Intel XTU, Throttlestop seem to have different thresholds

Discussion in 'Hardware Components and Aftermarket Upgrades' started by kiteman112000, Jun 9, 2020.

  1. kiteman112000

    kiteman112000 Notebook Enthusiast

    Reputations:
    0
    Messages:
    38
    Likes Received:
    3
    Trophy Points:
    16
    Sager laptop with i7 9750h. I was using TS forever at 125mv undervolt (this unit came at 50mv already). But playing Warzone mostly I would get Dev Error 6328 a lot and the game would crash. Many people pointed to TS as being the fault, so I switched over to Intel's tool. But now even at 125mv I get blue screens occasionally. I didn't even get BSOD with TS. But on XTU I also don't get dev error crashes in Warzone, so I'm not really sure what to do.

    If XTU can't undervolt as well as TS, the temps get way too hot on XTU. Just not sure why there would be any difference. On XTU the only setting i change is the undervolt. In TS i go a little further and enable speed shift.
     
  2. unclewebb

    unclewebb ThrottleStop Author

    Reputations:
    7,180
    Messages:
    5,986
    Likes Received:
    5,399
    Trophy Points:
    681
    ThrottleStop was not the problem. Your undervolt settings that were not properly tested was the problem. Lots of 9750H CPUs are not stable when the cache is set to -125 mV. You should have used TS to fix that problem.
    There are times when XTU does not apply the voltage adjustment that you have requested. It looks like it is working but if you check the voltage registers in the CPU, it is not working. XTU has proven to be completely unreliable.
    That is true, especially on the 9750H. Some of these CPUs perform much better when the core offset voltage is set significantly lower than the cache offset voltage. XTU does not allow you to do that so you end up with a hotter running CPU that does not perform as well.

    Spend some time doing some testing with Cinebench R20. A laptop with a 9750H should be up in the 3000 to 3200 point range. For stability testing, try running 1 or 2 threads of the TS Bench test or 1 or 2 threads of Prime95. Your full load voltage might be OK. Many laptops will BSOD when lightly loaded so it is important to do some light load testing.
     
    Last edited: Jun 10, 2020
    Arondel and tilleroftheearth like this.
  3. kiteman112000

    kiteman112000 Notebook Enthusiast

    Reputations:
    0
    Messages:
    38
    Likes Received:
    3
    Trophy Points:
    16
    Thanks for this. I started playing around and left cache at -100. I did core from -150, -175, -200. All 3 got about 3100 on Cinebench, with no noticeable difference in heat between the three. Does it make sense to keep going down in core until I see some difference in scores/temps, or just the opposite and keep going up until I see the same, for better or worse in either direction? Thanks for the help.
     
  4. kiteman112000

    kiteman112000 Notebook Enthusiast

    Reputations:
    0
    Messages:
    38
    Likes Received:
    3
    Trophy Points:
    16
    Any thoughts?
     
  5. miloaisdua

    miloaisdua Notebook Enthusiast

    Reputations:
    23
    Messages:
    44
    Likes Received:
    25
    Trophy Points:
    26
    you can keep going and watch your max temps and power consumption. you can stop when you don't see any further improvement while still maintaining stability. This cpu is locked so once you remove any power and thermal limits it'll just run at 4ghz all core so you can't really push your cb20 score further once you're at that point.

    on my i7 9750h, i can go -600 on TS and it makes no diff but that's because in reality i think it maxes out at -200mV. i get 3140-3190 on cb20 with -200mv core offset and -110mv cache offset. on a cb20 run from cold/idle, cpu will draw up to 78w to run at all core 4ghz and my max temp would hit around 80c. (from an idle temp of approx 38-40 today). Power draw increases as temps increase.

    you can also try the blender benchmarks. bmw would be fine, classroom takes much longer. On Blender asks for much more power out of the cpu and on mine it pulls up to 86w which i can't sustain due to voltage regulator thermal limits on my laptop. i think my best score is 4m22s for bmw and something like 13.5min for classroom. https://opendata.blender.org/


    screenshot below for scores and for ts screen after some consecutive cb runs.
    [​IMG]
     
    Last edited: Jul 11, 2020
Loading...

Share This Page