CPU Overclocking, Vcore MAX myths and truth

[Gillbot]

I would love to but the high end gear is off limits to my shadetree hands.

 

[Rubycon]

Originally posted by: Idontcare


CPU-z does not sample the Vcc at high enough frequency to capture the evolution of a transient with or without LLC

No software monitoring can do this as the demands are beyond the sensor providing the information. Testpoints taken right off the board to the right equipment would reveal this nicely I gather.

 

[dflynchimp]

Originally posted by: Gillbot
I'm done replying to you because you obviously fail to comprehend even the simplest of Ohms Law. Keep your misinformation from my thread.

this post FTW XD

 

[imported_RJR]

PLEASE don't let this thread DIE!!!! I've read through it at least a couple times now. I did a pencil mod on my board a couple years ago that completely eliminated voffset/vdroop (as per CPU-Z) and have been waiting with bated breath for a conclusion here. Someone out there has to have access to a scope so this can finally be indisputably answered for good. PLEASE!!!!!

Have pity on me, I'm an old man and don't want to go to my grave with this unanswered question hanging over my head.

 

[BonzaiDuck]

Originally posted by: RJR
PLEASE don't let this thread DIE!!!! I've read through it at least a couple times now. I did a pencil mod on my board a couple years ago that completely eliminated voffset/vdroop (as per CPU-Z) and have been waiting with bated breath for a conclusion here. Someone out there has to have access to a scope so this can finally be indisputably answered for good. PLEASE!!!!!

Have pity on me, I'm an old man and don't want to go to my grave with this unanswered question hanging over my head.

Well, old man, I'm an old man, and you remember Paul Newman in "Hombre!" facing down that nasty old Richard Boone? "We're all gonna die sometime . . . . it's just a matter of 'when' . . . "

But seriously. I came back to this thread just to pose what may even be a naive question, and it relates to your sunset-years pencil-modding. I have a 0.03V vDroop with an eVGA 780i motherboard. That's not very much, as I understand it, but I think I can lower my vCore setting a notch (if a pencil mod closes the 0.03V gap) and still have stability at my highest CPU speed configuration.

And I'm wondering how the effects of a simple pencil-mod would differ from LLC="ON" for motherboards which provide LLC. My 780i board doesn't have any such thing.

 

[imported_RJR]

Don't know for sure, that's why I'm waiting for some wise and generous person with a scope to map it out for all to know for sure. I've read this article http://www.anandtech.com/printarticle.aspx?i=3184 a while ago and would love to know how accurate the diagrams are. My board had a very bad voffset(vdrop) of around .08v and another .02-.03v vdroop, so the pencil mod was a blessing for me, just wish I knew for sure the affects of said mod or LLC for that matter.

 

[toslat]

Nice discussion.

My understanding of the issue is that with LLC you are essentially trading a higher V_load (and lower V_idle), for higher transient overshoot. I expect the actual calibration to be a one time even, but that does not exclude the occurrence of transients with load changes.

For illustration, using Anandtech graphs

Case 1: Following intel guidelines,
VID= 1.25V
V_idle=1.23V <=>V_offset=0.02V
V_load=1.19 <=> Vdroop =0.04V
V_max=1.25V
V_min=1.17V

Case 2. V_offset= 0(!?)
VID= 1.25V
V_idle=1.25V
V_load=1.21V <=> Vdroop =0.04V
V_max=1.27V
V_min= 1.19

but note that this is the same as if I had kept the V_offset=0.02V and increased VID=1.27V (or with V_offset =0 and VID = 1.23V, we would have Case 1).
The only thing different is the labeling of the graph.
The argument of exceeding VID is irrelevant as the moment you bump up your voltages, you are essentially defining your own VID.

Case 3: LLC enabled and V_offset =0.02V
VID=1.25V
V_idle=V_load=1.23V
V_max=1.29V
V_min=1.17V

Compared to case 1, we have increased the maximum overshoot to 1.29V (by 0.04V), but have also increased the V_load to 1.23V (by same 0.04V) making a higher overclock possible. To get same V_load using the intel system would require VID=1.29V.

