high performance computing on graphics processing units: hgpu.org

Graphical Processing Units (GPUs) are massive floating-point stream processors, and through the recent development of tools such as CUDA and OpenCL it has become possible to fully utilize them for scientific computing. We have developed an open-source CUDA-based acceleration framework for 3D Computational Fluid Dynamics (CFD) using Smoothed Particle Hydrodynamics (SPH). This paper describes the methods used in our framework and compares the performance of the implementation to previous SPH implementations. We implement two different SPH models, a simplified model for Newtonian fluids, and a complex model for Non-Newtonian fluids, which we use for simulation of snow avalanches. Having implemented two different models, we investigate the performance characteristics of SPH simulations on the GPU and find that despite the larger bandwidth-requirements of the complex model the GPU scales well. Our simulations are rendered interactively and in "real-time". Using an NVIDIA GeForce GTX 470 Fermi-based card we achieve 215.4, 122.2 and 64.9 FPS for the simple model and 69.6, 37.4 and 19.1 FPS for 64K, 128K and 256K particles respectively.