Quadro 4000

[Mouldotron]

Hi!

Got a few Quadro 4000 cards here, and some spare slots to put them in. They are quite old hat now but would be nice to see them contribute even if in just a small way.

They are strange beasts - Windows 10/64 doesn't want to know and can't seem to get Nvidia drivers to do anything with them. FAH client recognizes them as Quadro 4000 Mac Edition (?) but shows them Disabled (they are whitelisted in the GPU list).

Has anyone got one of these cards working in Windows? I don't believe my ones are mac versions, they have HP part number stickers on for one thing.

[JimboPalmer]

Welcome to Folding@Home!

Your Quadro 4000s are not a supported GPU. A GPU must support 64 bit (Double Precision) Floating Point math (which they do) and the OpenCL 1.2 or above Protocol, which they do not.

https://www.techpowerup.com/gpu-specs/q ... ition.c899

Your 2011 cards are a architecture called Fermi, which has very spotty support in general. I think the first Fermi cards (stating with a GF10) are unsupported, but the second set (which start with a GF11) are semi-supported. Your card is GF100.
https://en.wikipedia.org/wiki/Fermi_(microarchitecture)

Sorry.

https://www.nvidia.com/Download/driverR ... 0177/en-us
May be the last driver for the Quadro 4000 (non Mac edition)

[Mouldotron]

Ooooh ok. Well that's too bad! But I've another system with a 1050ti in, that only does 2D work, maybe I could put one of these old 4000's in that and swipe the 1050ti for folding.

Really useful to have the link to that driver! I'll try it tomorrow. I couldn't find a driver anywhere.

Yes they are pretty old cards now. Wasn't expecting much from them even if they did work!

Really appreciate your help, thanks so much

[bruce]

The fundamental problem is that the NVidia code distributed with FAHCore_22 has some limitation saying Fermi is not supported. I know John has communicated with somebody at nVidia but apparently there's no support for any Fermi, even Fermi-2. Unless there's some future support from nV driver support folks, there's really nothing FAH can do.

Are you able to give one to somebody with a Mac and can they use it?

[Mouldotron]

Thank you Bruce, very grateful for your wisdom on the matter. They are small beans anyway.

I guess the newer K4000 is still viable, I'm hoping to acquire some soon and could put them to use.

Edit: yes I could give them to a mac user, if it will work for them. I'll ask about

[bruce]

Ooooh ok. Well that's too bad! But I've another system with a 1050ti in, that only does 2D work, maybe I could put one of these old 4000's in that and swipe the 1050ti for folding.

That's a good plan provided you can still do that 2D work with the Quadro.

Be sure to
1) install the hardware.
2) Download and install the driver from nvidia.com. (I use the latest studio driver)
3) Reinstall FAH's client.

The 1050Ti does a reasonable job. I have several of them.

[bruce]

yes I could give them to a mac user, if it will work for them. I'll ask about

I don't know if it will fold, but it might. Let me know if you find out one way or another.

[Viehwadschdiggl]

I just registered to let you know that the Quadro 4000 is not dead, yet. It folds without issues under Linux (OpenSUSE Leap 15.2) using driver 390.141.

As you already mentioned, FAH 7.6.21 recognizes it as Mac Edition but I think that's only a naming issue in GPUs.txt because some time ago, it was recognized correctly as 'Quadro 4000'. In recent GPUs.txt the 'Quadro 4000' entry is missing and there's only the 'Quadro 4000 Mac Edition'. But that doesn't impact folding at all.

The issue of not being recognized by Windows 10 is something I cannot confirm, too. My workstation at work is also equipped with a Quadro 4000 and is running Windows 10 x64 without any issues in CAD applications. I just didn't try folding on this machine but I can't imagine why it shouldn't work. Driver version is 377.83, if I remember correctly, I can recheck on Monday.

