CUDALink allows the Wolfram Language to use the CUDA parallel computing architecture on NVIDIA graphics processing units (GPUs) on Windows and Linux.

To start using CUDALink, the compatible CUDA toolkit needs to be installed by the user. Up to Wolfram Language 12.1, the appropriate CUDAResources paclet needed to be installed instead.

If CUDALink is not working correctly after this setup, the most common root causes are:

• Your GPU is not CUDA enabled or does not have sufficient compute capability. Please ensure these hardware requirements are met.
• You are using a Wolfram product of Version 12.2 or later on macOS. CUDALink is no longer supported on macOS starting with this Wolfram Version as Apple dropped macOS support of CUDA.
• The NVIDIA-distributed GPU driver on your machine is out of date. Please update to the latest driver for your specific GPU:
  
  Windows
  Linux
  macOS up to Wolfram product Version 12.1
  

  Note that your operating system’s automatic driver updating tool may not always have access to the latest version, so it is essential to check NVIDIA’s website directly.

• If your Wolfram product is of Version 12.1 or earlier, the CUDAResources paclet on your machine may be out of date. Update it to the latest version by evaluating from within your Wolfram System:

  PacletSiteUpdate /@ PacletSites[];
  Needs["CUDALink`"];
  CUDAResourcesInstall[Update -> True]  

  Note that this functionality requires internet access.

The CUDALink setup tutorial provides further in-depth troubleshooting information if necessary.

If the issue still persists, please contact Wolfram Technical Support. In your support request, please include a notebook containing the output of the following code:

<< CUDALink`
CUDAQ[]
CUDADriverVersion[]
CUDAResourcesInstall[]     (* Up to Wolfram product version 12.1 *)
CUDAResourcesInformation[] (* Up to Wolfram product version 12.1 *)
CUDAInformation[]
SystemInformation[]