RADV vs. AMDGPU-PRO Vulkan Performance vs. OpenGL In May 2017

[pal666]

Looks like propriety Vulkan on linux needs some serious attention going on the fact that the oss GL is quicker in some cases.

why blame driver and not ports?

[nuetzel]

@Marek:
RX 580 regression: Kernel (4.12), firmware or driver? - Any idea at hand?
Wondered 'cause most tests (I ran since last Saturday under 4.11) I had expected more speed. ;-)

[jrch2k8]

Just saying except for Dota which it seem to be hitting a slow path or a bug and the rx 580 that obviously have issues, RADV got pretty damn close to AMDGPU-pro, I would dare To sat that in 2 months from now maybe useless to open up the closed driver beyond AMD patching certain stuff here and there.

Great work RADV guys

[marek]

@Marek:
RX 580 regression: Kernel (4.12), firmware or driver? - Any idea at hand?
Wondered 'cause most tests (I ran since last Saturday under 4.11) I had expected more speed. ;-)

My wild guess is: reclocking (it can be firmware)

[airlied]

Has Michael no i5 CPUs, not all gamers are able to afford best CPU on the market, so these benchmarks fail to show users what vulkan is good at. Lowerend cpu benchmarks please!

[dungeon]

He, he, Michael should just setup AM1 Athlon 5350 (4 cores but slow singlethread) with disabled integrated GPU, plus RX 470 that would be CPU bound at any time on FullHD, but will show threading API differences... Like Playstation 4 PRO simulation, only with half amount but the same type of CPU cores... GNM flies there

Does generic lower level Vulkan API (in comparison to GL) can do something similar is yet to be seen

Maybe not on Phoronix, as Michael seems likes his Kaby and 4K screen so much

Or Ryzen 7 downclocked to something like 2.4Ghz and disabled HT, turbo... that should do a job too

[humbug]

Just saying except for Dota which it seem to be hitting a slow path or a bug and the rx 580 that obviously have issues, RADV got pretty damn close to AMDGPU-pro, I would dare To sat that in 2 months from now maybe useless to open up the closed driver beyond AMD patching certain stuff here and there.

Great work RADV guys

RadV is progressing very well.
but all except one benchmark here was run at 4K. This makes it look closer to the amdgpu pro vulkan implementation.

[boxie]

Has Michael no i5 CPUs, not all gamers are able to afford best CPU on the market, so these benchmarks fail to show users what vulkan is good at. Lowerend cpu benchmarks please!

A really good idea here Michael - Can we have a Low/Middle/High PC test range?

[pete910]

why blame driver and not ports?

There is that I admit but as Vulkan on NV seems inline with NV windows I was thinking may be driver related

Other thing is a game with Vulkan from the start rather than a "wrapper" of sorts.

[higuita]

Vulkan only has lower CPU overhead, but GPU performance is unaffected if both GL and Vulkan program the GPU in the same way. If GL uses different/better GPU programming, it can be faster.

yes, i understand that, but the game code is the same in this case and i get better performance in the vulkan, Michael gets better performance in the opengl

what changes is the CPU, the mesa version (17.1.0 vs 17.2-dev) and kernel (4.11.0 vs 4.12.0-git). GPU, we have both a RX480

Looking to the data, Michael have a lot better performance in the OpenGL than me, for vulkan my and his performance is similar.

I would say that kernel do not make that difference, unless there was a miracle patch somewhere
mesa-dev, maybe the new multithread patch adds more performance, but that is a too huge difference to be just that, so my bet is the CPU
Michael is using a I7 7700 and i'm using a A10 7980k... he get better single thread performance and more threads than me. So my opengl must be taking a huge hit due to the weaker cpu

I will try the mesa 17.2-dev and see if this changes in any way