cancel
Showing results for 
Search instead for 
Did you mean: 

ASUS Z390 Load Line Calibration Discussion

joppiano
Level 7
Hi all.

As title says, I would like to discuss the LLC (Load Line Calibration) from ASUS Z390 series.

I have been reading a lot about it recently, and it's quite confusing at times, therefor this thread in hopes of getting even more clarification on the subject.

As of recently, I figured that there was no such thing as "overcompensation" of vdroop due to LLC level. (Source: Buildzoid - video linked below)
Prior to this video, I was under the understanding, that a to high level of LLC (depending of board manufactor) would cause your VCORE to overshoot(vboost) to much. When in reality, there is no such thing, since you would need a negative mOhms resistence, which doesn't exist.

The reason why you're seeing a higher vcore in windows than set in bios, is due to the reading of the Super I/O chip, which is very bad and inaccurate. I think on the MAXIMUS boards, you get the VR VOUT reading, which is way more accurate, but not on the Z390 boards.

On my specific board ASUS ROG STRIX Z390-F, my LLC level goes from 1-7. 1 being more vdroop and 7 being 0 vdroop.
Since my board is using the Super I/O vcore readings, I have been using LLC 5 for my overclocks.

Example 1:
LLC 5
VCORE @ BIOS: 1.360v
VCORE @ P95 SMALL FFTS LOAD(HWinfo): 1.323v

Example 2:
LLC 6
VCORE @ BIOS: 1.360v
VCORE @ P95 SMALL FFTS LOAD(HWinfo): 1.368v

Example 1 shows a vdroop of ~37mv, but I assume that vdroop is actually MORE than 37mv due to the reading inaccuracy.
Example 2 shows that according to HWinfo, I have an overshoot of 6mv during P95 loads. But since the Super I/O readings is wrong, that is not actually my real vcore. It is actually causing a vdroop, but isn't shown in Windows.

Now I am starting to believe that LLC 6 might be the better option, in the goal of reaching for a stable overclock.

Thing is, I had the understanding, that the "perfect LLC level" was causing you to have a "FLAT" transition from idle to load - however, that seems not to be the case after all.
As I understand it, you need a vdroop to counter the over & undershoot which occur in miliseconds upon transition from idle to load and load to idle.
But what is the perfect vdroop then? I guess that all boils down to each individual CPU & motherboard.

Here is a spreadsheet of my OC testing and playing around with LLC etc.

Changes that are made from each test is highlighted with underline.

I know that changing the LLC level, determine how much resistence (in mOhms), which determines how much vdroop you will encounter. But I do not know what each level of LLC represent in mOhms, do you?

I tryid to calculate the mOhms resistence of LLC 5 & 6, but my board doesn't show how many amps is being used, so i tryid calculating this myself.
To find amps used, you can divide watts with voltage, but I do not know if you do "watts / voltage under load" or "watts / voltage set in BIOS".

I am using i7 9700K on a ASUS ROG STRIX Z390-F.

I would like to know your experience with LLC on Z390 boards.
What is your setup, your experience and findings ?
------------------------------------------------------------
Buildzoid LLC Explained: https://www.youtube.com/watch?v=bUaP0r5-xhY

Note: If any of the above information is inaccurate, untrue or such. Please feel free to correct me. Only made this thread in hope to get more clarification on LLC @ Z390 boards.
17,902 Views
13 REPLIES 13

Arne_Saknussemm
Level 40
You need an oscilloscope to study this stuff accurately...it's all about transient response.

LLC 5 or 6 is a good place to be for a happy OC...get away with lower LLC if you can....Vdroop is a good thing if you want your CPU to live long and prosper....

Falkentyne
Level 12
This is a good post.

The "best" LLC depends purely if you are using fixed vcore or adaptive/offset vcore (with c-states) as this affects your idle voltage, so the CPU can downvolt at idle, which it can't do on fixed (it can downclock but not downvolt). With offset or adaptive, you can get by with a much lower LLC than with fixed. Keep in mind the AC Loadline being used has a large impact on offset and adaptive modes. The lower the AC Loadline/DC Loadline value, the more LLC you usually need, so if ACLL is set to 0.01 mOhms, you are going to get nowhere with Level 1 or Level 2 LLC; you may need at least level 4 or 5. With maximum Intel spec AC Loadline/DCLL (1.6 mOhms), you can use level 1 or level 2 LLC, but voltage swings will be very wild based on current.

ACLL isn't important on fixed voltage, and I am not going to discuss DCLL being set different than ACLL because I do not know if DCLL affects vcore on offset mode or not. I know it affects VID, but Intel's spec sheets say DCLL is used to calculate CPU Package Power while ACLL is the CPU voltage supply. I will not discuss this topic.

