ascent-system-timing | SUNDIALS and hypre: Exascale-Capable Libraries for Adaptive Time-Stepping and Scalable Solvers

Figure 5. Nyx weak scaling on NVIDIA GPU systems vs. CPU OpenMP version (Summit CPUs), speed-up 13 – 14.5× over CPU. Performance on Perlmutter (Phase 1, subject to improvement) is now 100× over the pre-GPU (MPI+OpenMP) November 2021 version. (Figure and results courtesy of Jean Sexton.)