LLC confusion and problems

[taq8ojh]

I thought I understood how to overclock but apparently not. In fact, after some tests this morning I probably don't have the slightest idea
My board is Gigabyte Z77X-UD3H, CPU is 3770K.

Recently I read several guides or postst that insisted people should and have to use LLC. I originally read that to find the proper level of that, one should see what the unchangable displayed vcore value (this) in BIOS settings say and compare it to what actually gets pumped into the CPU under load in Windows. Based on that I shouldn't need anything higher than lowest LLC value: BIOS says 1.085-1.090V and that's what I get with zero offset. It fluctuates a tiny bit of course, I guess based on the level of load, but I assume it's normal. Very occasionally it drops to 1.080V, but usually stays at 1.092V.

Anyway...
Recently I managed to get almost stable (crash after 21 hours) overclock of 4.4GHz with lowest LLC and an offset of +0.040V. Vcore is in 1.116-1.128V range then.
Just for the record, I can boot into Windows and run Prime95 for up to a few hours even with zero offset at 4.4GHz. Yes it's far from stable and I get WHEA errors, but it sort of works (see below why I mention this).

This morning I found some post somewhere saying the way to do this is to set static vcore, boot into Windows, start Prime95 (or whatever one uses), and compare. This is when the pile of WTF comes in, because 1.120V wouldn't even boot properly or just flat out crash Prime95 right away, and I had to set LLC to medium to keep it going for at least a little while. What's more important though, vcore was reported to be 1.080V with that setting. I didn't see any such huge drop under load (in fact none at all) when using offset.
What the hell is going on? Could someone explain?

 

[Borealis7]

to me, LLC is like the female mind - I haven't the slightest clue how it works.
I'd like an explanation as well.

 

[taq8ojh]

I think I understand the theoretical concept, but exactly what you say - no idea how it actually works at all

 

[Ed1]

Maybe this shine some light .

The reason they say start with manual is it is easier to find and get it stable (or it should be as your dealing with x value voltage)
reading your post and what you have done I would of started with higher manual and IMO its better to start with higher value, check for stability and errors in event viewer than move down on voltage (this just my opinion I am sure others would do it differently ).

Now on LLC , the idea of load line calibration is being able to adjust the surge amount of voltage when a load goes onto the CPU . but even if it drops down as long as you compensate with slightly more v core it should still work (you just got to find what voltage it is stable at a given load).

So lets for example say you set vcore at 1.20 and you boot into windows and run prime .if you looked real close at voltage when load is applied and it either went like 1.18, then 1.19 and stayed or just a lower voltage than bios or light load ,raise LLC up a notch till you get little drop .
I am not sure you should try for no voltage drop as you don't want surges of voltage either (this depends on how good our monitoring software is ). you don't want it overshooting , going higher .

Then once you find stable manual voltage for given speed . you find your VID (use core temp) for that speed and apply this formula and it should get you close to your offset

manual Vcore - VID = offset voltage

But since you already started or have good idea of offset voltage I would just keep on going as you were .

here another post on it

http://forums.anandtech.com/showthread.php?t=2220036

OP : When you were using offset and was stable what was the LLC set to ? Auto ?

 

[Idontcare]

The problem with LLC, or any voltage tweaking for that matter, is that the voltage changes faster over time than any software app is going capture and report...but the CPU itself will crash the moment the voltage drops below a critical threshold value.

Voltage readouts you are getting from your BIOS or from CPUz may seem like they are updating in realtime but they are only sampling the actual voltage in steps over time.

CPUz is even worse because it doesn't even show you the observed voltage, it quantizes the voltage in 0.008V increments, and to make matters even more worse it rounds down the number when reporting to you.

So if the BIOS is reporting "1.287V" for Vcore then CPUz will take that number and report to you that the voltage is "1.280V"...a full 0.007V lower than what it actually is.

So let's say your CPU really does need 1.286V and not a drop lower, but CPUz is telling you it is running along just fine with 1.280V. (which is not true, because it would crash at 1.280V) So you are led to think your chip is fine at 1.280V.

Now you start up prime95, and so long as LLC keeps the Vcore at a value between 1.280V and 1.288V then CPUz is going to keep telling you the Vcore is 1.280V.

Well lets imagine the actual voltage is 1.287V, just a smidge above the 1.286V minimum threshold. CPUz is still saying it is 1.280V but the CPU is stable nevertheless. Now a momentarily change in the Prime95 load causes a transient voltage to drop that Vcore from 1.287V to 1.285V...boom, CPU is now unstable.

