PNY GT 610 2GB

[tssmith2002]

This is more of a general question and can certainly post the 100 line log if needed. I have multiple Nvidia Fermi cards folding with 2GB of VRAM and when I discovered PNY GT 610 2GB card in the e-waste pile at work; I grabbed it for another machine. The CPU folds fine but this GT 610 card is not getting any WU and is getting the typical 192.0.2.1 address which is normal for when there are no WU available. However, my other Fermi cards are getting WU at regular intervals.

From the GPU.txt; it doesn't look like GT 610 are blacklisted but it is strange that the same configuration that works for multiple Fermi cards doesn't work with this one.

Any thoughts before starting the log dump?
Todd

[JimboPalmer]

https://www.techpowerup.com/gpu-specs/g ... t-610.c821

While this GPU does Double Precision floating point math,(FP64) it only supports OpenCL 1.1

At a minimum, we think F@H needs FP64 and OpenCL 1.2.

Should the New CUDA Core_22 0.0.13 allow the GT 610, it will be 8 times slower than the lowest 'current' card, the GT 1030. So if someone allows it to run, expect to not complete WUs within deadlines.

[tssmith2002]

JimboPalmer

Thanks for the reply but I have 3 Quadro 4000 cards running which have OpenCL 1.1 and 2GB of VRAM. I understand that these cards are low-powered but they are still crunching WU and easily finishing Wu before the deadline. Yes, this GT 610 card is weaker than those but running flat-out dedicated folder should still finish before the deadline.

Todd

[Joe_H]

The GT 610 is an oddball card. It is a rebranded GT 520 which uses the GF119 Fermi series GPU chip. Using FP64 actually was officially supported by the OpenCL 1.2 standard, these Fermi based cards have hardware support that is more OpenCL 1.1+. Some of the rest of the differences in available operations between 1.1 and 1.2 was supported through the drivers by emulating 1.2 instructions. What support the drivers provided has changed with different versions, and the latest versions only officially support OpenCL 1.1 on these older GPUs. So exactly what each Fermi based card will do and work on varies running under the OpenCL 1.2 linked folding cores.

Within the Fermi line, GF10n chips have less support for the changes between OpenCL 1.1 and 1.2 than the GF11n chips.

Your Quadro has 256 shaders running at 950 MHz compared to 48 shaders at 810 MHz on the GT 610, there are few WUs that will complete on time with 48 shaders. That may change under their reorganization of how work is created and assigned to different hardware. But that is a work in progress at this time, and what the minimum level of hardware and software support will be has not been finalized.

Currently about the minimum card that meets deadlines is the GT 630, and the Kepler based variants branded as "GT 630" (192 or 384 shaders) are more likely to complete within deadlines than the Fermi based GT 630s with 96 shaders.

[tssmith2002]

@Joe_H

Thanks, it seemed odd that the older Quadro 4000 cards were working and the slightly newer GT 610 wasn't. Not a total loss since the box has CPU that are working.

Todd

[JimboPalmer]

It looks like you need at least Kepler to use CUDA, so your card is blocked from both CUDA, and OpenCL.

Drat.

https://en.wikipedia.org/wiki/Fermi_(microarchitecture)

https://en.wikipedia.org/wiki/List_of_N ... 600_series

[Joe_H]

One thing I will add, for the different values of "n" in the GF10n and GF11n series of Fermi, slightly different sets of the transitional operations that were included from the changes between OpenCL 1.1 and 1.2. So your Quadro 4000 may work on many projects, while the GT 610 mightt not.

This all came to light a couple months ago as some of the COVID-19 related GPU projects were created. They are doing some different calculations than had been commonly used by F@h before, so on some projects some Fermi based GPUs would error out.

[MeeLee]

Even a GT 730 has trouble making the deadlines, if not folding 24/7.
The newest DDR5 model does make the deadline.
But it's kind of a waste. Despite the 730 using less than 25W, to have a complete PC running 24/7 will waste about 120-200W.
The 1030 uses the same power, and like mentioned, is 8x faster (I haven't measured it, but it wouldn't surprise me if it was true).

A 730 takes more than a day, to complete a WU that more modern GPUs finish in 30 min to an hour.

[tssmith2002]

Truly, I understand that this is an older card and not really efficient as these things go. I am interesting in crushing COVID by any means necessary and as I was looking in the e-waste pile at work; I saw it and knew that it was a Fermi card that supported DP. I thought that I could add one more card to my farm but if the differences in Fermi cards and OpenCL 1.1/1.2 cause it not work then it is not a complete loss. I fired up another machine and if I see a cheap GPU or maybe a Xeon Phi card floating around then I can work on getting that in my farm.

Thanks for the help. It would have been nice to get it work but that is what happens when you use older equipment.

Todd

[PantherX]

...maybe a Xeon Phi card floating around then I can work on getting that in my farm...

AFAIK, F@H doesn't support Xeon Phi so you can avoid going down that road.

[Joe_H]

...maybe a Xeon Phi card floating around then I can work on getting that in my farm...

AFAIK, F@H doesn't support Xeon Phi so you can avoid going down that road.

Not explicitly supported, but persons have tested the CPU folding core in the past on some models of the Xeon Phi. Depending on the size of the WU in atoms, there are limits on how many threads can be used before there is no further improvement in processing speed. Domain decomposition problems can be encountered at certain thread counts.

[bruce]

If you have a multi-socketed system, GROMACS can only use one chip (or one GPU, if we enable that option). I suspect that's the reason that the Windows client is limited to 32 threads.

[Neil-B]

The A8 gromacs core happily uses two CPUs in Windows (full analysis proving this shared with PantherX) and works fine for Linux multi cpu so it isn't a gromacs or core issue - simply something in the way the client has been coded which nerfs the ability to send the correct flags to the core when firing it up.

[MeeLee]

The xeon Phi comes in 2 variants,
1- as a PCIE card (like a GPU that needs a blower).
That one isn't supported, as it needs special drivers and OS to make it work.

2- As a CPU fitting in some Xeon sockets.
This kind, FAH will recognize, as it's nothing more than a multi core Atom CPU.

[PantherX]

Oh, I didn't realize that there were 2 different variants. Thanks for informing me

I wonder if the new dGPUs from Intel (https://en.wikipedia.org/wiki/Intel_Xe) would use information from previous products and produce a product that F@H can easily support.