Advancing Physics-Based Ground-Motion Simulations with GPU-Accelerated Hercules

Wenyang Zhang

Submitted August 30, 2026, SCEC Contribution #15564, 2026 SCEC Annual Meeting Poster #210

Physics-based ground-motion simulations at frequencies relevant to engineering applications require very large meshes and many thousands of timesteps, making computational cost a major barrier to regional simulation ensembles and model validation. We present a GPU-accelerated implementation of Hercules, an octree-based finite-element code for earthquake wave propagation. The implementation ports the effective-stiffness, BKT viscoelastic attenuation, and explicit displacement-update operators to CUDA, maintains the principal wavefield arrays on the GPUs, and exchanges boundary data using CUDA-aware MPI. On early-access NVIDIA GB200 systems, Hercules maintains 95.6% weak-scaling efficiency from 1 to 64 GPUs while increasing the problem size from 67.1 million to 4.29 billion elements. At an equal scaling-unit count, one GPU is 5.4–7.8 times faster per timestep than one 144-core NVIDIA Grace–Grace node. We demonstrate the production capability with a 2-Hz simulation of the 2013 Mw 6.4 Ruisui, Taiwan, earthquake. The model contains 1.584 billion elements and advances 60 s of ground motion in 40,000 timesteps. Solver execution requires 1,007 s on 32 GPUs and 558 s on 64 GPUs, representing speedups of 5.86 and 10.58, respectively, relative to 32 CPU nodes. Unfiltered three-component station histories from the GPU and established CPU implementations have a relative L2 difference of 1.215 × 10−5 and minimum component correlation greater than 0.9999; wavefield snapshots also show close spatial agreement. These results establish a numerically consistent accelerated Hercules workflow and substantially expand the practical scale and throughput of physics-based regional ground-motion simulations.

Key Words
Hercules, Ground Motion Simulation, High-Performance Computing, GPU Acceleration

Citation
Zhang, W. (2026, 08). Advancing Physics-Based Ground-Motion Simulations with GPU-Accelerated Hercules. Poster Presentation at 2026 SCEC Annual Meeting.

Related Projects & Working Groups
Research Computing (RC)