Hacker News Reader

That's still NVIDIA-only, plus you can't even tell a P100 from a K80 for scheduling purposes.

This is the reason node-feature-discovery exists. Kubernetes needs an extensible way to detect and label hardware features: See: https://github.com/kubernetes-incubator/node-feature-discove...

I'd just use a separate NodePool and match your jobs using a selector as appropriate.

So, is gci going to come with the right drivers?