VASP 6 on AMD GPUs

[Dankomaister]

Hi,
At the beginning of next year our group will get access to a new supercomputer with plenty of AMD Instinct MI250X GPUs. These are monsters with 128GB of GPU memory and I wanted to ask what is needed in order to run VASP 6 on AMD GPUs. As far as I understand OpenACC is supported on AMD Instinct GPUs, but that doesn’t necessarily mean that VASP 6 will work out of the box, right?

[fabien_tran1]

Hi,

OpenACC in VASP has been implemented and tested using NVIDIA GPUs. I guess that nobody of us has tried with other GPUs. So, we don't know if some changes in the code would need to be done for other GPUs, but actually in the code there are a few lines which seem to be specific to NVIDIA.

[Dankomaister]

Alright, I saw that the new version (6.3) includes optimizations for AMD CPUs.
Any plans to give the ACC version the same treatment for AMD GPUs?

/Daniel

[Dankomaister]

Hi,

OpenACC in VASP has been implemented and tested using NVIDIA GPUs. I guess that nobody of us has tried with other GPUs. So, we don't know if some changes in the code would need to be done for other GPUs, but actually in the code there are a few lines which seem to be specific to NVIDIA.

Since its been almost one year since I asked this I wanted to check in if there has been any progress on making VASP run on AMD GPUs.
This is very relevant since most of the new HPC clusters (Frontier, LUMI, Dardel, Adastra, etc) will/are running AMD GPUs.

[fabien_tran1]

Hi,

Unfortunately, porting VASP on AMD GPUs is not going to happen very soon, in particular since there is currently no OpenACC compiler that would work.

[Dankomaister]

Hi,

As far as I can understand the upcoming GCC 13 release will have support for OpenACC 2.7+ as well as support for the Instinct MI200 series devices (gfx90a) such as the AMD Instinct MI250X GPUs that are used in the LUMI supercomputer. This together with AMD’s HIP will make it possible to port VASP to AMD GPUs, see for example: https://www.lumi-supercomputer.eu/prepa ... ons-to-hip

I suggest making it a priority to port VASP to AMD GPUs as current and upcoming HPC systems will use AMD GPUs. Any VASP users who feel the same should reply to this post to show their interest.

[mmarsman]

Hi,

We are currently exploring several routes to support AMD GPUs.
Hopefully we will soon be able to compile and run the OpenACC port on AMD using either the Cray compiler or possibly GCC-13.
OpenACC support in these compilers might not be completely mature yet, so I do not expect things to work out-of-the-box for some time.
Another route we are exploring actively is the use of OpenMP offloading. This is still work in progress though.

[Dankomaister]

Ah great looking forwarded to see progress on this front!

[Dankomaister]

Has been a year now.
Any updates on supporting AMD GPUs?

[fabien_tran1]

Hi,

This is still work in progress. AMD has implemented (partial) OpenMP offloading in VASP, that works with their GPUs and the Cray compiler. We plan to integrate their work in our branch, however this will take some time.

[leszek_nowakowski]

Hello, I'm here solely to express my keen interest in porting VASP to AMD accelerators. I'm keeping my fingers crossed for the progress of the work!

[leszek_nowakowski]

Hello,

It's been half a year now since last news and new VASP6.5 version is about to be released, will it work with AMD GPUs?

Regards, Leszek

[michael_wolloch]

Hello Leszek,

thanks for your interest in the AMD port of VASP.

Unfortunately, there is not much news to tell you. We are still working hard on porting the code to AMD GPUs and making progress. However, there is no release date yet for a working port. Additionally, there is also no fixed date yet for the next VASP release.

Cheers, Michael

[fabien_tran1]

Hi,

Unfortunately not very soon, but a few of us are working quite intensively on that (using the OpenMP offloading).