Arch Linux

Forgot to add: This is some kind of regression, it worked in the past also with the packaged drivers by Arch.

This really does not give a lot of information, and foldingathome isn't an official package... is there any way to provide more debugging information for foldingathome in order to see what the nvidia package might not be providing that the prebuilt driver does?

Yep, have to stay on nvidia-dkms 384.90 because of this. foldingathome beta (7.4.16) is also affected.

It is interesting that prebuilt driver works, I thought that was a regression in 387.34, not distro-specific. I'll try the prebuilt driver now.

The Arch package for 387.34 appears to work for me running folding@home. I've never used it before but I see it show up in nvidia-smi as using the GPU.

I've tried nvidia-dkms 387.34-1 from Arch repos once again, and I've got to admit, foldingathome works fine with it, my previous problems were caused by running this script on startup (downclocking GPU memory):
nvidia-settings -a "[gpu:0]/GPUMemoryTransferRateOffset[2]=-2000"
This command worked on 384.90, but running it leads to immediate freeze on 387.34 or 390.12 (beta from AUR).

So, foldingathome works for me both on extra/nvidia-dkms 387.34-1 and aur/nvidia-beta 390.12. I've got some other issues on beta (sometimes I get a black screen after X startup), but that is possibly related to my multi-GPU multi-monitor setup and folding still works in this case.

Dietrich Hallentforden (aufkrawall):
AFAIK, you need to have both opencl-driver (opencl-nvidia) and opencl-icd-loader (ocl-icd) installed to make folding work on GPU. Also, some people say you need to generate xorg.conf to make foldingathome detect your GPU correctly (https://bbs.archlinux.org/viewtopic.php?id=170260).

Hmm, okay. Sounds like it isn't a good idea to downclock the GPU memory. ;)

Is this something you believe should still work (and is hence a generic nvidia bug) or is this -ENOTABUG?

I don't think that's a bug, nvidia just have some standard overclocking limits for all cards (from -2000 to 2000 MHz), they've never said all values in that range are safe for all GPUs. I just didn't expect that driver update would affect stability at same clock speed (but I should), now I'll be more careful. Looks like 1080 Ti is completely fine at -2000 memory downclock, while 1070 freezes past -1500. And memory downclocking is kinda silly anyway, it doesn't matter much in terms of temperature or power consumption.

Thanks for your efforts.
I am not sure if ocl-icd was installed the last time I tested, it's not listed as a dependency of foldingathome (AUR) or the Nvidia driver packages (so likely it wasn't installed). Dmitriy, could you recheck if it really makes a difference?
If so, it would make sense to add it as a dep of opencl-nvidia, I guess. :)

I unfortunately can't test this currently since I don't have an Nvidia card anymore. However, proprietary AMD opencl package (opencl-amd, AUR) definitely has it as a dep.

Hmm, I've rechecked that and apparently ocl-icd isn't needed to make foldingathome work with Nvidia GPU on Arch. I must have mixed it up with similar problem on Debian, where you had to install opencl-icd.

Is there anything to be done here from a packaging perspective?

Well, the reason must be in the packaging, since it works with manual driver installation. However, due to the lack of hardware, I can't provide any more feedback for quite some time. So I wouldn't mind if this got closed.