On shadertoy there are many examples of GPU shaded audio synthesis. My favorites are:

https://www.shadertoy.com/view/4ts3z2 which is a perfect example for ambient sound

https://www.shadertoy.com/view/ldfSW2 for an example of acid techno music

https://www.shadertoy.com/view/lsSXzD a doom "port" rendered in a single shader including music

Your links include the closing quotes for those clicking in a browser. Fixed links:

Which is a perfect example for ambient sound [1]

For an example of acid techno music [2]

A doom "port" rendered in a single shader including music [3]

[1] https://www.shadertoy.com/view/4ts3z2

[2] https://www.shadertoy.com/view/ldfSW2

[3] https://www.shadertoy.com/view/lsSXzD

Sorry/Thanks... apparently one should not quote URLs on HN.

Slightly unrelated, but I find that slide decks posted only at face value miss a lot of contextual information, discussion and audience interest that comes from watching a presentation on a video or in person with commentary.

What makes a good slide deck for a presentation makes for terrible reading. Likewise, what makes for good reading makes for a terrible PowerPoint presentation.

GPUs have single cycle sin/cos LUT implementations, so they should be good for this kind of thing - competitive with an FFT for a relatively small number of overtones.

But there are some technical issues. Older GPUs don't have full IEEE 754 floating point support, and compilers are optimised for graphics, not for audio, so you won't necessarily get the same output as you would from DSP code running on an x86.

If shader compilers included some audio-friendly tweaks and perhaps a standard audio API for co-processing, they could see a lot more use in music editing and synthesis.

It’s worth pointing out that there is a closed form for the sum of a finite geometric series used for summing harmonic sine waves into bandlimited impulse trains (BLIT) which are filtered to create alias-free classic analog oscillators:

https://ccrma.stanford.edu/~stilti/papers/blit.pdf

It may just be a limitation of the slides rather than the full presentation, but this is almost entirely about the audio synthesis side. There is nearly nothing specific to GPUs or shaders that I can see.

So fair warning: if, like me, you already know the fundamentals of audio and synthesis (and it starts with "what's a sine" and "how does 440Hz map to a piano), but were interested in the shader part, there's not much here for you.

The reason is there isn't much special about his use of shaders. The first two slides cover shader specific details (write to a 1D render target where each pixel is a 44khz sample). The audience is people who are familiar with fragment shaders, and specifically familiar with shadertoy, but with limited exposure to additive synthesis.

Very nice work. This could be used to efficiently re-implement 8-bit- and 16-bit-era sound systems, which are mostly emulated on the CPU at present. I wonder if the same techniques can also be used to enable DSP-like workloads of a more general sort on GPU compute.

Most of those are either very straightforward subtractive synthesizers, or use digital FM synthesis. It's much more efficient (and accurate) to start from sampled basic waveforms or reverse-engineered data from the actual systems, rather than building them using additive synthesis techniques.