cancel
Showing results for 
Search instead for 
Did you mean: 

Maximus XI Hero and i9-9900K -> LLC vs CPU Voltages

cniedzi
Level 8
Hi,

At 5GHz I'm able to pass 1h of ROG RealBenchmark for two sets of min. stable CPU Voltage and LLC.

1. LLC 6, CPU BIOS = 1.310V, CPU Idle = 1.296V and CPU Load = 1.225V (the first screen)
2. LLC 7, CPU BIOS = 1.300V, CPU Idle = 1.296V and CPU Load = 1.261V (the second screen)

On LLC of 8 I need at least CPU BIOS = 1.3 V, Idle=1.3V and Load 1.3V to be stable? I haven't try to test for long time due to high temps...

The question is why on LLC of 7 am I stable with much higher minimum CPU Load voltage than on LLC 6? With CPU BIOS=1.295V, Idle=1.288V and Load=1.252V ROG RealBenchmark failed. Maybe CPU Load voltage reading in CPUZ is incorrect?
419 Views
4 REPLIES 4

Falkentyne
Level 12
This has already been answered.
And your load graph you posted is basically useless.
You need an oscillioscope for these kind of tests. It has to be at least 50 mhz and (40us or ms (??) sampling rate).
Raja or Shamino can tell you more.

https://rog.asus.com/forum/showthread.php?109211-Question-about-Transient-Response-(to-Shamino-and-R...

It's transient response. It's because your minimum voltage at HEAVY current with LLC7 drops far below the minimum load voltage registered under LLC7 (dropping to about the LLC6 minimum read level). You cannot rely on onboard sensors to show these things. You need an oscilloscope.

In the intel spec documents, they only reference "overshoot - virus mode" and overshoot -microseconds" but there is no reference to undershoot.
That's because Intel specifies that Loadline Calibration is designed to be at 1.6 mOhms, so all undershoot is fully handled by the large vdroop and cushioned to allow the MOSFETS to return properly from a load to idle state, and an idle to load state causes no extra VMIN drops, because it happens on the large (1.6 mOhms) of vdroop.

When you use loadline calibration, you lower this vdroop, but at heavy loads, you get transient response penalities, which in effect acts like the vdroop you "removed" actually comes BACK for a few microseconds, with vSPIKES happening in the opposite way on load to idle changes.

When you are not doing stress tests, e.g., just playing games with light to medium loads, LLC7 will indeed make you more stable with a lower bios voltage than LLC6. That's because the current load isn't enough to cause the voltage oscillation to be as bad anymore.

If you REALLY want to test this scientifically WITHOUT an oscilloscope, you need to reduce the voltage and clock speed to safer levels to protect against transient spikes.

Try this setting:
4.6 ghz core, 4.3 ghz cache, voltage (BIOS):= 1.175v or 1.180v, LLC=5.

Download prime95 29.7 build 1.

https://mersenneforum.org/showthread.php?t=24094

Extract, install, then do a small FFT test with FMA3 enabled (better to use this one than the older versions).
You only need to do about 15 minutes per test cycle here. FMA3 is required to put as much load as possible--this will cause the most transient response penalities. Also this voltage is safe enough to not degrade your CPU on transient spikes.

First test: Try 1.180v @ LLC5.
Record in HWInfo64, the vcore idle and full maximum load minimum voltage.
If a thread crashes, increase voltage by 10mv until you are stable. Record the final bios and load voltage shown where you were stable.
I am taking a guess that your load voltage should be about 1.070, if you were able to do 1.180+LLC5 successfully.
If you were stable at the VERY FIRST attempt, reduce voltage 10mv and try again. Find the point where threads start crashing on you. Then increase voltage +10mv higher and record that as your VMIN (Min voltage required for stability) at 4.6 core, 4.3 cache, LLC5. Record this value.

ASSUMING your load voltage stable was 1.070v, you want to now use LLC6 and a LOWER BIOS VOLTAGE to match the LOAD VOLTAGE you got previously.
If it was indeed 1.070v load, I am guesstimating (I've done these tests myself), that now you would need a bios voltage of 1.155v + LLC6 to get 1.070v load.

