I'm currently running some workloads in a container that has the NVIDIA GPU driver installed. I'm wondering if I also need to install the driver on the host system, or is it sufficient to only have it in the container?
2 Answers
You definitely need to have the NVIDIA driver installed on the host system since it's part of the kernel. The container itself runs on the host's kernel, so you'll want that driver there to ensure everything works properly.
With NVIDIA's container toolkit, you typically only need the driver on the host. It gets mounted into your container, so you can have other things like CUDA runtimes in your container without needing duplicate drivers.
That sounds useful! Just to clarify, do I need to worry about version compatibility between the host driver and what I have in the container?