A part of the players on one of my servers is complaining about random freezes/hitching while moving in the area where most players are (Los Santos). My guess is that these hitching could be related to the game spawning/removing the player’s ped but I may be wrong. I hope you can help me solve this issue.
I have attached a video and a trace of where this issue happened for a player.
No big event / heavy events were triggered during the experiment, pmms:sync is a redundant event from a public script recently added it is not the issue here.
Seems to be GPU memory running out, and the NVIDIA driver being slow at paging these days (leading to the D3D11 query call in rage::grcSetup::BeginDraw being slow). What are the graphics settings here (and is ‘extended texture budget’ set to the default of 0% as it should never be any higher for 4 GB of VRAM)?
One says that removing NVE (he did not tell me he had that) and reducing the extended texture budget issue fixed the issue
The other one with a 2060 with 6GB of VRAM says he still has small slutters with the texture budget sets to
I’m updating this thread as I’m getting the same issue right now on canary with Extended Texture Budget at around 40% (with a 3080) so it could be something else, here are the traces: 535.79 MB file on MEGA
It’s again the game waiting for the GPU driver completing the last frame. Seeing the amount of background processes (including ‘NVIDIA Broadcast’ and quite a few Electron apps that would be holding on to GPU allocations) and assuming a ‘RTX 3080’ is a card with still ‘only’ 10/12 GB VRAM, it’s likely the exact same issue again: running out of VRAM and paging stuff back in since your budget setting (+ any NUI usage, perhaps?) still exceeds VRAM allocation.
Traces don’t seem to contain exact details in that regard but the commit size for the game process is way over what it should be and it could be large part of this is GPU allocations (and, similarly, in-process NUI GPU might trip an edge case in the NVIDIA driver that causes larger/longer hitches, the NVIDIA driver really does not appear to be tested with multiple GPU contexts).
For numbers’ sake, a 40% budget slider (vid_budgetScale being 8) seems to make the grcResourceCache cap for ‘very high’ texture quality around 6575 MB (strmem shows this), adding other things in the game that might use GPU memory here - mainly NUI - and other heavy background apps such as likely some 1-2 GB consumption by ‘NVIDIA Broadcast’ and you’re likely to run into weird paging behavior.
cl_drawperf nowadays has a VRAM column that’ll show VRAM usage, for example on my system right now it’s as follows:
… which means ‘1.5 GB used by game D3D11 device, 5.6 GB used on the entire system, 23.8 GB total’, also indicating that the workload I’m running now is in fact using 4.1 GB of VRAM on background apps (+ NUI), checking Task Manager (Performance → GPU) a bit after exiting the game shows this is correct as my idle non-game workload is using ~3 GB, with top consumers (Details with the ‘Dedicated GPU Memory’ column visible) being chrome.exe, dwm.exe, Code - Insiders.exe, mstsc.exe and WindowsTerminal.exe.
To reduce this lag we actually briefly had some behavior in game that’d make it not allocate anything over the available VRAM, but since some servers actually have so much oversized content that the required/pinned content (network players, weapons, and vehicles) would already fill people’s entire budget on low-end GPUs this had to be reverted.
I have an interesting note regarding a detail we found when trying to remedy ‘stutters’ for players on my server. We found that actually installing NVE basically eliminated all stutters, and other players were reporting that QuantV also had the same effect. Any idea why graphics modifications can contribute to basically eliminating these hard freezes?
These hitches were also an issue for me when running just vanilla GTA5, so possibly worth seeing if it’s FiveM specific.
Personally I ended up running DDU and disabled some stupid windows thing called superfetch - can finally run two clients at mid-high settings again, whereas I was struggling to run one at low settings for like a year.