Workshop: Thirteenth Workshop on Accelerator Programming and Directives (WACCPD 2026)
Authors: Marshall Ward (NOAA - Geophysical Fluid Dynamics Laboratory); Edward Yang (ACCESS-NRI); Jorge Gálvez Vallejo (NCI Australia); and Utheri Wagura (SAIC, NOAA - Geophysical Fluid Dynamics Laboratory)
Abstract: We describe a performance-portable GPU implementation strategy for the MOM6 ocean model, where portability is realized through a single codebase built on language standards while maintaining competitive performance on target CPU and GPU platforms. The implementation uses standard Fortran ``do concurrent`` as the kernel-level expression of parallelism, with OpenMP support for device-data residency and GPU-aware MPI communication.
Performance results show GPU acceleration of the dynamic core up to 3x relative to a single-socket CPU baseline while preserving bitwise-identical CPU and GPU results. Fixed-domain tests show scaling across multiple H100 GPUs. Module-level timings show that solver taxonomy guides the porting strategy but does not by itself predict realized speedup.
Back to Thirteenth Workshop on Accelerator Programming and Directives (WACCPD 2026) Archive Listing
Back to Full Workshop Archive Listing