Hello. The problem is that there are 2 video cards in my system: Nvidia Tesla K80 and GTX 1650 super. The second one is used for graphics output, since there is no video core in the processor. From the very beginning, I could not make them work without a driver error (yellow triangle), but in the end I was able to use the standard 390.85 driver for Tesla and 427.84 DCH driver for 1650 super. I installed CUDA 11.4, when suddenly there was a problem with the 1650 super driver again. I realized that the driver was being updated and at the next installation (after installing 390.85 and 427.84) turned off the video driver installation. As a result, the system determines cuda (on nvidia-sli does not work, only from the folder C:\Program Files\NVIDIA Corporation\NVSMI), however, the entire computing load goes through 1650 super. Tensorflow does not see the video card at all, and Adobe Premiere Pro slightly loads 1650 super, and Tesla loads 200 MiB and 70% of the power (why …). By the way, nvidia-smi writes the following about Tesla: “Not available in WDDM driver model”.
So, briefly and clearly. I need two video cards not to conflict with each other: the picture was, tensorflow determined, the cuda load was. Can you please tell me what might help me in my case? I will be glad of any help!
Brief information about the system:
GPU1: Nvidia Tesla K80 24GB
GPU2: MSI GTX 1650 super GAMING X 4G
OC: Windows 11 64 bit
You can only use 1 GPU driver per system. The driver support for K80 has ended: new drivers released by NVIDIA don’t support K80. You would need to find the latest available driver for K80 and use that. It should support your GTX 1650. This is the driver to use: Data Center Driver for Windows | 474.44 | Windows 10 64-bit, Windows 11 | NVIDIA
And you cannot use any CUDA toolkit newer than 11.4 with that driver.
From the very beginning I tried to install this driver. It gets up on the Tesla K80, but not on the 1650 super:
"This device is not working properly, because Windows cannot load the necessary drivers for it. (Code 31)
The input/output device is incorrectly configured or the driver settings are incorrectly set."
In addition, I was unable to adjust tensorflow to this driver (I may have tried a little, I will try in the future), and Premiere Pro takes a little less than a gigabyte, stretching the export (with the hardware acceleration parameter enabled) for tens of minutes (16-25).
I’ll try to mess with drivers that are lower than 474.44 and cuda versions. I hope that my raccoon fingers will succeed.
Maybe so far I’ve been able to make them friends with 474.44. In fact, it was trite (for those who might see this topic and encounter such a problem):
Goes to the properties of the problematic graphics card → Driver → Update driver
Find the driver on this computer → Select the driver from the list of available drivers on the computer → Uncheck “Compatible devices only”
Go to NVIDIA → Find your video card with the driver version, the same as the other video card (in my case, it is 30.0.14.7444) and install.
If I can make friends with them with CUDA, tensorflow and adobe, then I will report all driver versions.
As usual, Rocket genius was able to make friends with CUDA & tensorflow. Information for those who suddenly need:
Video driver: Data Center Driver for Windows | 474.44 (although it may not be necessary, because when installing CUDA, it installs the video driver it needs)
Thanks to Robert_Crovella for the tip! I hope that if someone encounters such a problem, they will be able to solve it quickly thanks to this topic without losing a couple of days :D