Consider supporting older GPUs about open-gpu-kernel-modules

[StayBlue]

By "hardware limitations" do we actually mean "licensing/legal limitations?" Because I don't know any other reasons from an engineering standpoint that would preclude sharing source code.

It's because this driver uses the new GPU System Processor which is not present on GPUs older than Turing. Please refer to the blog post from today for more info.

I think the main question people have is if there will ever be a way provided by Nvidia to allow for reclocking support on Maxwell and Pascal without the need to use the GPU System Processor.

[wyatt8740]

By "hardware limitations" do we actually mean "licensing/legal limitations?" Because I don't know any other reasons from an engineering standpoint that would preclude sharing source code.

Anyway, that's quite disappointing to hear. If GPU prices ever become sane again, maybe I'll be less disappointed.

[wyatt8740]

It's because this driver uses the new GPU System Processor which is not present on GPUs older than Turing. Please refer to the blog post from today for more info.

Thank you for the response. I understood that; it wasn't quite my question. My question is if there's any intention to make the non-GSP driver open source, or even release documentation/provide help with authoring a free (as in 'freedom') one for non-GSP GPU's.

It seems the answer is "no," at least for now. The intent appears to be to ignore any non-GSP platforms for the free drivers.

I think the main question people have is if there will ever be a way provided by Nvidia to allow for reclocking support on Maxwell and Pascal without the need to use the GPU System Processor.

This is a huge part of my question, yes. I'd like to not have my card turn into a pumpkin whenever Nvidia deems it time to stop releasing updates in the (hopefully distant) future.

Maxwell and Pascal are in an awkward position between being "too new" to work well with Nouveau -with the possible exception of the 750 (Ti) - and being "too old" to get to use the free drivers currently in this repository. So while Kepler cards will still possibly be usable with new kernels in 15 years, these cards seem destined to become e-waste.

[wyatt8740]

@nonetrix it's possible something of value could come out of this for, say, the Nouveau project. But they'd still need to get binary firmware signed for everything Maxwell and newer (except for the GTX 750/750 Ti Maxwell cards). And even then, they use different userland software, so the applicability might still be somewhat limited. Mesa and X.org instead of Nvidia's proprietary libgl/X server, for example.

At the very least it can't hurt, probably.

As for repurposing this code for older GPU's, that seems less likely but I have yet to deep dive into these sources.

[PAR2020]

Thanks for your feedback.
Today's release supports data center GPUs for Turing and Ampere architectures. There are hardware limitations in our older GPUs which negates support of open kernel modules for pre-Turing architectures. Please refer to aritger's comments at #19 for his detailed explanation.
Also see further responses regarding Tegra devices in #39

[aaronp24]

By "hardware limitations" do we actually mean "licensing/legal limitations?" Because I don't know any other reasons from an engineering standpoint that would preclude sharing source code.

It's because this driver uses the new GPU System Processor which is not present on GPUs older than Turing. Please refer to the blog post from today for more info.

[Daasin]

That's a shame for GTX10 Series cards on the Pascal architecture, open-source resources could've helped with a lot of issues especially important in this shortage

like there were some on DXVK/VKD3D that could've benefited from open-code or resources to look at... Not to stray off-topic but it is a shame since not all of us could upgrade 💔😓

[woachk]

Interesting.

I know about tegra, but did not know about that repository. How new is that?

Anyway, it seems like that might be somewhat useful, still not a complete solution, though.

All the Tegra since the K1 use it to drive the GPU block, so since quite a while now. And yep, not a complete solution, used for Tegra iGPUs almost exclusively by the public*... and GPLv2 only instead of being dual-licensed

[woachk]

There was some conversation about this one in: #19

[nonetrix]

However it might still effect older GPUs indirectly (just guessing I don't know much about GPUs) if they are similar enough that they code could be hacked together to work on older GPUs? Am I correct in assuming this?

[woachk]

@wyatt8740

For code for older GPUs you might want to look at this codebase: https://nv-tegra.nvidia.com/r/gitweb?p=linux-nvgpu.git;a=summary which does support some older GPUs.

That said, it doesn't come with firmware for it either, so that issue still remains.

(nvgpu is the GPU driver used for Tegra today, across Linux, QNX and Nintendo Horizon)

[ptr1337]

Actually I did from the current nvidia proprietary driver (dkms) and that is what I not understand. It uses the dma buf, yes.

Here if some is interested in.

https://www.toptal.com/developers/hastebin/ralegehixi.diff

[Hezuikn]

https://www.toptal.com/developers/hastebin/ralegehixi.diff

is down

[wyatt8740]

Interesting.

I know about tegra, but did not know about that repository. How new is that?

Anyway, it seems like that might be somewhat useful, still not a complete solution, though.

[ptr1337]

https://www.toptal.com/developers/hastebin/ralegehixi.diff

is down

https://gist.github.com/ptr1337/2e361f8f87abd57b1f6c1ea443f87f46