Bump nvidia driver for CUDA 12.1 support
ddelange opened this issue · comments
/kind feature
1. Describe IN DETAIL the feature/behavior/change you would like to see.
Bump the nvidia driver for CUDA 12.1 support.
- kops source currently on
nvidia-headless-515-server
- ubuntu repo latest available version currently 550
We are currently running our 1.26 cluster configured with DriverPackage nvidia-driver-535
(with CUDA 12.0 support).
Note that we moved away from nvidia-headless-XXX-server
because it does not install nvidia-smi
and some other binaries on the host system, which caused issues for us with some cuda docker images that rely on the host having them available. It's barely any additional disk space used, so that was a quick and effective fix.
When we tried bumping to nvidia-driver-550
on EC2, the nodes stopped registering in the cluster. It was reverted before I could pull logs from the host system I'm afraid. I hope that your CI catches this.
2. Feel free to provide a design supporting your feature request.
Unfortunately, our CI does not run on GPU nodes. We don't have credits to run such tests at the moment...
just double-checked: nvidia-driver-535
already supports CUDA 12.1, which runs stable in our cluster