NVIDIA/AMD Linux Gaming Performance For Hitman 2 On Steam Play

[drlamb]

The more I look at your results the weirder they become.

Vega64 at 1440p:

On Windows: 51.2FPS
On Linux: 65.1FPS

This can't be right. Linux cannot run games faster than Windows in emulation (two layers of emulation actually: Win32->Posix, Direct3D11->Vulkan and not to mention that the graphics stack in Windows works close to the kernel and in X.org works as a user process in Linux).

Or maybe it just means you can't directly compare two independent tests done on different hardware? 8700K vs 9900K.

[birdie]

Keep in mind no idea what CPU was used in the Windows results you are citing where as my system was with 9900K, among other factors.

TPU uses Intel Core i7 8700K which is not too much slower than 9900K in absolute most games.

I strongly suspect DXVK disables certain Direct3D 11 features for AMD GPUs that's why they end up running faster in Linux than in Windows.

[Brisse]

This can't be right. Linux cannot run games faster than Windows in emulation (two layers of emulation actually: Win32->Posix, Direct3D11->Vulkan and not to mention that the graphics stack in Windows works close to the kernel and in X.org works as a user process in Linux).

It is not impossible, only unlikely. Imagine a case where Windows performance was crippled due to driver regressions or poor optimization and it worked perfectly well on Steam Play. The latter could win the performance race in such a case. Actually, this might be the case with HITMAN 2. Look at Radeon cards on Windows compared to Nvidia and you will probably notice how poorly Radeon performs in this game on Windows. The Steam Play results are more on par if you compare red and green team.

Also, I hate to be that prick, but Wine is not an emulator (but you already knew that I guess)

[betam4x]

TPU uses Intel Core i7 8700K which is not too much slower than 9900K in absolute most games.

I strongly suspect DXVK disables certain Direct3D 11 features for AMD GPUs that's why they end up running faster in Linux than in Windows.

I haven't tried hitman 2, but it is well known that DXVK can be faster on Linux than on Windows. I know this based on my own gaming with my 1080ti. Sometimes there are bugs, yes, but it's also Vulkan and a shader recompile likely happens, so there is a ton of room for optimization when doing the translation. Disclaimer: I haven't looked at the source code myself. However, I know that with proton Valve is working hard.

I am a bit perplexed by his frame time chart though. If I am reading it right, it should be inverted.

[Etherman]

Dxvk spoofs an AMD GPU by default because compability reasons (nvapi) so at least the games code path should be the same.

[humbug]

Great benchmarks. Thanks for including the R9 290 Michael, lots of us still using Hawai GPUs. It's unfortunate the development effort has stopped to move the amdgpu work past experimental.

But still fantastic to see it as a 5 year old GPU performing so close to the gtx 1060. The aftermarket models are even quicker as they maintain higher clocks without thermal throttling.

[birdie]

It is not impossible, only unlikely. Imagine a case where Windows performance was crippled due to driver regressions or poor optimization and it worked perfectly well on Steam Play. The latter could win the performance race in such a case. Actually, this might be the case with HITMAN 2. Look at Radeon cards on Windows compared to Nvidia and you will probably notice how poorly Radeon performs in this game on Windows. The Steam Play results are more on par if you compare red and green team.

Also, I hate to be that prick, but Wine is not an emulator (but you already knew that I guess)

Wine is not an emulator but if you took a quick look at Wine sources you'd discover that very few Win32 APIs translate directly to POSIX calls: most Win32 calls require a lot of shim code to run on Linux which doesn't make this translation/emulation/whatever free.

Until someone who owns this game and compares its screenshots taken in Windows and Linux using Photoshop's layer difference, there isn't much to talk about. I could have carried out the comparison but I don't have any AMD GPUs in my possession. Michael has but I'm not sure he'd be willing to do that since it's relatively time consuming.

[Masush5]

TPU uses Intel Core i7 8700K which is not too much slower than 9900K in absolute most games.

I strongly suspect DXVK disables certain Direct3D 11 features for AMD GPUs that's why they end up running faster in Linux than in Windows.

Hitman 2 running faster on DXVK + radv compared to native AMD DX11 was already discovered months ago by reddit users:

https://old.reddit.com/r/linux_gaming/comments/alpcza/hitman_2_possibly_faster_in_proton_than_windows/

(there are more posts than this one)

It's not a case of DXVK cutting corners, it renders everything correctly, it's just that AMD's DX11 driver is typical AMD cpu-boundware and ends up getting beating by DXVK + radv in some games, even with the translation layer overhead. This doesn't happen on nvidia because their dx11 driver is very efficient.

[birdie]

Hitman 2 running faster on DXVK + radv compared to native AMD DX11 was already discovered months ago by reddit users:
https://old.reddit.com/r/linux_gamin..._than_windows/
(there are more posts than this one)

It's not a case of DXVK cutting corners, it renders everything correctly, it's just that AMD's DX11 driver is typical AMD cpu-boundware and ends up getting beating by DXVK + radv in some games, even with the translation layer overhead. This doesn't happen on nvidia because their dx11 driver is very efficient.

I'll wait for screenshots. I can't imagine AMD Windows drivers to be this bad. We are talking up to 20% difference which is just huge.

[Michael]

FWIW, in the next week or two hope to do a Windows vs. Linux Steam Play comparison with the same systems/GPUs.