R600 Gallium3D Loses To AMD Catalyst Legacy Driver On Windows 8

[verde]

Useless comparison... Apples to oranges...

Fantastic comparison i would say. These are Windows drivers against Linux drivers (after the support abandoning for Radeon 4000 series and below in both platforms that hurt Linux more). Now with Steam i think AMD developers should target the optimization on Radeon opensource drivers especially for the 4000 series and below because those users have no other choice.

Of course DPM was first priority. Optimization is the second and in my opinion feature parity is the third.

[Bitiquinho]

Useless comparison... Apples to oranges...

From my point of view, I think this is useful for showing there's still much to (and that can) be improved in the whole Linux graphics stack.

[CrvenaZvezda]

Ho-hum... Question, what is the bottle neck causing low FLOSS performance? Request, performance in 2D and 3D open vs. closed AMD-drivers in a FLOSS world.

[vadimg]

Vadim, did you expect better performance out of selected test-cases? Lately I had the impression we are *very* close to catalyst.

I think performance can be better with some simple tuning like turning off the compositor and using performance governor for cpu, at least I typically have better results with weaker CPU&GPU than Michael. It would be more clear if Michael also tested catalyst on linux, otherwise it's hard to say whether the difference in performance is caused by the driver or by the OS + compositor + cpu governor + moon phase.

Anyway, if you'll look at the gpu-limited applications in this benchmark - Unigine demos - you'll see that we are close. With other apps (mostly cpu-limited) higher cpu overhead of mesa/r600g probably shows itself.

[jrch2k8]

The result of Triangle benchmark is above 100 thousands on my 5750 with r600g and I don't think that 4870 renders single triangle 7 times slower. It seems something is wrong with Michael's results, maybe some problem with compositor/Ubuntu?

michael could you post your dmesg? i suspect your 4870 card is not scaling properly and could you make a script to loop to a file this command cat /sys/kernel/debug/dri/64/radeon_pm_info when running the tests[could be the slower one only]?

this should give information enough to verify if the asic is droping to lower power states while in the test

[droste]

The result of Triangle benchmark is above 100 thousands on my 5750 with r600g and I don't think that 4870 renders single triangle 7 times slower. It seems something is wrong with Michael's results, maybe some problem with compositor/Ubuntu?

How did you do this? But the test itself is kind of weird here (juniper 5770).

With compositing: ~32000
Without compositing: ~7000 (no there's no zero missing at the end)

[jrch2k8]

How did you do this? But the test itself is kind of weird here (juniper 5770).

With compositing: ~32000
Without compositing: ~7000 (no there's no zero missing at the end)

try again with radeon.dpm=0 and set profile to high and set your CPU governor to performance, my theory is that if your CPU can't feed fast enough the GPU DPM kicks in and switch back to a lower powerlvl until it got enough activity to raise the power level

[agd5f]

But how does this driver compare to the still-maintained Catalyst Legacy driver for Windows 8?

The Linux catalyst driver is still supported and maintained as well as long as you are using a supported distro (kernel and xorg version). It's the same as on windows. The only difference is that windows doesn't change the kernel and display server as often as Linux distros do.

[log0]

If what you say is true, then how come the problem started to happen like 6-8 months ago with a mesa update ? It sounds like a regression to me.

Well, you didn't share this info in your original post. So it was a guess based on my own experience with blender and linux. Your issue seems to be something else. Report this regression to mesa bugtracker if you care.

[verde]

The Linux catalyst driver is still supported and maintained as well as long as you are using a supported distro (kernel and xorg version). It's the same as on windows. The only difference is that windows doesn't change the kernel and display server as often as Linux distros do.

The problem is that no major distro uses both the last supported Xorg and Kernel version. Not even the long term support Ubuntu 12.04 LTS.

This model of Catalyst Legacy releases is not working for Linux.