GPU audio processing

Question

To begin with, it so happened that I am a guitarist with little programming knowledge (to be honest, I'm more interested in this than I can do anything). And so the other day I was studying the work of neural networks, which led me to research on the use of video cards for mathematical calculations. There is always some delay in programs like Guitar Rig and others. In my understanding, transforming sound using such programs is simply mathematical operations performed with sound (yes, it may be said too simply and inaccurately, but the essence, I think, is correct). And I wondered if there was no way to implement something like this, where gpu would be used for computation and not cpu? Considering that, for example, when rendering in video games, the latency on video cards is minimal, despite the fact that this is a more resource-intensive operation than audio processing.

I know there are various libraries for harnessing the power of graphics cards, but not sure if this is what I need. I found only two materials related to this in one way or another, but as I said earlier, I'm not sure if they work completely well. If you know good libraries, please share this. https://archive.codeplex.com/?p=cudafy

https://www.codeproject.com/Articles/1116907/How-to-Use-Your-GPU-in-NET

Also, I thought I would most likely try to implement this in C #.

Answer 1

Sorry, going to disappoint you straight away. I have tried using NVidia CUDA (the native library) for audio processing, using neural networks. It's what my company does for a living, so we're pretty competent. We found that the typical NVidia card has too much latency. They're fast, that's not the problem, but that means they can do many million operations in a millisecond. However, the DMA engine feeding data to the card typically has latencies that are many milliseconds. Not so bad for video, bad for audio - video often is 60 Hz whereas audio can be 48000 Hz.

Answer 2

I totally agree with the other answers in that GPUs will (probably) have too much much latency for online audio processing. They are designed to work on relatively large chunks of data.

For custom processing of streamed data (like your audio signals), you may want to look into dedicated hardware like DSPs or an FPGA if you really want all the customization. But then you're looking into working on hardware, rather than software - that's a rather different beast.

That being said, I've had a little experience with Alchitry's FPGA development boards and it's not quite as horrible as some people claim.