Stutters in games/system even if 1 GPU sensor is enabled (monitoring)

Discussion in 'Videocards - NVIDIA GeForce Drivers Section' started by Donater, Mar 30, 2024.

Thread Status:
Not open for further replies.
  1. Donater

    Donater New Member

    Messages:
    4
    Likes Received:
    1
    GPU:
    GameRock RTX 4090
    Unwinder, Алексей, не закрывайте, пожалуйста, мой пост (вопрос). Может быть конечно я тупой и правда не прав, но я вроде бы чётко указал на свою проблему, которая ухудшалась со временем, как будто-бы из-за Windows в том числе (и я понимаю что проблема на моей стороне как бы, я не обвиняю программы мониторинга): включение мониторинга даже одного любого датчика с видеокарты через HWiNFO (например Hot Spot раз в 20 секунд) вызывает на моей системе такую проблему и я хочу понять почему, может быть придёт человек с такой же проблемой и напишет, что решил её, условно, заменой видеокарты, я соответственно буду дальше работать этом направлении. Спасибо.
    ----------------------------------------------
    Stutters in games/system with monitoring programms enabled (almost any other video card monitoring). Is the problem in the components or in the system?
    STUTTERS screenshots
    Monitoring at least one of any GPU sensor through MSI Afterburner, HWiNFO, GPU-Z programs... causes constant stutters that coincide with the monitoring interval (whether it is 1 sec or 10 sec). There is no such obvious problem with FanControl.
    I also observe the problem of stutters without monitoring in some games and cannot confirm that everyone has it (is the problem on my side?)

    Maybe the problem is in the components and it all didn't make sense, but here's what I tried to turn on/off/update/change the settings in different variations/…
    • Mobo BIOS (GPU BIOS haven't tried)
    • iGPU
    • ReBar
    • fTPM
    • Spread Spectrum
    • Windows 10 fresh install (11 haven't tried)
    • chkdsk dism sfc scannow
    • Windows Power Plan (USB selective suspend...)
    • VBS - HVCI
    • HAGS
    • NVCP settings
    • DDU + NVCleanstall + 537.58
    • G-Sync V-Sync RTSS fps cap
    • etc
    CPU - AMD Ryzen 7 7800x3d
    MOBO - MSI MPG B650I Edge WIFI
    DRAM - Team Group 6000 cl38
    GPU - Nvidia RTX 4090 Palit GameRock OC
    PSU - PS-750SFX

    I saw several posts with similar problems, here is an example (in some games my fps/frametime lines looks about the same) - POST1 _ VIDEO _ POST2
    rtss stut.jpg
     
    Last edited: Mar 30, 2024
  2. Unwinder

    Unwinder Ancient Guru Staff Member

    Messages:
    17,272
    Likes Received:
    7,070
    And one more time: an approach of using web browser for stutter testing is fundamentally broken by design. Browsers are not games, they are not intended to handle relatime tasks and cannot handle smooth animation when any background processes are doing some periodic processing simultaneously with it. Even Marc at your own links tells you so.
     
  3. Donater

    Donater New Member

    Messages:
    4
    Likes Received:
    1
    GPU:
    GameRock RTX 4090
    check IMGUR screenshots, i used RTSS, CapFrameX etc. UFO Test is just an example, but very unfortunate, I agree.
     
  4. Unwinder

    Unwinder Ancient Guru Staff Member

    Messages:
    17,272
    Likes Received:
    7,070
    You're making a mess of completely different apps/factors/tests and trying to generalize the reasons, but in fact you're searching for a black cat in a dark room.
     
    pegasus1 likes this.

  5. Unwinder

    Unwinder Ancient Guru Staff Member

    Messages:
    17,272
    Likes Received:
    7,070
    Also, it is not a question of enabling any single GPU sensor like you wrongly claim. Different sensors are not equal from polling time point of view. Just one specific sensor can eat more performance than combination of 20 other sensors. That's why different monitoring software provide sensor performance profiling features, like I explained in the link I already provided in your previous thread. HwInfo allows you to diagnose and nail down the slowest sensors. MSI AB is allowing you to diagnose and nail down the slowest sensors:

    upload_2024-3-30_14-11-33.png

    upload_2024-3-30_14-11-47.png

    Monitor your system wisely!
     
