RADV vs. NVIDIA Vulkan/OpenGL Performance For Serious Sam 2017

[philips]

(posted earlier but messed up)
Michael, could you check if the lights are working in OpenGL mode?
I've gone through all settings (low to ultra) and there are no shadows in OpenGL on my system while working fine in Vulkan mode. This could explain lower Vulkan results.
I see you also have no shadows in the screenshots you posted and shadows should be present even on the low settings.

[ua=42]

I really wish these NVIDIA vs AMD articles listed the price of the cards.

Hey Micheal. You could probably set it up to automatically set up to display the price of it on amazon and have it be a referal link so you can get money if someone purchases the card using the link.

[haagch]

AMDGPU-PRO doesn't work yet on Ubuntu 16.10+....

You could test only Vulkan from amdgpu-pro. You only need the Vulkan blob (amdvlk64.so and amd_icd64.json) and I think a libdrm with these patches: https://github.com/libcg/archlinux-libdrm-amd
I heard it only runs well with mainline 4.10 or higher though.

(I keep them updated for latest libdrm git at https://github.com/ChristophHaag/libdrm - thanks to notaz for updating them at one time when libdrm git had a lot of conflicts)

It can probably be kept separate from the system by using LD_LIBRARY_PATH for libdrm and VK_ICD_FILENAMES for the vulkan blob.

[Niarbeht]

Well, in my opinion nowhere they run as they should. When an RX 480 is about as fast as a GTX 980Ti or slightly below a GTX 1070 one can say they run as fast as they should.
The state of RADV and the overhead and multi threading issues of OpenGL prevent this.

1070 isn't right for an RX480 comparison. It should be on par with, or slightly faster than, a GTX 1060 6GB.

[log0]

I'm a bit surprised Vulkan is slower than OpenGL with NVIDIA proprietary driver (about at more 5 FPS lesser).
So the RADV driver seems very mature (about at more 5 FPS lesser), that's good work for this Open Source driver.

It probably says more about the quality of the Vulkan port than the driver.

[efikkan]

Well, in my opinion nowhere they run as they should. When an RX 480 is about as fast as a GTX 980Ti or slightly below a GTX 1070 one can say they run as fast as they should.
The state of RADV and the overhead and multi threading issues of OpenGL prevent this.

What? A GTX 1070 is >50% faster than RX 480.

How is "multi threading issues of OpenGL" preventing this? There is clearly no CPU bottleneck.

@phoronix

Why no AMDGPU-PRO benchmarks? You should already know radv isn't so good optimized yet.

A driver is in really bad state when it needs to be optimized for every game.

[Guest]

AMDGPU-PRO doesn't work yet on Ubuntu 16.10+....

That's a shame.

[oooverclocker]

1070 isn't right for an RX480 comparison. It should be on par with, or slightly faster than, a GTX 1060 6GB.

Why? Because AMD couldn't sell it for higher prices? Because the card isn't busy in most games? Where did you get the idea that the RX 480 currently performs optimal on Windows? And this performance is the only argument for the final price tag...

A GPU should perform with it's maximum theoretical performance. Otherwise many shaders don't get stressed which means that the internal workload distribution is not optimal or the drivers or the applications are not optimal. An RX 480 has >25% more computing power than a GTX 1060. This theoretical calculation is proven practically by few highly optimized applications and games.
When you sum up all shader performance like 64/32/16... bit ALUs in the card and get the same performance out like a weaker card with less theoretical performance, why should you be happy with the result? And you can even stress the card when you push calculations on unused shaders. With enough workload you would mostly land exactly on the level that the theoretical performance tells you.

We have also seen Futuremark tables that show more overhead on AMD cards and so on. There is no doubt that a RX 480 is the better hardware but with worse drivers and games that aren't optimized to stress the high shader count and use all CPU cores to fire draw calls it just performs on the same level in average. Which means sometimes it performs extremely bad in comparison to the GTX 1060, when the game engine is crap, and it performs much better than the GTX 1060 when the game engine is a masterpiece.

[bridgman]

@phoronix

Why no AMDGPU-PRO benchmarks? You should already know radv isn't so good optimized yet.

A driver is in really bad state when it needs to be optimized for every game.

I don't think anyone is talking about per-game optimizations; more that radv is just starting to be optimized in a generic sense.

[bridgman]

An RX 480 has >25% more computing power than a GTX 1060. This theoretical calculation is proven practically by few highly optimized applications and games.

I don't think so. If you look at typical boost clocks for the two chips the difference is less than that.