Test again until you find which bios and load voltage doesn't crash a thread in 15 minutes. Record this.
You will see that your load voltage will need to be at LEAST 15mv higher at the bare minimum. Maybe more. Post these results.

Now try LLC7. Again we use 1.070v as the original load target. So now we can try lets say, 1.135v set in Bios LLC7.
Test until you find which bios and load voltages are required to not crash a thread in 15 minutes.

and finally LLC8.
You try Bios voltage 1.075v and load =1.075v (0 mOhm loadline = LLC8), you are probably going to insta BSOD. Or crash.
Keep going up.

If LLC5 + 1.070v (LOAD voltage) was fully stable, meaning you needed a bios idle voltage of 1.180v here, I am willing to be you $100 dollars that you will need to use LLC8 + Bios voltage 1.180v=load voltage 1.180v to be stable.

As you can see you basically gain nothing compared to 1.180v + LLC5=1.070v heavy load, due to voltage transient drops and oscillations at LLC8.

I hope you can understand this post. Sorry if it's confusing.

Falkentyne wrote:
This has already been answered.
And your load graph you posted is basically useless.
You need an oscillioscope for these kind of tests. It has to be at least 50 mhz and (40us or ms (??) sampling rate).
Raja or Shamino can tell you more.

https://rog.asus.com/forum/showthread.php?109211-Question-about-Transient-Response-(to-Shamino-and-R...

It's transient response. It's because your minimum voltage at HEAVY current with LLC7 drops far below the minimum load voltage registered under LLC7 (dropping to about the LLC6 minimum read level). You cannot rely on onboard sensors to show these things. You need an oscilloscope.

In the intel spec documents, they only reference "overshoot - virus mode" and overshoot -microseconds" but there is no reference to undershoot.
That's because Intel specifies that Loadline Calibration is designed to be at 1.6 mOhms, so all undershoot is fully handled by the large vdroop and cushioned to allow the MOSFETS to return properly from a load to idle state, and an idle to load state causes no extra VMIN drops, because it happens on the large (1.6 mOhms) of vdroop.

When you use loadline calibration, you lower this vdroop, but at heavy loads, you get transient response penalities, which in effect acts like the vdroop you "removed" actually comes BACK for a few microseconds, with vSPIKES happening in the opposite way on load to idle changes.

When you are not doing stress tests, e.g., just playing games with light to medium loads, LLC7 will indeed make you more stable with a lower bios voltage than LLC6. That's because the current load isn't enough to cause the voltage oscillation to be as bad anymore.

If you REALLY want to test this scientifically WITHOUT an oscilloscope, you need to reduce the voltage and clock speed to safer levels to protect against transient spikes.

Try this setting:
4.6 ghz core, 4.3 ghz cache, voltage (BIOS):= 1.175v or 1.180v, LLC=5.

Download prime95 29.7 build 1.

https://mersenneforum.org/showthread.php?t=24094

Extract, install, then do a small FFT test with FMA3 enabled (better to use this one than the older versions).
You only need to do about 15 minutes per test cycle here. FMA3 is required to put as much load as possible--this will cause the most transient response penalities. Also this voltage is safe enough to not degrade your CPU on transient spikes.

First test: Try 1.180v @ LLC5.
Record in HWInfo64, the vcore idle and full maximum load minimum voltage.
If a thread crashes, increase voltage by 10mv until you are stable. Record the final bios and load voltage shown where you were stable.
I am taking a guess that your load voltage should be about 1.070, if you were able to do 1.180+LLC5 successfully.
If you were stable at the VERY FIRST attempt, reduce voltage 10mv and try again. Find the point where threads start crashing on you. Then increase voltage +10mv higher and record that as your VMIN (Min voltage required for stability) at 4.6 core, 4.3 cache, LLC5. Record this value.

ASSUMING your load voltage stable was 1.070v, you want to now use LLC6 and a LOWER BIOS VOLTAGE to match the LOAD VOLTAGE you got previously.
If it was indeed 1.070v load, I am guesstimating (I've done these tests myself), that now you would need a bios voltage of 1.155v + LLC6 to get 1.070v load.

