Load Line Calibration and You

When 45 nanometer CPUs were first released by Intel there was significant speculation about how durable they would be, it was feared that the drop from Intel’s bulletproof 65nm process to their new 45nm process would result in fragile CPUs.  As overclockers worked with them, they found that for the most part 45nm CPUs were not nearly as fragile as feared.  It was discovered that they did not like high VTT values at all, but beyond that they were found to be solid.

Around this same time motherboard manufacturers released motherboards with a setting generally listed as “Load Line Calibration”, or “LLC”.  The instruction manuals and BIOS help screens were less than helpful, describing the setting as “Improves VCORE directly”.  Ok, thanks, that tells me a lot.

It turned out that what this setting did was prevent the CPU VCORE from dropping under load, a situation known as “vdroop”.  Normally if the idle VCORE was 1.300v and you put a heavy load on the CPU the VCORE would drop to 1.28 to 1.24, sometimes even more then .1v!  What LLC does is watch for that drop and supply more power to prevent it, keeping the VCORE at a steady level and helping avoid crashes caused by the suddenly lower VCORE.

Overclockers were thrilled, this meant they didn’t have to set the VCORE at levels that caused idle VCORE to be through the roof in order to get full load VCORE where they wanted it.  Then came an article at Anandtech that used some graphs and logic to prove that LLC actually caused voltage spikes when going from full load to low load.  In theory, the voltage regulators on the motherboard could not react fast enough to the sudden change in CPU load, and kept delivering full load power for an instant when the CPU did not need it, resulting in a brief jump in VCORE.

There was quite a bit of argument about this, but nobody was able to put it to rest solidly, nor to prove it solidly.  Theories abounded.

This brings us to today, when I decided that I’d had enough and set out to test the theories with my handy Snap-On MODIS’s oscilloscope function.

Test Methodology:

Using an Intel E5200 CPU on my Asus P5Q Pro, I connected the scope input to the output side of one of the inductors that feed the CPU.  The scope was set to freeze the screen upon detection of any voltage higher then normal.  Being a benching team member, I had plenty of kneaded eraser to insulate the motherboard with.  If you look closely you can tell the motherboard has also been insulated with conformal coating, as it gets down to -70 ° C or so on occasion.

LLC testing, probe on inductor output of the P5Q Pro

LLC testing, probe on inductor output of the P5Q Pro

The CPU was loaded with two threads of Prime95′s Large FFT test, and while watching the scope I stopped the load test.

Please excuse the reflections in the following pictures, glossy screens in a bright room lead to dubious pictures.

Testing and Results:

First up, the VCORE was set to 1.3625v in bios, this is Intel’s maximum recommended VCORE and seemed like a good level to test.

Screenshot of vcore and LLC setting

Screenshot of Vcore and LLC setting

It should be noted that on the P5Q Pro, a LLC setting of Auto means LLC is enabled.

I then rebooted and checked the BIOS’s voltage readout, to get an idea how accurate the sensor was.

Bios voltages with LLC enabled and 1.3625vcore set

Bios voltages with LLC enabled and 1.3625Vcore set

The scope had a different view on things.

Scope pic of 1.3625vcore LLC on in bios

Scope pic of 1.3625Vcore, LLC on in BIOS

In case you’re wondering, that is the VCORE jittering from 1.4v to 1.34v, with an average of 1.35 or so.  It should be noted that the jitters are exactly the same with LLC disabled.

Then it was time to fire up windows, and check CPUz and its interpretation of the VCORE.

LLC on, 1.3625Vcore, Windows Idle CPUz

LLC on, 1.3625Vcore, Windows Idle CPUz

The bios and CPUz don’t agree very well, do they?

Next, some load!

LLC on 1.3625vcore load cpuz

LLC on, 1.3625Vcore load CPUz

Under load the voltage recorded by CPUz goes up.

It looks like this on the scope:

LLC on 1.3625vcore load scope

LLC on 1.3625Vcore load scope

Pretty much the same as idle, the highest peaks are .01 higher, and the average voltage is a bit higher.

Now it came time for the actual test, removing the load suddenly by stopping the test.  With baited breath I hit the button while watching the scope like a hawk.  The result was disappointing, the scope saw nothing except a gentle drop in VCORE back to the normal idle level.  Figuring the scope must have missed it, I lowered the trigger and made it as sensitive as possible, restarted Prime95 and stopped it again.  Nothing.

I did this five more times with different scope settings, and found absolutely nothing in the way of spikes.

