There and Back Again: The Practicality of GPU Accelerated Digital Audio

General-Purpose GPU computing is becoming an increasingly viable option for acceleration, including in the audio domain. Although it can improve performance, the intrinsic nature of a device like the GPU involves data transfers and execution commands which requires time to complete. Therefore, there is an understandable caution concerning the overhead involved with using the GPU for audio computation. This paper aims to clarify the limitations by presenting a performance benchmarking suite. The benchmarks utilize OpenCL and CUDA across various tests to highlight the considerations and limitations of processing audio in the GPU environment. The benchmarking suite has been used to gather a collection of results across various hardware. Salient results have been reviewed in order to highlight the benefits and limitations of the GPU for digital audio. The results in this work show that the minimal GPU overhead fits into the real-time audio requirements provided the buffer size is selected carefully. The baseline overhead is shown to be roughly 0.1ms, depending on the GPU. This means buffer sizes 8 and above are completed within the allocated time frame. Results from more demanding tests, involving physical modelling synthesis, demonstrated a balance was needed between meeting the sample rate and keeping within limits for latency and jitter. Buffer sizes from 1 to 16 failed to sustain the sample rate whilst buffer sizes 512 to 32768 exceeded either latency or jitter limits. Buffer sizes in between these ranges, such as 256, satisfied the sample rate, latency and jitter requirements chosen for this paper.