Radeon Linux Benchmarks: Catalyst 15.3 Beta vs. Linux 4.0 + Mesa 10.6-devel

[iamnotabot]

Generally that's not the case. Performance tuning in some areas of the driver (eg shader compiler & memory mgmt) affect workload-driven frame rates while other areas of the driver limit the ability to run very high frame rates. It's not that simple but you also can't assume that performance deltas at high FPS will translate into a similar delta at higher res or higher workload.

Work is still ongoing in the workload-dependent areas of the driver, I'm just saying that I doubt much time goes into looking at why we only run 250 FPS instead of 300.

I wouldnt mind if we get a constant 60 fps in games where fglrx gets 100 or more. As long as it applies to new hardware intesive games as well.

Though im wondering about powerconsumption/heat generation/fan speed. If radeon gets 60 fps in a game and fglrx gets 100 do both require the same power? Or will radeon draw less? For example fglrx draws 100 and radeon draws 60 watt?

[deft666]

Something is wrong with that benchmark. I'm on Ubuntu 15.04 64bit with a Radeon HD 5870 and the radeon/mesa driver is waaaaay faster than the propiertary fglrx fuckup. It starts with 2D performance (gnome shell lags like hell on fglrx), goes over 'no alt-tab' in fglrx and ends in ~20 to 30 fps with fglrx in CS:GO. With mesa I got smooth framerate.

Either there is something happening in the 4.0 kernel I'm not aware of, or there is something seriously wrong with that benchmark.

Tell me how to install the driver amd 15.03 in ubuntu 14.04?

[edoantonioco]

The mesa driver is full of regressions ,I use an old card (mobility radeon hd4250) I remember when I used mesa 9.2, and TF2 in normal settings runs great, I even delete the game from my windows partition. Now in mesa 10.5 it's almost unplayable on low settings. I remember the performance difference happens from one version working fine to the other working a lot slower.
Related to the benchmarks, I wonder if what we all want will ever happen.

[bridgman]

The mesa driver is full of regressions ,I use an old card (mobility radeon hd4250) I remember when I used mesa 9.2, and TF2 in normal settings runs great, I even delete the game from my windows partition. Now in mesa 10.5 it's almost unplayable on low settings. I remember the performance difference happens from one version working fine to the other working a lot slower.
Related to the benchmarks, I wonder if what we all want will ever happen.

Are you fairly sure the regression is in mesa rather than, say, the kernel driver ? In other words if you go back to the older mesa does your performance come back ?

[bridgman]

Though im wondering about powerconsumption/heat generation/fan speed. If radeon gets 60 fps in a game and fglrx gets 100 do both require the same power? Or will radeon draw less? For example fglrx draws 100 and radeon draws 60 watt?

My recollection from earlier tests was that radeon doesn't load the GPU as much (ie doesn't overlap CPU & GPU processing quite as much as fglrx) so power draw when busy would be a bit lower. Probably won't be linear like the example you used since things like shader compiler optimization also affect how much GPU time (and hence energy) it takes to render a frame,

[peppercats]

My recollection from earlier tests was that radeon doesn't load the GPU as much (ie doesn't overlap CPU & GPU processing quite as much as fglrx) so power draw when busy would be a bit lower. Probably won't be linear like the example you used since things like shader compiler optimization also affect how much GPU time (and hence energy) it takes to render a frame,

unrelated, but since you're here
do you know if there's any plans to get codeXL working with mesa?

[drSeehas]

Again Pitcairn/Cura?ao twice but no Bonaire :-(

[dungeon]

Again Pitcairn/Cura?ao twice but no Bonaire :-(

You can watch Yaroslav's yt channel, lot of videos with Bonaire

https://www.youtube.com/channel/UCh1UanySk_Ht0WRfo4O9STw/videos

And same lot of 2-3 times slow cases for any (nine, wine, native) when compared with Win D3D

[smitty3268]

I was just asking brigman if there has been any progress in identifying why we see this trend of hitting a frame rate wall and not scaling in newer openGL 4 games.

It's overhead/validation in the GL driver calls. NVidia spent a great deal of time and effort optimizing that down to almost nothing. AMD decided to create Vulkan instead, to do the same thing. Unfortunately there's not one magic call that is causing all the problems, which would be easy to fix. It's that there are thousands of calls each frame, and each one is a hundredth of a ms slower.

There's a reason NVidia hasn't been nearly as gung-ho about Vulkan as all the other vendors. It's basically bringing everyone else back up to the same speed they can provide.

Seeing it happen in GL4 games is a bit of a misdirect, i think. It has more to do with those games just being much more intensive and calling into the driver a lot. Older games didn't have to draw nearly as much into a scene.

And of course, all the porters are focusing all their development time on getting things working well on nvidia. Adding AMD support generally happens at the very end when they know they have things generally working, and are just trying to wrap it up and push it out the door.

[frosth]

i expected more

Unfortunatelly i can't run this tests at 1600p and i don't have CS:GO but other games/apps looks good for my HD7770:
http://openbenchmarking.org/result/1...DE-FGLRXRADE43