high performance computing on graphics processing units: hgpu.org

The computation of virtual acoustics for physical modelling synthesis using the finite difference time domain is a computationally expensive process, especially at audio rates such as 44.1kHz. However, the high level of dataindependence is well suited to parallel architectures such as those provided by graphics processing units. This paper describes the use of the latest Nvidia Kepler cards to accelerate the computation of three-dimensional schemes. The CUDA language and hardware architecture allow many possible approaches to computing even a basic model. Various techniques are considered, such as full tiling, iteration slicing, and the use of shared memory. A standard simulation was used to measure the performance of these different approaches. Benchmark times were compared for the latest Nvidia Tesla K20 GPU against the previous generation cards. Results show the continuing maturity of the hardware, especially in terms of data caching, which allows basic code designs to perform as well as more complex shared memory versions.