For the sake of completeness, I repeated the test with LLC disabled, the results were essentially the same though the VCORE dropped under load instead of going up.  No spikes.

Conclusion:

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

Bonus Thoughts:

It’s worth noting that the P5Q Pro used in this test is a high quality enthusiast motherboard, it has quite a few nice beefy capacitors to smooth out voltage spikes, and eight phase power to begin with.  It is entirely possible that a lower grade motherboard with fewer capacitors may get spikes but I don’t have any I can test.  It is also possible that my scope’s 50 microsecond refresh is too slow to capture the spike but I don’t think that this is likely myself.

When I first saw the jittery voltage going into the CPU, I was rather disturbed, but after thinking about it I think it’s ok.  There are two main types of capacitors (ok, so there are actually a ton, but as far as VCORE is concerned there are two important ones), big ones and little ones.  Big capacitors can absorb big voltage spikes and drops but react slowly and can’t catch high frequency jitter/noise.  Little capacitors can’t deal with big voltage spikes and drops, but can smooth out jitter and noise.  If you look on the motherboard, you’ll see a lot of big capacitors around the CPU socket, those take the 3v to 13v(!!!) coming out of the inductors and smooth it down into whatever the VCORE is.  There aren’t any little capacitors in the circuit here.  Now if you look at the bottom of the CPU you’ll see that it has a big patch of little tiny surface mount things, most of those are capacitors.  At least some of them are in the power circuit, and smooth out the jitters and noise.  I can’t test to find out how many of them for two reasons:  1) It’s different on every CPU.  2) My multimeter sends 2v down the line when I check continuity, the CPU would not appreciate this much.  In any case, my bet is that by the time the voltage actually gets to the cores it has become much smoother.

I would love to see more data. I don’t really recommend prodding your motherboard with metal clips and wires as it can cause instantaneous destruction of the motherboard, the CPU, your PSU, and possibly the world.  If however you want to risk it, please let me know what you find out.

In the mean time, enable LLC and enjoy the stable VCORE!

-Bobnova (OC Forums Benching Team)

Tags: , , , , ,

37 Comments:

EarthDog's Avatar
Excellent article Bnova! Informative!

One thing I didnt see (if I missed it I apologize) is the flip side of using LLC is that under moderate/light loads where the the LLC isnt active/compensating for vdroop it could potentially cause reboots/failures due to not enough voltage being supplied to the CPU in an idle or light load state. This condition becomes more apparent when LLC is enabled and the load voltage goes UP (vRaise?) as opposed to keeping the idle/set voltage and the user lowers the voltage in the bios to compensate.

I would, and do still use it, however I enable it after I have my overclock set and stable. Sometimes, albeit it rare, LLC/Vdroop control has been known to cause instability in some situations.

I am trying simply regurgitate some things I remember, if I am mistaking, please let me know...
thideras's Avatar
I agree with EarthDog. Can we see some results going from idle to full load? That is usually where the spike occurs. Besides that, very well done.
EarthDog's Avatar
The problem as I understood it is simply in the user compensating too much for the spike (as in goes above idle voltages) that does occur with LLC. The spike is there, its just the user trying to make it the same as the idle voltage is where this problem comes in to play.
MIAHALLEN's Avatar
Very interesting read Bob....thanks for taking the time
Bobnova's Avatar
I kept watch for spikes when the load hit as well, nothing there either.

I haven't seen any odd vcore with LLC enabled at light load unless I've had EIST and/or C1E enabled, in which case vcore in any situation is anybodies guess.

I had fun doing this, when I have a 1156 setup again I'll test it, as well.
EarthDog's Avatar
Maybe I wasnt clear...

Its not an 'odd vcore' its that people tend to compensate for the overvoltage LLC gives you on load and they lower the vcore to compensate causing instability at idle/low loads. I wouldnt imagine this happening to those in the know, but to others as I have seen it happen.
Bobnova's Avatar
Ahh ok, that makes sense. I haven't run into that personally, but my board doesn't gain a whole lot.
EarthDog's Avatar
Maybe its just the 'noobs' then? I havent run into this issue either, but then again, I dont touch the vcore after LLC is enabled either later in my 24/7 stabel o/c process.

Again, great article and dont think I was picking at it either.
Bobnova's Avatar
Oh I don't mind. I ask plenty of annoying questions myself, I'm happy to answer other people's questions.
kayson's Avatar
Nice to see some real investigation into this. But if I had to guess I'd say there's a very good possibility that the voltage spikes are <50us. When you say your scope "refresh rate" is 50us, do you mean your scope takes a sample every 50us (i.e. 20 kSamples/s)? Or that each time division is 50us? In either event, that's probably too slow to see a voltage spike that occurs on a nanosecond scale. I have some nice scopes at work that can go down to nanoseconds, but getting a computer in here to test them would be rather difficult.

