Help stabilize 9900KS overclock - water cooled!

[Alexauwa]

Ok. Then simple: Use VR VOUT for your voltage reading.

Just tired Prime95 or real bench stress test. VR Out is only 1.27 vcore. It drops down when stress test is running. On IDLE it is higher at around 1.31 vcore. Why is that and what vcore shall I set in BIOS? I tired 1.290 vcore and Real bench crashes. That is weird. Can someone explain this?

[Falkentyne]

There is such a thing called vdroop.
Vdroop is exactly equal to "VRM Loadline mOhms * CPU Amps current"

On the 9900k, LLC High (= LLC5 on Asus) is 0.8 mOhms and LLC Turbo (=LLC6 on Asus) is 0.4 mOhms.
The lower the mOhms value, the less the vdroop. But the transient response gets worse (voltage dips below the "Average" that is shown on sensors gets larger--large transients dips hurt your overall stability).

The Aorus Master functions best on LLC High. The same as most of the Asus boards (Gene, Hero, Code, Formula do better on LLC5 than LLC6 because of smaller transient drops).
VR VOUT and Vcore (Maximus XI on-die sense) cannot show transients. They happen in microseconds.

Transients can be seen here.
https://elmorlabs.com/index.php/2019...ne-visualized/

Avoid Extreme and Ultra Extreme on Gigabyte boards. Turbo is also not as good as high.
You can see the transient differences here by looking at the Prime95 AVX 16K tests.

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

[Alexauwa]

There is such a thing called vdroop.
Vdroop is exactly equal to "VRM Loadline mOhms * CPU Amps current"

On the 9900k, LLC High (= LLC5 on Asus) is 0.8 mOhms and LLC Turbo (=LLC6 on Asus) is 0.4 mOhms.
The lower the mOhms value, the less the vdroop. But the transient response gets worse (voltage dips below the "Average" that is shown on sensors gets larger--large transients dips hurt your overall stability).

The Aorus Master functions best on LLC High. The same as most of the Asus boards (Gene, Hero, Code, Formula do better on LLC5 than LLC6 because of smaller transient drops).
VR VOUT and Vcore (Maximus XI on-die sense) cannot show transients. They happen in microseconds.

Transients can be seen here.
https://elmorlabs.com/index.php/2019...ne-visualized/

Avoid Extreme and Ultra Extreme on Gigabyte boards. Turbo is also not as good as high.
You can see the transient differences here by looking at the Prime95 AVX 16K tests.

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

Thanks mate.

I have not set my vcore to 1.325 in BIOS. So far I do not have any issues at stress at all. RealBench gave me a crash at 1.320 so I raised to 1.325. LLC is turbo. Would high help me there? Turbo gives me slightly higher vdroops, right? Meaning in case I run high I might have to raise my vcore again?


Here is a screenshot under load. I'd just like to understand VID now as it is much higher than my vcore. 1.461 is pretty high. Does the cpu actually run at that voltage? Isnt it too high? Apart from that the 9900KS has a max vcore of 1.52. So looking at my min, max and avg vcore out my voltages are totally fine as well as my temps. Thanks for you help.

I also dont understand my it crashes with vcore of 1.320 if my max vcore out is only 1.316?

[Falkentyne]

Thanks mate.

I have not set my vcore to 1.325 in BIOS. So far I do not have any issues at stress at all. RealBench gave me a crash at 1.320 so I raised to 1.325. LLC is turbo. Would high help me there? Turbo gives me slightly higher vdroops, right? Meaning in case I run high I might have to raise my vcore again?


Here is a screenshot under load. I'd just like to understand VID now as it is much higher than my vcore. 1.461 is pretty high. Does the cpu actually run at that voltage? Isnt it too high? Apart from that the 9900KS has a max vcore of 1.52. So looking at my min, max and avg vcore out my voltages are totally fine as well as my temps. Thanks for you help.

I also dont understand my it crashes with vcore of 1.320 if my max vcore out is only 1.316?

Vdroop is how much your voltage drops BELOW the voltage set in BIOS. Your "minimum" voltage reading (if C-states, power saving, speedshift, etc are all DISABLED) is your maximum vdroop. So your voltage dropped down to 1.266v when you were drawing about 147 amps of current.

LLC Turbo has *less* vdroop than LLC High. (a lower 'mohms' value for resistance means less droop. It's part of Ohm's law).
Maybe this chart will help explain this more?

https://www.overclock.net/forum/6-in...l#post27860326

The problem with REDUCING the vdroop is the voltages "oscillation" you can NOT see on sensors, which is called "transient voltages".
Transient voltages (minimum) is what determines your "random" stability issues or random crashes--if transients drop too low, you might be stable for a short time then random crash. In this screenshot, you can see how "Stronger loadline calibration (higher LLC)" causes larger transients as well. "You can't have your cake and eat it too".

Vdroop is "Average voltage level that drops compared to BIOS setting",--VR VOUT shows this perfectly.

