Integrate ZLUDA for AMD CUDA

[GermanAizek]

Issue type

Feature Request

Describe the feature

@chokobole,

https://github.com/vosen/ZLUDA

it can give feature run CUDA program on amd video cards if cuda code is optimized, and speed up work on amd.

[chokobole]

Nice! Thanks for sharing! We are now working on replacing Halo2 with Tachyon and optimizing Halo2 provers. After this, we'll upstream scroll zkevm circuits and Halo2. Maybe we can continue on GPU things after all these.

[peter941221]

Hi @chokobole — I saw the earlier note that GPU work here was deferred behind the Halo2/Tachyon and prover work, which makes sense.

If this direction is worth revisiting now, I have access to a local RTX 5090 setup and would be happy to help with a small first step rather than treating it as a full feature push.

Would a first external contribution be more useful as a feasibility / build-validation pass around the ZLUDA path, or as a very narrow runtime experiment once you feel the timing is better?

[chokobole]

Hi @peter941221 Thanks for the interest! Recently I've been focusing on compiled approach for gpu code generation with a different angle. Unfortunately, i or the team no longer maintain this project.. Instead, if you still are interested, could you take a look at https://github.com/fractalyze/?

[peter941221]

@chokobole hi, sure, the new project looks really interesting! I'll take a look at that.