Test again until you find which bios and load voltage doesn't crash a thread in 15 minutes. Record this.
You will see that your load voltage will need to be at LEAST 15mv higher at the bare minimum. Maybe more. Post these results.

Now try LLC7. Again we use 1.070v as the original load target. So now we can try lets say, 1.135v set in Bios LLC7.
Test until you find which bios and load voltages are required to not crash a thread in 15 minutes.

and finally LLC8.
You try Bios voltage 1.075v and load =1.075v (0 mOhm loadline = LLC8), you are probably going to insta BSOD. Or crash.
Keep going up.

If LLC5 + 1.070v (LOAD voltage) was fully stable, meaning you needed a bios idle voltage of 1.180v here, I am willing to be you $100 dollars that you will need to use LLC8 + Bios voltage 1.180v=load voltage 1.180v to be stable.

As you can see you basically gain nothing compared to 1.180v + LLC5=1.070v heavy load, due to voltage transient drops and oscillations at LLC8.

I hope you can understand this post. Sorry if it's confusing.


Thx @Falkentyne for your reply. Due to my technical education I was dreaming about such comprehensive answer :). I will perform tests you depicted and I will report back but when my new RAM will arrive.

PS. My graphs was only to show temps much better for LLC 6; I would like to stay with LLC 6.

Regards
Czarek

cniedzi
Level 8
Rockford wrote:
Very high LLC is wobbly and not very good for long term stability, a handbrake is better than a trampoline many times (High LLC is good for extreme benchmarking with high overall voltages, where its about reaching maximum performance during a short period of time) = extreme voltage power draw (delivery over/quality)

Example:

FIRST find out how much voltage the CPU needs to remain stable at a given clock under MAXIMUM load (takes maximum 3 minutes to test), check... 1,380v
Then add some extra voltage for good times (assuming the cpu temperature is good and well within recommendations)

Now voltage target is 1,385v, set as close to 1,385v in bios as possible, and make sure it sticks there during maximum load (LLC bios section have many different tools for this) dial in..

Everytime the computer is rebooted the system is going to act slightly different than the last time, motherboard components (CPU/RAM,etc.) is physics, and nothing is straight in the laws of universe

The higher the overclock, the more the boundaries is pushed/physics (meaning that a seemingly stable high CPU overclock can still fail the following day, week or month)

Good luck.

PS: I am not english, and can be difficult to understand ever so often
sign here ** to join the movement against obsessive Prime 95 disorder (OP95D)


Thx @Rockford for reply 🙂
I am few days @ 5GHz (all cores), LLC 6, CPU BIOS = 1.310V, CPU Idle = 1.296V and CPU Load = 1.225V (sometimes 1.207v) - it looks like rock solid, temps are great. Currently I'm with only one RAM stick (8GB) - waiting for new 32GB 3866MHz kit. If 4 new stick will not negatively influence my stability, a will remain with LLC 6 and my current voltages.

PS. I turned off all CPU power saving functions but I can see one of the cores (random core number) drops ocassionally to 4,7 GHZ on idle - is it normal? I have AVX offset set to 3, but I in Windows desktop ???

cniedzi wrote:
Thx @Rockford for reply 🙂
I am few days @ 5GHz (all cores), LLC 6, CPU BIOS = 1.310V, CPU Idle = 1.296V and CPU Load = 1.225V (sometimes 1.207v) - it looks like rock solid, temps are great. Currently I'm with only one RAM stick (8GB) - waiting for new 32GB 3866MHz kit. If 4 new stick will not negatively influence my stability, a will remain with LLC 6 and my current voltages.

PS. I turned off all CPU power saving functions but I can see one of the cores (random core number) drops ocassionally to 4,7 GHZ on idle - is it normal? I have AVX offset set to 3, but I in Windows desktop ???


Yes, it's normal. Even one background thread running AVX will trigger the AVX offset on all cores. Windows uses AVX occasionally and some games use it quite heavily (BF5, Apex legends).