On fixed voltage, what you want, measured with an *OSCILLOSCOPE*, is the LOWEST PEAK TO PEAK voltage, with the lowest AVERAGE load voltage. That's your goal.

Let's pretend 9900K sample "A" needs a minimum voltage floor of 1.21v @ 5 ghz for heavy AVX and it crashes if it goes below 1.21v, at 5 ghz.
Let's say your heavy AVX draws 193 amps, max intel spec. So there are 3 ways to get that:

1) 1.520v bios set, Level 2 LLC (1.6 mOhms LLC), absolute max safe bios voltage run at Intel spec loadline, this gives 1520mv - (193 * 1.6) = 1212mv load, with 10mv transient dips to 1.202v. Your average load will be 1.212v. Transient spike above 1.520v will be 10mv, so 1.530v. So that's a 330mv peak to peak with god tier transients.

2) 1.380v bios set, Level 5 LLC (0.8 mOhms LLC), that gives load at 193 amps of 1380mv - (193 * 0.8)=1225mv, with 20mv transient dips to 1.205v. And 30mv transient spikes to 1.410v. So that's 200-205mv peak to peak, with 1.225v average load voltage.

3) 1.310v bios set, Level 8 LLC (0 mOhms LLC, no droop). This is 1310 mv - (193 * 0)=1.310v, but transients are TERRIBLE, 100mv dips to 1.210v and 100mv spikes to 1.410v.

All three settings give the same stability at max AVX current.

It's obvious #3 is beyond horrible. The heat and power draw at 1.310v is going to be much higher because the average voltage is higher, and it needs to be higher to compensate for the terrible dips. BTW the power draw will also be much higher than 193 amps if the same prime95 test is done since the load average is higher. I just kept it at 193 to keep it simple.

#2 looks very reasonable. Your bios voltage is not too bad. 1.380v LLC5 is not too bad, and your load voltage drops down quite a bit but the transient below the load voltage is still pretty good. Solid choice. Your peak to peak is 205mv--same as LLC8 but the AVERAGE LOAD VOLTAGE IS MUCH LOWER TO MAINTAIN STABILITY!! Remember what i said--a balance of low peak to peak combined with low average voltage needed.

( NOTE: LLC5 IS NOT 0.8 MOHMS. it's 0.8 mOhms on my Gigabyte board so I'm using my board as an example. I have no idea what it is on Asus. Other tests on overclock.net showed it slightly higher, almost the same as Gigabyte's LLC4).

#1 is terrible. You get the best transients, sure. But you need such a high bios voltage to compensate for the vdroop that your CPU is running at the bare borderline of Intel safe max spec all the time, that could cause the same rate of slow degradation at both idle and load (if temps are kept in check). Because your peak to peak is so freaking high. The peak to peak (>300mv) is not worth the 20mv improvement on a transient. Offset mode with C-states would be much better with LLC1 to avoid a high idle voltage so the voltage drops at idle. Adaptive mode may be even better. (again I can NOT HELP WITH THIS).

But your conclusions are close to correct: on FIXED VOLTAGE, LLC5 tends to be the best. (LLC6 has substantialy more transient ripple).

You need an oscilloscope to find peak to peaks with ripple as the ripple minimums determine your stable voltage floor during heavy loads. VR VOUT only helps find averages.

https://elmorlabs.com/index.php/2019-09-05/vrm-load-line-visualized/

Falkentyne wrote:

On fixed voltage, what you want, measured with an *OSCILLOSCOPE*, is the LOWEST PEAK TO PEAK voltage, with the lowest AVERAGE load voltage. That's your goal.


Thanks for the reply.

So, if it's the "lowest peak-to-peak at the lowest average load voltage" that is the goal, and that can only really be monitored with an oscilloscope - is there any rule of thumb that can be used when only monitored in windows?

Also as said, my previous understanding of LLC was to get as close to a flat transition as possible - but that turned out not to be true, since you would just pick the LLC setting with 0 vdroop, and as stated - you WANT vdroop to take care of the transient respond that is occuring when going from load to idle, idle to load.

Personally, if I go with LLC 5, I will get some decent vdroop. If I go with LLC 6, I will according to windows get a vboost, but as confirmed above, I do get a vdroop, it is just not showing in windows. But my avg voltage is lower - but I have NO IDEA if my transient are better or not.
As shown in the spreadsheet, I can get along with lower average voltage with LLC 6 than with LLC 5, which is good - But if transient dips & spikes are worse, that is worse, due to peak-to-peak is bigger all of the sudden. Or at least can be.

