[gsliepen]

The simplest form of string synthesis is perhaps the Karplus-Strong algorithm (https://en.wikipedia.org/wiki/Karplus%E2%80%93Strong_string_...). Instead of simulating the motion of all parts of the string for each timestep, it simplifies it to a perfect transmission of motion, except at the end of the string where some function will be applied that captures the losses from the motion, typically by performing a simple low-pass filter. Computationally it is very efficient: just a few FMA operations per sample. The drawback is that you need a buffer to store the state of the string, which can be quite large (especially for low frequencies which correspond to long strings), and you need to read and write to different parts of this buffer. Thus, if you are simulating multiple strings, it's more likely that memory bandwidth will be the bottleneck instead of raw compute power.

[DennisP]

So doing it on a big GPU would be pretty helpful. I wonder if anyone has used CUDA for this, or if any hardware synths have GPUs.

[Blackthorn]

The time it takes to get data to/from the GPU is problematic given the sample rate you need to deliver.

Anukari used the GPU but eventually abandoned it for the CPU. They have some good and informative blog posts about it.

[gsliepen]

I actually tried doing it as a Vulkan compute shader, continuously simulating 128 strings (all possible MIDI notes). It didn't work very well. While the latency was tolerable, it was slower than doing the same on the CPU (although this was with an integrated GPU, not a dedicated card).