How do you overclock?

[biostud]

Funny. I thought I'd been looking for such a thread in past weeks and months: it escaped me. That's the sort of information sharing I'd been hoping for, as I do EARLY planning for an E-build next year.

Lepton87 had a thread here and at Cases/Cooling on his 5820K build, and we discussed this.

The AVX2 features shoot the temps through the roof with standard and commonly used stress-tests.

For some reason, I stumbled back into using OCCT. I'd had it installed three years ago, but the free version has a 1-year use limit, and you have to download the new improved OCCT to get it back.

The sparse info offered by the OCCT author (who could be French or Russian but who knows? and what does it matter?) -- states that the "OCCT:CPU" tab will catch errors faster in OC settings than will the Linpack tab (which is essentially the same as IBT or LinX). He notes that it distinguishes thermal stress from "other stress" which would turn up those errors.

The OCCT:CPU test seems to give temperatures peaking at maybe 4+C below the LinPack option.

I just ran a short test of OCCT and the temps does peak at 85C, but not at all cores and only for a short time. Most are below 80C. In linX my CPU temperature reaches 95C and stays there. But I'll do a proper run later.

 

[BonzaiDuck]

I just ran a short test of OCCT and the temps does peak at 85C, but not at all cores and only for a short time. Most are below 80C. In linX my CPU temperature reaches 95C and stays there. But I'll do a proper run later.

This type of discussion is going on at this very moment on another thread or another forum. I had read -- today -- a post by IDontCare on these issues, probably made within the last 12 hours or so. It points to the issues of stability testing and temperatures, and I pretty much agree with him.

But if the author of OCCT is providing accurate information, I'm willing to assume that we don't need to push the thermals to their very, very highest to discover whether or not an OC is stable. I'm even willing to consider that tests forcing the CPU to throttle aren't likely to damage anything, but complicate the process and the issue. Here's why I tentatively say that.

These new 22nm processors, their IHS fabrication process with the TIM, the new instruction-set additions, and the TDP's of processors like the E chips almost seem to be designed by Intel with expectations of higher thermal effects. Even for the 4790K, TDP 95W if I'm not mistaken, gets very hot.

I'm a statistician (or was), but I attended my freshman/sophomore physics course covering electrical concepts back in 1966/67. Not only do I "not remember," but it was rudimentary. IDC, on the other hand -- knows his S***.

So unless someone disagrees and proves otherwise, here's what I understand.

There are calibration errors for the core sensors of these processors. Maybe four years ago, they were estimated at +/- 6C. You can use RealTemp (or CoreTemp?) to calibrate them yourself within the software, but I don't think it's really recommended.

Testing offers simultaneous "small sample" and "large sample" observations: with a quad-core or hexa-core, you have a small sample of "cores," and with HWMonitor, Aida64, CoreTemp/RealTemp and even ASUS AI Suite "Sensor Recorder," you can build a file of observations taken every second -- whatever interval you choose to set.

So for the sensors themselves, you can take an average (central tendency) and range (variation), even from the "maximum values" that get snagged in the recording. You can also take averages by core over an hour's stress-testing, and you can average those averages across the cores.

So with the cores, which are likely all "out of whack" to some degree -- some below and some above -- the average is the best measure, whether of maximums or "average-of-averages."

Suppose the highest core reading rivets your attention? It is much less likely that all the core calibration errors are in one direction, and more likely there is a distribution above and below.

A persistent problem of information-sharing on these forums and especially in threads like the one you linked for which I'm very grateful, is that folks apply their own choices of how these things are read and measured. Another example of that is "what is meant by load voltage."

Since we've come far in the ability to OC with EIST and C1E enabled, I've come to realize that there are two values of voltage you may see during a stress-test besides the EIST minimum. There is a drooped load voltage (probably the one that counts the most), and an unloaded "turbo" voltage -- captured by monitoring software momentarily after a stress-test ends and before the processor settles down to EIST speed and volts. So often, it is uncertain what someone is reporting, and that would be true even with EIST/C1E disabled to give a fixed voltage setting. LLC settings also affect this, so it is possible that there's something like "negative droop." Worse, it's possible that VCORE could exceed VID at some point with high LLC. I always make sure my LLC setting gives me 20mV of droop.