So, about the calculation mOhms of each LLC level on ASUS, how do I calculate how many amps I am pushing. I know you divide watts with voltage. But is it voltages under load you use, or voltages at idle (set in bios) ?
(Watts / voltage under load or Watts / voltage set in bios)

Because if I can figure out that calculation, I can calculate the mOhms of resistence on each individual LLC level on ASUS Z390.

Also, how did you determine the transient dips & spikes in your 3 examples?

Arne Saknussemm wrote:
You need an oscilloscope to study this stuff accurately...it's all about transient response.

LLC 5 or 6 is a good place to be for a happy OC...get away with lower LLC if you can....Vdroop is a good thing if you want your CPU to live long and prosper....


So when you say "lower LLC" I assume you mean LLC 4 or maybe even LLC 3 - so MORE vdroop and not less, yes?

joppiano wrote:
Thanks for the reply.

So, if it's the "lowest peak-to-peak at the lowest average load voltage" that is the goal, and that can only really be monitored with an oscilloscope - is there any rule of thumb that can be used when only monitored in windows?

Also as said, my previous understanding of LLC was to get as close to a flat transition as possible - but that turned out not to be true, since you would just pick the LLC setting with 0 vdroop, and as stated - you WANT vdroop to take care of the transient respond that is occuring when going from load to idle, idle to load.

Personally, if I go with LLC 5, I will get some decent vdroop. If I go with LLC 6, I will according to windows get a vboost, but as confirmed above, I do get a vdroop, it is just not showing in windows. But my avg voltage is lower - but I have NO IDEA if my transient are better or not.
As shown in the spreadsheet, I can get along with lower average voltage with LLC 6 than with LLC 5, which is good - But if transient dips & spikes are worse, that is worse, due to peak-to-peak is bigger all of the sudden. Or at least can be.

So, about the calculation mOhms of each LLC level on ASUS, how do I calculate how many amps I am pushing. I know you divide watts with voltage. But is it voltages under load you use, or voltages at idle (set in bios) ?
(Watts / voltage under load or Watts / voltage set in bios)

Because if I can figure out that calculation, I can calculate the mOhms of resistence on each individual LLC level on ASUS Z390.

Also, how did you determine the transient dips & spikes in your 3 examples?



So when you say "lower LLC" I assume you mean LLC 4 or maybe even LLC 3 - so MORE vdroop and not less, yes?


I can tell you already right now.
LLC1=2.1 mOhms and LLC2 is 1.6 mOhms and LLC6 is SOMEWHERE CLOSE TO 0.4 mOhms (either 65% or 70% reduced vdroop) and LLC8 is 0 mOhms.

Someone on OCN calculated LLC5. I want to say it's 1.05 mOhms but don't mark my words on that. 1.05 is 50% of 2.1 mOhms right? And gigabyte's LLC "High" is 50% reduced vdroop, and since their baseline (8 cores) is 1.6 mOhms, that's 0.8 mOhms LLC.

What I don't know are the other values on your board. LLC is a fixed percentage of a "baseline" value. On Asus Z390 boards, the baseline value is LLC1 (but the processor default SKU loadline for 8 core processors is LLC2). So every value is a percentage of 2.1 mOhms. On Gigabyte boards, LLC6 is 75% reduced vdroop or 25% of 1.6 mOhms as 1.6 mOhms is the "baseline" on 8 core processors on Gigabyte boards. Asus and Gigabyte have different baselines. Asus probably uses 2.1 mOhms because that's the value for 4 and 6 core CFL processors.

Less LLC=more vdroop. But too much vdroop is bad on fixed vcore because you need a too high idle voltage to compensate.

Rule of thumb is, medium levels of LLC (not too low, not too high) are good on fixed vcore.
Offset modes get tricky because the default AC Loadline heavily affects the voltage. With 0.01 mOhms ACLL, you can use the same rule as fixed vcore--medium levels of LLC. With higher ACLL, you want less LLC (more vdroop).

The more vdroop you have, the more accurate the onboard sensors will be with respect to ripple or average voltages, because there will be less transient ripples.
The more vdroop you remove, the less accurate the onboard sensors will be (even VR VOUT or Maximus XI) because the ripple will be all over the map. Look at the LLC8 example on elmor's website and imagine that happening constantly but your VR VOUT shows 1.20v all the time.

Falkentyne wrote:

Less LLC=more vdroop. But too much vdroop is bad on fixed vcore because you need a too high idle voltage to compensate.

