NVIDIA R550 Linux Driver's Open Kernel Modules Performing Well On GeForce GPUs

[phoronix]

Phoronix: NVIDIA R550 Linux Driver's Open Kernel Modules Performing Well On GeForce GPUs

With the recent NVIDIA 550.54.14 Linux driver release the R550 series is now out as stable. One of the prominent changes with the NVIDIA R550 Linux driver is bringing the GeForce and workstation GPU support up to "CERTIFIED" quality when using NVIDIA's open kernel modules that are distributed as part of their driver package. Previously the open-source (out-of-tree) kernel modules were just certified for their data center GPUs while now they are basically acknowledging that they are in good shape too for GeForce and workstation products. In this article are some benchmarks of the open and proprietary kernel driver options of the NVIDIA R550 Linux driver.

https://www.phoronix.com/review/nvidia-r550-open

[ptr1337]

Seems very good, but since the "nvidia-utils" are also required for using it properly, there is not a big difference.
The modules can be accessed since all time, when extracting the .run file.

I think it would be cool to directly integrate the kernel modules with a patch into the kernel, but would also bloat it quite much. Would increase the size for around 40 MB probably.

I need to take a look and diff the open source module and the ones packaged in the .run, i think there are just some differences due older gpus.

[avis]

It will be interesting to see if/when NVIDIA ends up transitioning to the open kernel driver by default for consumer hardware and if future NVIDIA GPUs only end up being enabled along the open driver route.

AFAIK it only exists as a reference implementation to ease nouveau's implementation efforts.

[dragon321]

Is there any good reason to use it? It still requires proprietary user space and it's still out of tree module that needs to be recompiled for every installed kernel.

[aufkrawall]

Has anybody tried HDMI 2.1 4k 144Hz with the nvidia-open driver?

[pong]

Recently (this version or maybe the previous one) it was noted that to use some new CUDA features (unified heterogeneous memory) it was a requirement to be using the open kernel drivers. I don't know why or how that would be so, but it was mentioned in the new feature info release notes.

It that was so, I suspect it could be possible that other disparate features could exist between the open kernel drivers installation and the other.

Is there any good reason to use it? It still requires proprietary user space and it's still out of tree module that needs to be recompiled for every installed kernel.

[bug77]

vRAM is not preserved across power management yet

What?

[pong]

So what's the known outlook for prominent linux distributions / repos that already routinely package the nvidia drivers, are there conspicuously known plans to offer / use open kernel module based drivers package as an option, by default, not for the near future, whatever?

The relevance of that ties into whether there are actually possible to be new / different features / better UX (sysadmin, development, ...) when using the new open modules interface vs. the historically available schemes. I know at least one recently new CUDA feature (heterogeneous memory) was documented (release / feature notes) to require the open kernel modules based driver to be used so that may indicate there could be / become other relevant use case differences between them.

[ssokolow]

What I'd like to see is a test run including an RTX 3060 12GB (or really any RTX 3000-series), since that's what I run my CUDA on.

[Beryesa]

I think this could be useful if it could work together without clashing the needs of NVK, so that we could use the open userspace for our desktops and maybe the proprietary userspace for specific flatpaks etc.
But not sure if that's actually possible.
Maybe there can be a nouveau-patched open-kernel-modules?