On Low-End GPUs, Nouveau Speeds Past The NVIDIA Driver

[MùPùF]

Hey Michael, did you upclock the cards? The 9800GT looks strange :s

[cruiseoveride]

When using a low end card, I suppose CPU time makes up a lesser part of the overall processing time for each frame, but on a high end card, the CPU time for each frame is comparatively is much larger.

i.e if for each frame it takes 1 unit of CPU, and on a low end GPU it takes 5 GPU units, then the CPU time is only 1/6 of the time for each frame. But say on a high end GPU, it takes 1 GPU unit per frame, then the CPU time is 1/2 for each frame.

If that logic is correct then the problem with nouveau for higher end cards is that nouveau is using too much CPU.

[MùPùF]

When using a low end card, I suppose CPU time makes up a lesser part of the overall processing time for each frame, but on a high end card, the CPU time for each frame is comparatively is much larger.

i.e if for each frame it takes 1 unit of CPU, and on a low end GPU it takes 5 GPU units, then the CPU time is only 1/6 of the time for each frame. But say on a high end GPU, it takes 1 GPU unit per frame, then the CPU time is 1/2 for each frame.

If that logic is correct then the problem with nouveau for higher end cards is that nouveau is using too much CPU.

Well, nouveau is capable of going way faster than that On my laptop (3 years old, quadro nvs 140m 128MB of vram + core duo @ 1.73GHz), I get 125fps with open arena and arround 55fps with nexuiz (not everything at the maximum, I lack vram for that).

[agd5f]

Holy crap!!
I am becoming to think that nvidia GPUs, are somewhat easier to command(program) than Radeons are. I have no other explanation for this performance boost without any nvidia support, or documentation. Maybe some one of the Radeon developers will clear the situation.

Older chips have more mature open source drivers because they've been worked on longer. The open source r300g driver beats catalyst in some tests as well:
http://www.phoronix.com/scan.php?pag...expanded&num=5
http://www.phoronix.com/scan.php?pag...pflipper&num=3

[calim]

Holy crap!!
I am becoming to think that nvidia GPUs, are somewhat easier to command(program) than Radeons are. I have no other explanation for this performance boost without any nvidia support, or documentation. Maybe some one of the Radeon developers will clear the situation.

They do have some helpful features, like command checking in hardware. They tell you what you did wrong instead of "crashing". Most of the time anyway.

[smitty3268]

On the radeon/nvidia side, nouveau people are in the lucky position of having not freeze their API which means they can change the way the communicate with the kernel. It's not the case for radeon and if i was to write it again today i would do it completely differently.

Can't this be done by using versioning?

[Artemis3]

Today I learned that my 8800GT is considered a low end GPU.

The GPU can get very hot but the only game that does so is OilRush. I know this due to a very noisy fan.

The 9800GT (8800GT and GTS 250 are the same) is NOT a low end.

For the temperature, try forcing the fan with nvclock. What i usually did to these cards, was change the firmware fan settings, so that at worse it stays at 70c instead of the 100c it will happily go if left alone. And yes, the one with the smaller (reference design) fan is noisy.

[bridgman]

Can't this be done by using versioning?

Yes, but then code for all the different versions has to be updated and tested, which significantly increases the workload and reduces the time available to spend on other improvements. Another nasty tradeoff.

[darkbasic]

Is this true ? If yes then the test is useless.

It's not the first time. That's why I DON'T use PTS to bench nexuiz.

Can you repeat it after disabling all not supported features ?

No, they already told him the apache benchmark was completely useless without an external machine which generates requests but he never cared. They still publish USELESS apache benchmarks, he only cares about using PTS, it doesn't matter if the benchmark is a complete FAKE.

[crazycheese]

The reason there is advantage is dead simple. Obfuscation requires CPU time.
But don't jump out of joy, amd loves proprietary same as nvidia.