Speaker
Description
Hermes-3, a multi-fluid plasma simulation framework built on BOUT++, is currently being ported to GPUs using the RAJA performance portability layer and CUDA. This work builds upon the GPU infrastructure previously developed at Lawrence Livermore National Laboratory (LLNL) for BOUT++, enabling the reuse of established programming models and optimization strategies.
The porting effort is being validated using the Annulus plasma turbulence model and three-dimensional edge plasma simulations of the TCV tokamak. We present the current status of GPU-enabled kernels and preliminary performance results, highlighting key challenges related to memory management, solver portability, and data movement. The roadmap toward full GPU capability, including further physics module acceleration and optimization for future exascale systems, is also discussed.