Nvidia 3080 in ARM (Ampere) Server

[thaddeusf]

All,

I'm running Ubuntu 24.04.2 LTS with CUDA driver 570.1254.06. FAH client is reporting as expected but I'm never getting past 'Assign Wait' on the Nvidia card. The CPU is running. In the logs I see:

03:09:45:E :OUT20:HTTP_SERVICE_UNAVAILABLE: {"error":"No matching core: [\"184.x.y.z\",\"8.4.9\",4194648133750027833,\"ARM64\",\"LINUX\",\"Thaddeus\",255437,1478361088,\"FAH\",\"any\",0,0,0,\"ARM\",0,8,1,1,4318,8710,\"8.6\",\"12.2\",\"\",\"NVIDIA\",8,\"158.130.118.23\",18238,\"\",1744772985,1744772985,\"justinjm@seas.upenn.edu\",\"highland1.seas.upenn.edu\",443,38923,\"0.0\",\"0.0\",36,0,0,\"6.8\",\"RWfIA4dbHKRArrPNd6z5DcSa7Ekp94P5MY6FpdaREE8\",\"v6iQPO2a4S+tEOBoSQof9wN3FDOurmSmKDmIhkpcrEQ=\"]"}

Any idea why?
Thanks

[Joe_H]

Unfortunately there are no ARM based GPU folding cores for the client to download and use on your Ampere server. Currently only CPU folding is supported on ARM systems, the GPU cores are only compiled for 64-bit Intel/AMD for Windows and Linux.

[arisu]

There are no cores for GPU folding on ARM yet. You can fold on the CPU but not the GPU.

[calxalot]

https://github.com/FoldingAtHome/fah-cl ... issues/139

[thaddeusf]

That's disappointing. I was pretty excited about pushing this server.

Is this something being worked on at all? I had thought FAH was fully supported on MacOS ARM...

[arisu]

In theory it would be something that could be done within a few minutes. Barring any unexpected problems, it would take nothing more than recompiling the GPU core for ARM. Unfortunately the cores are not open source for reasons that I find silly (it's meant to make cheating harder ostensibly, but anyone who would know how to cheat from the source code would be able to cheat without the source code).

I wonder if QEMU's architecture emulation would help here, or if the performance impact would be too great even for the component that runs on the CPU. You could give it a try. Performance could range from normal to abysmal. Or a static binary translator would probably retain sufficient performance, but that would involve modifying the core binaries which is against the EULA.

Folding on ARM CPUs is fully supported btw. It's just folding on GPUs that are on ARM machines that is not.

[Joe_H]

Is this something being worked on at all? I had thought FAH was fully supported on MacOS ARM...

If you define "fully supported" as CPU folding only, then yes F@h is fully supported on Apple M-processor Macs. But no GPU folding available, and looking unlikely for the foreseeable future.

I could go into the history, but it involves a number of problems and changes in direction by Apple. Originally F@h intended to have support for GPU folding on OS X, now macOS, but that failed to get out of beta testing due to a bug in Apple's OpenCL code that was nit fixed for about 2 years.

[arisu]

GPU folding on an ARM server is a lot easier (I'd say trivial for whoever has the source code) than GPU folding on ARM Apple, though.

[thaddeusf]

Thanks for the insight. Please put my vote down for ARM Server with GPU support - and I'm sure NVIDIA would be a big fan too.

Thanks!

[muziqaz]

Thanks for the insight. Please put my vote down for ARM Server with GPU support - and I'm sure NVIDIA would be a big fan too.

Thanks!

Nvidia wouldn't