Folding Forum

> Hopefully the developers figure out a way to multitask large work units on large GPUs in the folding core so we won't have to manually use virtualization for that. They've already started discussing it on the OpenMM GitHub.

We're indeed looking into this!

~ John Chodera // MSKCC

John, the 3090 has a out 2,5x the core count of a 2080Ti, and about 7x the core count of a 1660Ti.
Nvidia isn't finished yet, they announced in their video they want to see holograms possible in a few years. That means, they're going to up the tensor core counts drastically, using a matrix design.. I know fah doesn't use tensor cores (half precision), however, if those matrices can still somehow be used in fah, they could seriously speed up folding.
Perhaps using tensor cores to calculate relative position of atoms that matter least in a folding process, and focus full or double precision calculations on the areas of interest of the cluster?
I don't know.. but I think those tensor cores contain massive computational power, and it would be nice if fah was able to harvest it!

I was under the impression that Ampere tensor hardware has some FP32 and FP64 units. Or am I confused?

Geekbench score:

NVIDIA GeForce RTX 3080 Topples TITAN RTX in New OpenCL Benchmarks - The FPS Review

markdotgooley wrote:I was under the impression that Ampere tensor hardware has some FP32 and FP64 units. Or am I confused?

They have. And they're rated as such (30Tflops on a 3080).
However, that same 3080, produces almost 240 Tops on the Tensor cores (163 Tops on the 3070).
The 3090, being only 6Tflops faster than the 3080 at 36 Tflops, delivers a whopping 285 Tensor Tops.
It seems Nvidia keeps their tensor cores at a ratio of 8:1 with FP32 cores on those RTX 3000 series GPUs.

If you could compare the workload that can be done on a single 3090, to a modern day 1060, the 3090 is faster than 8x 1060 GPUs.
If you'd just use the 3090's Tensor cores, it's faster than 65x 1060 GPUs using only cuda cores.
And that's without their regular CUDA cores active!

I have never even seen a bitcoin mining farm of 65x 1060 GPUs!

I know FAH said they don't use 8 and 16 bit (half and quarter precision), but I'm wondering if 32 and 64 bit precision is really necessary for an entire molecule, running only a fraction of a second of video footage.
Like with COVID, researchers were only interested in a few specific focal points of the virus.
All the rest is just blobbing around, and could easily be calculated at a lower precision IMHO...

MeeLee wrote:...but I'm wondering if 32 and 64 bit precision is really necessary for an entire molecule, running only a fraction of a second of video footage.
Like with COVID, researchers were only interested in a few specific focal points of the virus.
All the rest is just blobbing around, and could easily be calculated at a lower precision IMHO...

I would actually argue that we need more precision as that's a more realistic simulation. In the early days, F@H used single precision as that was "good enough" for their work. As time moved on, using double precision was too costly since consumer GPUs don't have that level of performance and professional GPUs are too expensive. Thus, a balance was struck between the single and double to ensure that we get the best of both worlds, i.e. consumer GPUs can be used while still maintaining the scientific accuracy required (https://www.sciencedirect.com/science/a ... 5512003098).

Of course, I am not a researcher but that's an educated guess based on what I have read so far and my understanding (happy to be corrected and learn more about it).

gunnarre wrote:Ryan Smith of AnandTech reports that all GA10x processors have SR-IOV enabled. If that's correct, then all 3000 series GPUs will actually support vGPU. In other words, you'll be able to segment the GPU into several virtual machines - not just pass through the whole card to a virtual machine.

Unfortunately, while the RTX-3000 cards support SR-IOV in hardware, it will be disabled in the driver. They had enabled SR-IOV by mistake. Nvidia will only support vGPU in the enterprise cards. Source https://twitter.com/RyanSmithAT/status/ ... 0140964864

Passing through the whole GPU will still work though.

Is anyone here buying a 3080 early on? What would be needed to get it to fold?

It needs driver support, and it has to be added to GPUS.txt by the developers/server managers after we find out what the device IDs for the 3080 cards are. Presumably Nvidia already has support for the card in the drivers. If we're lucky enough that the developers have a 3080 under Non Disclosure Agreement then you'd just have to delete the GPUS.txt file and restart the client. If not, please find the device ID for your new 3080 like this:

viewtopic.php?f=83&t=26208#p262894

and submit a whitelist request here:

viewforum.php?f=83

For the 3080 to become an efficent folder, though, the OpenMM project needs to figure out a way to run several WUs on the same card in some way. Otherwise the 3080 is going to be under-utilized on lower atom count molecules. If you want to play CP2077 and fold at the same time that should work fine though.

markdotgooley wrote:Is anyone here buying a 3080 early on? What would be needed to get it to fold?

Get it added to the whitelist.

viewtopic.php?p=342723#p342723

toTOW wrote:

rbpeake wrote:Off topic, but how long will it take to whitelist the GeForce RTX 3080 Ampere after it is released on September 17th?

I already got the required data from NV, but they requested that FAH follows the same NDA as reviews, so the new GPUs will be supported by FAH on release.

FAH is ready : viewtopic.php?f=83&t=36117

Question is if there will be good WUs available for that GPU.

And if we can predict Nvidia's next move is either a minor optimization (kind of like the Super GPUs were on the 2000 series), and after that, 33-45% higher performance/core count when they drop to 7nm or 6nm.

Well... I dabbled at trying to get a 3080 ordered this morning. Yeah, that didn't work out with the websites crashing and apparently low supply... o-well TIme is on my side for getting one at some point.

https://www.youtube.com/watch?v=nJBggUfYozY

very good video explaining how the capability of a core and therefore what can be called a core has changed since Fermi.

basically in the beginning a CUDA core could do either INT or FP, then it could sort of do both, now the INT path can also do FP so each core has two FP paths, so NVidia have called what used to be one core two cores. This is not SMT. Each physical core has two separate parallel paths which can now both do FP32 hence why they are being double counted.

I can't wait to see if this is actually beneficial and whether FAH can make use of the immensely wide chip that Ampere appears to be.