Message Boards Message Boards

GROUPS:

Driver issue with CUDA on Linux?

Posted 6 months ago
1238 Views
|
2 Replies
|
0 Total Likes
|

Hello,

I have Mathematica 12.3 installed on a linux machine (I am using arch-linux). My GPU is a Quadro P620. to work with CUDA I installed the following packages (using pacman as usual)

community/cuda 11.3.1-1 community cuda-tools 11.3.1-1

If I print the environment variables I get the following relevant PATH

NVIDIADRIVERLIBRARY_PATH=/usr/lib64/libnvidia-tls.so.465.31 CUDALIBRARYPATH=/usr/lib64/libcuda.so.465.31 CUDA_PATH=/opt/cuda PATH=/usr/local/bin:/usr/bin:/bin:/usr/local/sbin:/opt/cuda/bin:/opt/cuda/nsightcompute:/opt/cuda/nsightsystems/bin:/usr/lib/jvm/default/bin:/usr/bin/siteperl:/usr/bin/vendorperl:/usr/bin/core_perl

To me it looks everything is well-settled. However, when I try to use CUDALink I get the following

In[1]:= Needs["CUDALink`"]
CUDAQ[]
CUDAToolkitCompatibilityInformation[]
CUDADriverVersion[]

During evaluation of In[1]:= CUDAQ::nodriv: A valid CUDA driver was not found. Refer to CUDALink System Requirements for system requirements.

Out[2]= False

Out[3]= Success["CUDALinkCompatible", 
Association[
 "MessageTemplate" -> "CUDALink is compatible with CUDA Toolkit \
`installedVersion`.", 
  "MessageParameters" -> Association["installedVersion" -> 11.3], 
  "CUDAToolkitVersionMajorDetected" -> "11", 
  "CUDAToolkitVersionMajorCompatible" -> "11", 
  "Data" -> Association[
   "CUDAToolkitVersionMajorDetected" -> 11, 
    "CUDAToolkitVersionMajorRequired" -> 11, 
    "CUDALinkCompatibilityQ" -> True]]]

Out[4]= "465.31.0"&[Wolfram Notebook][2]

I attach a notebook. Somebody can help me?

Attachments:
2 Replies

Frankly speaking, CUDA support is currently in very bad shape on all platforms. There is no, generally reliable, approach for Linux and Windows.

Thank you. I finally solved my problem (I post here the solution since it could be useful for somebody):

I discovered that I did not have any driver problem. My problem was just that in my laptop I have two GPUs: the integrated GPU and the Nvidia GPU. Given the two GPUs, Mathematica was not able to make use of the Nvidia GPU. As far as I understand, the problem should be fixed with some suitable packages (like bumblebee) but in my case I found a easier solution: I noticed that in my BIOS I can turn off completely the integrated GPU. Hence, whenever I want to use CUDA, I simply reboot the system turning off the integrated GPU.

Reply to this discussion
Community posts can be styled and formatted using the Markdown syntax.
Reply Preview
Attachments
Remove
or Discard

Group Abstract Group Abstract