  6. aufkrawall2

    aufkrawall2 Ancient Guru

    Messages:
    4,798
    Likes Received:
    2,018
    GPU:
    7800 XT Hellhound
    Hm, I wonder if this is normal: Afterburner reports my GPU (7800 XT Hellhound) temperature polling time as ~17ms, whereas hwinfo reports ~4ms GPU total (including temperature and everything else).

    Edit: If I disable temperature, then GPU usage (which then is the top entry) is displayed as ~17ms instead.
    hwinfo isn't showing individual GPU sensor latencies for me, only GPU total category latency.
     
    Last edited: Mar 30, 2024
  7. Unwinder

    Unwinder Ancient Guru Staff Member

    Messages:
    17,272
    Likes Received:
    7,070
    Yes it is normal for AMD ADL8 API design, and yes it was asked and answered before. Unlike NVIDIA, which provide independent polling APIs for different subsets of sensors so you can choose what do you want to poll, AMD's ADL2_New_QueryPMLogData_Get poll all sensors at once in single call. So you see cumulative value for the first sensor and the rest simply reuse the same cached data (so polling time for them is effectively 0).
    NV also group some sensors and poll them in single call, e.g. GPU/VRAM clocks are polled in single call, both fan tachometer readings are grouped, GPU usage / bus usage / fb usage calls are combined too.
     
    Last edited: Mar 30, 2024
    aufkrawall2 likes this.
  8. aufkrawall2

    aufkrawall2 Ancient Guru

    Messages:
    4,798
    Likes Received:
    2,018
    GPU:
    7800 XT Hellhound
    Thanks, very illuminating. Will try to make up for my already asked question when other users ask this again. :D

    Hm, though it's not yet clear to me why hwinfo total polling is ~4ms vs. ~17 in AB.
     
  9. Unwinder

    Unwinder Ancient Guru Staff Member

    Messages:
    17,272
    Likes Received:
    7,070
    ADL's native GPU activity moniotring is bugged for decades for old GPUs, which could result in reporting false/jigsaw GPU activity values if you don't force the thread to sleep prior to polling ADL sensors. So polling time for ADL includes additional artifical sleep delay applied immediately before each polling iteration. AFAIR it was controllable and could be disabled somewhere in ADL related section inside MSIAfterburner.cfg
     
  10. aufkrawall2

    aufkrawall2 Ancient Guru

    Messages:
    4,798
    Likes Received:
    2,018
    GPU:
    7800 XT Hellhound
    When I set ActivityMonitoringDelay = 0 , the polling latencies in total are 1-2ms now. Would these be realistic numbers?
     

  11. Unwinder

    Unwinder Ancient Guru Staff Member

    Messages:
    17,272
    Likes Received:
    7,070
  12. aufkrawall2

    aufkrawall2 Ancient Guru

    Messages:
    4,798
    Likes Received:
    2,018
    GPU:
    7800 XT Hellhound
    Thanks a bunch!
    I guess the sleep delay is harmless, despite the higher numbers reported? Then it wouldn't really matter if the workaround was dropped or not. At least I haven't noticed any stutter issue introduced by Afterburner polling on that card. (Same goes for hwinfo. GPU-Z used to be more critical, having its sensor tab opened used to turn vsynctester.com into a stuttery mess. But that seems to be gone.)
     
