Skylake OC not going so well.

[Deders]

Your rig is making me drool, looks a bit like an upgraded version of what I have running.

Nice setup.

:thumbsup::thumbsup:

It has the same effect on me. I should probably add the droolbucket to my sig, pretty essential when running this kind of kit.

 

[tolis626]

First things first. Uninstall Prime95. Completely. Never use it again. I'll elaborate later.

The cooler is probably fine for what you're trying to do. I'd try lowering the voltage. 1.34V seems a tad on the high side for just 4.4GHz. If you're crashing then bad luck, but I think you can get lower. As Deders said, increase your LLC (load-line-calibration) and see what other power delivery options there are in your BIOS. Also disable any power-saving features as they can mess with your stability. Generally, you want everything set for better performance/overclocking, not low power consumption.

Other than that, general good practice when overclocking is to always set your voltage to manual when trying to determine a stable overclock. Really, you can't go wrong with that, as you remove the possibility of software messing with you from the equation. Also, set your RAM speeds to the default set by the CPU. If it's DDR4, that's 2133MHz. If it's DDR3, it's 1600MHz. Only put your RAM speeds back up (Assuming you have a faster kit) after finishing with the CPU overclock. Memory can cause weird issues that are really hard to diagnose. Been there, done that, never doing it again. While you're at it, you might want to test lowering your cache clock and voltage. That can sometimes help tremendoulsly with modern Intel CPU overclocking and the performance you lose is negligible.

Install OCCT and Asus RealBench. I've found these two to be the most useful stress testing tools by far. OCCT finds most kinds of instability without risking to kill your CPU and RealBench does the same but using a really heavy, but realistic load. You might also want to try out AIDA64, it's the most complete suite of testing and monitoring software out there. But let me tell you again, never use Prime95 again. It runs your CPU in a completely unrealistic scenario that uses every subsystem at the same time at 100%. The result is that modern Intel CPUs will throw more voltage in there to keep everything stable under that crazy AVX load. Some Haswell CPUs would even increase the core voltage by 0.1V or something crazy like that. Mine would do a 0.05V increase when I had already set it to 1.315V. And the craziest thing? It will still crash most of the time. So just get rid of it and use the other ones.

If you still have issues after doing all that... Then yeah, it's the IHS. I doubt it's THAT bad though.

Hope I helped. If you want more help, I'll be around.

 

[pantsaregood]

Going to skip on switching the coolers for now. It's annoying and I'm lazy.

I'm going to pull the fans off the heatsink and let it run passive. Heat build up should be easily observable that way. Heatpipes and fins should heat up nicely.

 

[ehume]

First sentence of OP: "Upgrading from an i5-2500K running at 1.33v/4.4 GHz for 4 years stable using a Hyper 212+. Runs below 85C unless I throw it under the bus with AVX-capable LinPack."

I'd say his Hyper 212+ is a tried and true heatsink.

As I said earlier, the TIM is the most likely culprit. My 4770k ran into the same issue: it's stable, but it runs out of cooling before it runs out of stability. My 1st 4790k was better -- it ran into instability before it ran out of cooling.

So you have confirmed that Skylake is like the 1st gen Haswells, and like the Ivy Bridges before them. Good to know.

 

[Deders]

First sentence of OP: "Upgrading from an i5-2500K running at 1.33v/4.4 GHz for 4 years stable using a Hyper 212+. Runs below 85C unless I throw it under the bus with AVX-capable LinPack."

I'd say his Hyper 212+ is a tried and true heatsink.

As I said earlier, the TIM is the most likely culprit. My 4770k ran into the same issue: it's stable, but it runs out of cooling before it runs out of stability. My 1st 4790k was better -- it ran into instability before it ran out of cooling.

So you have confirmed that Skylake is like the 1st gen Haswells, and like the Ivy Bridges before them. Good to know.

Generally it is supposed to be the same as Devil's Canyon as most reviews have confirmed, there might be an issue with the application in his but like I mentioned before I also had a tried and tested cooler, 60c max with a 1GHz overclock and it wasn't enough to tame my Skylake.

 

[Zodiark1593]

Generally it is supposed to be the same as Devil's Canyon as most reviews have confirmed, there might be an issue with the application in his but like I mentioned before I also had a tried and tested cooler, 60c max with a 1GHz overclock and it wasn't enough to tame my Skylake.

If the load power use of the chip is the same, or even higher than DC, the smaller physical size of the 14nm chip may cause the same TIM used in DC to hit it's limit sooner.

I'm curious though. Was it Ivy they started using the TIM, or Haswell? Unless there is some physical reason where Solder becomes far more costly on smaller dies (or 22nm and smaller is more sensitive to heat, requiring more expensive low-temp solder), I don't quite understand the reasoning for Intel using the TIM at all.

