Development of a GPU-accelerated, Finite Element based Dynamical Core for Sea Ice

Sea ice is an import part of Earth's climate system. Yet, an accurate, highly resolved simulation of sea ice dynamics remains challenging. As the development of faster processors has slowed down, a turn to more specialized hardware is needed to achieve more accurate simulations at higher resolutions. Graphics processing units (GPUs) offer an order of magnitude higher floating-point performance and efficiency compared to CPUs. However, their full utilization often also requires significant engineering effort. Therefore, several frameworks have emerged in recent years which aim to simplify general-purpose GPU programming. In particular, heterogeneous compute frameworks such as SYCL and Kokkos make it possible to develop a unified code base that works accross GPUs and CPUs. Similarly, machine learning frameworks like PyTorch combine an easy to use interface with highly specialized backends that can make it possible to transparently exploit new hardware features to accelerate large-scale linear algebra workloads. Furthermore, their use provides a simple path-way to integrate machine learning components into simulations.

In this talk, we compare available options for the GPU-parallelizaton of the novel sea-ice code neXtSIM-DG. Its dynamical core is based on higher-order finite elements for the momentum equation and discontinous Galerkin elements for the advection. This makes the code highly parallezible. We discuss characteristics of our discretization and its consequences for the GPU implementation. For the full port of the dynamical core we use Kokkos as, based on our assessement, it combines usability with good performance. With moderate changes compared to the OpenMP-based CPU code, the new implementation achieves a sixfold speedup on the GPU while being as fast as the reference on the CPU.