  13. Unwinder

    Unwinder Ancient Guru Staff Member

    Messages:
    17,272
    Likes Received:
    7,070
    Sleep delay is harmless, CPU idles during this time.
     
    aufkrawall2 likes this.
  14. aufkrawall2

    aufkrawall2 Ancient Guru

    Messages:
    4,798
    Likes Received:
    2,018
    GPU:
    7800 XT Hellhound
    Great. Looks like AMD did something right, with TBP being pollable at almost no latency (just took them over a decade vs. ASIC-only power...).
    Afair Nvidia offers to poll slower (but more exact, or AMD's doesn't smooth out spikes in a meaningful manner) power sensor even in GeForce Experience overlay. And since it can't show polling latency, lots of users might be affected by stutter due to this without knowing. Just recently witnessed this on a MSI 4090 Suprim, the stutter it introduced in CP77 was really absurd. :confused:
     
    Last edited: Mar 30, 2024
  15. Unwinder

    Unwinder Ancient Guru Staff Member

    Messages:
    17,272
    Likes Received:
    7,070
    Intel Arc are the most effective ones from sensor polling efficiency implementation point of view. They also poll all sensors in single call and cumulative time for all of them is insanely good, less than 1 ms.

    upload_2024-3-30_20-56-10.png
     
    aufkrawall2 likes this.

  16. aufkrawall2

    aufkrawall2 Ancient Guru

    Messages:
    4,798
    Likes Received:
    2,018
    GPU:
    7800 XT Hellhound
    What's weird is that in hwinfo, total category polling latency doesn't go down, even when disabling monitoring and logging for all sensors it includes. I guess that means you practically can't disable troublesome sensors with hwinfo, unless disabling the whole category (unlike with AB). :(

    It's freaking 15-20ms on that 4090 Suprim...
     
  17. Unwinder

    Unwinder Ancient Guru Staff Member

    Messages:
    17,272
    Likes Received:
    7,070
    20ms is more than normal value for majority of systems. For default 1000ms polling period it is 2% of single core CPU time. For CPU with 32 logical cores it is 0,0625% of total CPU load consumed by polling.
    Also, do not forget that CPU execution times are CPU clock dependent, so 20ms polling time measured when CPU idles and let's say run at lowest 1GHz, become just 4ms when CPU is clocked at 5GHz.
     
  18. aufkrawall2

    aufkrawall2 Ancient Guru

    Messages:
    4,798
    Likes Received:
    2,018
    GPU:
    7800 XT Hellhound
    For just the power sensor, hwinfo also reports just ~4ms. I have no idea where most of the latency for the 4090 graphics card sensor category in hwinfo is coming from. It seems it can't be linked to any individual sensor. :eek:
    But will check again.
     
  19. Donater

    Donater New Member

    Messages:
    4
    Likes Received:
    1
    GPU:
    GameRock RTX 4090
    For now “problem” seems to have been fixed in some scenarios. I reinstalled the Nvidia driver again: DDU + regular clean install 551.86 (without NVCleanstall). Some games still show terrible performance, but apparently the problem is with them (for example, Midnight Ghost Hunt, which I rechecked again on another computer and similar stutters appeared there). I will continue to check these games and find out from other users if they have encountered such problems. Here are screenshots that show what used to happen in Cyberpunk 2077 even during the pause (only MSI Afterburner + RTSS + FanControl were enabled) and that's what's now, a clean, smooth graphs without obvious problems.
    And another dumb question, how do I get the CPU line back above the GPU. I used to do it, but now it doesn't work.
    cyberpunk.png
    ---------------------
    cyberpunk1.png
    -
    Алексей, спасибо Вам большое за полезную информацию и такие ультимативные программы, ещё раз всё перепроверив по вашим советам проблема как-то решилась, на сенсорах аномальных задержек не было замечено до решения проблемы... возможно я что-то упускаю.
     
    Last edited: Mar 31, 2024
    Unwinder likes this.
  20. RealNC

    RealNC Ancient Guru

    Messages:
    5,494
    Likes Received:
    3,750
    GPU:
    4070 Ti Super
    I don't understand how polling can cause stutter. It's performed by a separate process. Is the GPU blocked from rendering when polling specific sensors?
     
Thread Status:
Not open for further replies.

Share This Page