Folding Forum

This is interesting, since the new 21.40.1 drivers for Linux seem to be quite different.
https://www.amd.com/en/support/kb/relea ... ux-21-40-1

But the installation instructions say that they support OpenCl.
https://amdgpu-install.readthedocs.io/e ... all-script

I can try them out later for an RX 570 on Ubuntu 20.04.3.
Hopefully someone can get to them earlier.

Asghan wrote:With the RadeonPro 21.40 Beta driver AMD now confirmed that OpenCL will be dropped in favor of HIP.
21.40. does no longer include OpenCL. (see gpuopen.com)

Please provide a specific link, I suspect you are misreading a notice for a different software package than the drivers and OpenCL support.

I think they are referring to this article - https://gpuopen.com/blender-cycles-amd-gpu/

It seems like it could be misread as AMD dropping support for OpenCL.
It's not.
The way I read it, Blender uses a rendering engine called Cycles X. The article is claiming an upcoming version of Cycles X will be dropping support for OpenCL and is switching to HIP.

Blender is not even an AMD product.

The 21.40.1 drivers work for me on an RX 570 (Ubuntu 20.04.3). I am folding on P18431 at 356 kPPD, so that is OpenCl enough.

I guess you might be right, let me check this again.

But from my side the even more important question (since OpenCL performance on AMD is very unreliable) :
Are there plans for an AMD HIP migration? This should deliver more performance on the same hardware (and hopefully more CONSTANT performance than AMD delivered the last months. Since the driver after 21.3.2 the OpenCL performance is just fucked up.)

Just as light reading ... https://techxplore.com/news/2021-04-ope ... uters.html ... https://github.com/openmm/openmm/pull/3338 ... I guess once AMD HIP is fully incorporated into OpenMM then this will then be a possibility for FaH - but from what I understand it will be for OpenMM to incorporate this as a first step.

Neil-B wrote:Just as light reading ... https://techxplore.com/news/2021-04-ope ... uters.html ... https://github.com/openmm/openmm/pull/3338 ... I guess once AMD HIP is fully incorporated into OpenMM then this will then be a possibility for FaH - but from what I understand it will be for OpenMM to incorporate this as a first step.

Might be easy, as they can take their CUDA Code and use the AMD HIP converter AMD is already providing

I can envision the legal teams of AMD and NV in long, long uncomfortable meetings. NV uses CUDA as a powerful proprietary money-maker.

Asghan wrote:

Neil-B wrote:Just as light reading ... https://techxplore.com/news/2021-04-ope ... uters.html ... https://github.com/openmm/openmm/pull/3338 ... I guess once AMD HIP is fully incorporated into OpenMM then this will then be a possibility for FaH - but from what I understand it will be for OpenMM to incorporate this as a first step.

Might be easy, as they can take their CUDA Code and use the AMD HIP converter AMD is already providing

The second of those links I posted is the openmm dev github ... looks like work is ongoing ... there was another thread in there about initial work from a while back ... makes me feel it isn't that simple/quick ... but time will tell