Development of a GPU-Based Discrete Ordinates Transport Sweeper in the Griffin Reactor Physics Application

The computing landscape is experiencing one of its most rapid and disruptive evolutions, driven by the advancement of graphics processing units (GPUs). It also offers a significant potential for making neutron transport calculations computationally more affordable, which is important for advanced reactor development. This work presents the development of a GPU-accelerated, discontinuous finite element method–based discrete ordinates (SN) sweeper in Griffin, which is the flagship reactor physics application of the MOOSE framework.Kokkos was adopted as the programming model to achieve cross-architecture portability while maintaining a single codebase. The proposed sweep algorithm is a block-Jacobi algorithm, and it maximizes parallelism by exploiting concurrency not only across directions and wavefront elements, but also across energy groups and local degrees of freedom. Furthermore, a wavefront angular flux storage scheme was introduced to overcome GPU memory constraints, allowing for full parallelization over directions and ensuring scalability for large-scale problems.A comprehensive performance analysis demonstrated significant speedups in the transport sweep. For the two-dimensional C5G7 benchmark, a single GPU achieved performance comparable to more than 400 CPU cores, and a grind time analysis confirmed that the proposed sweep algorithm attains sufficient parallelism. In the three-dimensional (3D) Empire microreactor benchmark, four GPUs outperformed 1120 CPU cores by more than a factor of 2.The block-Jacobi sweep algorithm adopted for GPUs was validated, showing only marginal degradation in convergence rates, and the indispensability of wavefront angular flux storage was proven by successfully fitting the full 3D core problem into four GPUs. These results demonstrate the potential of making SN transport calculations practical for routine reactor analysis through GPU acceleration.