Difference "nvidia-driver-" vs. "cuda-drivers-" ; Issue with NVLink

Hello together,

while installing the driver for two A5000s with NVLink. I discoverd that there are two different drivers. The normal one and the Datacenter Driver.

Ubuntu 20.04, always fresh install in all methods.

My problem is, i don’t get nvlink to work with the “cuda-drivers-515” install method like in the documentation of the Cuda Toolkit. (sudo dpkg -i cuda-keyring_1.0-1_all.deb) After installation it shows all Links as Inactive in nvlink -s, p2p test shows bad performance.

If i install the driver like a normal RTX. With
add-apt-repository ppa:graphics-drivers/ppa and apt install nvidia-driver-515.
NVLinks works fine and shows 14 GB/s per Link. p2p test works fine.

So, what is the difference between the packages “nvidia-driver” and “cuda-drivers”?

Do i have to install something else withe the cuda-driver to get NVLink to work?

Thanks for the help!

cuda-drivers is just a metapackage for nvidia-driver + nvidia-modprobe
So nothing related to nvlink. Keep in mind you changed the repo so the real problem might be with the cuda repo.

Hi,

I do have the same problem on Ubuntu 22.04 I always install via cuda drivers to install everythign and is the first time I have that issue with Nvlink not working, I tried multiple GPUs and the problem persist.

If I try with driver 470 on UBuntu 20.04 it works fine.

Did you find the solution to this problem, I think is an issue with driver 515 via cuda installation.

Please if you solved it let me know.

Thanks