Ahead Of Radeon RX Vega, AMDGPU+RadeonSI Is Offering The Most Competitive Performance Yet Against NVIDIA On Linux

[ConvexEd]

Good shit, I'm glad AMD is finally great on Linux. Now if only RX Vega had SR-IOV enabled, then I would grab it day 0.

[Michael]

Well very good results for the Fury card, is that with AMDGPU or RadeonSI driver running on the Fury, never can tell. Anyway I hope RADV and those benchmark optimizations come fast, I would really love to see Vulkan performance excellently like it should for AMD hardware under Linux. AMD really should be kicking NVIDIA's ass with Vulkan!

AMDGPU is the DRM kernel driver, RadeonSI is the user-space OpenGL driver. There is no AMDGPU vs. RadeonSI itself (only AMDGPU vs. Radeon when it comes to kernel drivers, but that is only a 'or' with GCN 1.0/1.1 hardware, newer is only supported on AMDGPU). The only other driver option for Fury is AMDGPU-PRO, but that "-PRO" wasn't mentioned at all in the article.

[theriddick]

The thing is that a teraflop is a teraflop, so a 7 teraflops radeon is as strong as a 7 teraflops geforce. The only difference is the circumstances under those the full strength of each is visible.

Don't get too far ahead of yourself, keep in mind that under Windows the 980Ti still wins vs a Fury X in DX11 results, only in SOME Vulkan/DX12 applications can it really shine but those are far and few between. This VOID of FPS and actual Teraflop performance is going to be ever so present with the Vega once its released, touting 13TFLOP of processing but still well behind the 1080ti which last time I check was a 11.3TFLOP card.

Lets also not forget that NVIDIA hasn't optimized OGL for a long time, not like how MESA is. They have had the luxury of developers doing all the work for them by making their games work best on NVIDIA and NOT AMD hardware. (under Linux). This happens purely by them not bothering to test against MESA/AMDGPU-Pro drivers, or optimize for them.

[theriddick]

It would be nice to see some benchmarks comparing AMD and NVIDIA pure TFLOP compute performance without a graphics payload. I know there are certain coin mining types that can do this (however some will favor memory performance or configuration so you gotta be careful what you look at). The other option would be a OpenCL 2.2 render job?

[birdie]

Most of these results run contrary to the situation that we have with Windows except for ... native Unigine benchmarks.

What I think could go wrong
1) some regression with Nvidia drivers
2) Mesa doesn't properly render something
3) there's some kernel/x.org regression or bottleneck which affects Nvidia cards

Unless I see these results being reproduced in Windows I'm not gonna believe them.

[LeJimster]

Ive been a bit underwhelmed by the Vega launch so far. But if the drivers keep improving, it may just be that Vega is the king on Linux.

I'm eager to see what the 64 and 56 models perform like.

[Michael]

Most of these results run contrary to the situation that we have with Windows except for ... native Unigine benchmarks.

What I think could go wrong
1) some regression with Nvidia drivers
2) Mesa doesn't properly render something
3) there's some regression or bottleneck which affects Nvidia cards

Unless I see these results being reproduced in Windows I'm not gonna believe them.

Didn't notice any rendering problems with any of the tested configurations for this article.

[Michael]

Ive been a bit underwhelmed by the Vega launch so far. But if the drivers keep improving, it may just be that Vega is the king on Linux.

I'm eager to see what the 64 and 56 models perform like.

Sadly, not sure I'll be lucky enough to test both models.... Currently saving up for the Radeon RX Vega 64 so unless AMD ends up stepping up to the plate or plenty more tips, won't be able to get the 56 too.

[pal666]

fury destroying 1080 in bioshock and xonotic is priceless
so are cries of nvidiots

Unless I see these results being reproduced in Windows I'm not gonna believe them.

i see how it hurts. wtf are you doing here if you are only interested in windows results?

[theriddick]

I think he is talking about windows OpenGL results, since comparing to DirectX is pointless atm because we all know DX11 is HUGELY more optimized with games these days.

In saying that there are unlikely to be MESA type performance fixes under Windows for AMD's drivers, so you can't compare until AMD does something about unifying the Windows/Linux driver payloads, like going open-source on the windows driver also (that would be really awesome to get parity).