But what I'm trying to say with this explosive diarrhea of words, is that even if we were to agree to some standard of reporting, recording and so forth, a lot of folks might not follow to the letter. We're always going to find these confusions and ambiguities.

 

[biostud]

This type of discussion is going on at this very moment on another thread or another forum. I had read -- today -- a post by IDontCare on these issues, probably made within the last 12 hours or so. It points to the issues of stability testing and temperatures, and I pretty much agree with him.

But if the author of OCCT is providing accurate information, I'm willing to assume that we don't need to push the thermals to their very, very highest to discover whether or not an OC is stable. I'm even willing to consider that tests forcing the CPU to throttle aren't likely to damage anything, but complicate the process and the issue. Here's why I tentatively say that.

These new 22nm processors, their IHS fabrication process with the TIM, the new instruction-set additions, and the TDP's of processors like the E chips almost seem to be designed by Intel with expectations of higher thermal effects. Even for the 4790K, TDP 95W if I'm not mistaken, gets very hot.

I'm a statistician (or was), but I attended my freshman/sophomore physics course covering electrical concepts back in 1966/67. Not only do I "not remember," but it was rudimentary. IDC, on the other hand -- knows his S***.

So unless someone disagrees and proves otherwise, here's what I understand.

There are calibration errors for the core sensors of these processors. Maybe four years ago, they were estimated at +/- 6C. You can use RealTemp (or CoreTemp?) to calibrate them yourself within the software, but I don't think it's really recommended.

Testing offers simultaneous "small sample" and "large sample" observations: with a quad-core or hexa-core, you have a small sample of "cores," and with HWMonitor, Aida64, CoreTemp/RealTemp and even ASUS AI Suite "Sensor Recorder," you can build a file of observations taken every second -- whatever interval you choose to set.

So for the sensors themselves, you can take an average (central tendency) and range (variation), even from the "maximum values" that get snagged in the recording. You can also take averages by core over an hour's stress-testing, and you can average those averages across the cores.

So with the cores, which are likely all "out of whack" to some degree -- some below and some above -- the average is the best measure, whether of maximums or "average-of-averages."

Suppose the highest core reading rivets your attention? It is much less likely that all the core calibration errors are in one direction, and more likely there is a distribution above and below.

A persistent problem of information-sharing on these forums and especially in threads like the one you linked for which I'm very grateful, is that folks apply their own choices of how these things are read and measured. Another example of that is "what is meant by load voltage."

Since we've come far in the ability to OC with EIST and C1E enabled, I've come to realize that there are two values of voltage you may see during a stress-test besides the EIST minimum. There is a drooped load voltage (probably the one that counts the most), and an unloaded "turbo" voltage -- captured by monitoring software momentarily after a stress-test ends and before the processor settles down to EIST speed and volts. So often, it is uncertain what someone is reporting, and that would be true even with EIST/C1E disabled to give a fixed voltage setting. LLC settings also affect this, so it is possible that there's something like "negative droop." Worse, it's possible that VCORE could exceed VID at some point with high LLC. I always make sure my LLC setting gives me 20mV of droop.

But what I'm trying to say with this explosive diarrhea of words, is that even if we were to agree to some standard of reporting, recording and so forth, a lot of folks might not follow to the letter. We're always going to find these confusions and ambiguities.

Yeah, I don't judge other people on how to do the "correct" stability testing. I'm, just looking for input for "best" practice and a way to do it as easily as possible.

 

[Deders]

But if the author of OCCT is providing accurate information, I'm willing to assume that we don't need to push the thermals to their very, very highest to discover whether or not an OC is stable. I'm even willing to consider that tests forcing the CPU to throttle aren't likely to damage anything, but complicate the process and the issue.

I'm not disagreeing with you but surely the temperature will affect how the electrons flow, it it possible that an overclock might be stable at low temps but not as stable at higher?

