Originally Posted by Jpmboy
Voltage is meaningless regarding "kill", current kills (and only current can increase temperature). Voltage is only the potential that current (amps) is delivered at, when a "work request" is made. Sure, LLC does not kill a cpu - but transient spikes degrade them over time, This is why LLC continues to be programmed into bios - only on those rails subject to this effect (it's just the physics of current change at constant voltage). With the VR on x99, we have access only to vccin in this regard, all other voltages are stepped appropriately on the die.
Likewise, I've never had a cpu or overlcock for which I could only achieve the needed voltage by defeating vdroop. Unfortunately what we see as "flat" is not in fact, flat.
I just have a different approach: Idle voltage is meaningless, so when needed I just raise the voltage (VCCIN for this generation) and allow vdroop to mitigate the transient-induced voltage spike. Intel describes the effect in their spec sheet for this generation.
Certainly there is "art" or personal preference in this. I prefer to allow some vdroop.
No suggestions if it is running the way you want it to.
Sorry, your understanding of what kills mosfet transistors is not correct. There are a few ways you can kill the FETs on our cpus. Yes, too high current either on gate or source will burn out your transistor by *heating* and effectively melting the transistor.
However, most transistors die due to excess voltage, either on the gate or source. Gate overvoltage will damage the transistor by effectively 'poking a hole' in the gate insulation, causing the gate to always be 'on' regardless of gate bias(permanent gate biasing). While excess voltage on the source will effectively ruin the electron doping that creates your electron band by simply pushing all the electrons to the far side of the transistor, making it impossible to bridge the gap between source and drain(migration of charge carriers). This is why Intel prints vcore and vccin maximums in documentation in terms of maximum allowable voltage, and not maximum current rating. Although there are some noted maximum current ratings. https://en.wikipedia.org/wiki/Failure_of_electronic_components#Semiconductor_failures
In fact it's ONLY in overclocking that current related heat deaths are anything other than an anomaly caused by material defect.
Intel has pulldown circuits to protect from transient voltage spikes. This was already true back in the Core2 days when Anandtech originally published the article. To put it another way, protections for vCore overshoot were already in place. It's only if the overshoot exceeded the protection circuit's capabilities that you would have a problem.
Here are the datasheets for both the x99 HW-E platform and the old qx9650 platform discussed by anandtech:
The first document you'll find the overshoot spec for Vccin on page 55, and the second doc you'll find the same exact specs for Vcc(classic vcore) on page 25. The spec for Vccin is identical to the classic Vcc at 50mV tolerance over 25 micro seconds.
Anandtech's article was seriously flawed to begin with in several ways. Even in their ABSOLUTE WORST case scenario overshoot was only 60mV on 1.25v core, nowhere even close to dangerous. They incorrectly state that Vcc isn't supposed to exceed VID(wrong as you can see from the datasheets). They also don't state the duration of the overshoot, what they were seeing could very well have been under the 25 micro second tolerance. The community recently got a crash course on transient load spikes because of the RX480's PCIE slot power problems. It's not a direct relationship, but the same principles apply.
Additionally, Anandtech ignored the processor power delivery specifications. Which states that an OVP violation is actually 1.6v+200mV(or 1.8v) on page 41:
This document applies to LGA775, LGA1366, and LGA1156.
The guys at anandtech generally do very good work, but this section of this particular article is extremely misleading and caused the false stigmatization of LLC. You can't damage your CPU with an overshoot of 60mV at 1.25v like anandtech measured. You have to exceed 1.6v absolute max on the older cpus to cause damage. And if you DO exceed that, the OVP protection would trigger and cut Vcc entirely, preventing CPU damage. I think they knew this which is why they primarily focused on the possible stability problems, and why they only mentioned the 1.6v+ problem extremely briefly as a 'user considered maximum safe'. Also, it was common knowledge at the time to overclockers that 1.6v was flat out too much voltage for these 45nm chips. So the entire premise of damage due to LLC induced overshoots is false.
This section about LLC in the anandtech article was written because Kris Boughton was upset that he had stability issues when he turned on LLC, and he didn't understand that it's a setting that will be necessary for SOME to achieve high overclocks. Anandtech has always been known to be extremely conservative when it comes to overclocking, so LLC would never be applicable in their testing.
The issue with my 5930k is that the vcore has an unusually large droop unless LLC is turned up to 8 and 1.9v Vccin. To put it another way, the IVR on my 5930k sucks and it can't regulate vcore correctly without tightly controlled, and raised, Vccin. I might have been able to mitigate this by increasing the switching frequency, but generally you see higher heat loads by increasing the switching frequency than you do by simply controlling the voltage more tightly using higher levels of LLC. This 5930k is by far the WORST cpu I've had in the last 6 years as far as voltage control is concerned. Vcc droops beyond intel spec at stock speeds with no LLC. This doesn't cause instability at stock speeds, but when overclocking it causes big stability problems.
Again, just like any other overclocking option it all depends on the silicon lottery. Increasing LLC is not necessary for all CPUs, and is (usually) not necessary for mild overclocks like those tested at anandtech. LLC is one of the options available for making lemonade out of the sourest of lemons like my 5930k.