Workshop: High Performance Computing for Environmental and Earth Sciences (HPC4EES)
Authors: Wenyang Zhang and Yinzhi Wang (Texas Advanced Computing Center, University of Texas at Austin)
Abstract: Physics-based earthquake ground motion simulations, in order to achieve a desirable maximum frequency for practical applications, require billions of finite elements and tens of thousands of explicit timesteps, making the element-local force operators, inter-rank exchange, and output path all important to end-to-end performance. We present a modern GPU-resident implementation of Hercules, an octree-based finite-element earthquake simulator. The implementation maps effective stiffness, BKT viscoelastic attenuation, and explicit displacement update to CUDA; retains the primary field histories on the accelerator; exchanges boundary data through CUDA-aware MPI; and limits host transfers to data required by non-accelerated paths and output. On early-access NVIDIA GB200 resources, a fixed 67.1-million-element problem attains a 25.02X speedup on 64 GPUs, while a weak-scaling sweep maintains 95.6% efficiency from one to 64 GPUs and grows to 4.29 billion elements. At equal scaling-unit count, one GPU is 5.4–7.8 times faster per timestep than one 144-core NVIDIA Grace--Grace node. A 2-Hz simulation of the 2013 Mw 6.4 Ruisui earthquake contains 1.584 billion elements and 40,000 timesteps. It completes the solver in 1007.05s on 32 GPUs and 558.06s on 64 GPUs, speedups of 5.86X and 10.58X over 32 CPU nodes. Unfiltered station histories agree with the CPU implementation to a relative L2 error of 1.215e-5 and a minimum component correlation above 0.9999. These results demonstrate that GPU residency and accelerator-aware communication make Hercules practical for high-resolution production ground motion simulations without changing its established numerical model.
Back to High Performance Computing for Environmental and Earth Sciences (HPC4EES) Archive Listing
Back to Full Workshop Archive Listing