I chatted with a co-worker who has experience in power supplies and this is what we came up with as far as the voltage spikes go: it is far more likely to see a large spike going from heavy to light loads as opposed to light to heavy. Switching regulators on a motherboard have a feedback loop that makes sure the voltage is what its "supposed" to be. When the load increases, the feedback loop has a fixed slew rate that determines how fast it can react, and it will definitely be slower than the time it takes to go from a light load to a heavy load. The light-to-heavy spike comes from an overshoot in the feedback loop where it continues ramping the power up beyond what is necessary to maintain the voltage. But the feedback loop keeps chugging along and brings it back to normal very shortly after.

The voltage spike going from heavy-to-light is worse. Switching regulators work by storing energy in an inductor then passing it to the load. The output voltage is controlled by the duty cycle of a pulse-width-modulation signal used to do the switching. If the current goes beyond a certain limit, that voltage will start to droop because energy isn't being transferred fast enough. The solution is to increase the duty cycle to supply extra current. But if all of a sudden you drop the current draw (going to light load), the duty cycle is still longer, and since the regulator isn't strained the voltage will spike. The feedback loop comes into effect and lowers the duty cycle, but you'll still see a large spike before it reaches equilibrium.
I.M.O.G.'s Avatar
Kayson, I think this article started from what happens in theory compared to what can be proven and tested for in application on actual hardware with a reasonable attempt at observing the theoretically expected behavior. Your comment while informative and well stated, brings us full circle - Yay for symmetry!
kayson's Avatar
Oh I got the point of the article. I'm just providing some more theory as to what should be seen and suggesting that someone take a look with a higher-speed scope just to be sure.
I.M.O.G.'s Avatar
Good points, just having a bit of fun.
Bobnova's Avatar
I agree, it'd be nice to use a finer grade scope with this, if this article can inspire someone to do that I will be very pleased.

Thanks for the details on how the power supply works, too.
infam0ussteven's Avatar
Thanks for this article, Bob.
Jmtyra's Avatar
Excellent info, I'm glad this was enabled in my BIOS. Thanks for doing the research! =)
Bobnova's Avatar
Inspected the scope carefully, for those who care:
50uS per screen (50 microseconds), the finest point it can detect/display appears to be 90-120 nanoseconds, somewhere in that area.

And you're welcome!
bcsizemo's Avatar
Kayson would be correct in his description of how the switch mode setup should work. 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 your scope is measuring around 100 ns, that would be roughly a 10 Mhz signal. That should be fine even for a cpu psu. (I don't know the frequency that the switching signal runs, but I'd think it would be in the 1-5 Mhz area.)

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 (yeah I already said that, but.) Most of these are based on programmable switch mode supply control ics. It's likely someone did a poor job of implementing the over voltage detection, and is just letting the unit cycle and use the feedback loop like normal. Most of these chips should have a dedicated over volt input which should cause the circuit to cycle preventing an over volt situation. (Kind of like applying the brakes while cruise control is on, it disables and you stop.)

That's just my theory based on my experience with larger scale switch mode ic chips.
retrogreq's Avatar
very good read bnova, thanks for the writeup
johnm's Avatar
one of the better reads thanks to Bobnova
thanks to the oc benchmarking team for all their good posts.
Would be cool to have a scope on for a hour or so but that
might hurt your scope.
CompuTamer's Avatar
Nice article! I was never aware of that possible spike, but i'm glad that it's no present on higher end boards... i would assume that if those spikes really did happen at some point (Maybe before power regulation got so good?) the caps would smooth it back out before it hit the CPU.
gahz73's Avatar
i was wondering if you could do this test using IBT. ibt seems to have higher power requirements from my system so i'm wondering if LLC is being affected more from ibt than prime95.
i normally use ibt after running prime for 24/hrs. for some reason if i cut it close on voltages i can pass prime but not ibt.
ihrsetrdr's Avatar
Thanks for covering this topic Bob, I've always left LLC off since I hadn't seen anything definitive about it, until now.
Bobnova's Avatar
A quick followup:
50us is the width of the screen, the minimum spike time to capture is something like 50ns if I recall correctly, very short indeed.

