Multi-GPU Discontinuous Galerkin Solver for Maxwell’s Equations

Modern supercomputers increasingly rely on Graphics Processing Units (GPUs), a trend likely to continue. GPUs offer high performance and low cost per GFLOP, making them ideal for scientific computing. In computational electromagnetics, GPU use began with finite-difference time-domain (FDTD) and discontinuous Galerkin (DG) methods. DG is especially suited for GPUs, combining the flexibility of the finite element method mesh with explicit FDTD time-stepping and parallelism. This work presents a multi-GPU implementation of a time-explicit DG method for Maxwell’s equations, optimized to efficiently exploit modern heterogeneous clusters with hundreds of GPUs. On that hardware, the solver achieves good strong and weak scaling performance up to around a hundred GPUs.