Rule of thumb is, medium levels of LLC (not too low, not too high) are good on fixed vcore.


Okay that explains alot then.
That is why LLC 6 is seemingly stable on lower voltages, than it will be on LLC 5. (Obviously more thorough & longer testing is needed)

On LLC 5 I will archieve more vdroop - but I need a higher average voltage to compensate for that.
On LLC 6 I will archieve less vdroop - but I can use a lower average voltage.
But LLC 6 isn't "medium" hmm.

However, looking at elmors website ( https://elmorlabs.com/index.php/2019-09-05/vrm-load-line-visualized/ ) I would assume that LLC 4, 5 & 6 would be the best, if you just look at the graph of the transient.
But according to your previous statement, it seems that LLC 3 is the best option in that specific test case. That has the lowest peak-to-peak, but also a lot of vdroop. It actually seems that LLC 1 to LLC 3 is the best in that test case.

But if LLC 5 is around 1.05 mOhms on Asus z390, that seems like A LOT of vdroop that is not visable in windows what so ever.
Using elmors spreadsheet https://docs.google.com/spreadsheets/d/1nKdrj_eoxxXdsAkL6lj-EOdHj06khw1tN0fxVOR3mlc/edit#gid=0 and putting in 1.310v, it should droop to 1.210v at 100amps. That seem extremely low ?

On my board (Z390-F), LLC goes from 1-7. So assuming LLC 7 is 0 mOhms, LLC 6 is probably around the 0.4 mOhms area.

joppiano wrote:
Okay that explains alot then.
That is why LLC 6 is seemingly stable on lower voltages, than it will be on LLC 5. (Obviously more thorough & longer testing is needed)

On LLC 5 I will archieve more vdroop - but I need a higher average voltage to compensate for that.
On LLC 6 I will archieve less vdroop - but I can use a lower average voltage.
But LLC 6 isn't "medium" hmm.

However, looking at elmors website ( https://elmorlabs.com/index.php/2019-09-05/vrm-load-line-visualized/ ) I would assume that LLC 4, 5 & 6 would be the best, if you just look at the graph of the transient.
But according to your previous statement, it seems that LLC 3 is the best option in that specific test case. That has the lowest peak-to-peak, but also a lot of vdroop. It actually seems that LLC 1 to LLC 3 is the best in that test case.

But if LLC 5 is around 1.05 mOhms on Asus z390, that seems like A LOT of vdroop that is not visable in windows what so ever.
Using elmors spreadsheet https://docs.google.com/spreadsheets/d/1nKdrj_eoxxXdsAkL6lj-EOdHj06khw1tN0fxVOR3mlc/edit#gid=0 and putting in 1.310v, it should droop to 1.210v at 100amps. That seem extremely low ?

On my board (Z390-F), LLC goes from 1-7. So assuming LLC 7 is 0 mOhms, LLC 6 is probably around the 0.4 mOhms area.


It varies motherboard to motherboard. And each motherboard has different transient response. The evga z390 dark for example is literally in a league on its own. But a board with that good transient response without any die-sense vcore readings...I guess they designed this board for world records...

Falkentyne wrote:
It varies motherboard to motherboard. And each motherboard has different transient response. The evga z390 dark for example is literally in a league on its own. But a board with that good transient response without any die-sense vcore readings...I guess they designed this board for world records...


I see.

But looking at elmors website https://elmorlabs.com/index.php/2019-09-05/vrm-load-line-visualized/ - would you agree that LLC 3 is the best option, since that has the lowest peak-to-peak voltages ?

In his case, even LLC 1 & 2 outperforms LLC 4 & 5, since peak-to-peak voltages are lower.
But LLC 1 & 2 would require a higher average voltage than LLC 4 & 5, to remain stable at same workload right?


- UPDATE:
When talking about stability.
Can instability take place, if low peaks or high peaks is to low/high ? Or is it mainly the voltages under load that causes it.

If we assume LLC 5 is around 1.05 mOhms and we assume LLC 6 is around 0.4 mOhms.

Example 1: (1.365v BIOS & LLC 5)
180W / 1.365v = 132 amps.
132 amps * 1.05 mOhms = 139mv Vdroop.
0.139v - 1.365v = 1.226 @ load.

Example 2: (1.315v BIOS & LLC 6)
180W / 1.315v = 137 amps.
137 amps * 0.4 mOhms = 55mv Vdroop.
0.055v - 1.315v = 1.260v @ load.

Example 1 & Example 2 is running the same load with P95, but how come Example 1 is passing 2 hours of P95, but Example 2 is failing after 30 minutes with BSOD & Worker Stopping ?
Example 2 has a higher vcore @ load - so is this instability caused by bad transient response ?

joppiano wrote:
I see.