The above analysis is based on the anandtech graphs and is not a reflection of true performance ( the Voltage relationships would likely be non-linear in real life). I just used it to illustrate that the LLC is not as bad as some people paint it.

To have a true assessment, one would need data on the transient response of the system with and without LLC enabled. Of interest would be the relationship between increase in V_load, and, the consequent increase in overshoot. Another thing is the effect of a low V_idle on heat generation.

 

[imported_RJR]

Originally posted by: toslat
To have a true assessment, one would need data on the transient response of the system with and without LLC enabled. Of interest would be the relationship between increase in V_load, and, the consequent increase in overshoot. Another thing is the effect of a low V_idle on heat generation.

Please tell me you have a scope handy, or know someone with one that could do some real world PC tests for us. My new system is even worse than my last with offset/droop a little over .12v, so I would really be interested in knowing for sure the effects of droop vs mods vs LLC.

 

[bob5568]

First of all, thanks to Idontknow and Gillbot for helping me better understand vdroop and vdrop, and the role of LLC. I'm interested in a comment early in the threat made by Idontknow to the effect that sometimes it could be ok to enable LLC. To ask the question in a specific manner, I took some data on my current oc project. ----------------------------Hardware: i7 860/ga-p55a-ud4p/prolimatech megahalems/g.skill ripjaw pc1600 (4x2gb)/gforce 8800gts (g92)/corsair 520/lancool pc-k60b/wd 640 black (x2 in raid 0). .................2 versions of oc, both survive the one hour OCCT 3.1.0 test with no errors. First,(21x181) LLC enabled; vcore (bios) 1.29375v; vcore (cpu-z) idle is 1.280; vcore (cpu-z) 100% load is 1.296...................Second, (21x181) LLC DISABLED; vcore (bios) 1.346v; vcore (cpu-z) idle bounced back and forth between 1.31 and 1.33; vcore (cpu-z) 100% load went back and forth between 1.28 and 1.30. Without knowing the amplitude of the ringing, how could one assess the first condition was safe? And is there any reason to assert that condition one or condition two were better? Temps were similar, worst core topping out at 74C in both cases, best core around 68C. And if it matters, ram is underclocked at 1448mhz, with vdimm at 1.63v; vtt at 1.21v; vpll at 1.86v and vmch elevated one bios notch, I didn't write the value down. Best,Bob

 

[bob5568]

i7 860 @ 4.0 GHz @ 1.32v/quote]

RJR, is your overclock of 4ghz using LLC? Is your 1.32v Vcore as set in bios? Based on my current testing my board would have drooped too far to allow 4.0 stable with no LLC, and our systems are quite similar.

Best,
Bob

 

[Makaveli]

This was a very good read. I agree to keep this alive.

 

[imported_RJR]

i7 860 @ 4.0 GHz @ 1.32v/quote]

RJR, is your overclock of 4ghz using LLC? Is your 1.32v Vcore as set in bios? Based on my current testing my board would have drooped too far to allow 4.0 stable with no LLC, and our systems are quite similar.

Best,
Bob

Hey Bob,

I've had a few discussions with you on another forum, Yes LLC is enabled, the Bios is set to 1.3v and under load is 1.312 (so I rounded up a little).

This whole discussion with LLC/Mods to reduce vdroop/vdrop and said effects of transient overshoot are real, BUT the only thing I haven't been able to find anywhere is the duration of said overshoot that would tell everyone if in fact the overshoot is dangerous. A very short transient overshoot will not do any damage, but a longer one will. That's the only reason I'm still begging someone with a scope to map it out for all of us to know.

 

[bob5568]

Hi RJR...I wonder how high your bios setting would need to be to acheive 1.32 at the core under load without LLC. Furthering that line of thought, is it possible to understand how safe a voltage setting would be with LLC enabled to learn what alternate settings would need to be without llc to acheive the same vcore under load. Essentially, as I understand it, with llc disabled, we are establishing a max vcore by entering a number into the bios. It could be that this was figured correctly, and therefore that bios setting may be the value actually realized by the system even if that system is then reconfigured with LLC enabled, and the bios setting changed to more or less match the realized vcore under load.

