AMD vs. NVIDIA Vulkan & OpenGL Linux Performance With The New Drivers

[nocri]

This is extremely odd graph -- 640x480 ? And what else, like low details, low geometry, no shadows and windowed mode ? and on the horizontal axis it is not FPS ... someone in the marketing dep. went over the top to prove something

[AJSB]

NVIDIA cards DON'T have truly Async Compute in HARDWARE like AMD, be it in Dx12 or Vulkan
This is why you get those results.

This is why also in Dx12 you notice a large increase of performance for AMD cards and that kinda of crap for NVIDIA.

These modern APIs arrived in the right moment for AMD

[kwahoo]

This is extremely odd graph -- 640x480 ? And what else, like low details, low geometry, no shadows and windowed mode ? and on the horizontal axis it is not FPS ... someone in the marketing dep. went over the top to prove something

Test system:
NVIDIA TITAN X 356.45
i7-3770k @ 3.50GHz
Test settings:
Resolution: 640x480 (CPU Perf)
Highest Rendering Quality
Vulkan/GL/DX9/DX11

https://www.youtube.com/watch?v=xXyZ4YaktyU

[altarius]

NVIDIA cards DON'T have truly Async Compute in HARDWARE like AMD, be it in Dx12 or Vulkan
This is why you get those results.

This is why also in Dx12 you notice a large increase of performance for AMD cards and that kinda of crap for NVIDIA.

These modern APIs arrived in the right moment for AMD

well ... the moment these things really count we are close to the generation after the already announced new gpu's. so when most software supports this feature, there are enough gpu's of all vendors supporting it.

[iBloodLust]

You never did mention it, but what were the settings in this benchmark?

[ThrowAway3000]

If you have VSYNC turned on, that low GPU usage is probably because it can draw 60fps with only half the power of the GPU. You are likely not THAT cpu bound.

Of course I don't use Vsync, also Source engine is known for using CPU a lot.
CPU: Intel Core2 Duo E7200 @2.53GHz.
GPU: AMD HD 7790 (Bonaire), RadeonSI drivers with git Mesa and git LLVM versions.
I'm getting 66 FPS in Dota 2 on low settings, 1080p. Radeontop shows about 50% GPU usage.
I can load my GPU more and it will probably be something like 50 FPS on High with GPU still not used fully, so it's definitely CPU bound.
Same 66 FPS on Windows 8 with DX9, but the game is more responsive on Linux and plays much better (Xmonad WM).
AMDGPU + RadeonSI driver doesn't seem to work well on Bonaire yet, it starts at 65 FPS then drops to 35 over time, so I'm back to using RadeonSI.
I expect to have >100 FPS with Vulkan on this system on Low Settings.

[artivision]

GPU Queues are asynchronous in Vulkan, just like DX12.
‘Asynchronous shading’ just refers to a concurrent usage pattern of said queues, so support will be pretty much the same for games supporting both APIs.

As far as these benchmarks go, it seems fairly consistent that right now (maybe upcoming optimizations from AMD will change that) the Fury is pretty equivalent to the 980, the 290 to the 960 and the Tonga to the 950, with memory bound workloads (Xonotic 4K) favoring AMD because of their generally wider buses (especially the Fury's HBM).

That is totally wrong! Async_Compute is a hardware future with advanced threading units that allows you to do Graphics+Compute or simply topical compute on specific shaders and not on the entire field, aka very complex shaders. Nvidia can only do Graphics+Graphics or Compute+Compute and that is all. As the time passes AMD GCN gpus will become stronger like +50% when activate many FX and get only a 10% drop, wile Nvidia ones will become weaker. Even if Nvidia does AS_CO for future gpus they will not benefit, because today they replace game shaders with their own version via their driver and guess what, they all ready blend some FX in the shaders during the build.

[smitty3268]

In the presentation with these numbers, Valve was also very clear that their engine in many ways is still based around DX9 capabilities, and that there is a lot of stuff they can still do in Vulkan that they haven't gotten around to yet.

So it's really in a similar situation to Talos, where the engine has been ported to Vulkan but isn't using it fully yet.

[cusa123]

A test data of the thalli principle in which conditions of quality jurisdiction made.
My question is whether 3840x2160 in ultra?

[rabcor]

Didn't know the vulkan driver over at talos overtook the opengl performance by such a decent margin yet. Well done croteam! keep it up!.

Also Michael; you really need some up to date 300 series AMD cards

Besides the Fury you're a generation behind.

Also folks, don't forget that this was just a Fury, not a Fury X.

But I will say that the Fury line was kinda disappointing because of extremely low memory clock speeds compared to nvidia's cards (ah; now I know why my 670's VRM heats up like a furnace! it has apparently 6008mhz effective memory clock speed to Fury X's 1000mhz and it has 3x the actual memory clock speed of Fury X according to gpuboss; how could AMD even make this mistake? This might be the biggest reason why 980-Ti usually outperforms the fury despite being a weaker card in all other aspects)

Anyhow, nice to see AMD performance catching up. But now that AMDGPU has caught up to Catalyst, the real question is; will it catch up to Nvidias tier of performance, or will we always be behind like we have been up to this point on AMD on Linux; will AMDGPU now stop slpeeding up and continue at snailing speeds comparable to catalyst of the past? Or will the driver actually be improved to catch up (or go beyond) nvidia's driver performance levels?