Forward-pass performance on 4× Tesla V100-PCIe-32 GB GPUs as a function of cubic grid size. Open symbols show the wall-clock time for a single forward evaluation after compilation, while filled symbols indicate the one-time setup/compilation cost for the corresponding kernel. The FFT-based solver (black diamonds) and multigrid solver (MG; red circles) have comparable compilation overheads at small resolutions, but FFT compile time rises more steeply with grid size, consistent with larger plan construction and distributed communication setup. In evaluation mode (open symbols), MG is consistently faster than FFT over the full resolution range. Parameters for MG were chosen at each resolution to require <0.1% accuracy over all pixels for net force calculation. Absolute timings vary with system configuration and interconnect.