Folding Forum

MeeLee wrote:

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.

Unless you are deploying a pure eco folding system, using 100% renewable energy, that is not grid tied in any way shape or form. Basically a fire and forget folding solution that generates its own power and doesnt generate any waste besides maybe heat (and even that could be recycled on some level). it might not be the fastest solution, but if it doesnt destroy the planet while doing it, it may appeal to some people.

bruce wrote:I doubt the Xavier is going to work with FAH but I don't understand exactly what it's intended to do. FAH does not use AI processing capabilities. It uses the compute capabilities of the geometry engine in a GPU. It's designed specifically to use the Floating Point resources of a CPU and/or GPUs. Those GPUs that are intended to do AI processing have resources that FAH does not need or use.

The Xavier is not a CPU nor a supported GPU.

To be specific, Xavier is a SBC containing a Tegra SoC that include an 8-core Carmel ARM64 CPU and a Volta GV11B GPU with 512 CUDA cores. While one of the purposes of the device is low power AI processing, a Jetson AGX Xavier is a complete, standalone computer, like a Raspberry Pi on steroids. The GPU has feature-complete graphics capabilities, and along with the CUDA cores, its a baby derivative of the Quadro GV100. The AI functionality coming from the tensor cores and DLAs I agree is not relevant. I am pretty sure now that the GPU is on NVLink, which is why it has no PCI identifier for FAHClient to find.

So far the FahCore_a8 has been folding nicely on the Xavier's CPU cores. I am just happy to give this little thing something to do when I am not doing dayjob work with it.

It's not clear to me how FAH will be able to use the 512 CUDA cores. FAHCore_a* are designed to use traditional CPUs. FAHCore_2* are designed to use CUDA cores on identifiable GPUs.

v00d00 wrote:

MeeLee wrote:

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.

Unless you are deploying a pure eco folding system, using 100% renewable energy, that is not grid tied in any way shape or form. Basically a fire and forget folding solution that generates its own power and doesnt generate any waste besides maybe heat (and even that could be recycled on some level). it might not be the fastest solution, but if it doesnt destroy the planet while doing it, it may appeal to some people.

2 sides of the coin.
A 300W AMD Vega 64 may heat the same room in the same way, as a 300W 3080. But the 3080 will do twice to trice as much of work.
The pricing gap between a VEGA 56, RTX 3070, VEGA 64, and RTX3080 isn't too big

xnupanic wrote:

bruce wrote:I doubt the Xavier is going to work with FAH but I don't understand exactly what it's intended to do. FAH does not use AI processing capabilities. It uses the compute capabilities of the geometry engine in a GPU. It's designed specifically to use the Floating Point resources of a CPU and/or GPUs. Those GPUs that are intended to do AI processing have resources that FAH does not need or use.

The Xavier is not a CPU nor a supported GPU.

To be specific, Xavier is a SBC containing a Tegra SoC that include an 8-core Carmel ARM64 CPU and a Volta GV11B GPU with 512 CUDA cores. While one of the purposes of the device is low power AI processing, a Jetson AGX Xavier is a complete, standalone computer, like a Raspberry Pi on steroids. The GPU has feature-complete graphics capabilities, and along with the CUDA cores, its a baby derivative of the Quadro GV100. The AI functionality coming from the tensor cores and DLAs I agree is not relevant. I am pretty sure now that the GPU is on NVLink, which is why it has no PCI identifier for FAHClient to find.

So far the FahCore_a8 has been folding nicely on the Xavier's CPU cores. I am just happy to give this little thing something to do when I am not doing dayjob work with it.

I think the problem isn't the CUDA cores.
But a rewrite of the FAH cores, to be able to use an ARM CPU to feed the internal GPU.
The GPU is between a GT730 to a GT1030 (which is extremely slow btw), and the GPU isn't yet registered in the whitelist.
But if that same GPU was fed by an x86 CPU, it would have worked just fine.
Current A8 core is only able to find work for ARM CPUs.
For an ARM CPU to GPU workload, a whole core rewrite has to be done.

While the Xavier and Jetson Nano boards aren't the fastest boards, they do have potential.
Especially if in the near future, Nvidia might decide to upgrade their nanos to include Neonverse CPUs (A70 series) with larger graphics cards.

Right now, Nvidia is mostly focusing on AI (lowering from 8bit to 4bit precision even).
Once they got the gist of it their next goal will be to focus on deep learning and neural networks (16bit) on those tiny boards (eg: driverless automation, automated drones, artificial intelligence in a sense of a robot answering you like a person, etc...),
And only after that, could it jump to 32 (FPP) tasks (like folding, bitcoin, repetitive calculations, or whatever else they may be inventing).