If you really want to get your Quadro 4000s to fold, there are some pros and cons in my opinion:
(+) They're not as slow as one might think regarding their age, e.g. I have no problems with finishing before timeout even when folding only for a few hours per day
(+) They have a quite low GPU clock (as almost all Quadro cards) which means there's a quite high overclocking potential (mine are running at around 140 to 145 % stock clock rate)
(-) Compared with your 1050ti they are much slower and less energy efficient
(-) Their single slot cooling design is not very powerful, so they will run quite hot and / or loud, especially if not regularly cleaned. You can fit aftermarket cooling solutions as long as you check compatibility with a GTX 480 (done that myself). The only question is if you want to buy a cooler that costs the same or more than the market price of the card itself..

[bruce]

@Viehwadschdiggl Thanks for the report.

My guess is that Linux/MacOS have different drivers than the ones that are being distributed to Windows. Look up the version numbers on both and compare. Hopefully there might be an older version that works on Windows.

Unfortunately that won't solve a cooling problem but consider the linked pictures here viewtopic.php?f=16&t=36967&p=350258#p350246

[vivatobst]

A few months ago I had three of these running for FAH. All of them were recognised as Mac Edition, even though they obviously weren't. The cards support OpenCL 1.1+ (as far as I know), which used to work for FAH last summer. For each of these cards I got 15000 PPD - even with Passkey. On average they needed about 1.5 days to finish one WU. Note that you need a good cooling solution, because these things run HOT. They are freaking loud and are NOT SAFE TO TOUCH WHEN RUNNING. I recommend you to clean the coolers first before running.

EDIT:As far as I know you cannot mount another fan onto the PCB. If you want to research some things yourself: I used to have 2x a PNY Quadro 4000 Rev. 1 (the black ones) and one Rev.3 (the silver one). https://www.ebay.de/b/NVIDIA-Quadro-400 ... n_25572961.
All of them were run on a Windows 10 OS that was released back in October2020. They were fed by a Phenom II x4 965 CPU on a AM2+ Motherboard if that matters.
As JimboPalmer mentioned in his post I used the latest available driver 377.83 WQHL. It was kind of a hassle to set up the dual GPU configuration, as Windows sometimes didn't want to start. (Did a full DDU and tried with newer and older drivers, but one time it did work with 377.83.
If I recall correctly, I read in FAH Client that it DID support CUDA, but I'm not a 100% sure, because these cards were slow as crap

[Viehwadschdiggl]

@ bruce: Yes, you're right, the version numbers are different. But the supported features of the cards should be the same in Linux and Windows. One of the Linux machines I use for folding is a dual boot config with Windows 7. The Windows 7 driver for the Quadro 4000 seems to be almost identical to the one for Windows 10 - at least version numbers match. If that helps, I can try folding on Windows 7, at least to install and configure FAH. For security reasons I don't want to have Windows 7 connected to the internet longer than necessary.

And thanks for the link. That's a good example for a recommended cooling design. It uses low rpm axial fans instead of the tiny stock high rpm radial (or blower type) fan. For my cooling upgrade I chose the Arctic Accelero Xtreme III which uses a similar design (3 x 92 mm axial fans). For sure it's quite a bit of overkill (a 300 W capable cooler on a card with 142 W TDP) but the results are amazing. Even overclocked the card is inaudible and stays cooler under full load than it was before in idle.
The only downside is the increased size so that it now needs three slots instead of one.

@ vivatobst: Thanks for your input. The Quadro 4000 seems to support OpenCL 1.1 including some features of 1.2. At least this is how I interpret the output of the Linux utility 'clinfo', where it mentions 1.1 capabilities as well as 1.2. I have not enough OpenCL knowledge to go into details but I can post the output, if someone's interested. What I can say for sure, is that the card doesn't support CUDA in FAH.

I agree that the card is hot and slow at least in stock condition but both aspects can be improved. As stated above, you can definitely mount other coolers. Of course there are no official upgrades and there's no guarantee that every GTX 480 compatible cooler fits without issues, since the PCB layout is quite different. One of my machines with Quadro 4000s is also a dual GPU system (together with a 560ti). Under Linux it required a bit of tweaking to set up, but runs very stable.

