Files

AGU2026.pdf
  • Open Access
  • Adobe PDF
  • 4.79 MB

Details

Authors
Abstract
We present a high-performance GPU implementation of SLIM, a coastal ocean model that solves the hydrostatic Boussinesq equations on unstructured 3D meshes using the Discontinuous Galerkin method. The model employs a split IMEX scheme to separate fast barotropic and slow baroclinic processes, enabling efficient time-stepping. Our GPU architecture leverages the vertical extrusion structure of coastal meshes to maximize effective memory bandwidth usage and computational throughput. We introduce novel memory layouts and matrix-free solvers for key vertical processes, enabling more than an order-of-magnitude speedup compared to CPU implementations. The model demonstrates strong and weak scaling across multiple GPUs via domain decomposition with GPU-aware MPI and overlapping communication. Benchmarks on AMD and NVIDIA GPUs show performance exceeding 1,500 CPU cores per GPU for large-scale coastal simulations. We also present examples of regions where the model is currently being applied, demonstrating its readiness for scientific deployment.
Affiliations

Citations

De Le Court, M., Lambrechts, J., & Legat, V. (2026, February 23). Designing an Efficient Multi-GPU ​Coastal Ocean Model. OSM26, Glasgow. https://hdl.handle.net/2078.5/277691