Radeon DRM Linux 4.4 + Mesa 11.1 + DRI3 vs. AMD's Proprietary Driver

[phoronix]

Phoronix: Radeon DRM Linux 4.4 + Mesa 11.1 + DRI3 vs. AMD's Proprietary Driver

On Friday I posted benchmarks showing Nouveau's re-clocked performance relative to NVIDIA's proprietary driver for showing the performance potential of NVIDIA's GeForce GTX 600/700 series with the performance state code there beginning to work. That article was followed by AMDGPU driver tests on Linux 4.4 against Catalyst for the newest AMD GPU tech that uses this newer Direct Rendering Manager driver. The third test now is comparing the Radeon DRM performance on Linux 4.4 against AMD's binary blob when using older AMD GCN GPUs as well as a Northern Islands GPU for reference.

http://www.phoronix.com/vr.php?view=22488

[duby229]

Cool. That was a pretty good read. The only other thing that would be interesting is to make sure you have representatives from each gpu architecture. On this benchmark you have a VLIW5 and GCN1.1. I kind of would like to see a GCN 1.0 on this as well. I can refer to your earlier article about amdgpu to see a GCN1.2 so that much is ok.

The only reason I mention this is because people assume that benchmark results can be extrapolated between different cards and that's just not true. If every architectural generation was represented on every article then that would become plainly obvious to anybody that reads benchmarks.

EDIT: I hate it when I read people say their card performs so much better than that, etc. That's all I'm saying.

[Michael]

Cool. That was a pretty good read. The only thing I that would be interesting is to make sure you have representattives from each gpu architecture. On this benchmark you have a VLIW5 and GCN1.1. I kind of would like to see a GCN 1.0 on this as well. I can refer to your earlier article about amdgpu to see a GCN1.2 so that much is ok.

The only rason I mention this is because people assume that benchmark results can be extrapolated between different cards and that's just not true. If every architectural generation was represented on every article then that would become plainly obvious to anybody that reads benchmarks.

Too bad I don't have more AMD cards... The HD 7950 was busy on another test system while running this article.

[duby229]

I agree and I understand the innuendo fully. You should get the cards you need for your articles.

[raom]

It seems the schmucks at AMD have been busy making profiles specifically for the benchmarks Michael tests for. I'm getting significantly better performance on the open source driver on nearly every steam linux game I own, compared to catalyst 15.9, using a HD6870.

[duby229]

I have read some articles you wrote that were not particularly nice to AMD. I think you should put your public relations gloves on when you type AMD related topics though. It would help a lot I think.

[mannerov]

Not sure if the problem with the triangle test is memory bandwith. Perhaps the problem is rather in the vertex shader (perf of vertex shader doesn't matter much for normal apps), since radeonsi is known to not do some optimizations there. That doesn't explain the r600 card performance of course.

[Michael]

I have read some articles you wrote that were not particularly nice to AMD. I think you should put your public relations gloves on when you type AMD related topics though. It would help a lot I think.

AMD PR did invite me out to California next week... But only told me about it Friday when I already have other commitments to make next week...

[extraymond]

AMD PR did invite me out to California next week... But only told me about it Friday when I already have other commitments to make next week...

Told them you are going to NVDIA's, and you'll get some special gift later XD

[marek]

Some comments/questions:
- R9 290 is too slow in openarena and tesseract. This is worth looking into.
- Why did Valley look different compared to Catalyst? Did you forget to remove ~/.drirc and replace /etc/drirc with fresh drirc from Mesa? Packages should take care of the latter, but not the former.
- Regarding GpuTest: Triangle, the open source driver is ~2x slower in fullscreen, but if you test the windowed mode instead, the open source driver is ~2x faster than Catalyst in the same test.