high performance computing on graphics processing units: hgpu.org

This paper presents a study of the computational enhancement of a Graphics Processing Unit (GPU) enabled 2D flood model. The objectives are to demonstrate the significant speedup of a new GPU-enabled full dynamic wave flood model and to present the effect of model spatial resolution on its speedup. A 2D dynamic flood model based on the shallow water equations is parallelized using the GPU approach developed in NVIDIA’s Compute Unified Development Architecture (CUDA). The model is validated using observations of the Taum Sauk pump storage hydroelectric power plant dam break flood event. For the Taum Sauk flood simulation, the GPU model speedup compared to an identical CPU model implementation is 80x-88x for computational domains ranging from 65.5 k to 1.05 M cells. Thirty minutes of event time were simulated by the GPU model in 2 min, 15 times faster than real time. An important finding of the analysis of model domain size is the GPU model is not constrained by model domain extent as is the CPU model. Finally, the GPU implementation is shown to be scalable compared with the CPU version, an important characteristic for large domain flood modeling studies.