Until Re-Clocking Is Figured Out, Nouveau Remains Very Slow Against NVIDIA

[phoronix]

Phoronix: Until Re-Clocking Is Figured Out, Nouveau Remains Very Slow Against NVIDIA

Up for sharing today are our benchmarks comparing the very latest open-source Nouveau graphics driver code (Linux 3.15 + Mesa 10.3-devel) against the proprietary NVIDIA Linux graphics driver to see how the two NVIDIA Linux drivers compare.

http://www.phoronix.com/vr.php?view=20402

[Serafean]

How about a fair comparison

Hi Michael,

As a frequent reader, I really, really know that "Until Re-Clocking Is Figured Out, Nouveau Remains Very Slow Against NVIDIA", but would be very interested in a clock for clock comparison.
I don't have any nvidia cards anymore, but when I last had the honor, I remember it was possible to set clock speeds manually (heck I did it as a manual power saving feature...). Plus a quick google search tells me it might be pretty well scriptable : http://guilleml.wordpress.com/2011/0...izer-on-linux/

So how about it? I think it would be a pretty good performance comparison of the blob vs nouveau, even though not of a real world usecase.

Serafean

[jakubo]

Hi Michael,

As a frequent reader, I really, really know that "Until Re-Clocking Is Figured Out, Nouveau Remains Very Slow Against NVIDIA", but would be very interested in a clock for clock comparison.
I don't have any nvidia cards anymore, but when I last had the honor, I remember it was possible to set clock speeds manually (heck I did it as a manual power saving feature...). Plus a quick google search tells me it might be pretty well scriptable : http://guilleml.wordpress.com/2011/0...izer-on-linux/

So how about it? I think it would be a pretty good performance comparison of the blob vs nouveau, even though not of a real world usecase.

Serafean

+1

besides, when it is "silly" to compare these 2 on same low clock, then why bother to compare when the one has a considerably higher clock? thats just close to bullying.

[Michael]

Hi Michael,

As a frequent reader, I really, really know that "Until Re-Clocking Is Figured Out, Nouveau Remains Very Slow Against NVIDIA", but would be very interested in a clock for clock comparison.
I don't have any nvidia cards anymore, but when I last had the honor, I remember it was possible to set clock speeds manually (heck I did it as a manual power saving feature...). Plus a quick google search tells me it might be pretty well scriptable : http://guilleml.wordpress.com/2011/0...izer-on-linux/

So how about it? I think it would be a pretty good performance comparison of the blob vs nouveau, even though not of a real world usecase.

Serafean

FTA: "For those that then suggest to run tests of the NVIDIA binary blob with the graphics cards intentionally under-clocked, that is rather silly too, given that no one would intentionally do this in practice and that when the GPU core and video memory frequencies are severely restricted we're not seeing the driver's full optimization abilities in full."

[prazola]

Idea

You can try to put clocks always at 3D frequency by BIOS editing, no?

[Calinou]

You can try to put clocks always at 3D frequency by BIOS editing, no?

This is possible, but:

- VBIOS flashing is always risky (unless you have multiple VBIOS chips on your graphics card, like some overclocking-targeted cards),
- you don't benefit from any power management (high noise and power usage all the time).

Not to mention potential warranty issues (if you send your graphics card to warranty).

Heck yeah, we have Coolbits back. Let's stop this "Until Re-Clocking Is Figured Out, Nouveau Remains Very Slow Against NVIDIA" bullcrap. We all know that since ~2011.

We've always had CoolBits (eg. for fan control), what came back is overclocking support.

It's a sad reality, Nouveau is mostly only suitable for people who do little to no 3D stuff (which may count more people than you may think).

[zanny]

FTA: "For those that then suggest to run tests of the NVIDIA binary blob with the graphics cards intentionally under-clocked, that is rather silly too, given that no one would intentionally do this in practice and that when the GPU core and video memory frequencies are severely restricted we're not seeing the driver's full optimization abilities in full."

The point is that, while reclocking is the elephant in the room, it is one elephant - and it dominates the conversation. And obviously the fact it still is not fixed means those working on Nouveau have hit a brick wall and repeating the same thing over and over won't give them the information to actually implement it. It would be new and interesting information to see how the Nouveau pipeline would perform in the absence of that elephant, though, because as the radoen driver demonstrates there are a lot of other performance characteristics besides clock speed.

[prazola]

This is possible, but:

- VBIOS flashing is always risky (unless you have multiple VBIOS chips on your graphics card, like some overclocking-targeted cards),
- you don't benefit from any power management (high noise and power usage all the time).

Not to mention potential warranty issues (if you send your graphics card to warranty).

We are looking for architecture supporting progress, power management will be tested when freq switching support arrive.

VBIOS flash is not risky if you keep parameters between default ranges. A bad BIOS flash isn't a problem if you don't power off while writing the ROM.

[Serafean]

FTA: "For those that then suggest to run tests of the NVIDIA binary blob with the graphics cards intentionally under-clocked, that is rather silly too, given that no one would intentionally do this in practice and that when the GPU core and video memory frequencies are severely restricted we're not seeing the driver's full optimization abilities in full."

No, no one would do that intentionally in practice. But you're missing the point of why we ask : the interesting information not present in your benchmarks is how well does the 3D driver work. And only the 3D driver, not other hw features (no one is asking for vdpau performance comparisons, and you're not doing them )
The information from this would be "once recklocking works, we can expect about 60% of the blob's performance".
I'm not telling you what to do, nor to do it often, just that it's somethin I (and apparently some others) would find interesting.
And I maintain that not testing against an under-clocked blob is as silly as testing against and underclocked nouveau.

Serafean.

[mufasa72]

FTA: "For those that then suggest to run tests of the NVIDIA binary blob with the graphics cards intentionally under-clocked, that is rather silly too, given that no one would intentionally do this in practice and that when the GPU core and video memory frequencies are severely restricted we're not seeing the driver's full optimization abilities in full."

As previously stated, we're intereseted in the "what-if" scenario. What is the overhead of nouveau compared to the properiety, which would be answered by clocking the properiety driver down? If the nouveau driver performance scaled linearly with clock rate, what performance could we expect?

Just imagine all the website traffic you would generate by doing this... there is a LOT of interest in seeing a clock-for-clock comparison between the two drivers.