 

[ehume]

I'm curious though. Was it Ivy they started using the TIM, or Haswell?

Ivy. That's the one
IDC delidded. Intel used solder on Sandy Bridge.

 

[Deders]

If the load power use of the chip is the same, or even higher than DC, the smaller physical size of the 14nm chip may cause the same TIM used in DC to hit it's limit sooner.

I'm curious though. Was it Ivy they started using the TIM, or Haswell? Unless there is some physical reason where Solder becomes far more costly on smaller dies (or 22nm and smaller is more sensitive to heat, requiring more expensive low-temp solder), I don't quite understand the reasoning for Intel using the TIM at all.

I can only think that Intel wanted to limit people's overclocking. Seeing as pretty much any decent tim shows a 10-15c improvement after de-lidding, they must have gone to extra special lengths to make there's less conductive.

As for limits, people are getting 4.6 easily, 4.7 is quite common and 4.8 is possible. All with higher voltages than Haswell. Like I said before I have managed 4.7@1.425v so far and the max temp is 77c with Realbench stress tester.

Am still working on the OC.

 

[Kenmitch]

I can only think that Intel wanted to limit people's overclocking. Seeing as pretty much any decent tim shows a 10-15c improvement after de-lidding, they must have gone to extra special lengths to make there's less conductive.

As for limits, people are getting 4.6 easily, 4.7 is quite common and 4.8 is possible. All with higher voltages than Haswell. Like I said before I have managed 4.7@1.425v so far and the max temp is 77c with Realbench stress tester.

Am still working on the OC.

It's not really the TIM that is the issue. It's more the bonding agent they use to secure the heatspreader creates too much of a gap. Once the compound is removed the gap tightens up between the core and the heatspreader. Sure results will very dependant on your choice of TIM but the largest gain is from the gap being tighter.

 

[tolis626]

I can only think that Intel wanted to limit people's overclocking. Seeing as pretty much any decent tim shows a 10-15c improvement after de-lidding, they must have gone to extra special lengths to make there's less conductive.

As for limits, people are getting 4.6 easily, 4.7 is quite common and 4.8 is possible. All with higher voltages than Haswell. Like I said before I have managed 4.7@1.425v so far and the max temp is 77c with Realbench stress tester.

Am still working on the OC.

1.425V? Well, that can't be good. My 4790k sometimes boosts its voltage to 1.325V under heavy load and I crap my pants. If I ever saw 1.4+... I'd probably use a fire extinguisher or something.

 

[Deders]

1.425V? Well, that can't be good. My 4790k sometimes boosts its voltage to 1.325V under heavy load and I crap my pants. If I ever saw 1.4+... I'd probably use a fire extinguisher or something.

Recommended max voltages for Skylake are 1.42v for heavy AVX and FMA Prime type loads, and 1.45v for gaming and X264 encoding type loads.

iirc it defaults to about 1.325 at stock speeds.

 

[richierich1212]

As for limits, people are getting 4.6 easily, 4.7 is quite common and 4.8 is possible. All with higher voltages than Haswell. Like I said before I have managed 4.7@1.425v so far and the max temp is 77c with Realbench stress tester.

Not a good thing for 14nm.

 

[Kenmitch]

1.425V? Well, that can't be good. My 4790k sometimes boosts its voltage to 1.325V under heavy load and I crap my pants. If I ever saw 1.4+... I'd probably use a fire extinguisher or something.

Does sound high to me also for 24/7 use at least.

On another note I've gone MUCH higher in the past

That chip was a beast to say the least. Not a popular chip but it's still the #1 spot on hwbot.

 

[tolis626]

Recommended max voltages for Skylake are 1.42v for heavy AVX and FMA Prime type loads, and 1.45v for gaming and X264 encoding type loads.

iirc it defaults to about 1.325 at stock speeds.

Maybe they've locked vcore and you're playing with vccin. That would mean that your real applied vcore would be in the 1.2-1.3V range. 1.425V would result in some really massive power consumption and, of course, to really high temps. Check it with a kill-a-watt if you are curioous. I'd buy a 6700k to test, but financials are still tight.

 

[Deders]

We've got Vcore and VID. If I just use an offset voltage of -0.02 @4.6GHz then the Vcore is much higher than the VID, that's with or without LLC level 5.

If I use Adaptive voltage and set it to max out at 1.425 and switch up the LLC to 6 then they are both the same.

Just tested with Killawatt.

89W idle
225W Prime95 v2.66 blend.

If you take the PSU's efficiency into account that works out at:

80.1W Idle
207W load

So a difference of 126.9.

