high performance computing on graphics processing units: hgpu.org

The computation of large-scale virtual acoustics using the 3D finite difference time domain (FDTD) is prohibitively computationally expensive, especially at high audio sample rates, when using traditional CPUs. In recent years the computer gaming industry has driven the development of extremely powerful Graphics Processing Units (GPUs). Through specialised development and tuning we can exploit the highly parallel GPU architecture to make such FDTD computations feasible. This paper describes the simultaneous use of multiple NVIDIA GPUs to compute schemes containing over a billion grid points. We examine the use of asynchronous halo transfers between cards, to hide the latency involved in transferring data, and overall computation time is considered with respect to variation in the size of the partition layers. As hardware memory poses limitations on the size of the room to be rendered, we also investigate the use of single precision arithmetic. This allows twice the domain space, compared with double precision, but results in phase shifting of the output with possible audible artefacts. Using these techniques, large-scale spaces of several thousand cubic metres can be computed at 44.1kHz in a useable time frame, making their use in room acoustics rendering and auralization applications possible in the near future.