Is it possible to port this project to ROCm for AMD GPU users using tools like hipify?

[rafrafek]

Hello 👋,

I am not an expert in CUDA/HIP porting, so please excuse me if this is a naive question! Is there a straightforward or automated way (e.g., using tools like hipify) to make ppf-contact-solver work on non-NVIDIA GPUs without requiring heavy manual rewriting?

I am using an AMD Radeon RX 9070 XT (RDNA4) graphics card. It handles general GPU workloads like Ollama LLMs well, so I was wondering if hipify or ROCm translation tools could work for this project's CUDA kernels.

Thanks for your time and for open-sourcing this project!

[ryichando]

At the moment, I haven’t tried it myself because AWS does not offer the latest AMD GPUs, but I guess it will need some manual work.

We only work on AWS, and we do not have plans to ship for AMD devices simply because we don’t have the need ourselves.

If AWS provides AMD options, and if they turn out to be much more reasonable than NVIDIA hardware, we may change our mind.

[rafrafek]

Thank you for the quick and clear response! That makes total sense regarding AWS infrastructure and internal priorities.

[ryichando]

@rafrafek Hi AMD (ROCm) now shipped.

[ryichando]

Let me add one thing: ROCm compiled is surely shipped but they are not tested on real AMD Radeon GPUs because I do not own AMD GPUs. I tried my best but it's more likely than NVIDIA CUDA that something can go wrong.

[rafrafek]

Hi! I'd be happy to test this on my RX 9070 XT. Do you have an existing way of benchmarking GPU performance so that the results are comparable with Nvidia GPUs?

[ryichando]

Thank you! We have a table for NVIDIA L4 here at

https://github.com/st-tech/ppf-contact-solver#-budget-table-on-aws

It's possible it might not even launch. In that case posting any possible logs help.

[rafrafek]

Thanks! I tried running the "trapped" example so I can compare the results, but I can't find the output in the same format. I can only find this table in the JupyterLab:

Frame	Time/Frame	Time/Step	#Contact	#Newton	#PCG
300	376ms	113ms	249	1	32

Maybe I'm running it wrong. I downloaded the ppf-contact-solver-2026-09-27-20-44-win64.zip, unpacked it, started with start.bat then opened the localhost 8080 and then opened the trapped.ipynb and just clicked Run -> Restart Kernel and Run All Cells...

I monitored the GPU usage in the Windows task manager and it was between 55 to 65 % during the process. I think it produced all 300 frames correctly. So it works :D

[ryichando]

Sounds like it's working anyway. Thank you for the try!