NVIDIA Gaming/GPU Performance: Windows 11 vs. Ubuntu Linux Benchmarks

[Luke_Wolf]

... seeing the results of the comparison on amd linux vs windows:

https://www.phoronix.com/review/radeon2022-windows-linux

I suspect that there is something odd/wrong in Michael setup or his general approach to benchmarking, specifically the requirement to run all automatically via scripts may hide some quirks or differences that would not pass unnoticed if testing is done with higher degree of manual review (see the comment above about differences in quality in Portal Linux vs Windows).

Talking AMD result with mesa, the BeroTech youtube channel has shown that the performance are very close linux vs windows:

https://www.youtube.com/@beronori/videos

In the case of Marvel’s Spider-Man: Miles Morales, Linux beats Windows by 10 FPS on average (83.9 vs 73.6) on rx6700xt and 5800x -- minute 3.06 in the bleow:

https://www.youtube.com/watch?v=B7OzNQ9IUbg

Phoronix's results show completely different outlook of performance (with windows winning hands down) across AMD and Nvidia

Something does not add up in what shown on Phoronix (unfortunately)

There's definitely something wrong with his setup, Cyberpunk 2077 is broken for him on the RX 7900 XTX but works perfectly fine for me on mine, indeed I've actually seen no issues with any games on my system that are specific to this card.

[bug77]

Considering that Nvidia is not a prime example of good driver support on the Linux side I'm not very surprised to be honest. Its how I was expecting it. But I think its no secret that for gaming in Linux AMD and Intel is recommended. Especially with mesa.

I'm not sure exactly why you were expecting that. That "poor support" comes down to Nvidia using basically one driver across all operating systems, with as small a shim as possible to adapt it to each platform. With that in mind, the default expectation would be to see the same performance across the board, unless the shim itself becomes a bottleneck. Or if there are bugs.

[piotrj3]

I'm not sure exactly why you were expecting that. That "poor support" comes down to Nvidia using basically one driver across all operating systems, with as small a shim as possible to adapt it to each platform. With that in mind, the default expectation would be to see the same performance across the board, unless the shim itself becomes a bottleneck. Or if there are bugs.

Not issue on Nvidia side. Notice in computing Nvidia on ubuntu is even faster then windows (and that is pretty much everywhere).

Where is key is that Windows DWM (and in general scheduling) is much more optimized for full screen aplications.

Run something fullscreen? Run something borderless? Windows will basicly skip all the composing and allow basicly render directly content of game to screen. The way windows works is like gamescope or (at minimum) Kwin's Alt-Shift-F12.

Linux also doesn't work that well 1-1 with explicitly synced Nvidia driver. There are oversynchronization issues.

in case of AMD on Linux vs Windows, keep in mind AMD has totally diffrent open source driver to windows one. Some time ago openGL driver on Windows from AMD was totally garbage and AMD on linux could easly outperform here windows. Now things improved tho.

[middy]

I'm surprised the gap between Windows and Linux on NVIDIA is that large – I thought it was only 3-4% on average at this point (with DXVK sometimes outperforming native on Windows).

stuff like wine/proton/dxvk/vkd3d/etc still adds overhead. but the fact it has overhead and the gap is just small linux is doing great. the gap isn't big enough to really matter outside benchmark epeen where people declare a winner when something is 0.5% faster. just like the amd linux vs windows results where amd did really great, especially with the rdna2 results being a small 6% gap, nvidia is doing good too.

and remember, sometimes is the exception, not the rule. never take exceptions and try to present them as a rule. even the developers of dxvk will tell you the majoirty of time, games running in dxvk will perform slightly worse because of the overhead. dxvk states at best, a 5% gap. typically the few times you see a difference is because of really bad driver optimizations for something. see amd and opengl on windows for the longest time. or because hacks cause something to run faster because its not doing something correctly. like its not rendering somethings in the background or some sort of "feature" that's performance heavy doesn't work.

[JoshuaAshton]

Portal and Portal 2 don't ship GL at all on Windows. This was likely benchmarked against D3D9, which is shown in the window title.

[theriddick]

Pretty good overall, with some CPU overhead issues showing up at 1080p and 1440p for several titles, but that can be expected and hopefully resolved one day.

Really I don't know why people hate on NVIDIA for gaming under Linux so much except for the obvious limitation with open-source stuff. But if you only care about gaming performance does it matter?

Also does DLSS3.0 work under Linux yet? I know previous versions work(ed).

[Linuxxx]

Michael

Would it be possible for You to re-run the "Strange Brigade" benchmark with any CPU other than Intel's hybrid x86 architecture?

My gut feeling tells me that this is yet another case of the Linux kernel placing the game threads onto the E-cores instead of the P-cores.

Notice how the difference between Linux & Windows kept shrinking with inreased render resolutions, when the bottleneck shifted from the CPU to the GPU.

Another reason why I think the Intel 13900K is a poor choice on Linux is the fact that even though the performance governor was used, it still didn't manage to beat Windows' default "balanced" power-plan, even though doing so has always resulted in Linux easily beating Windows on non-hybrid CPUs in the past.

Therefore the conclusion of this particular benchmark should be that hybrid CPUs are best avoided on Linux...

(Which begs the question how Android is actually faring with these hybrid big.LITTLE ARM SoCs -- granted, no Windows there to compare.)

[ntropy]

Now I am able to run Cyberpunk 2077 on my 3070 Ti laptop gpu @70 fps in 2k, while in windows it would hit 74 fps... what a shame .

The compute results are nice tho, speak for Linux for upcoming blender devs.

[MrCooper]

Where is key is that Windows DWM (and in general scheduling) is much more optimized for full screen aplications.

Run something fullscreen? Run something borderless? Windows will basicly skip all the composing and allow basicly render directly content of game to screen. The way windows works is like gamescope or (at minimum) Kwin's Alt-Shift-F12.

Any Wayland compositor can do the same as gamescope, if the drivers can allocate scanout capable buffers.

Linux also doesn't work that well 1-1 with explicitly synced Nvidia driver. There are oversynchronization issues.

If so, that's still an nvidia driver issue. It mostly bypasses the infrastructure for implicit synchronization in the kernel (which results in issues such as xwayland glamor renders incorrectly on nvidia), handling synchronization itself instead.

[piotrj3]

Any Wayland compositor can do the same as gamescope, if the drivers can allocate scanout capable buffers.



If so, that's still an nvidia driver issue. It mostly bypasses the infrastructure for implicit synchronization in the kernel (which results in issues such as xwayland glamor renders incorrectly on nvidia), handling synchronization itself instead.

Can != does. Michael tests default settings (like he should) and by default composing is there even in full screen exclusive aplications. What is more (correct me if i am wrong) linux DEs doesn't predict turning off composing in borderless like situations.

Looking around internet i can see that Windows on exclusive full screen games generally (comparing to normal windowed mode) on same resolution has almost the same performance (like 1-3 fps range). On linux it can go easly 10fps+

2nd issue is how you would explain that AMD's results in geometric mean https://www.phoronix.com/review/rade...indows-linux/7
https://www.phoronix.com/review/nvid...dows11-linux/8
are proportionally very similar to Nvidia's. And Nvidia is even faster in compute on linux. But somehow both AMD and Nvidia everytime they interact with window system they seem to have suprisingly same drop in performance what is more even interesting we have same drop in performance in both X (nvidia on ubuntu is on X11) and AMD was on Wayland.
This suggest we have significant overhead in linux DEs.