First CPU upgrade since SB...I have a few questions.

[TiredEngineer]

I ran a i7 2600k at 4.5 GHz (1.4v) for nearly 4 years. I did have to drop it down to 4.2 recently as my system seemed to be losing stability. I purchased a 6700k with a Gigabyte Z170X-UD5. I did just fail a stress test at 4.5 GHz and 1.25v. My temps peaked around 67C (and were usually below 60C). I will probably bump it up to 1.275v next.

1) I remember load line calibration controllers were sketchy in the past. While gaming, CPU load goes up and down often, which makes me nervous with unreliable controllers. I have 3 LLC options...auto, standard, high. Auto and standard are crap. My vcore drops all the way down to 1.1Xv when set to 1.25v. The high setting seems to do pretty good, but are there burst voltage spikes occurring that are damaging the CPU due to overshoot?

2) I never used XTU in the past. Is there a reason CPU-Z and HWinfo are reporting core voltage with what I am expecting based on BIOS settings, but XTU is reporting something like 1.37v (instead of 1.25v).

3) C states were always disabled while overclocking back in the day. Is the sentiment still the same? Will the voltage actually even drop when I have a hard coded voltage (havent wanted to test this yet while working on general stability)?

4) It seems people are using voltage much higher than mine and claiming it is safe. It makes me nervous even at 1.25v on this 14nm process. What has changed since the previous generations to make people believe 1.35v+ is safe for long term use (I am hoping to get another ~4 years out of this)?

 

[pantsaregood]

I've never bothered with LLC because you could potentially end up pushing more voltage than intended during AVX loads. Even without it, AVX/AVX2 stability tests act as power viruses.

I've never used XTU.

C-states are fine to use and have been for a while. The last PC I had that wouldn't properly overclock with power saving features enabled was an Athlon 64 5200+ on an AM2 nForce 6100 board. My Core i5-2500K has had EIST and C-states enabled at 4.4 GHz for almost five years at this point - it would've easily done 4.7 GHz, but I liked the idea of a 33.3% overclock.

"Safe" voltage is a tough question. Even at stock voltage, your processor will eventually fail thanks to electrons deciding to randomly jump around when they aren't supposed to. As transistors shrink, this becomes more of an issue. More voltage increases the probability of this happening, as well. That said, if you're worried about it, just buy Intel's tuning plan and be done with it.

Also, I've killed motherboards before I've killed CPUs before - I pushed 1.7v to a Phenom X4 9750 and blew out the power delivery system on the motherboard I was using. The CPU itself was fine despite having been repeatedly stress-tested at levels between 1.5v and 1.65v.

 

[TiredEngineer]

Thanks for the reply. Apparently voltage has to be done through offsets with a Gigabyte board to get the C states to work. I tried doing this, and while it was hovering around 1.2Xv when at low loads, as soon as it was loaded (this was at the minimum offset that the mobo would post as an FYI)...the voltage jumped up between 1.4-1.5v. Scared the crap out of me (I doubt damage was done in less than 30 seconds of these high voltages...I hope).

Anyway, back to hardcoded voltage where C states are worthless.

I still have LLC at "high" since the C states don't work...the amount of idle voltage required to get a stable overclock with maximum vdroop would be a lot (I would probably need 1.35v to get 4.5 GHz stable, whereas with vdroop on high I am at 1.275v and it looks good so far).

 

[Yuriman]

Maybe I'm misinterpreting, but you probably should not be using fixed voltage when overclocking, unless it's for a very specific reason / short term test. It wasn't even a good idea in the Sandy Bridge days.

On Haswell and newer CPUs, when you use AVX2 instructions, the CPU adds extra voltage. To my knowledge, this can't be disabled. Most people will stress-test their overclock using a non-AVX2 load (e.g. a very old version of Prime95) to simulate what they'll be doing with your PC 99% of the time and then check to make sure temperatures and voltages are still safe/acceptable with a brief AVX2 stress test.

