I'm frequently testing this game and there's not much room for optimization here, certainly nothing that explains 25% losses that can be attributed to DXVK rather than GPU-specific driver issues and some general engine deficiencies that we can't do anything about.
Resizeable BAR is required, but that seems to be enabled on your system already anyway.
At least since your Windows+DXVK results also seem significantly better than DXVK+linux, I would suggest that you report it over at https://gitlab.freedesktop.org/mesa/mesa as a driver bug. You could also double check with AMDVLK on linux, which should be basically the windows vulkan driver (though now abandoned).
Tested AMDVLK as well and interestingly it's a little worse than RADV by about an additional 2%. Of course it hasn't been updated since Q2 but it seems there's something majorly different Windows side. Mesa issue has also been opened (https://gitlab.freedesktop.org/mesa/mesa/-/issues/14305)
Does it run any better with RADV_PERFTEST=nggc?
Does it run any better with
RADV_PERFTEST=nggc?
No difference unfortunately.
The Linux performance maybe have improved a bit with current mesa main
The Linux performance maybe have improved a bit with current mesa main
Seems that way, haven't done any long runtime testing but a quick check with mesa-git as of this comment (25.2.6/26.0.0)
ANH Start: 93/101 (+8%, -7% vs WinVK, -18% vs D3D11)
ANH End: 68/72 (+6%, -5% vs WinVK, -14% vs D3D11)
Halved the gap to the proprietary driver although translation still seems the heaviest piece of the chain.
Small update, this game is affected by a RADV limitation where fast clears didn't work for some images, which is partially fixed in mesa main as of https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39394 . Not sure off the top of my head to which extent this ever affected GFX12, but it does hurt the older GPUs.
The game may also see a small improvement from #5455. I'm also working on single pass mipgen right now, which further shaves quite a few microseconds off the frametime.
Not sure if all of this is enough to explain a >10% delta, but especially the slow clear thing can be bad.
Tested again with mesa-git & Proton-GE 32 (should include the above changes).
Numbers for the "End" scenarios won't be comparable unfortunately as I've had to use a different save point.
| Scenario | Stage | D3D11 | WIN_VK | LIN_VK | %- Win | %- Lin |
|---|---|---|---|---|---|---|
| ANH | Start | 122 | 111 | 113 | 9.02% | 7.37% |
| End | 88 | 78 | 78 | 11.37% | 11.37% | |
| LA | Start | 126 | 109 | 114 | 13.50% | 9.53% |
| End | 98 | 86 | 88 | 12.24% | 10.30% |
The situation RADV side seems to have been resolved as the performance is effectively the same regardless of driver.
Closing because there's nothing immediately actionable here, just a bunch of game-side inefficiecies that do hit us harder than native but no obvious issue on the DXVK side anymore.
There is an up to 25% performance drop-off when using Vulkan translation in Frostpunk, substantially above what I would usually see for a wider variety of titles. When testing across native D3D11, DXVK on Windows and DXVK on Linux, the stepwise performance loss is about 10% at each stage.
Tested with two of the in-game scenarios: A New Home (ANH) and The Last Autumn (LA) at both the start of the scenario and at a save some way into each (game becomes more CPU-bound as the population and map populates).
Related to https://github.com/doitsujin/dxvk/issues/2520 but not isolated to Nvidia/Windows environments.
ANH (Start)
D3D11:
Vulkan (Windows):
Vulkan (Linux):
Software information
Frostpunk (2018)
All graphical settings set to their highest values.
Borderless Windowed w/ VSync disabled.
System information
Apitrace file(s)
Unsure if correct, game took over 10 minutes just to reach gameplay.
https://mega.nz/file/51sn3LbR#bqbFV1ZitanL2pn7_Pw4tOaD0nMcwiy1AvEyDrKWRFQ
Log files
steam-323190.log