: SoC Solutions to PCI Bandwidth Limitations

Traditionally, with high performance GPU-based computing, the PCI memory bandwidth has been a limiting factor on performance. To avoid this problem, System-on-a-Chip (SoC) devices are explored as a potential solution. The main benefit of these devices is unified memory, which allows zero-copy algorithms to be implemented, eliminating the need for any PCI memory transfer. These devices are analyzed in both performance and cost and are compared to a modern discrete GPU, an Nvidia V100. To compare these, benchmarks from the SHOC benchmark suite were used to analyze performance on different commonly used algorithms for scientific computing in physics.