But CPUz is going to keep on reporting to you that the voltage is still 1.280V.

In short, CPUz sucks and in ways that fool people into thinking their system, or parts of it like LLC, are not acting rationally when infact they are. It is CPUz that is problem as it is generating crap data for you.

 

[Ed1]

What he said ^

you do much better job of describing it . And I agree totally its the polling rate of the monitoring app that doesn't really show whats going on cause it is just not fast enough to report and display .

You can see this type of affect with temp monitoring if you can adjust polling (you see a different max peak) it same thing though I don't think any apps can control voltage polling and it wouldn't really help as were talking ms

 

[taq8ojh]

Well based on what you said ANY software is crap at this... However I am using HWiNFO for readings of all kinds and I guess it gives me good enough idea.

However... I am more curious about the mysterious vcore drop that ONLY happened with fixed value. There was no drop at all when using offset.

Oh btw, the universal question is: higher offset and lower LLC or the other way around?

 

[Idontcare]

However... I am more curious about the mysterious vcore drop that ONLY happened with fixed value. There was no drop at all when using offset.

How do you know there was no drop when using offset?

 

[taq8ojh]

How do you know there was no drop when using offset?

Hm, because the reported vcore under load was the same as the value showed in BIOS.

 

[2is]

I've noticed this with my gigabyte board also. It seems that it applies some sort of LLC on its own it seems. As far as the software is concerned in get very little to no droop. It behaves like my ASUS board does at its lowest LLC setting.

 

[Idontcare]

Hm, because the reported vcore under load was the same as the value showed in BIOS.

Exactly, reported by some software program for which you have no way of determining neither the accuracy nor the precision.

In other words, you really don't know if the value dropped, went up, or went nowhere. All you have is some questionably unqualified program telling you something happened, and what it is telling you isn't quite adding up (hence your motivation to create this thread).

You have two different values, measured by two different programs, observed under two different operating system environments (BIOS vs Windows), and you are attempting to reconcile the two...when the reality is that neither number is correct, so taking note of the differences between them is going to make for confusing conclusions.

If you want to measure the voltage on your processor then you need a physical voltage measuring device, preferably one that samples fast enough as to capture the transients that might otherwise hide themselves in the time-window averaging algorithms of software and multimeters but will cause your processor to become unstable nevertheless.

 

[taq8ojh]

What do you suggest I do then? I don't have such tools at hand and I wouldn't know how to use them anyway.

I never said whichever program was reporting exactly perfect values of any kind. BUT if it reports exactly the same numbers under the same BIOS settings, I believe I can use it to get an idea.

 

[Ed1]

Exactly, reported by some software program for which you have no way of determining neither the accuracy nor the precision.

In other words, you really don't know if the value dropped, went up, or went nowhere. All you have is some questionably unqualified program telling you something happened, and what it is telling you isn't quite adding up (hence your motivation to create this thread).

You have two different values, measured by two different programs, observed under two different operating system environments (BIOS vs Windows), and you are attempting to reconcile the two...when the reality is that neither number is correct, so taking note of the differences between them is going to make for confusing conclusions.

If you want to measure the voltage on your processor then you need a physical voltage measuring device, preferably one that samples fast enough as to capture the transients that might otherwise hide themselves in the time-window averaging algorithms of software and multimeters but will cause your processor to become unstable nevertheless.

to bad not many MB have option to get direct reading off the MB (I think you posted some pics with DMM) .
It would be interesting to see output with oscilloscope of various MB and compare smoothness of output .

 

[taq8ojh]

The funny thing is my MB does have that . But I am just an average simple Simon who learned how to punch some numbers in a BIOS, thus unable to benefit from it...

 

[Phynaz]

The secret to LLC:

Don't touch it.

 

[Idontcare]

This is when the pile of WTF comes in, because 1.120V wouldn't even boot properly or just flat out crash Prime95 right away, and I had to set LLC to medium to keep it going for at least a little while. What's more important though, vcore was reported to be 1.080V with that setting. I didn't see any such huge drop under load (in fact none at all) when using offset.

What do you suggest I do then? I don't have such tools at hand and I wouldn't know how to use them anyway.

I never said whichever program was reporting exactly perfect values of any kind. BUT if it reports exactly the same numbers under the same BIOS settings, I believe I can use it to get an idea.

You don't need perfect programs, you just have to know the limitations of the ones you are using.