A typical behavior will resemble this:
-Idle - voltage sitting in the ~1.0v or less range, ~800mhz?
-Partial load - CPU will scale to a percentage of max frequency, and voltage will scale with it to a percentage of max load voltage, following a voltage table built into the CPU
-Full, non-AVX2 load - CPU will scale to max frequency and peak voltage
-Any time AVX2 instructions are used - CPU will automatically add ~0.10v

I think I see most people staying at or below 1.40-1.45v under AVX loads on Skylake, but we have no long-term data to support the safety (or lack thereof) of these voltages. These voltages are assumed safe because Skylake K-parts have higher stock voltages than previous generation CPUs.

 

[rgallant]

I ran a i7 2600k at 4.5 GHz (1.4v) for nearly 4 years. I did have to drop it down to 4.2 recently as my system seemed to be losing stability. I purchased a 6700k with a Gigabyte Z170X-UD5. I did just fail a stress test at 4.5 GHz and 1.25v. My temps peaked around 67C (and were usually below 60C). I will probably bump it up to 1.275v next.

1) I remember load line calibration controllers were sketchy in the past. While gaming, CPU load goes up and down often, which makes me nervous with unreliable controllers. I have 3 LLC options...auto, standard, high. Auto and standard are crap. My vcore drops all the way down to 1.1Xv when set to 1.25v. The high setting seems to do pretty good, but are there burst voltage spikes occurring that are damaging the CPU due to overshoot?

2) I never used XTU in the past. Is there a reason CPU-Z and HWinfo are reporting core voltage with what I am expecting based on BIOS settings, but XTU is reporting something like 1.37v (instead of 1.25v).

3) C states were always disabled while overclocking back in the day. Is the sentiment still the same? Will the voltage actually even drop when I have a hard coded voltage (havent wanted to test this yet while working on general stability)?

4) It seems people are using voltage much higher than mine and claiming it is safe. It makes me nervous even at 1.25v on this 14nm process. What has changed since the previous generations to make people believe 1.35v+ is safe for long term use (I am hoping to get another ~4 years out of this)?

I did not like the shoots on my 3770k @ 4.6 as it cycles on off every mil second [ low c state downclock to a highish overclock]
so I turned off c states
only left on EIST
maxed power options in win7

normal idle speed and volts to a somewhat high oc [can run @4.8 with a small volt increase]

giving me a idle speed at 3.5 1.175v per coretemp
and after ramp up to a stable 4.6 clock under load @ 1.296 v [cpuz]
not sure if your system can do this.

also I don't care about power or temps being on water

 

[TiredEngineer]

Maybe I'm misinterpreting, but you probably should not be using fixed voltage when overclocking, unless it's for a very specific reason / short term test. It wasn't even a good idea in the Sandy Bridge days.

Nearly confirmed at this point 4.5 GHz stable at manually setting vcore to 1.3v and LLC to high.

Leaving everything on auto it shot my vcore up to 1.392v for the same frequency just running the benchmark in XTU.

I am wanting to stay below 1.3v as my chip seems stable at that vcore, but I also want the reduced voltage afforded by the C states. It seems my choice is 1.3v 24/7 or C states with spikes up to ~1.4v.

 

[Yuriman]

Nearly confirmed at this point 4.5 GHz stable at manually setting vcore to 1.3v and LLC to high.

Leaving everything on auto it shot my vcore up to 1.392v for the same frequency just running the benchmark in XTU.

I am wanting to stay below 1.3v as my chip seems stable at that vcore, but I also want the reduced voltage afforded by the C states. It seems my choice is 1.3v 24/7 or C states with spikes up to ~1.4v.

I'd probably take the latter, personally.

 

[TiredEngineer]

I'd probably take the latter, personally.

Not sure how accurate package power is in hwinfo, but with all C states enabled and auto selected, it seemed like I was usually hovering around 15 watts. With a manual voltage of 1.3 and keeping the lower C states enabled (disabled the high latency C states), at idle I typically consume 19 watts.

I would guess at idle most MOSFET gates would be open, so the only voltage damage should be from leakage, and it seems like this must be pretty low based on the power consumption difference.

 

[Denithor]

I'd probably take the latter, personally.

Ditto. The only time you're going to get those "power spikes" to ~1.4V is under AVX2 load, right? And how often do you inflict those on your CPU? Very likely, only during stability testing. AVX2 just doesn't have widespread use yet...