Try to enable NVidia GPUs on Ubuntu 24.04.1 LTS for Folding

[natprop]

Hello All,

I am new to both folding at home and Linux. I have some old hardware of my own added to some that my ISP was giving away to the community. I want to put this hardware to work and not turn into e-waste.
I tried using VMs on Proxmox but and got folding at home on Ubuntu working but failed to get PCI pass through to work and failed to get containers to work so I installed onto bare metal.

I have successfully install folding at home on , and broken my Ubuntu install 3 times now trying to get the Nvidia GPUs to show up in the web control app following different online tutorials, enabling cuda toolkits, manual Nivida driver installs and more these end up getting me locked at 640X480 and having to use the onboard VGA display and unable to recover. On a fresh install of Ubuntu 24.04.1 LTS I am able to display on all of the GPUs them all and they appear in lspci as working.

I have run
sudo apt install ocl-icd-opencl-dev
and followed this guide
https://docs.nvidia.com/cuda/cuda-insta ... ux/#ubuntu
--This breaks my install and I do not know how to recover back to old driver fix the 640X480 and 1920X1080 with out just reinstalling

I am running
Ubuntu 24.04.1 LTS
Folding at Home fah-client_8.3.18_amd64.deb
--Computer set up is a Supermicro X8DAH+-F-LR motherboard with dual Intel Xeon W5590 processors 96 GB RAM
I have 3 GPUs installed all listed on the GPU white list
2 NVidia Quadro 4000 installed in the 16X PCI slots
1 Quadro FX 580 installed in an 8X PCI slot

0x10de:0x06dd:::GF100 [Quadro 4000 Linux/Mac Edition] 486
0x10de:0x0659:::G96 [Quadro FX 580]

I am able to fold on the 16 CPU threads but can not get any of the GPUs to show and am looking for suggestions on what to try next.

Thank you for your time and help

[Nicolas_orleans]

Hello

Quadro 4000 = GF100 = Fermi generation
Quadro FX 580 = Quadro FX Tesla generation

I don't know if all pre-Maxwell graphics card are still supported by the project's current cores, see viewtopic.php?t=36972 that refers to GF10x cards not being supported by core22, then folding (2021 thread) with some possible whitelisting implications. Could be worth to see if their PCI id numbers are whitelisted.

Best regards

Nicolas

[Nicolas_orleans]

In practice I could suggest to check if your devices are whithelisted in the /var/lib/fah-client/gpus.json file

For example I do see in my local json file:
- entry 776, GF100 [Quadro 4000 Linux/Mac Edition] 486, PCI Vendor Id 4318 / PCI Device Id 1757
- entry 758, G96 [Quadro FX 580], PCI Vendor Id 4318 / PCI Device Id 1625

BUT I don't know how to know if they are supported / unsupported (type value 0,1,2 ?)

[natprop]

Nicolas,

Thank you for the help looking at the other thread I may just be out of luck on this one. Was a fun problem just kept hitting walls, I have a couple of GTX 560 TI cards will give those a try next.

Cheers

[Joe_H]

I don't know if all pre-Maxwell graphics card are still supported by the project's current cores,

Kepler based Nvidia cards are still supported, but only use OpenCL instructions for folding processing. The current GPU folding cores include CUDA library support that covers Maxwell up to currently released Nvidia GPUs. It would require creating and maintaining a separate folding core to be able to support CUDA on the older Kepler cards.

The current requirements for Nvidia are support for OpenCL 1.2 or later, and I forget the minimum CUDA version. But Maxwell based cards are the oldest that can use the CUDA pathways in the folding core. Tesla, Fermi, and older only support OpenCL 1.1.

In practice I could suggest to check if your devices are whithelisted in the /var/lib/fah-client/gpus.json file

Easier to read than the .json file is the GPUs.txt file still maintained for use by the v7 client. It can be downloaded from - https://apps.foldingathome.org/GPUs.txt. The .json file use decimal values, many utilities report the hexadecimal PCI ID values. In the case of the two cards mentioned here:

Code: Select all

0x10de:0x06dd:::GF100 [Quadro 4000 Linux/Mac Edition] 486
 
0x10de:0x0659:::G96 [Quadro FX 580]

The absence of numbers between the colons separating the fields after the PCI ID and before the description field indicate the card is unsupported or blacklisted. An example of a supported Kepler based card looks like this:

Code: Select all

0x10de:0x0fc8:2:3:GK107 [GeForce GT 740] 762

[natprop]

Joe,

Thank you for the help. The info about the missing numbers between the colons, and non-compatibility is good to know. I thought they just needed to be on the list. Looks like I need to find some different cards.

Cheers
Nathan

[Nicolas_orleans]

Many thanks Joe for sharing all this knowledge !
Nicolas

[toTOW]

To be safe in the near future and to optimize efficiency, never aim below Pascal chips.

[natprop]

Update:
Thank you all for your input and help. I pulled out the two Quadro 4000 and the Quadro FX 580 cards and replaced with two Quadro K5200 cards I picked up the pair for under $80 and it just works. Took a month of banging my head to get here but we got here, Happy Folding All!