I have an RTX Pro 4500 and am attempting to install it in the NAS after upgrading to 25.10.3.1 which shows as the latest (stable) version:
When I attempt to follow the instructions here, their screenshot shows a checkbox to install the drivers between the “Address Pools” and the “Check for docker image updates” checkbox… my version doesn’t have the checkbox to install the NVIDIA Drivers like the instructions and instruction screenshot shows… so I did RTFM… it was just wrong?
Full details on your hardware may help and give clues. Is this the only video card? Was this a swap of video cards, etc.
Apologies.
AMD Ryzen 9 5900X
32 GB DDR4 3200
1 Samsung NVMe Boot
1 Samsun NVMe Apps
4x Seagate 12 TB data pool
Asus Asus prime b550-PLUS motherboard
Yes, that’s the only GPU. When I’m running 25.04.2.6 with my GTX 1070, it boots fine, the GPU is visible, can be applied to containers, etc.
When I replace the GPU with the RTX Pro 4500 and upgraded to 25.10.3.1 it wasn’t visible via smi, the checkbox to install the drivers doesn’t appear. I know the old GTX isn’t compatible because of the new open drivers, but according to the release notes, the blackwell and RTX PRO lines are supposed to be supported.
I just put the old GTX 1070 in, reverted back to 25.04.2.6 and it boots, recognizes the card, all pools are healthy and docker containers are running with access to the card.
@HoneyBadger What is needed to be done with changing NVIDIA cards and upgrading and Docker?
Well, FML. I think the devil is in the details.
When comparing the Nvidia driver packaged in this version according to the documentation (570.172.08) with a newer version (570.211.01) the cards are listed differently. In 570.172.08 the support is listed for only “RTX 4500 Ada Generation“ vs “NVIDIA RTX PRO 4500 Blackwell Workstation Edition” in the 570.211.01 version.
So while the documentation says (emphasis mine)
“The open GPU kernel drivers are compatible with Turing architecture and later GPUs, which includes GTX 16-series cards and all RTX series cards. Newer architectures like NVIDIA Blackwell (RTX 50-series, RTX PRO Blackwell) require the nvidia-open driver to function, as this driver leverages the built-in NVIDIA GSP (GPU System Processor).”
The RTX Pro Blackwell is not actually supported on this version, just the RTX ADA versions.
It does look like they’re upgrading the drivers for v26 (590.44.01) however, so I can either try the beta once it’s stable or just wait and sit on a $3k video card that I was hoping to put to use with some LLM work and training containers.
Sorry for the confusion there. That paragraph was intended as a general description of how the open driver works and example of newer architectures that will never work with the previous closed driver and the next one does go on to direct you to check the particular installed driver for compatibility, but I see how you could have interpreted it otherwise.
The open GPU kernel drivers are compatible with Turing architecture and later GPUs, which includes GTX 16-series cards and all RTX series cards. Newer architectures like NVIDIA Blackwell (RTX 50-series, RTX PRO Blackwell) require the nvidia-open driver to function, as this driver leverages the built-in NVIDIA GSP (GPU System Processor).
To verify your GPU is supported, check the installed NVIDIA driver version in Software Component Versions. Click the driver version number to view the NVIDIA release page, which includes a “Supported Products” tab listing all compatible GPUs for that driver version.
That makes sense. I installed the 26 beta 1 as it’s supposed to have a compatible kernel and driver but it freezes at initializing ramdisk. I just gave up and put my B580 in it and put the RTX in my PC and am running model training in WSL.
The Battlemage drivers seem to be working alright though. The containers are able to leverage the GPU
Not sure if you saw this or not but a very kind user created a package to support the older cards on the newer TrueNAS builds…
As far as I’ve followed the thread the issue is that the card needs a newer driver not the older one