AI suite measures 11.9W Idle and 104W load. There are also 6 fans to take into acocunt. 3 of them go from about half speed to max, the other 3 from about 80% to max. I would presume the fans, memory, and motherboard would account for the 22w difference?

 

[tolis626]

We've got Vcore and VID. If I just use an offset voltage of -0.02 @4.6GHz then the Vcore is much higher than the VID, that's with or without LLC level 5.

If I use Adaptive voltage and set it to max out at 1.425 and switch up the LLC to 6 then they are both the same.

Just tested with Killawatt.

89W idle
225W Prime95 v2.66 blend.

If you take the PSU's efficiency into account that works out at:

80.1W Idle
207W load

So a difference of 126.9.

AI suite measures 11.9W Idle and 104W load. There are also 6 fans to take into acocunt. 3 of them go from about half speed to max, the other 3 from about 80% to max. I would presume the fans, memory, and motherboard would account for the 22w difference?

That 22W figure sounds about right. Maybe a little high, but definitely not off the mark.

At 100-150W for the CPU, I'd think there is no way that it's runnig at 1.425V. That would mean 200+W for the CPU alone, if I'm correct. Hell, I get between 200 and 250W total system consumption on my 4790k at 1.305V during Prime.

What I think is more likely is that motherboard manufacturers just name the setting vcore. I don't think you're setting your vcore to 1.425V, but rather the vccin (Input voltage for the CPU package) and then the vcore is taken from that. I think a vcore of between 1.25-1.35V is where it's at, and probably at the lower end of that.

Unless always there's something completely different about Skylake and I'm dead wrong. That's a possibility.

 

[Deders]

That 22W figure sounds about right. Maybe a little high, but definitely not off the mark.

At 100-150W for the CPU, I'd think there is no way that it's running at 1.425V. That would mean 200+W for the CPU alone, if I'm correct. Hell, I get between 200 and 250W total system consumption on my 4790k at 1.305V during Prime.

What I think is more likely is that motherboard manufacturers just name the setting vcore. I don't think you're setting your vcore to 1.425V, but rather the vccin (Input voltage for the CPU package) and then the vcore is taken from that. I think a vcore of between 1.25-1.35V is where it's at, and probably at the lower end of that.

Unless always there's something completely different about Skylake and I'm dead wrong. That's a possibility.

What would the VID be for Haswell?

I'm reading the core VID from HWinfo which has to be calibrated for each individual motherboard. Raja from Asus has confirmed to me that for the z170 deluxe, the voltages in HWinfo are correct as tested with a multimeter.

The VID is in the CPU section and has 1 reading for each core. so 4 total.

In the motherboard section there is a single Vcore reading that corresponds exactly with the Vcore reading in AI suite. Raja has confirmed he has measured this to be correct.

If I leave LLC at default or set it to 5, then the Vcore is significantly higher than the VID. I have LLC set to 6 for my current 4.6 overclock and the Vcore is slightly higher at 1.408 than the VID at 1.404. Both readings are rock solid, no variation at load.

Strangly the Killawatt meter still reads 225 @Prime load, but the CPU reading in AI suite reads 96W. Not sure what is happening there.

I think I'm going to stick with 4.6GHz, maybe see what the BCLK can improve on. The DPC latency is much better at this speed than at 4.7, and the extra voltage and additional latency just isn't worth it for 100MHz.

I'm sure I can lower the voltage from 1.4 and still have it stable.

 

[ClockHound]

There's no FIVR on skylake-s...so it's more like uh...whoa..Sandy?

Here's the 6700k datasheet link: http://www.intel.com/content/www/us/en/processors/core/6th-gen-core-family-desktop-brief.html

Page 115 of the pdf states max Vcc of 1.52V. Definitely more like Sandy.

 

[Absolute0]

bad TIM

1) delid and repaste IHS
too daunting for most people but I did this on a few AMD Opterons 1 decade ago and each dropped something like 12c under load mounting on the bare core. I imagine your TIM is so crappy that you'd see a much better improvement
2) better cooler... cooling capacity of a 212 is pretty meh to say the least. the hyper 212 is not an impressive cooler if you're gunning for 4.5+ Ghz. seems this can only help so much tho... hence #1
3) sell/buy a new one - if 30+ dollar loss is worth a few hundred Mhz and some buyer wants to run stock
4) back off to 4.0-4.2 or so Ghz and see how low you can get the Vcore and temps should be OK

 

[Deders]

I should probably mention that my statement from before about Raja confirming HWmon voltage, It was actually Aidi, and it was specifically the Deluxe model that he tested with a multimeter.

I confused HWmon and HWinfo and have been reading from HWinfo which has the same readings for me as both Aida and AIsuite.

Results may vary for different motherboards as the program needs to be individually calibrated for each board model.