This is some research I’ve been doing on LLC and overclocking, I don’t think it is a particularly well explained or well understood topic so I hope a few people find it useful. I certainly have been learning whilst experimenting (and my pc still works too – bonus)!
At this particular moment I am running my 3570K at 4.6GHz. If I were to say that I had the choice between running the Vcore at 1.32v or at 1.40v (and both being stable) I’m sure that most people would know which one is the obvious choice and would therefore be surprised to find that I chose 1.4v. Why? Well, believe it or not, manipulating the LLC gets me lower temperatures at 1.4v than at 1.32v.
Setting a voltage in BIOS is only a fraction of the story.
If you set 1.32v into BIOS then that doesn’t mean that the CPU simply runs at 1.32v. In fact 99.9% of the time the CPU will be running on voltages which are completely different yet still very much related. What you’ve set is more like a target which the motherboard doesn’t want to exceed.
When the CPU is running it demands power from the motherboard. The motherboard carefully supplies power as a reaction to this demand but, because it is a reaction, the actual change in power delivery is always a bit late. This means that the CPU is very briefly supplied with more power than it needs
- a power spike. You won’t see these on any hardware monitor because they happen so fast but they do appear on an oscilloscope.
Power spikes are bad. In order to protect itself from any potential damage the system will run the CPU underneath this limit.
Therefore when these spikes happen they also remain underneath the limit that has been set. This is called the Voffset.
When the CPU is working hard and demanding more power the size of these power spikes will increase so Intel added a second function called Vdroop
which simply increases the offset, by another name, when the CPU is under load so that these larger spikes still remain under the limit. Here’s a graph:
Not too long ago enthusiasts noticed that they weren’t getting the volts that they asked for and the market response was to create Load-Line Calibration AKA Vdroop Control.
This does exactly what it says on the tin – the greater the Vdroop control, the less able the motherboard is to reduce volts under load. On the face of it overclockers were happy because their hardware monitors were now reporting higher voltage which meant less Volts set in BIOS for the same overclock. What wasn’t realised (because you can’t see them) is that these voltage spikes still happen, only now the large spikes are more likely to break the limit because the overall Voffset is much less without Vdroop:
There is also another drawback.
On some small level LLC reduces the responsiveness of the power supply and that means instability. As I understand it, by limiting voltage control you limit one of only two methods (volts and amps) which the motherboard can use to control power.
Here’s a practical example.
I setup my rig to run a 4.6GHz clock on fixed volts. I then selected a Vcore in BIOS which gave me a stable system within two conditions – LLC off (level 7) and LLC on fully (level 0). To get some reliability I stipulated the Vcore had to be the minimum required (to the nearest 0.005v) to run 10 passes of IBT on Very High stress. This gave me 1.4v/LLC7 and 1.32v/LLC0. I then logged the sensors whilst running IBT for a second time:
As you can see the LLC0 condition supplied a much more varied amount of volts due to the free Vdroop. For those who are concerned about volts it’s interesting to note that the 1.4/LLC7 condition supplied noticeably few volts under load but that really isn’t the total picture:
You can see that, even though the voltage supplied was vastly different between the two conditions, the actual Watts was very similar. However there is a noticeable increase in watts supplied to the CPU under the LLC0 condition in the later runs. The control that the LLC is having over the voltage is affecting the control that the motherboard has over the power supplied to CPU resulting in a tiny increase in power which is reflected in the temperature graph.
The resulting cost in temperature was 2C. The important point is that the temperature mirrors the wattage and not the volts.
The moral of the story?
High volts aren’t bad, high watts are – and watts make heat.
Vdroop is a good thing – at least you know what the upper voltage limit is.
If your default is to run with LLC on then try it the other way around, or somewhere in the middle – you might get better temps or even a better clock.
That's as far as my research has got me at the moment. If anyone has anything to add it would be great to hear it!
Credit due here.