 

[imported_RJR]

Finding the correct load vcore isn't the issue (by the way my vcore would have to be set at 1.44v in bios W/vdroop/vdrop) the issue is if the transient overshoots that increase when you enable LLC/volt mod which is reported to be a bad thing and bring about an over/under voltage condition. The only problem is that the duration and length of said transient overshoots (as far as I know) are not known. If someone had a scope and said here is a scope pic of an i7 860 at 1.35v vcore with LLC enabled and placed/removed a load on said configuration to show the exact spikes and duration of said spikes we would all know if this indeed is a concern or not.

My feeling are that the spikes are of too short a duration to matter, but once again, I can't prove anything without a scope.

 

[bob5568]

Hi RJR, thanks for your response. Bear with my thinking, I don't think I was clear.

In your case, you determine that 1.44v in bios acheives 1.32vcore under 100% load, with LLC disabled. Right? Given what I'm understanding, that would mean that your system will not produce transients that exceed 1.44v ~ ever. Vdroop and Vdrop assure that.

The interesting leap I'm suggesting is this.....could that ALSO mean that you'll not exceed 1.44v transient ~ ever...if you now ENABLE LLC and set your bios at 1.35 or whatever you are using. If so, than our interest in seeing the o'scope trace may be unnecessary.

Best,
Bob

 

[imported_RJR]

In your case, you determine that 1.44v in bios acheives 1.32vcore under 100% load, with LLC disabled. Right? Given what I'm understanding, that would mean that your system will not produce transients that exceed 1.44v ~ ever. Vdroop and Vdrop assure that.

The interesting leap I'm suggesting is this.....could that ALSO mean that you'll not exceed 1.44v transient ~ ever...if you now ENABLE LLC and set your bios at 1.35 or whatever you are using. If so, than our interest in seeing the o'scope trace may be unnecessary.

Best,
Bob

The unknown IS the length and duration of said spike. At 1.35v (W/LLC) it could go to 1.6v (who knows) and at a duration that would cause problems or it may be very voltage dependent so reducing the voltage slightly could have a big impact on the length of the spike. These and many others are the variables that I don't know for sure. The only way I can come up with to answer these questions would be with a scope.

 

[Gillbot]

The unknown IS the length and duration of said spike. At 1.35v (W/LLC) it could go to 1.6v (who knows) and at a duration that would cause problems or it may be very voltage dependent so reducing the voltage slightly could have a big impact on the length of the spike. These and many others are the variables that I don't know for sure. The only way I can come up with to answer these questions would be with a scope.

A short spike can still be damaging.

 

[imported_RJR]

As per Intel's data sheet:

http://download.intel.com/design/processor/datashts/320834.pdf

It states at 2.11.3 " Overshoot events that are <10 NS in duration may be ignored", that's why I am thinking that duration has something to do with it.

 

[extra]

In reality using LLC is a matter of personal preference and you can safely use it. Using it while also having a really retardedly high vcore like 1.55v or something will probably increase your risk of damaging your chip

 

[BassBomb]

I will admit, I did not read the whole thread, but I read the OP..

I can set my vcore from anywhere within the CPU's set range in BIOS (1.5V top end) and regardless my motherboard monitor (in BIOS) and all windows monitors hover around 1.3x V... what is it actually running at? LLC is on
Do it just go by what I set (1.5v) as the TRUE voltage?

 

[Gillbot]

To get true voltage, yo uneed to physically measure it. Don't trust software readings.
http://forums.anandtech.com/showthread.php?t=257485

I know that is for power supplies, but the same principles apply.

 

[BassBomb]

To get true voltage, yo uneed to physically measure it. Don't trust software readings.
http://forums.anandtech.com/showthread.php?t=257485

I know that is for power supplies, but the same principles apply.

So setting it to 1.5 may or may not be 1.5?

 

[mutz]

So setting it to 1.5 may or may not be 1.5?

yeah, probably, yet it doesn't matter for you,
it only represent a number, a mark,
unless you want to find the real values for measurements.


/post

this thread holds gems,
very good read! though that article is a bit ~.

 

[chinobino]

I have read this thread a few times in the past (some very good info) and was finally prompted to register with the forum to share a link to an article with LLC testing with an oscilloscope,
'Load Line Calibration and You'.

