Ultimate AMD GPU Linux FAH Guide for OpenCL (FAH-Client v8 only)

[foxpy]

In this case, I am once again confused about F@h strategy of allocating hardware resources. There are lots of small CPU work units, which can be completed well before deadline even by an old x86 CPU with theoretical performance of just 100 F32 GFLOPS or 50 F64 GFLOPS (that's i5-3210M which I fold on), while an iGPU with a theoretical performance of 2000 F32 GFLOPS or 125 F64 GFLOPS (that's the iGPU I was talking about in this thread) is already considered outdated. I clearly don't understand something

[muziqaz]

The projects which CPUs are folding cannot be run on any GPUs, that's why they are being run on CPUs. If any project could be run on GPUs, there wouldn't be any CPU projects at all, since GPUs are infinitely more powerful than CPUs.
iGPU job is to display a desktop, nothing else.

[foxpy]

If any project could be run on GPUs, there wouldn't be any CPU projects at all, since GPUs are infinitely more powerful than CPUs.

That's exactly what I was thinking about. Thank you for clarifying!

[arisu]

The projects which CPUs are folding cannot be run on any GPUs, that's why they are being run on CPUs. If any project could be run on GPUs, there wouldn't be any CPU projects at all, since GPUs are infinitely more powerful than CPUs.
iGPU job is to display a desktop, nothing else.

Are projects actually "unable" to work on GPUs even if the researchers built the WU for OpenMM? Or is it that some projects are so poorly suited for GPU architecture that they would actually be slower on GPU than CPU? I always assumed it was the latter but I never found an answer to what would make a particular project better suited for a CPU than a GPU.

[muziqaz]

Unable

[arisu]

What would make a project truly unable to use a GPU?

[muziqaz]

GPU not being able to do certain simulations

[arisu]

How so? Maybe it would be better for me to make a thread about that if the technical answer is complex.

[muziqaz]

How so? Maybe it would be better for me to make a thread about that if the technical answer is complex.

Ask OpenMM and Gromacs on their respective websites.
You are asking as if you think you will be able to solve decades old issues that people dedicated to that cannot solve.

[arisu]

I just like to learn. I don't have any thought that I could change it, but I was curious what could make a simulation physically incapable of running on a GPU. A shader core has a Turing-complete ISA after all, and in theory you could even emulate an entire x86 CPU, SIMD and all, on a single shader (but it would probably fold worse than a Pentium II). That's why "unable" surprises me. Unable to do so efficiently sure. But totally unable? I feel like I'm back in CS!

Please don't take my questions as arrogant. I ask "why" a lot because it helps me learn, not because I think I could find a solution.

[foxpy]

I assume that anything could be simulated on a GPU, but if the task is serial in its nature, performance would be so terrible there is no point to bother with a GPU anymore.

[muziqaz]

There has been some development in Gromacs, where it can do CUDA or nVidia OpenCL (can't remember) now. Which allows CPU workloads to be done on GPU only or on GPU and CPU combined. There are still Gromacs workloads which cannot be done on GPUs, but situation is "improving".
However on our tests it was unstable on nVidia GPUs and Gromacs does not support AMD OpenCL, let alone HIP. So for now fahcore_a9 is CPU only.
OpenMM is used for GPU workloads only in FAH.
You can learn a lot from openMM GitHub and their website, same from gromacs.

P.S. the problem is not serialisation, it's just some molecular dynamics algorithms cannot be done on GPU hardware

[toTOW]

In Gromacs, the most blocking feature is the use of virtual sites, but it's not the only one and I forgot all the details ...

[Joe_H]

Yeah, there is a lot of overlap in the simulation features supported by GROMACS and OpenMM, but not completely overlapping. Many if not most F@h projects could be done with either, choice is up to the researcher and can depend on size and how fast they need results back. CPU folding is a bit more reliable with a lower error rate on returned WUs from what I have seen reported.

[muziqaz]

OK, AMD seem to have listened to me and added rocm-opencl-sdk to 6.3.x branch. I will have to test it though, as some users seem to have some issues with it.
AMD removed opencl-icd-loader from rocm repo, so in order for 6.3 have a chance to work with FAH user needs to install opencl-icd-loader separately.
Well, at least they added rocm-opencl-sdk

EDIT:
here is a solution for 9070(xt) and older cards (start from step 7) to work with latest rocm:
1st: use older card to set your Linux up (this is for 9070, as RDNA4 does not boot into linux yet).
2nd: download mainline kernel (6.13+)
3rd: upgrade amdgpu linux-firmware
4th: add kisak PPA repo for Mesa25 install
5th: install Mesa25
6th: install your 9070 card(s).
7th: add https://repo.radeon.com/rocm/apt/6.3.4 repo to your software sources (replace apt with your distro installer)
8th: install rocm-opencl-sdk only
9th: sudo ln -s /lib/x86_64-linux-gnu/libOpenCL.so.1.0.0 /lib/x86_64-linux-gnu/libOpenCL.so (this makes sure things work in multi GPU system, as well as single GPU) (damn it, step 10 is required every time you restart the PC, wtf)
10th: sudo systemctl restart fah-client
11th: set up fahclient to fold on the GPU(s)
Easy

For everything before 9070 installation step, use this as guide (if you are on older cards, ignore this):
https://forums.linuxmint.com/viewtopic. ... 6#p2603040
If you try doing things while 9070 is your primary GPU, you will not be able to boot into your OS. I heard Manjaro works quite well, and fedora is bringing out the firmware
That firmware link in the guide, you might need to try several times for you to download the firmware