But looking at elmors website https://elmorlabs.com/index.php/2019-09-05/vrm-load-line-visualized/ - would you agree that LLC 3 is the best option, since that has the lowest peak-to-peak voltages ?

In his case, even LLC 1 & 2 outperforms LLC 4 & 5, since peak-to-peak voltages are lower.
But LLC 1 & 2 would require a higher average voltage than LLC 4 & 5, to remain stable at same workload right?


- UPDATE:
When talking about stability.
Can instability take place, if low peaks or high peaks is to low/high ? Or is it mainly the voltages under load that causes it.

If we assume LLC 5 is around 1.05 mOhms and we assume LLC 6 is around 0.4 mOhms.

Example 1: (1.365v BIOS & LLC 5)
180W / 1.365v = 132 amps.
132 amps * 1.05 mOhms = 139mv Vdroop.
0.139v - 1.365v = 1.226 @ load.

Example 2: (1.315v BIOS & LLC 6)
180W / 1.315v = 137 amps.
137 amps * 0.4 mOhms = 55mv Vdroop.
0.055v - 1.315v = 1.260v @ load.

Example 1 & Example 2 is running the same load with P95, but how come Example 1 is passing 2 hours of P95, but Example 2 is failing after 30 minutes with BSOD & Worker Stopping ?
Example 2 has a higher vcore @ load - so is this instability caused by bad transient response ?


Oh, I forgot about that. However Elmor's chart isn't based on a heavy load. He took those graphs while idle in the BIOS.
At a very heavy load, the spikes and dips on both ends will be larger than anything shown on these graphs, but how much larger will depend on the VRM quality and the inductors and capacitors.

Here is a 100 amp load, prime95, done with a 1.50v bios set point, LLC1 and LLC8. You can see my LLC8 math way above is pretty accurate at least 😕

https://www.overclock.net/forum/6-intel-motherboards/1638955-z370-z390-vrm-discussion-thread-413.htm...

Pretty sure 193 amps will be quite worse.

Maybe Shamino can jump in on this? He's probably extremely busy though 😞

My math was just me guessing from my head randomly. I forgot that with a large loadline slope, your idle voltage will have substantial vdroop on that.
I forgot that with LLC1 or LLC2, the voltage you set in BIOS won't spike above that point (e.g. 1.20v will be 1.17v or 1.18v).

So technically yes you are correct. But the problem again is the high idle voltage. You don't want your computer to be sitting constantly at 1.48v because you set 1.520v in BIOS with LLC2 to do a 193 amp prime95 run (drops to 1.213v). It's better to do 1.38v + LLC5 in that case. Or consider offset or adaptive modes.

Falkentyne wrote:
Oh, I forgot about that. However Elmor's chart isn't based on a heavy load. He took those graphs while idle in the BIOS.
At a very heavy load, the spikes and dips on both ends will be larger than anything shown on these graphs, but how much larger will depend on the VRM quality and the inductors and capacitors.

Here is a 100 amp load, prime95, done with a 1.50v bios set point, LLC1 and LLC8. You can see my LLC8 math way above is pretty accurate at least 😕

https://www.overclock.net/forum/6-intel-motherboards/1638955-z370-z390-vrm-discussion-thread-413.htm...

Pretty sure 193 amps will be quite worse.

Maybe Shamino can jump in on this? He's probably extremely busy though 😞

My math was just me guessing from my head randomly. I forgot that with a large loadline slope, your idle voltage will have substantial vdroop on that.
I forgot that with LLC1 or LLC2, the voltage you set in BIOS won't spike above that point (e.g. 1.20v will be 1.17v or 1.18v).

So technically yes you are correct. But the problem again is the high idle voltage. You don't want your computer to be sitting constantly at 1.48v because you set 1.520v in BIOS with LLC2 to do a 193 amp prime95 run (drops to 1.213v). It's better to do 1.38v + LLC5 in that case. Or consider offset or adaptive modes.


Okay that make sense now.

But if you take a look at my previous post and those calculations - would you say that instability can be caused due to bad transient response ? As I believe is what happened in the 2 examples made above.

They have the same load and pretty much the same current, but the "LLC 6" example has a higher vcore under load. But still the LLC 6 example fails the test due to instability. So since it isn't vcore under load that is the culprit here, could it be the transient response causing the instability?

Also, it seems for my specific board & cpu, that LLC 5 is the most optimal load line.
Unless even more Vdroop is better. Only testing will show I guess.