NVIDIA vs. Nouveau Drivers On Ubuntu 13.04

[phred14]

Actually I thought these results looked pretty good. Nouveau was only a factor of 2 to 4 behind the proprietary blob. Seems like not that long ago the factor was more like 10, across the board.

I have to agree that there would be some very good information available by under-clocking the proprietary drivers to match the OSS ones. That would help unlayer the performance gap between clock control and "other features".

[kokoko3k]

older benchmark...

Somebody could explain why early benchmark leaded to so different results?

NVIDIA vs. Nouveau Drivers On Ubuntu 13.04 on April 11, 2013

Nouveau Re-Clocking, NVIDIA vs. Nouveau Drivers on October 08, 2012


Apart from the different ubuntu version, the system is the same.
Even the same card with the very same nvidia blob version performs so different.

[bridgman]

Michael did benchmarks for Nouveau vs Nvidia on clocks limited to defaults for Nouveau. And perf was quite good. Bribe Michael by sending some more hw. :P Than he will look into it for your APU's.

And I wonder how that can be hard. You just run 3rd run with clocks limited to specific values... (Only hard coded value being name of driver from which gets speeds for other drivers..)

Good point, it was late in the day and I didn't think of slowing down the GPU running the proprietary drivers

[przemoli]

Hax0rs usually only play Quake 3.

Newer FLOSS games at constant alpha or are not that good for automatic tests, or are just old tech anyway. So keeping to same old test set we can compare results by viewing different tests from the past.

And Steam games are still in BETA. Too many people ignore licenses that forbid benchmarking on them.


We will see some good stuff in the future (Valve games, SS3, Unvanquished, etc...) WHEN THEY ARE READY.

[przemoli]

Same reason as for benchmarking our APUs with the open source drivers right now (where performance is generally poor because the clocks are low), I guess. For better or worse, it tells you what to expect today.

The one thing that would be "nice to have" would be a third bar indicating "clock-adjusted performance" to give an idea how much of the performance delta is clock-related vs driver-related. The numbers would be imperfect (nothing scales completely linearly although clocking-up an underclocked GPU is as close as you'll get) and it would be non-trivial to add to an automated framework, but that would give a good indication of potential and a bit of guidance re: where to focus efforts.

Michael did benchmarks for Nouveau vs Nvidia on clocks limited to defaults for Nouveau. And perf was quite good. Bribe Michael by sending some more hw. :P Than he will look into it for your APU's.

And I wonder how that can be hard. You just run 3rd run with clocks limited to specific values... (Only hard coded value being name of driver from which gets speeds for other drivers..)

[Kivada]

He might aswell link to the previous test, what gives? Is it just for clicks?

We have a winner!

[Calinou]

Article sucked because of same stupid quake3 games being benchmarked over and over again.

Hax0rs usually only play Quake 3.

[DMJC]

Nouveau Progress

Realistically Nouveau is progressing extremely well. If you view their status matrix on their wiki of 19 MAJOR tasks most cards only have 4-6 tasks left to complete. That's huge progress. I mean really who needs TV-Out, SLI, and Video-input support? (I have a SLI setup and even I don't care that much about it I'd rather see reclocking get fixed.) Reclocking is becoming more of a major target for them so they will get it working in the next 6-12 months IMHO. The OpenGL support is fantastic and rapidly approaching full 3.3 and then 4.2 compliance. I can only see good things coming for this project. I know people are impatient to see the end goal, but once the OpenGL/OpenCL stuff catches up, and the old cards are all working the newer hardware will be a lot easier/quicker to support. The 99% of use cases is rapidly being closed in on and once it's done the work can shift to the fringe cases like SLI etc.

[XorEaxEax]

The 'point' of these still elude me. He (Michael) starts off the benchmark essentially saying that nothing has changed which could affect the performance difference in any meaningful way, and then goes on the present that (now) pointless data.

Why not wait until there are any worthwhile changes to actually benchmark? He's not attempting to use re-clocking so he knows the results will be the same as the last benchmark without re-clocking. Is it all about clicks or am I missing something?

Until nouveau re-clocking is available on the cards he is using it is pointless to continously redo this nvidia vs nouveau comparison, just as pointless as doing gcc vs clang/llvm openmp benchmarks while clang/llvm doesn't have support for openmp.

[bridgman]

So if there's been no change in terms of re-clocking in noveau, and he isn't doing any tests with cards where re-clocking works, WHY continue doing these tests until re-clocking is supported on the cards he is using for these tests?

Same reason as for benchmarking our APUs with the open source drivers right now (where performance is generally poor because the clocks are low), I guess. For better or worse, it tells you what to expect today.

The one thing that would be "nice to have" would be a third bar indicating "clock-adjusted performance" to give an idea how much of the performance delta is clock-related vs driver-related. The numbers would be imperfect (nothing scales completely linearly although clocking-up an underclocked GPU is as close as you'll get) and it would be non-trivial to add to an automated framework, but that would give a good indication of potential and a bit of guidance re: where to focus efforts.