DGX-2 (V100-SXM3/NVSwitch), last supported NVIDIA driver and CUDA version?

Accelerated Computing DGX Systems (DATA CENTER)
1

Hello,

I have a DGX-2 (16 Tesla V100-SXM3-32GB, NVSwitch) running DGX OS 4.14.1 (Ubuntu 18.04) with NVIDIA driver 470.182.03 and CUDA 11.4. I need a definitive answer on what NVIDIA driver and DGX OS releases are actually supported for this system, and whether a driver that supports a CUDA 13.x?

Could you also point me to the specific official pages that state, for the DGX-2 specifically?

Thank you for any guidance.

Best regards,
Shakhizat

2

HI @shahizat ,

You can find the supported DGX OS releases on DGX-2 at NVIDIA BaseOS - NVIDIA Docs . Specifically, Release Notes — NVIDIA DGX OS 6 User Guide is the latest version officially supporting the DGX-2, and uses GPU driver r580.

You can see the compatibility of r580 with CUDA 13 in Forward Compatibility — CUDA Compatibility .

When you say “supported” though, do you mean “with Enterprise Support from NVIDIA” (aka, you have a support entitlement for the DGX-2)? If so, please contact NVIDIA Enterprise Support (see the pinned post in this forum) for guidance on upgrading.

If you just mean “will work”, you will likely find success with following NVIDIA BaseOS 9 User Guide — NVIDIA BaseOS 9 User Guide to get a working recent Ubuntu installation.

ScottE

3

Hi @shahizat ,
Please note that the last DGX OS version supporting DGX-2 is DGX OS 6 and the last driver supporting the Volta architecture is R580.

4

Thanks @czankel, @ScottEllis for your reply. Yes, we are enterprise customers. We currently don’t have an active entitlement for the DGX-2 servers. My next question is how to upgrade it without reimaging, since we have multiple researchers working on it.

5

The DGX OS 6 User Guide linked above has a section on how to upgrade. Since you’re on DGX OS 4, it’ll take a few steps to upgrade (you shouldn’t jump to DGX OS 6 directly), but you should be able to get there without reimaging (just rebooting).

In my experience, you may be more pleased with the result by backing up data and starting “fresh”. It often results in a much cleaner final system. (and/or converting it to a system managed and configured with Ansible and/or Base Command Manager or similar)

ScottE

6

@ScottEllis, ok, we will try. Please share the steps on how to upgrade without reimaging. If it fails, we will reimage the machine.

I wasn’t able to find any documentation describing the process, so I used some AI assistance. Could you pls check it? Thanks

Step 1: bring DGX OS 4.14.1 fully current

sudo apt update && sudo apt -y full-upgrade
sudo reboot

Step 2: upgrade 4 to 5 (Ubuntu 20.04)

sudo apt install -y nvidia-release-upgrade
sudo nvidia-release-upgrade
sudo reboot

Step 3: upgrade 5 to 6 (Ubuntu 22.04)

sudo apt install -y nvidia-release-upgrade
sudo nvidia-release-upgrade
sudo reboot

Step 4: pull R580 driver + matching fabric manager on DGX OS 6

sudo apt update && sudo apt -y full-upgrade
7

That looks roughly correct @shahizat . The “Upgrading the OS” section in Upgrading the OS — NVIDIA DGX OS 6 User Guide (and the DGX OS 5 version as well) have the discrete steps.

8

@ScottEllis ok, I was doing this before, but I completely forgot a lot of the details. How did it work for the /raid mount drive? Can it be saved while the system is being reimaged?

9

Hi @ScottEllis, I was not able to run the following command on the DGX2:

sudo nvidia-release-upgrade
/usr/bin/nvidia-release-upgrade is not supported on this platform

Could you please help us with this? We cannot reimage the system at the moment.