Accelerating atmospheric physics parameterizations using graphics processing units.

This article focuses on porting two physics parameterizations from the Common Community Physics Package (CCPP)—the aerosol-aware Thompson microphysics scheme and the Grell–Freitas cumulus convection scheme—to Graphics Processing Units (GPUs) using the OpenACC directive programming language. The study demonstrates significant acceleration, achieving up to 120× speedup for the Thompson scheme and 90× for the Grell–Freitas scheme compared to single-core CPU runs, with realistic speedups around 10× when compared to multi-core CPUs. Multi-GPU implementations show good weak scaling efficiency, especially when using one GPU per node, due to the column-wise independence of physics computations that minimizes inter-node communication. The work highlights challenges such as the impact of automatic arrays on GPU performance and the necessity of porting entire numerical weather prediction models—including dynamics—to GPUs to realize overall speedups, as data transfer overheads can negate gains from accelerating physics alone. Validation against CPU results confirms the accuracy of the GPU implementations within acceptable tolerances, and ongoing efforts aim to integrate these accelerated kernels into operational weather models.