Installation of K40c with Quadro 600 in Z620

Accelerated Computing CUDA CUDA Setup and Installation
#1

Hello,

TLDR; question is : What is the procedure to install both drivers for the GPGPU Tesla K40 and for the display adapter Quadro 600 ?

We had a GPU computing workstation since 2014 running a Quadro 600 + Tesla K40c under Windows 7 enterprise. But we were forced to format and reinstall Windows 7, and now we are unable to make both card operating properly.

Our simulation software require to have CUDA 10.0 driver installed with the K40c. This is provided by the nvidia drivers 411.98. In the other hand, to make the old Quadro 600 card working, we also need the 377.83 graphics driver.

The main issue is the following: if we install drivers for the Tesla, driver for Quadro 600 is overwritten, and graphic card will not work properly. At this point, nvidia-smi will display an error message “blahblah”, and we can’t use our simulation tool that fails to discover the K40c.
If we install driver for Quadro 600, it “removes” the CUDA 10.0 driver and nvidia-smi will show on properly, but without cuda version. And our simulation tool will not run properly.

We don’t have directly the control on the source code of the simulation tool, so we can’t seek in this direction.

I already contacted nVidia online chat support, they were not able to answer to me.

Regards,
Maël

#2

This is not a nice solution but I solved the issue by replacing the old Quadro 600 by a GeForce 1050 Ti. This is a newer & mainstreem card as compared to the old Quadro pro one. So the drivers are less incompatible each between the others.

The procedure:
I sintalled GeFore 1050 Ti with 441.66 driver, then I re-installed Tesla driver 411.98. Now both cards are detected, recognized, working, and active drivers are 411.98 with CUDA 10.0.

I didn’t learn anything by doing this. But at least we can restart our activities. nVidia drivers incompatibilities are a pain in the… That’s incredible.

#3

Quadro 600 is no longer supported by any recent NVIDIA drivers or CUDA toolkits. The card is obsolete. Replacement is the recommended approach. You cannot load 2 separate NVIDIA drivers at the same time, whether on windows or linux.

#4

I understand that answer, but even if the card is “obsolete”, it do the job. We only need a graphic card in order to boot the computer, it’s never used locally, only using remote desktop. So:
1- it’s stupid for us to spend money in a graphic card that will never be used.
2- this setup (Quadro 600 + K40) was operational and running on the same computer from 2014 up to now. So a compatible driver was existing in the past. My only question is to know which driver version was it, and where can I download it.

#5

That 377.83 graphics driver should work with both your K40 and Q600. What it won’t do is support CUDA 10. The last version of CUDA supported by that driver I believe would be CUDA 8.

You can download drivers using the wizard at http://www.nvidia.com/drivers

Just follow the wizard path for the Q600. The driver the wizard offers should work with the K40 also.

Or for the highest confidence, just install CUDA 8. The bundled driver with CUDA 8 will work with both GPUs. And CUDA 8 is the last version that works with the Q600. You cannot use CUDA 10 with this approach, however. To use CUDA 10, as stated already, the recommendation is to remove the Q600 and replace it with a newer GPU, which you have already done.

#6

Hi,

I already know how to use the website’s wizard to get the drivers from the target card. But it only shows the latest compatible driver. In my case, the archive driver section at https://www.nvidia.fr/Download/Find.aspx?lang=fr don’t show enough possibilities to get older releases.

However, when you say “just install CUDA 8/10”, do you state for CUDA toolkit ? Because as far as I know, it doesn’t install drivers, but just a library of DLL usable for development purposes.
We don’t want to use CUDA with the Q600 (obviously, the card is too old to rely on for computational purposes), we only want CUDA 10 for the K40c, and a graphic driver for Q600. Both running together on the same computer.

Again, that’s possible because already did on the same computer, and running since 5 years ago. But I can’t find the way to reproduce this situation today.

#7

All CUDA installers at http://www.nvidia.com/getcuda, including those linked via the legacy releases page, include a bundled driver installer.

That is not possible. With respect to CUDA 10, it has never been possible. Support for Fermi generation GPUs has been dropped.

#8

You say it’s not possible, but I confirm it was possible and working fine during the last 5 years, on the same machine, with same cards, fully same hardware and same operating system version. Without any “non-recognized” devices or so on.

So let’s stop here, we’ll throw away the outdated Q600 and replace it with a the cheapest recent graphic card. Just to make the computer boot. For us it’s nothing else wasting money. At least if we can restart our simulations that’s all we want.