The simulation of fusion experiments, such as the ITER, JT-60SA or DEMO power plants, is extremely complex since it involves coupling fundamentally non-linear multi-scale multi-physics phenomena that require a huge number of degrees of freedom in order to be solved realistically. It is therefore crucial for fusion researchers to be able to fully harness the enhanced capabilities offered by the latest generations of supercomputers.
Within the framework of the EUROfusion consortium, the Advanced Computing Hub HPC centers of excellence actively engage in enhancing existing European fusion simulation codes, which are specifically designed for modeling plasmas within tokamaks and stellarators fusion reactors in order to accelerate their design.
Our presentation outlines the overarching strategy that the EPFL ACH has adopted for porting multiple fusion research codes onto GPUs. In particular, we focus on an approach which centers on maintaining a single codebase by employing unified OpenACC or OpenMP directives in order to achieve performance portability across diverse hardware architectures, including NVIDIA, AMD and Intel GPUs, while ensuring compatibility with various compilers.
In order to provide concrete insights into the depth of our strategy, we showcase results obtained using multiple fusion codes across various European supercomputing platforms.