*Warning - article spoiler below*

Conclusion:

If you have a decent quality motherboard LLC is a good thing, and not likely to cause any issues at all.

Also, some of the comments after the article make some very valid points.

I do a lot of strenuous oc stability testing at high vcore & high clock speed with LLC enabled (worst case scenario) and it seems that you are most likely to do the CPU damage from a voltage spike immediately after the stress test has finished.

bcsizemo describes what I believe is the potential 'LLC issue'...

Part of where the spike could come from is that time lapse between the cpu going from full load to light load in a matter of cpu cycles. Now if that really happened then the cpu PSU (the switch mode parts you are testing) would have a cycle time much longer and potentially supply a dangerous voltage.

If this outlines the potential LLC problem then what is the governing force that attempts to stop this happening?

I would think that a large part of this issue might deal more with the feedback circuit than the actual regulation itself. Most of the capacitors and inductors are sized to meet a goal load/regulation requirement. The capacitors would need to store enough juice to be able to go from light to heavy load quickly, but that size should be able to handle ripple in the sub 1 volt category easily.

I'd suspect if this issue is ever verified I would think the likely cause would be a poorly implemented feedback circuit...

So, in summary, a poor performing PSU (with lax tolerance/poor switching) coupled with a motherboard with cheaper components (or should I say being setup in a way that it's unable to compensate for the ringing in the feedback loop with LLC enabled) whilst running variable CPU load tests is the recipe for disaster.

Anyone wanna try kill a CPU? :biggrin:

Used sensibly LLC seems to be quite safe - even at moderately high voltages (i.e. 1.45V to 1.50V for 45nm Core 2 Duos) assuming you a have good quality PSU and motherboard (and your cooling is up to the task).

So if the spiking is a non-issue - we only have to worry about electromigration (and yes I realise that this test was run once by one person using one scope and one motherboard etc and really does need to be replicated for proof).

 

[fsardis]

I had to get my trusty old electrical engineering books out of the shelf to follow this thread but since I am no electrical/electronic engineer, I think I have missed about 50&#37; of the info here.

I understand that there is not way to prevent transients no matter what. That is just basics, it is the law, it is the universe, it is god's will.
I did understand that without LLC you get a a few mV droop and then the CPU (lets assume 100% load) will receive stable voltage slightly above the droop spike. Then when the load is off the CPU, you will get a spike up a few mV above the idle Vcore and then you will stabilise at the normal idle Vcore.

Now with LLC enabled, you will get the same initial spike at droop but then the regulation circuit will try to push the load Vcore closer to the idle Vcore. This means that when you remove the load from the CPU, the spike that you will get will be some mV above idle Vcore but this time it will be several mV over the idle Vcore compared to when you had no LLC obviously because the beginning of the spike is now closer to the idle Vcore.

So are people here saying that there is a prevention mechanism that limits the upward spike in LLC so you don't momentarily overvolt the CPU to a dangerous level? Or are they simply arguing that LLC is bad just because of that and it should not be used because it is not safe for the CPU?

I have been using LLC both on my Q6600 and my i920 but I have been using it with very slight increase in Vcore. So hopefully this huge upward spike caused by LLC does not go into dangerous Vcore levels. So based on all that, I do understand that running the CPU at a very high Vcore and enabling LLC on top is akin to putting your head in a lion's mouth. However, what if you have your Vcore running pretty much at factory spec and all you need is to limit the droop so you can get some extra GHz out of the CPU? Isn't that better than leaving LLC off and instead upping the Vcore by a few mV?

In my example, my i920 runs at stock Vcore but it was impossible to hit 4GHz stable without raising Vcore substantially so that the load Vcore will be enough. So instead I enabled LLC and presto, I have a stable CPU under load without increasing the factory Vcore. So the next logical question now would be: is the upward spike caused by LLC many more mV than I would have added to the Vcore to get it stable? If it is considerably more mV over the Vcore I would have added, does it have a worse impact in the lifecycle of the CPU compared to running it at an increased Vcore 24/7?

I am pretty sure some of the answers to these questions are in the thread but they are given in highly technical format so I can't get them.