ROCSten: The Accelerated Stencil Library for AMD Instinct™ GPUs
E. Morsi, M. Sabony · 2022
Summary ROCSten library is a highly tuned stencil patterns predesigned to tackle different memory problems. It is taking advantages of the quantitative improvements provided by AMD Instinct™ MI200 microarchitecture. As AMD Instinct™ MI250 can achieve a decent performance at reasonable power consumption due to features like unique interconnects between GCDs with the largest HBM2e memory per GPU to tackle the largest and most demanding problems. ROCSten is anticipated to provide the most used stencil patterns for many domains like Oil&Gas with some hints to couple the stencil computation with wave updates. Our preliminary results show a great potential on AMD Instinct™ MI250 compared to the latest competitive high-end accelerator. ROCSten can achieve ∼2x on AMD Instinct MI210 compared to Nvidia A100 80GB