I’m building a new pc with a 5070 ti. How are the drivers: proprietary and opensource? Any good?
When I was on an Amd card I was surprised when it was no longer supported but the open source drivers seemed to work just as well.
@Scott_S I have the rpms running on Leap 16.0 closed version on a Turning based GPU as primary and another couple of systems with Pascal, but they are both server based.
The new open driver should be better than closed as this is the development focus. I do use on Tumbleweed but prefer the run file install for the open drivers, rpms get out of sync…
I’ve also had pretty good experience with Nvidia’s new open drivers. Before that, the old proprietary ones were good too.
There were some hassles over the past year due to arrival of the new-open version along with Wayland churn and Plasma-6 churn. The driver wasn’t always the source of the trouble. Of late, things seemed to have settled down.
I’m now also running the cuda versions, which require a bit of locking to prevent the various sources of drivers from tripping over each other. I eventually got that sorted by carefully reading through
@mchnz that’s why I like the run file install, can be a bit more selective. It only takes a few minutes to rebuild/reboot… it just works for me… On Leap, the closed rpms work since they are old Pascal GPU’s, so I just rolled with the same on the Turing system as well…
@malcolmlewis normally I have always stuck with the OpenSUSE packing. I thought I should exercise the packager’s efforts and potentially provide feedback. So I’m not making a choice purely based on least-hassle or best-practice.
When I started using CUDA, I got myself into a bit of a mess and did eventually try the run-file installer - which did work well. After recent experiments with getting the zypper locks right, I felt OK with returning to the distro packaged versions and have had no trouble with them since. Now that SUSE (OpenSUSE?) is also allowed to package CUDA, I’m hoping the situation will improve further.
I would say the Nvidia website is currently better at guiding the user to the right run-file driver for non-Turing cards. I had to use locking to force Tumbleweed zypper to choose the G06 driver for a GT 1030, it first installed some old crufty driver. With locks, zypper did get me there in the end.
For me the two pascal systems are running k3s and nvidia-containers, so only need the driver present… the Turing system is running GNOME, so only want for display and nothing really more at this point. Primary desktop is where I do most stuff and that runs Tumbleweed The Quadro RTX4000 is just for offload, Intel ARC is primary GPU, the compute and all the good stuff mostly works OTB eg oneAPI, hardware encode/decode.