Take the statement of yours above which I quoted from your OP. The problem you are observing is not that the "average" voltage is high enough or too low, rather the problem is that you are blind to the transient voltage swings that are occurring during loading and unloading (which can happen on the timescale of micro-to-milliseconds).

The "maximum negative undershoot" value during idle and during loading is what causes your processor to become unstable, not the average voltage over the long-run at idle or under load.

^ those voltage oscillations are called "transients" and your BIOS and software-based voltage monitoring software are blind to them because they happen very very quickly (milliseconds) but when they happen they can still cause your processor to be borked.

 

[JAG87]

The secret to LLC:

Don't touch it.

This times a million. Just leave it alone. Increase only the vcore offset and that's it.

 

[Ajay]

Hmm, wonder if that is why mobo makers are still putting huge VRMs on Haswell motherboards - trying to reduce overshoot (or at least have an even better critically damped signal). If so, this could really help with Haswell overclocking. I would assume that the on-board VRM would be faster than the mobo based VRMs. Can't wait for review.

 

[Phynaz]

Hmm, wonder if that is why mobo makers are still putting huge VRMs on Haswell motherboards

Don't confuse voltage regulation with power delivery, they are two different things.

Also, wouldn't the rest of the components on the board still need voltage regulation?

 

[Ajay]

Don't confuse voltage regulation with power delivery, they are two different things.

Also, wouldn't the rest of the components on the board still need voltage regulation?

Huge as in many phases. I suppose there would be some need for VRM's, mainly for PCIe slots I would think. Less complex VRMs would suffice for most mobo components, IMO - there's neither the need for large amounts of power or for handling high speed current swings (as in the case of a CPU).

 

[Phynaz]

Huge as in many phases. I suppose there would be some need for VRM's, mainly for PCIe slots I would think. Less complex VRMs would suffice for most mobo components, IMO - there's neither the need for large amounts of power or for handling high speed current swings (as in the case of a CPU).

You're still confusing voltage regulation with power delivery. Voltage regulators don't have phases.

Maybe this will help http://www.hardwaresecrets.com/arti...e-Motherboard-Voltage-Regulator-Circuit/616/4

 

[Ajay]

You're still confusing voltage regulation with power delivery. Voltage regulators don't have phases.

Maybe this will help http://www.hardwaresecrets.com/arti...e-Motherboard-Voltage-Regulator-Circuit/616/4

OK, I always considered the MOSFET to be part of the VRM, since modern VRMs seem to be designed that way. I see your point though, since I've seem older VRMs w/o MOSFETS. Is this just a technical difference? Modulating the voltage is pointless w/o driving current into the CPU.

 

[pm]

If you want to measure the voltage on your processor then you need a physical voltage measuring device, preferably one that samples fast enough as to capture the transients that might otherwise hide themselves in the time-window averaging algorithms of software and multimeters but will cause your processor to become unstable nevertheless.

Even this doesn't get you an accurate measurement because it will give you only the coarse measurement that you can get off chip. For true accuracy, you need to measure it on-die.

We used this technique on a prior design that I was involved with:
https://www2.lirmm.fr/lirmm/interne/BIBLI/CDROM/MIC/2009/ITC_2009/Papers/PDFs/0008_2.pdf

I spent a bit of time about two years ago getting the ODDD and ODDI systems described in the paper work on a more recent microprocessor and the whole thing is really cool.

 

[Idontcare]

Even this doesn't get you an accurate measurement because it will give you only the coarse measurement that you can get off chip. For true accuracy, you need to measure it on-die.

We used this technique on a prior design that I was involved with:
https://www2.lirmm.fr/lirmm/interne/BIBLI/CDROM/MIC/2009/ITC_2009/Papers/PDFs/0008_2.pdf

I spent a bit of time about two years ago getting the ODDD and ODDI systems described in the paper work on a more recent microprocessor and the whole thing is really cool.

Cool paper! :thumbsup: Lovely to see all the real data rather than just simulated results.

 

[pm]

Cool paper! :thumbsup: Lovely to see all the real data rather than just simulated results.

Yeah, figure 3 is pretty amazing. That you have these detectors all over the place and you can get massively different readings depending on where you are... I mean, it makes intuitive sense, but seeing it - and seeing the magnitude of the delta - is something else. Previously I had this idea that a grid is a grid and pretty much all the points to pull power from move more or less together and droop was createdout on the external regulator. I don't think that I'd really internalized how much of a delta you could get on chip in a specific region (unless you messed up the grid by forgetting to contact it or something) but then you see how much of a delta there is across the die and you can see how large a delta you can get within a die over a short period of time.