Measuring the Performance of Realtime DSP using Pure Data and GPU

In order to achieve greater amounts of computation while lowering the cost of artistic and scientific projects that rely on realtime digital signal processing techniques, it is interesting to study the performance of commodity parallel processing GPU cards coupled with commonly used software for realtime DSP. In this article, we describe the measurement of data roundtrip time using the Pure Data environment to outsource computation to GPU cards. We analyze memory transfer times to/from GPU and compare a pure FFT roundtrip with a full Phase Vocoder analysis/synthesis roundtrip for several different DSP block sizes. With this, we can establish the maximum DSP block sizes for which each task is feasible in realtime by using different GPU card models.