Driver container + Container toolkit info

[nico88chessa]

Dear all,

I’m starting develop a new application using CUDA.
I’d like using docker as container technology (thanks its advantages); if I’m correctly understanding, I strictly need NVIDIA Container Toolkit installed on host side in order to use GPU inside a container.
After that, what other things do I need on host side?

Can I avoid to install NVIDIA driver and use NVIDIA Driver containers for that?

If I understand, using NVIDIA Driver containers should avoid having nvidia driver on host side; however, for use NVIDIA Container Toolkit, a Nvidia driver is require… so I may don’t understand some things.

Another question: is NVIDIA Driver container (R510 Drivers+) compatible with all Maxwell+ architecture, or just Tesla ones?

Thanks in advance,
Nicola

[nico88chessa]

Dear all,

I’ve read some more info about using nVidia GPU through docker container. I’ve understood that:

• NVIDIA Container Toolkit is mandatory to use GPU inside a docker container
• NVIDIA Driver Containers let NVIDIA driver provisioning thanks the use a NVIDIA Driver container; thanks to this, it is no longer necessary to install a NVIDIA driver on the host, right?

So my questions are:

1. should be NVIDIA Driver container base image OS be aligned to host one? For example, if image is Ubuntu22.04-v535.54.03, should be host OS be Ubuntu22.04 (due to linux-kernel-headers?)
2. watching release driver from this page, it seems that driver must be of the family Data Center / Tesla, so this means that GPU supported are these one:
   
   

Regards,
Nicola