Justin Kin Jun Hew: GPU-Accelerated Ice Sheet Modelling in ACCESS: Deploying ISSM on Gadi for Antarctic Projections
Continental-scale Antarctic simulations are limited less by our understanding of ice physics than by the cost of the stress balance solve, which dominates runtime and forces uncomfortable trade-offs between mesh resolution, ensemble size, and the length of transient experiments. We present ACCESS-ISSM, a Spack-deployed build of the Ice Sheet System Model on NCI’s Gadi, extended with GPU-accelerated linear algebra through PETSc and Hypre.
We compare two GPU preconditioning strategies for the SSA stress balance: PETSc’s native GAMG on cuSPARSE matrices, and Hypre’s BoomerAMG operating on device-resident Hypre matrices, and show that keeping the operator in Hypre’s own format avoids per-solve format conversion and substantially reduces time spent in solver kernels relative to GAMG. We discuss the practical costs that profiling exposes but wall-clock timings hide, particularly device memory churn during multigrid setup.
We demonstrate the capability on AIS3, a whole-of-Antarctica configuration, through a sensitivity study across basal friction laws and coupling coefficients spanning several orders of magnitude, an ensemble whose feasibility depends directly on per-solve cost.
Finally, we describe the Python-side changes required to make this accessible: GPU solver presets and GPU queue support in pyISSM, so that moving an existing experiment onto GPUs is a configuration change rather than a rewrite. The build is reproducible through the ACCESS Spack infrastructure and available as a module on Gadi.
We close with what remains unresolved: memory-management overhead, scaling beyond a single node, and which classes of ice sheet problem genuinely benefit from GPUs.