MeeLee wrote:

v00d00 wrote:
Unless you are deploying a pure eco folding system, using 100% renewable energy, that is not grid tied in any way shape or form. Basically a fire and forget folding solution that generates its own power and doesnt generate any waste besides maybe heat (and even that could be recycled on some level). it might not be the fastest solution, but if it doesnt destroy the planet while doing it, it may appeal to some people.

2 sides of the coin.
A 300W AMD Vega 64 may heat the same room in the same way, as a 300W 3080. But the 3080 will do twice to trice as much of work.
The pricing gap between a VEGA 56, RTX 3070, VEGA 64, and RTX3080 isn't too big

I'd raise that by asking how much power that solution uses? Could you run it from say a 3kw solar array based in California getting x amount of sunlight per day? If not and that system uses regular electricity generated by coal or nuclear how much damage to the environment is the Vega solution indirectly causing? How much carbon was created to make the product?

Its probably possible to extrapolate further and ask questions like if the work gets done quicker, but the environmental damage it causes indirectly creates a new world problem such as a new viral pandemic, can you justify it as a solution to the original/current problem?

I dont have answers for many of those questions and it also just boils down to the person running the project. I know people that fold on green power and i know most that don't. I prefer to if I can, but its not always possible.

Although the thought of harnessing geothermal energy to power a massive folding array is something i have thought of many times over the years, maybe based in Iceland where it is generally cool and possibly using a more low power, green solution aimed at a more harmonious approach to doing dist computing. With a much longer term view to completing work. I also wondered whether you could create a beowulf cluster of enough power from many smaller, low power cpus, possibly arm based that could still compete or even beat modern GPU's while remaining fairly carbon neutral. I dont know the answer to those questions either but the possible answers and methods of creating such a system does intrigue me.

Like mentioned, both GPUs use around 300W under full load. So there won't be any extra load on the environment choosing the bigger GPU.

Even if you are folding with only a 1030 (25W), the total amount used by your system will be ~125W, and a WU takes 20-24 hours to finish.
The RTX2080Ti / RTX3070 uses ~200W, with a total system power being closer to 325W (~3x more power), but it finishes most WUs in 1 to 2 hours (10-20x faster).
To get that same work done with a 1030, you'd need 10-20x more systems, each sucking 125W, which is anywhere from 1-2kW.
Maybe you won't pollute more, but together with all the other people combined, all thinking the same, they will pollute more to do the same work...

I'm new to this folding thing. My new Mac with M1 ARM SOC (in translated mode) uses very little power, makes very little heat, and seems to crunch well. Is there a way to grade the performance per day that is comparable? For example is 68,785 points per day on Mac Mini M1 per web view good? I haven't measured power but I understand at full CPU they use 25 watts.

The 68k per day on the M1 is ok for cpu units. Especially considering the power level and core count. I get closer to 125k give or take with 6 or 10 cores (can't remember exact number offhand) on desktop. But in the end all science completed is valuable and your contribution is appreciated.

MeeLee wrote:The GPU is between a GT730 to a GT1030 (which is extremely slow btw), and the GPU isn't yet registered in the whitelist.

Nvidia's code for "extremely slow" is GTnnn (without the X).

Does the GPU show up on lspci?

Slightly Offtopic for this thread, but does anyone know if you can use the Compute HAT on a Pi to create a cluster that FAH could use? Has anyone tried?

If you could create a 16 core cluster and have FAH use all 16 cores, i would be interested in seeing some numbers, assuming its even possible.

F@H doesn't support Beowolf cluster... instead, you can run an independent client on each "node" and see how they perform.

Here's few links that appear when search for Beowolf cluster... not ARM related but can give you some ideas:
viewtopic.php?f=61&t=20246
viewtopic.php?f=80&t=23279&start=30

PantherX wrote:F@H doesn't support Beowolf cluster... instead, you can run an independent client on each "node" and see how they perform.

I did this some time ago. It was boring.

yup, separate systems work, however the Pi being made at 28nm, isn't very efficient.
A cluster of Pis will use about 50W, and they'd perform not much faster than the slowest of GPUs.

So, if you want more power than a one-processor machine then a GPU is the best way to go?

Besides, I couldn't stream "Tenet" on my notebook machine without experiencing interruptions when the F@H was running a simulation. I have only one cable modem. I had to stop F@H before I could stream without interruption. How am I going to use 4 machines dedicated to F@H at one time? Won't they bog each other down?