The Open-Source NVIDIA Linux Driver Continued Evolving In 2015

[phoronix]

Phoronix: The Open-Source NVIDIA Linux Driver Continued Evolving In 2015

This year the open-source NVIDIA Linux driver (Nouveau) continued to evolve with improvements for re-clocking, the start of OpenGL 4 support, and other new functionality. Here's a recap along with some performance benchmarks showing how the OpenGL performance evolved over the past 12 months.

http://www.phoronix.com/vr.php?view=22626

[bug77]

I'm quite surprised that, despite the large difference in the available documentation, nouveau isn't lagging way behind AMD's driver. Reclocking is still experimental (at best) and 900 series is still a no go, but other than than that, AMD doesn't have much to show for. Unless I'm missing something, which is very possible, because I don't keep a close eye on the red camp.

[karolherbst]

bug77 radeon actually pushes the pcie link from 2.5 to something higher. Nouveau still stucks with 2.5, but there are already patches for nouveau to also fix this. For desktop gpus this is mostly boring and will only bring benefits in a handful of games (talos principle, wasteland 2).

But for laptop uses with prime offloading this will give a steady +5% improvement at higher framerates, because the offloaded data is faster transfered between gpus.

[schmidtbag]

I'm quite surprised that, despite the large difference in the available documentation, nouveau isn't lagging way behind AMD's driver. Reclocking is still experimental (at best) and 900 series is still a no go, but other than than that, AMD doesn't have much to show for. Unless I'm missing something, which is very possible, because I don't keep a close eye on the red camp.

Unless I'm missing something, I don't know what you're talking about. Last I checked, nouveau is significantly behind the open source AMD drivers (performance-wise) and even further behind the closed source nvidia drivers. A lot of the mesa specs that nouveau supports are because of the work done by AMD, Intel, Red Hat, and other 3rd party devs. Don't get me wrong - the nouveau team is doing a fantastic job considering the lack of help they get, but AMD is anything but lagging.

[SystemCrasher]

I'm quite surprised that, despite the large difference in the available documentation, nouveau isn't lagging way behind AMD's driver.

Actually, it lags behind. In terms of being actually usable for most workloads. Though it makes farily good progress despite of nvidia being utter bitches.

With AMD driver I can actually play some rather heavy games and in recent distros it even provides reasonable experience out of the box on quite many hardware. Nouveau isn't anyhow close to this point and only provides anyhow adequate experience only on some few GPUs, only after uncommon system level fiddling. It completely lacks compute at this point, so even if AMD/Intel opencl implementations are half-assy, nouveau lacks even half-assy implementation. It can also require some rather uncommon actions like extracting firmware blobs from blob driver. Killing like a half of point using open driver, since you have to DL huge blob anyway. So it good in terms of formal features in MesaMatrix. But it quite bumpy thing in terms of practical systems use. No, sure, if you're really inclined on getting it working, it could work. But its nowhere close to being ready for mortals who are not a hardcore GPU driver gurus. Its nowhere close to PnP experience. And lacks some major features like compute which would not magically appear tomorrow.

[imirkin]

A lot of the mesa specs that nouveau supports are because of the work done by AMD, Intel.

[removed the bit about RH/3rd party devs... RH *are* 3rd party devs -- it's not like AMD is contracting with them to develop things, nouveau is volunteer-only so *all* nouveau work is done by essentially 3rd party devs]

The same is true of a lot of what Intel and AMD support. A lot of the common tessellation work was done by Marek and I, almost all of the (common) fp64 work was done by Dave (who I would say generally doesn't affiliate with any particular driver team), I'm in the process of doing atomic/ssbo which will be largely reusable by AMD (although the core was sponsored by Intel). Lots of other examples, these are just a few recent ones.

While I haven't directly asked them, I think most core developers would agree that I put in my fair share of core work, not just leeching on the achievements of the other teams, as you're making it out to be.

[smitty3268]

While I haven't directly asked them, I think most core developers would agree that I put in my fair share of core work, not just leeching on the achievements of the other teams, as you're making it out to be.

I think nouveau has done great work on the Mesa side of things. It's the kernel side that is kind of sketchy, which makes perfect sense because the hardware side of things is what isn't documented and needs to be reverse-engineered.

[schmidtbag]

[removed the bit about RH/3rd party devs... RH *are* 3rd party devs -- it's not like AMD is contracting with them to develop things, nouveau is volunteer-only so *all* nouveau work is done by essentially 3rd party devs]

Well, if we're going to get nit-picky, I said "and other 3rd party devs".

The same is true of a lot of what Intel and AMD support. A lot of the common tessellation work was done by Marek and I, almost all of the (common) fp64 work was done by Dave (who I would say generally doesn't affiliate with any particular driver team), I'm in the process of doing atomic/ssbo which will be largely reusable by AMD (although the core was sponsored by Intel). Lots of other examples, these are just a few recent ones.

While I haven't directly asked them, I think most core developers would agree that I put in my fair share of core work, not just leeching on the achievements of the other teams, as you're making it out to be.

I'm well aware AMD and Intel take advantage of mesa work from other contributors. I'm well aware that many of the mesa specs have been done by nouveau devs. I'm also aware from other forum topics that you tend to have a bit of an attitude. I never implied that you or other nouveau devs leech off others' achievements. The phrasing I used was "a lot of the mesa specs". That doesn't even imply "most", and I also never implied that was a bad thing - this is why mesa is open source, after all.

I also said "doing a fantastic job" but apparently you chose to ignore that.

I understand the hip thing to do in 2015 is to get offended by everything and nothing, but c'mon, we're only a few days from 2016.

[bug77]

Actually, it lags behind. In terms of being actually usable for most workloads. Though it makes farily good progress despite of nvidia being utter bitches.

Of course it lacks in performance, working reclocking is still missing in action. I didn't mean to imply anything else. What was surprising for me was that, looking at the bigger picture, nouveau seems to tick almost all the checkboxes the amd driver does.
That said, nouveau will have a much harder time yanking users from the blob, compared to radeonsi/amdgpu, so this valiant effort will have little impact in the real world.

[Kano]

Maybe Nvidia provides more information for nouveau devs next year, they already used nouveau for the Pixel C instead of a binary kernel module but with binary userspace. Reclocking would be very important and of course the signed firmware files for the GM2 cards. If it would work like amdgpu for desktop cards it would be nice.