I am currently working on a workstation with Fedora 35. I have my own CUDA 11.7 installed from the official repo, and the CUDA 11.6 bundled with the HPC SDK (I installed the RHEL/CentOS rpm package using repo provided on the download page) sometimes messes things up when setting up the environment variables.
Is it possible to install the SDK without the CUDA bundle, that I configure it to use my own CUDA installation?
Is it possible to install the SDK without the CUDA bundle,
No, it will always come with at least the latest version of CUDA (or multiple versions depending on which package is installed)
that I configure it to use my own CUDA installation?
I’m not sure if we document this, but you should be able to set the environment variable “CUDA_HOME” to CUDA 11.7 SDK and the compilers will use that version.
Note that the upcoming NV HPC SDK 22.5 release will include CUDA 11.7.
Thanks for your kind answer! The CUDA_HOME variable works for me.
I see the 22.5 update today with CUDA 11.7. I do wonder how this CUDA will coexist with my current installation, as they are on the exact same version.
Would you mind helping me with this question?
We ship all the same bits as the CUDA SDK, though the CUDA math libraries get moved to a common location since we include a few more than what the CUDA SDK provides.
Thanks for your answer. Is it possible on the developers’ side to further split the installation package a bit and make cuda a dependency? It will be meaningless to keep two cuda 11.7 on a machine.
And a suggestion: adding gpg verification for the repos will be great.
The person who’s in charge of packaging is away on vacation this week but will ask once he’s back.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.