I've been meaning to repeat this with my P67a-UD4 and 2600K, maybe on dry ice if I'm really feeling hardcore.
I.M.O.G.'s Avatar
Seeing a followup would be cool!
johnm's Avatar
just a small follow up please
break loose and put that scope to more tests.
stevenb's Avatar
try with a smaller timebase as well as mv/div, how can you actually see precisely at 200mv?

Its very hard to see that kind of voltage overshoot, and also why measure at the inductor's output? Why not measure at an MLC capacitor as output? i mean in real life the CPu is fed by the caps, not the inductors. Well yes the inductors feed the caps, but the amount of Vpp you are seeing is what 600mv? that is way to freaking much, your overshoot wouldn't even be 600mv!!!! acceptable ripple on these dc/dc SMPS is under 150mv.
freeagent's Avatar
Soo what would the general consensus (sp?) be on this?

I know I have been useing it since the p5b deluxe wifi was the shiznit, and I am still useing it to this day, and always on the max setting of course

Works good.
kayson's Avatar
The general consensus I've seen recently (both here and at other places) is that it's not some super-destructive setting that will fry your CPU with voltage spikes. That being said, not everyone will need to use it, and TBH if you don't need it, there's no reason to use it. But if it helps your OC, go for it.
stevenb's Avatar
yea its not going to be harmful to your CPU in the shorterm, nor really in the long term. If you want your CPu to last as long as it can then it'd be wise not to use it, but otherwise its not going to cut the CPus life down noticeably shorter than your OC will.

In this day and age, SMPS design in terms of the PWM has come a long way since anadatech published that article.
micjmac's Avatar
Interesting article. I decided to play it safe and disable LLC, but it's requiring a 1.35 Vcore to run at my measly 3.6 GHz. My system wasn't stable at 3.8 GHz with 1.375 Vcore, so I know that the only way I'm going to possibly be able to push this system any further is by enabling LLC. Let's say that by enabling LLC, I am able to hit 4.2 GHz but I have to run at 1.4v. Is the benefit even worth the risk when I only use this PC for gaming?
Bobnova's Avatar
How do your games run at stock clocks? That's the real question.

Personally I have run LLC of some level 24/7 and benching since this article and have yet to experience CPU damage because of it.
Or, for that matter, for any other reason. These have not been friendly skies my CPUs have been flying in either!
micjmac's Avatar
Have you enabled it when your Vcore is at or above Intel's maximum spec (e.g. 1.4v for i5 750) on a 24/7 overclock setup?
iueras's Avatar
So here is a strange thing that I can't find much info on and have no idea why it happens: When I enable LLC (setting enabled as opposed to auto) on my Asrock 970 Extreme3 (BIOS version 1.40) with a 955 Phenom II in it, the comp boots and initial CPU voltage appears normal (in fact, it is closer to what I set in BIOS than with it off). but, when I put a load on (such as a prime95 test), the voltage starts to DROP.

There seems to be some kind of rate limit on how fast it can change in either direction, but the voltage goes down... and down... and down. I'm not talking about just a little, I mean on the order of a 0.25V drop. Before I BSOD with LLC on, the lowest voltage I have seen it hit was 1.12V from a BIOS set voltage of 1.3375.

The exact opposite happens when I set LLC to disabled in BIOS. My initial voltages are higher (setting 1.3375 yields a BIOS HW monitor reading of 1.352V), and when I load the CPU with prime95, the voltages go UP. BIOS setting of 1.3375V yields a max voltage under max load of 1.375V as reported by both AIDA64, CPU-Z, and Asrock's tuner utility. All the software has exactly matched what the BIOS HW monitor shows for CPU voltage.

This behavior seems to be exactly the opposite of what I would expect. Has anyone else seen behavior like this?

PS: With LLC off and the load voltages going up, I am prime95 stable after 6 hours (longest I have run it, I don't like testing hardware while I sleep just in case). With LLC on and the voltages dropping under load, I error out on core 4 after 5 seconds to 2 minutes max.

EDIT: I meant to post this in a new thread on the AMD mobo forums, my bad.
Bobnova's Avatar
That sounds like a messed up bios, plus a defective board. Voltage shouldn't slowly do anything. When load hits it should go up or down immediately, and then stay there.
iueras's Avatar
Voltage rate limit was a function of the monitoring software I was using. Tried some other software it is indeed instantaneous.
caddi daddi's Avatar
great writeup bobnova.
it shot down what i was thinking was causing my chv-f to freeze when closing programs when above 5.2 or 1.72 vcore.
also you made a fibber out of asus, the manual says that llc lowers the vcore under load to help control temps.
Leave a Comment