Transients are "voltage dips that drop BELOW the VR VOUT reading--you need an oscilloscope to see these." (There are also transient spikes also).

https://elmorlabs.com/index.php/2019...ne-visualized/

Compare the two links until you fully understand them.

In general, "LLC=5" or "Loadline calibration= "high" works best for Gigabyte boards. But you may have to increase "Bios" voltage a bit to compensate for the more vdroop you have.

[Alexauwa]

Vdroop is how much your voltage drops BELOW the voltage set in BIOS. Your "minimum" voltage reading (if C-states, power saving, speedshift, etc are all DISABLED) is your maximum vdroop. So your voltage dropped down to 1.266v when you were drawing about 147 amps of current.

LLC Turbo has *less* vdroop than LLC High. (a lower 'mohms' value for resistance means less droop. It's part of Ohm's law).
Maybe this chart will help explain this more?

https://www.overclock.net/forum/6-in...l#post27860326

The problem with REDUCING the vdroop is the voltages "oscillation" you can NOT see on sensors, which is called "transient voltages".
Transient voltages (minimum) is what determines your "random" stability issues or random crashes--if transients drop too low, you might be stable for a short time then random crash. In this screenshot, you can see how "Stronger loadline calibration (higher LLC)" causes larger transients as well. "You can't have your cake and eat it too".

Vdroop is "Average voltage level that drops compared to BIOS setting",--VR VOUT shows this perfectly.

Transients are "voltage dips that drop BELOW the VR VOUT reading--you need an oscilloscope to see these." (There are also transient spikes also).

https://elmorlabs.com/index.php/2019...ne-visualized/

Compare the two links until you fully understand them.

In general, "LLC=5" or "Loadline calibration= "high" works best for Gigabyte boards. But you may have to increase "Bios" voltage a bit to compensate for the more vdroop you have.

Got you mate. Thanks a lot. Thats why I would have to raise my vcore a bit for cinebench.

So looking at my VID now. Are these fine and safe? What does VID tell me now as there pretty much out of the frame and much higher?

[Alexauwa]

Are you setting manual voltage...or adaptive? And what power plan settings...
Any avx offset....latest Realbench and prime run avx and will pull a lot to run at high ghz.

I run into stability issues when rendering 4k proxies with Adobe media encoder. I am at 1.33 vcore and llc turbo (llc 6 on asus). Still crashes at 1.33. Is easing vcore the only option?

Also once I set avx = 1. I get immediately crash. An idea why?

[Arne Saknussemm]

V core is pretty much 1 to 1 for your OC these days. Of course general system stability helps...with SA and IO volts being key for RAM. But it is pretty much Vcore that gives you 100 mhz increase...

5.1 at 1.35 and 5.2 at 1.385 for non avx stability has been my experience for the average chip...this on ASUS boards all core

You'll have to translate that to GB board...

Not sure what AVX triggers on those boards...multiplier drop and maybe more droop? Hence instability?

[Alexauwa]

V core is pretty much 1 to 1 for your OC these days. Of course general system stability helps...with SA and IO volts being key for RAM. But it is pretty much Vcore that gives you 100 mhz increase...

5.1 at 1.35 and 5.2 at 1.385 for non avx stability has been my experience for the average chip...this on ASUS boards all core

You'll have to translate that to GB board...

Not sure what AVX triggers on those boards...multiplier drop and maybe more droop? Hence instability?

Thanks. I have my 4k render now stable at 1.30v. I reset my XMP. Still avx off and it manged the whole process of rendering my proxies. Adobe media encoder had all going, 80% Gpu, heavy cpu load and m2 writing and reading. Ofc I could now go back to all stress test, but if crash there, I start all over again and might have to raise vcore to a level I might not really need in my video editing. I see big drop in temps on 1.32. I hit 80 ° max. Water temps are between 31-34. 1.34 vcore might not sound big, but I am at 190 Watt and CPU Temps about 88 max. Kind of a sweet spot though?

What would you recommand for memory voltages? Have them fixed? At the moment they are on auto. Can I read them in hwinfo and are they reliable?

[Arne Saknussemm]

Yep! No need to conform to stress test opinions

If your system is stable for what you need it to do then all good...it's stable. Stress tests are only usefull in providing a a kind of top end catch all but they aren't the be all and end all of stability.

I personally tinker with all the main voltage...just to see if i can run less but auto should be fine. HWiNFO is as reliable as it gets for software...so that should provide you with good enough readings without having to go to poking about with multimeters,

[Alexauwa]

Yep! No need to conform to stress test opinions

If your system is stable for what you need it to do then all good...it's stable. Stress tests are only usefull in providing a a kind of top end catch all but they aren't the be all and end all of stability.

I personally tinker with all the main voltage...just to see if i can run less but auto should be fine. HWiNFO is as reliable as it gets for software...so that should provide you with good enough readings without having to go to poking about with multimeters,

Thanks for your help mate