Steam Client Stable vs. Beta Tests With Vulkan On AMDGPU-PRO

[davidbepo]

Ok I actually got a pretty good laugh on that one. Have you seen the bug tracker on github?

they are good at FIXING bugs but not at not commiting them

a lot of these bugs are caused because of using ubuntu 12.04 32 bit libs which seems stupid to me

[bridgman]

Seems to be the bottleneck of AMD anyway. Any idea about that, bridgman?

That's a fairly old post, but AFAICS the issue with DX11 was similar to the one with OpenGL - the standard calls for a lot of error checking to maintain consistent behaviour between vendors, and if you actually implement all that error checking it slows you down by increasing CPU overhead. You can work around that to a certain extent by doing more multi-threading in the drivers but you start to get into vendor-specific behaviour fairly quickly.

The next gen APIs pull multi-threading into the standard and move the error checking out to validation layers used during development rather than on every user's system, so now you're seeing more of an apples-to-apples comparison with DX12 and Vulkan.

I don't think that explains all of the differences but it's probably the biggest single factor.

[cRaZy-bisCuiT]

bridgman Thank you for the detailed information! I'm a dev myself, even if I'm not into GPU programming and low-level stuff I get your point. I was just wondering why AMD stepped into that bottleneck much more than nVidia did.

The problem with Vulkan is the not at all optimised code I assume. If the whole capability of Vulkan would be used, the problem should be solved I assume.

[VikingGe]

For what it's worth, similar problems can go with OpenGL.

But on a completely different level. OpenGL doesn't let you choose where to allocate memory for textures, buffers etc. - you may run out of memory, but that's about it.

At 1920x1080 it appears that AMD cards hit the CPU limit sooner than Nvidia, which seems weird.

We're talking about a difference of ~15% here. I'm not saying it is entirely irrelevant, but in this specific scenario (100+ FPS) it definetly is - it's not nearly as concerning as the Dx11 gap (although I think that has improved quite a bit)

[DonQ]

I've noticed from previous tests that the R9 290 seems to overall be right where it should be. It's performance is supposed to pretty balanced with the GTX 970 (better in some tests, worse in others), which at least the 2nd graph shows here.

What drivers? I have an R9 390 and an R9 380x, Debian Unstable kernel 4.7 rc2 and Ubuntu 16.04 with amdgpu-pro 16.30, and the 390 sucks. It's slower than the 380x and when it isn't, it crashes.

[johnc]

they are good at FIXING bugs but not at not commiting them

a lot of these bugs are caused because of using ubuntu 12.04 32 bit libs which seems stupid to me

Oh please. They don't even fix bugs on Windows either. They basically release a game, get your money, and then transition it to abandonware.

[bridgman]

What drivers? I have an R9 390 and an R9 380x, Debian Unstable kernel 4.7 rc2 and Ubuntu 16.04 with amdgpu-pro 16.30, and the 390 sucks. It's slower than the 380x and when it isn't, it crashes.

Just checking, you mean amdgpu-pro 16.20, right ? The 16.30 beta is just for SteamOS and Bonaire - you'll probably have more luck with the more broadly tested 16.20.

[schmidtbag]

What drivers? I have an R9 390 and an R9 380x, Debian Unstable kernel 4.7 rc2 and Ubuntu 16.04 with amdgpu-pro 16.30, and the 390 sucks. It's slower than the 380x and when it isn't, it crashes.

I haven't used the pro drivers, but I own an R9 290, used in Debian sid for a while, and it performed just fine. If you think it's lower than the 380X, you probably did something wrong.

[SaucyJack]

It's not that weird! There's a very lon thread in the german forum of the magazine "Hardwareluxx" about that issue (with DirectX, Windows):

http://www.hardwareluxx.de/community/f14/amd-und-die-bescheidene-performance-im-dx11-cpu-limit-1089150.html

Seems to be the bottleneck of AMD anyway. Any idea about that, bridgman?


@Michael: Thank you very much for retesting! Now we could see the 290 seems to perform not too bad! Strange thing: The 285 didn't gain that much of a performance improvement. By the way: The 390 should be working with AMDGPU as well, right? Since it's just a rebrand... If Polaris will perform well under Linux I'll definetely get one (480(x))!

Given Nvidia's driver history, they could have just thrown the standard out and cut a bunch of corners yielding lower CPU overhead.

[jaxxed]

You raise an interesting problem with Vulkan and DX12: All this low-level power also gives developers power to shoot themselves in the foot. Valve is pretty good about tracking down and fixing the issues, but by no means are all game developers as skilled as Valve.

It's my understanding that anyone can enable the debugging layers on their vulkan setup, and get understanding of what could be going wrong. Any dev company that makes stupid mistakes is going to have armchair support blowing out their Mail server.