Can't install right CUDA driver version for MATLAB

Hello,

I'm using MATLAB 2020a in a windows 10 VMclient. The machine has a virutal GPU, GRID P40-1Q, and its driver version (as seen in C:\NVIDIA) is 392.05.

When trying to run the 'gpuDevice' command in matlabs command window I'm getting the following error:

"Error using gpuDevice (line 26)

CUDA driver version 9.1 is not supported. Update CUDA driver to

version 10.1 or greater. For more information on GPU supprot, see GPU Support by Release."

I've installed the lastest CUDA Toolkit (currently 10.2) but the driver didn't update. I also tried to install only the driver (both from 10.2 and 10.1 toolkits) but it did'nt work. I tried to install Teslas newest driver but got compatibility error. I also installed Gefore Experience for auto detecting the needed driver version, and it recomanded on another version of the Tesla driver, but that did not work as well.

After my fault attemepts, I've been told by the IT Team who manages the VM that the newest CUDA driver version which supports the virtual GPU is 9.1.

Is there any way to use this GPU with MATLAB?

Thanks!

Answer 1

[Jason Ross]

To use a virtual GPU, the hypervisor must be running the same version of the driver as on the VM. So you will need to update the hypervisor's version of the CUDA driver to 10.1 or greater before the guests can update.

nVidia does have updated driver/virtualization software that supports CUDA 10.1 or greater (currently, "NVIDIA Virtual GPU Software v10.0 through 10.1 Revision 03"), as follows:

• The documentation for the latest version can be found here.
• The P40 is listed as a supported card in that documentation, reference is here.
• The list of driver versions that are available can be found here. You didn't mention which hypervisor is in use but they list the versions for each one there.

[Avital Kleiman]

Thanks for the information. I've talked with the IT team and they told me they are not planning to update the servers driver anytime soon. Is there any walkaround for using the GPU with MATLAB under the current installed driver?

[Jason Ross]

You would need to revert to an earlier MATLAB release, since the GPU/CUDA functionality (and GPUs) supported with each release is closely tied to the CUDA toolkit that was used/available at the time. We have a mapping of MATLAB release to minimum toolkit version published here. If you only have access to a 9.1 version of the driver the latest you can go is R2018b.

[Avital Kleiman]

Just wanted to let you know installing R2018b enabled working with the GPU.

Thank you for your help!