[Joe_H]

As for the cards being recognized as "Mac Edition", the entries in GPUs.txt are matched by PCI device ID numbers, the text field after the final colon is just descriptive and not used for anything else. Possibly the device ID was the same as a non-Mac version of the card, or there was an editing error in GPUs.txt at some pint in the distant past. Whatever was the first description received for a GU and its device ID is what usually appears. As far as I know the only difference is that Mac Edition cards had firmware to work with the Mac EFI during startup and POST. Some included firmware code for both the Mac and PC hardware environments.

As an aside, if you do have an actual Mac Edition Quadro 4000, there is a small market for them to Mac Pro owners with the 2009-2012 models.

nVidia OpenCL support for some of the Fermi cards used to list them as meeting 1.2. As best as I could find from various hints was that part of that support was in old versions of the drivers as emulated code for some missing instructions. That code was removed years ago from what little I could find. Basically depending on which Fermi chip, more 1.1+ instructions are supported by the hardware than on others, and GF 11n based cards support more than GF 10n cards. The 1.1+ instructions were added to support double precision on these Fermi cards, it officially is not part of the OpenCL standard unit version 1.2.

As for CUDA support, F@h only supports that on Kepler and later based GPUs. This is due to the way the CUDA support is present in the driver and OpenMM, they can't dynamically load a library specific to a given card. Instead they are using one that supports that range in Core_22, and Fermi is not included.

[bruce]

I don't think the emulation software was ever part of the drivers. I have a vague memory of it being embedded in the FAHCore but that required everybody into emulation mode and cost FAH a considerable slowdown. The infrastructure doesn't support selecting two versions of the FAHCores, one with emulation and one without. Thus the hardware has to support FP64

[zotric]

Ooooh ok. Well that's too bad! But I've another system with a 1050ti in, that only does 2D work, maybe I could put one of these old 4000's in that and swipe the 1050ti for folding.
Really useful to have the link to that driver! I'll try it tomorrow. I couldn't find a driver anywhere.
...

Shame about the 4000s. The 1050Ti is not at all a bad card. On my 'cheap' Linux Mint 20.1 box a 1050Ti gave around 300,000 ppd for COVID using a standard NVIDIA driver downloaded in February.
I should think it would be OK on Windows as well. On Windows I've again used the standard NVIDIA driver for an NVIDIA folding GPU (but not the 1050Ti) with no problems at all.
The 1050Ti is not all that power efficient in ppd terms, but on the other hand it does use little power and it is certainly cheaper to run than trying to make a CPU deliver that many points. My14 core Intel chip struggled to get 300,000 ppd though the PC did have an undiagnosed PSU issue at the time (it needed a bigger one!). Anyway, I abandoned wrecking the expensive CPU in favour of using a only a GPU for folding.
Maybe this helps a bit.

[Frontiers]

Basically both Quadro and Team Red pro cards designed for working in different thermal loads when cards working in pro applications: it's not continuous 95-100% chip load during hours of gaming, it's somewhat different mode say roughly 50% of time chip have high utilization and other roughly 50% of time have low or no at all utilization. And cooling systems tuned with such loads in mind for not being very loud and cooling down when card in low chip utilization phase. And looking to some pro cards with IR-vision device when card constantly loaded to 90-10% for hours - some pro models may be hotter than good gaming 3 top-flow fans designs with same but differently configured chip and different drivers.
So Folding with expensive workstation cards which at constant full chip utilization, when looking to space around their chips - may run hotter at constant full utitilization than 3 fans gaming cards of good designs - may be not the best idea if looking for long term hardware reliability.

On my 'cheap' Linux Mint 20.1 box a 1050Ti gave around 300,000 ppd

There is Covid19 Project 13477 with very low atom count. When it meets with Linux Mint 20.1 with 750Ti clocked to 1400 MHz with 2 fans tuned to have <70 C at GM107 chip - it gives very good for GM107 395000 PPD from only 45-50 watts per card. After some initial batch of two times slower projects - they may meet for weeks of only 13447 coming with short breaks for 1-2 days to slower projects. So if you have some 750ti's - you can try them for 13447.