Performance Benchmark of FMO Calculation with GPU-Accelerated Fock Matrix Preparation Routine

Abstract

GPU acceleration for OpenFMO, a fragment molecular orbital calculation program, has been implemented and its performance was examined. We have developed a GPU-accelerated Fock matrix preparation routine without atomic operation and implemented it into the worker module of an OpenFMO program, which is parallelized in a master-worker programming model. The GPU-accelerated program shows 1.5× speedups from CPU only FMO-HF/6-31G (d) calculation with a direct SCF method for 642 atomic protein on 8 nodes of HA-PACS base cluster.