AMD Kaveri: Open-Source Radeon Gallium3D vs. Catalyst 14.12 Omega Driver

[phoronix]

Phoronix: AMD Kaveri: Open-Source Radeon Gallium3D vs. Catalyst 14.12 Omega Driver

Last week when running Linux benchmarks of the AMD Catalyst 14.12 Omega driver that is much improved over older versions of Catalyst on both Windows and Linux, plus running a fresh 12-way AMD vs. NVIDIA Linux GPU comparison with the newest drivers, there were many Phoronix reader requests for some new AMD APU Linux numbers from this Omega driver. In this article are the Omega benchmark results for an AMD A10-7850K APU when testing both the open and closed-source drivers.

http://www.phoronix.com/vr.php?view=21250

[Nietzsche]

RAM precisions

Hi. While testing APUs, you should include more information about the RAM used (single or dual channel, frequency and CAS latency) since it's an important performance factor.

[coder111]

Open-source drivers are GOOD

I am just thinking to what the situation was ~5 years ago. We were happy to get 10-30% of the performance of fglrx with open-source drivers. Now it's closer to 60-80%. No proper video acceleration. No OpenCL. It has gotten so much better.

There is still room for improvement, but we are getting to a point where running 3D intensive stuff with open-source AMD drivers is really feasible.

And I think we are getting open-source OpenCL as well- I don't remember seing any benchmarks for that at Phoronix, but I got some open-source open-cl drivers pulled in in Debian/sid and they seem to work- at least they sped up some WPA key cracking app I used as a benchmark.

And video acceleration also kinda works.

[dungeon]

Hi. While testing APUs, you should include more information about the RAM used (single or dual channel, frequency and CAS latency) since it's an important performance factor.

For sure, i don't have other idea how to explain this results with A10-7800 results from 9 days ago

http://www.phoronix.com/scan.php?page=article&item=amd_kaveri_dec7&num=1

And cpufreq should be always on performance (not ondemand) when testing games

[tomtomme]

For sure, i don't have other idea how to explain this results with A10-7800 results from 9 days ago

http://www.phoronix.com/scan.php?page=article&item=amd_kaveri_dec7&num=1

And cpufreq should be always on performance (not ondemand) when testing games

Yeah.

Michael: Do you have any idea what happened there? How is it possible that your A10 7850K is about 70% faster than your A10 7800 on the same software?
Is it even possible that this is caused by cpu-freq and / or RAM settings?
Maybe some strange power-saving setting in the BIOS?

[dungeon]

Is it even possible that this is caused by cpu-freq and / or RAM settings?

Probably RAM diff plus maybe A10-7800 is in 45W cTDP mode or something like that

As for cpufreq, it is in ondemand for both so that si probably not a issue in comparison. But generally if in performance mode people should see much more fps rate in both those cases and generally on any APU while gaming or do benchmarking tests.... so really if not performance mode i can't say how those APU can be fast for gaming, only thing i know those fps rates is slower then they should be.

Well that is not cpufreq or linux only issue, people on Windows also knows and it is some kind of generall knowledge - that auto mode power saving what they get by default on APUs is never good for gaming.

[Nobu]

Xonotic: 1920x1080, effects: low... fps: min: 38, max: 167, avg: 119.7
Xonotic: 1920x1080, effects: high... fps: min: 2, max: 88, avg: 38.04
Xonotic: 1920x1080, effects: max... fps: min: 2, max: 80, avg: 19.71

Quite a range, there. What I'd like to know is how often it dips down to 2, or 38. If it stays in the 60+ range most of the time and only dips down when it's loading resources, I'd say that's pretty decent (though not really playable, except maybe on low/medium).

[dungeon]

Xonotic: 1920x1080, effects: low... fps: min: 38, max: 167, avg: 119.7
Xonotic: 1920x1080, effects: high... fps: min: 2, max: 88, avg: 38.04
Xonotic: 1920x1080, effects: max... fps: min: 2, max: 80, avg: 19.71

Quite a range, there. What I'd like to know is how often it dips down to 2, or 38. If it stays in the 60+ range most of the time and only dips down when it's loading resources, I'd say that's pretty decent (though not really playable, except maybe on low/medium).

Those are Ubuntu 14.10 results, so 3.16 kernel and mesa 10.3 results - which does not even have hyperz enabled (it is disabled by default), etc. does not worth to analyze that, better upgrade at least mesa

[przemoli]

It seam that PTS have trouble with first few frames. Hence those really small MIN values. They are real, ok, but on the graphs we can see that those are (almost) always first few. So their impact on real game play (or perceived performance) is small.

Also we really need percentiles of slowest frames (how long it took for eg. 5% of slowest frames), as its there where (in games) perceived slowness show its ugly head.

[przemoli]

@Michael
Would You consider awarding phoronix.com premium for adding option to skipp first few (lets say first 5) frames from results (and some mod to graphs so that its clearly labeled as such)? Or for adding those percentiles to results (when its possible to compute, so per-test option?)?

Or even better, make pull about "what improvements to phoronix do You want?", and give such bounty for tasks that are not Your priority.