GPU support?

[jaredjeya]

Hi,

I was just wondering if there are any plans for GPU support on the horizon e.g. through CUDA? (I've tried to test this myself, by adapting the input MPS & MPO to use GPU memory and hoping that would be enough, but it seems like a lot of internal code changes would be needed).

[mtfishman]

We don't have time to work on that right now but we'd be happy to accept a PR, it shouldn't be too hard to add support for that with judicious use of Adapt.jl.

[jaredjeya]

Thanks, using Adapt on my own fork, I've been able to get the iTDVP to run (and subspace expansion too using unified rather than GPU memory) but for the moment it seems to be much slower than the CPU - even at bond dimension e.g. 400, where I'd expect to see benefits. So I suspect the code right now is not written in a way that lets it take good advantage of the GPU, or perhaps since I'm not too experience with GPU programming, I've made some mistakes with my use of Adapt.

I suppose in the future I could look into this further.

[mtfishman]

Thanks for the summary of what you tried, that basically sounds like the approach we would use. That's too bad that you didn't see speedups, maybe some operations in the library aren't "GPU friendly". Maybe you could do some profiling to try to identify if certain operations are slowing down the code.

Anecdotally, we haven't had great success with unified memory, though I think when it hasn't worked it basically ends up being the same speed as CPU, so I don't know if that would account for the GPU code being much slower than CPU.