Folding Forum

Support for ARM has been announced - https://foldingathome.org/2020/11/24/ne ... m-support/. It is limited to CPU folding, I know of no plans to implement GPU folding on ARM. For hardware the processor needs to be a Raspberry Pi 4 or better, OS needs to be Linux. Some additional information was posted here by one of the persons who tested the ARM client - https://www.raspberrypi.org/forums/view ... 3#p1766983.

As far as processing power, I have no idea how the ARM CPU on a Jetson Xavier compares to a Raspberry Pi 4. So it may or may not be usable.

The Jetson Xavier GPU is not so powerful anyway. But maybe there are plans to use iGPUs and such slow GPUs as accelerator for CPU slots? Then also an ARM CPU FAHcore could use the Jetson Xavier GPU as accelerator?

1- The Jetson/Xavier series deveopment boards have GPUs that are slower than a GT 730. Essentially not (or barely) making the deadline.
2- There are driver issues. The only supported OS that has OpenCV (not openCL) enabled would be Nvidia's own OS, which is based on Debian (i believe).
Drivers don't seem to be ported to other Linux versions yet (like Ubuntu or so).
3- CPU folding will probably not be possible on these boards due to using the Cortex A50 series cores, not A71. (Raspberry Pi 3-like cores, not Pi 4-like cores).
Software between cores are incompatible with one another.

I think on board graphics are on the ARM, but I figured it has plenty of CUDA and Tensor cores so maybe they have something that'll work with this. Mainly because the wattage is so low if you could get some good results it would be worth running 24/7

I have a 3950X and even if I assign 12 threads to folding my 3070 takes a performance hit. I only fold on my GPU's now.

My point is that it seems the higher spec GPU's use a lot more than just one thread/core of the CPU.

Tuvar wrote:I think on board graphics are on the ARM, but I figured it has plenty of CUDA and Tensor cores so maybe they have something that'll work with this. Mainly because the wattage is so low if you could get some good results it would be worth running 24/7

FAH favors faster hardware.
Considering the cost of these devices, it's not going to be a good investment just for folding.
You're better off with a decent GPU.

Already have one on my desk is why I was wondering. Lately I've not had time to fool with it so I was hoping I could put it to work is all.

The ARM that nVidia went with is their own design and it's pretty fail. Works for it's intended purpose with is to power the GPU cores, but a RPi 4 is better. I happen to have one of those as well lol. Even have a Raspberry Zero, but that would be a laughable CPU.

Potentially the GPU on the Jetson Xavier has enough CUDA cores and processing power to be usable. The Tensor cores are not usable, F@h does not use FP16 calculations. Against that is the lack of support in the software F@h GPU folding cores are based on, and the mentioned limited driver support. A lot of development work would be needed before F@h could use the GPU, and if the return on that was limited to just a few systems might not be undertaken.

Joe_H wrote:Potentially the GPU on the Jetson Xavier has enough CUDA cores and processing power to be usable. The Tensor cores are not usable, F@h does not use FP16 calculations. Against that is the lack of support in the software F@h GPU folding cores are based on, and the mentioned limited driver support. A lot of development work would be needed before F@h could use the GPU, and if the return on that was limited to just a few systems might not be undertaken.

It only has 128 cores, and runs at a very low frequency (600-900Mhz).
Essentially 2,5x slower than a GT 730 GDDR5 in Cuda Core count.

MeeLee wrote:It only has 128 cores, and runs at a very low frequency (600-900Mhz).
Essentially 2,5x slower than a GT 730 GDDR5 in Cuda Core count.

You are thinking of a different Jetson processor, there are 3 different ones under the Xavier codename. The 2 lower end ones have 384 cores and would be comparable to the GT 730. The full chip has 512 cores, about the equivalent of a GTX 750.

Specs can be read here - https://www.techpowerup.com/gpu-specs/j ... -gpu.c3232, or here - https://en.wikipedia.org/wiki/Tegra#Xavier.

This is the one I have and at 15W if it had the ability to fold it would be nice. I think it's about to go on E-bay though. I just don't have the time to mess with it, and when I do have time I have my RTX 2080 as well as a RTX 3080 on order...... It might make it by April at this rate. At least my parts for the Folding Rig should all be here by tomorrow. Waiting on the CPU and RAM.

https://www.techpowerup.com/gpu-specs/j ... -gpu.c3642

Tuvar wrote:This is the one I have and at 15W if it had the ability to fold it would be nice. I think it's about to go on E-bay though. I just don't have the time to mess with it, and when I do have time I have my RTX 2080 as well as a RTX 3080 on order...... It might make it by April at this rate. At least my parts for the Folding Rig should all be here by tomorrow. Waiting on the CPU and RAM.

https://www.techpowerup.com/gpu-specs/j ... -gpu.c3642

Based on the relative performance comparison, if you got it folding it should meet all the deadlines and such, and I'd guess be worth 60-65K PPD. It's not much slower than the Vega 11 on my machine.

That being said, I don't mind letting this machine run 24/7 to fold. But if I had quick GPU's like you do, I think the energy would be better spend letting them fold more often. It should be even more power efficient in the long run, and why set up a slow rig to fold if it's not very efficient?

I've been folding on my 2080, I just though jt would be kind of nice to use a Jetson is all. I ebayed it though since I've had no time to use it for its intended purpose. That said I did get the multi gpu folding machine up with four 1060 mining cards. Ill add the last two today or tomorrow and see if they work or if the CPU's lack of threads tank them. Ill post all of what I find here when its done though. So far the four are putting out about 1.2M pdu at just under 300 watts at the wall.

Adding to the discussion, it is odd that FAHViewer hits the GPU but FAHClient does not. Simply rotating the protein in the viewer clearly shows the GPU activity via jetson_stats. I can't help feeling there is some magic configuration option eluding me.

Synchronic wrote:Adding to the discussion, it is odd that FAHViewer hits the GPU but FAHClient does not. Simply rotating the protein in the viewer clearly shows the GPU activity via jetson_stats. I can't help feeling there is some magic configuration option eluding me.

FAHViewer uses OpenGL libraries to display the graphics images, for GPU computing the client uses a folding core which is based on OpenCL or CUDA in the case of nVidia GPUs. They are different coding languages, for different purposes.

FAHClient itself does not do any WU processing, it takes care of downloading WUs and uploading the completed work, downloading CPU or GPU folding cores as needed, and starting and stopping processing of a WU using those folding cores.

For a Jetson to do GPU folding there would need to be a folding core coded to use it, and driver support to allow that usage for computing as opposed to video display.