Also I read on the Overclockers forums that every 10c of heat added halves the lifespan of chips and components. Can you remember anything from your course that would back this up?

 

[BonzaiDuck]

I'm not disagreeing with you but surely the temperature will affect how the electrons flow, it it possible that an overclock might be stable at low temps but not as stable at higher?

Also I read on the Overclockers forums that every 10c of heat added halves the lifespan of chips and components. Can you remember anything from your course that would back this up?

I wouldn't remember anything from any course. If I said something about a "course," it was about taking Physics in college 45 years ago.

Whether there is some rule-of-thumb about the effects of temperature, I would not know. The chip-makers would know about the thermal effects on silicon. Intel posts the TCASE temperature spec for the Haswell-E processors at 66C. For the 4790K, it is about 75C. For my 2600K -- 73C.

The history of "TCASE" suggests it can be as much as 10C lower than the core average.

so when we stress-test, those temperature can peak to anywhere between 70C and 95C depending on cooling and the degree of overclocking and the VCORE voltage. Obviously, these chips will not generate those temperatures when running at their stock settings, even though stress-tests of this or that CPU at stock settings can put those temperatures into the 60s -- even 70s with the relatively-new 22nm lithography of the Ivy Bridge and Haswell processor models.

And -- when we stress-test, it is to prove the stability of the OC settings.

And here's where it gets a bit dicey.

At some point, higher temperatures require higher voltage to maintain stability for a high-clock setting. It would seem that this could be a vicious circle: a voltage setting generates high temperatures under 100% usage or stress; the temperature requires higher voltage to maintain stability; the higher voltage increases temperature, etc. Just intuitively (and I don't quite trust my intuition), this would seem consistent with exponentially rising voltage necessary to get a higher clock setting. Or it seems consistent with the "wall" which is the upward-sloping part of a parabolic or exponential curve.

So yes -- it is likely an overclock might be more stable at lower temperatures. A member named IDontCare proves with his data and graphs that lowering a processor temperature by maybe 20C by de-lidding the processor cap may mean that something like 20mV less voltage is necessary to hold the stable overclock speed. (It could be 2mV, but I thought I recalled it was 20. IDC can correct me.)

However, lowering temperatures more and more by conventional means may not reduce voltage requirement much. We have also seen in benchmarking competitions that using liquid nitrogen as a cooling approach allows for clocks moving toward maybe 6Ghz, but the voltage setting to get there (as I recall) was also moving toward 1.6V. This -- when the VID limit of the processor is only 1.5.

 

[ninaholic37]

I underclock to 800mhz by turning off Speedstep in BIOS. This way the fan doesn't run at full speed when booting into MS-DOS.

edit: I found another way around it! Download this then add "sstep 6 18" (or whatever it says your lowest value is) to your autoexec.

 

[BonzaiDuck]

I underclock to 800mhz by turning off Speedstep in BIOS. This way the fan doesn't run at full speed when booting into MS-DOS.

edit: I found another way around it! Download this then add "sstep 6 18" (or whatever it says your lowest value is) to your autoexec.

Don't you mean "by turning it ON?"

And -- hey! -- MS-DOS? You're being funny, aren't you?:awe:

 

[mcbaes72]

3770K @ 4.2 with 1.14v = sweet spot.

On air, highest stable OC was 4.7 with 1.35v. Temps waaay too high though.

 

[CropDuster]

I shoot for the sweet spot. Six core 3.8ghz @ 1.20 under load, idles at .96 and 25*C. It will run happily at 4.2 as well but I don't think the extra voltage, heat and higher idle is worth it.

 

[BonzaiDuck]

3770K @ 4.2 with 1.14v = sweet spot.

On air, highest stable OC was 4.7 with 1.35v. Temps waaay too high though.

Did you create profiles in 100Mhz increments? And what are the temperatures under stress-test with that voltage and speed? With that processor, I might try pushing it as far as a load voltage of 1.28 to 1.30V would carry me. IF -- the average-of-core-maximums stayed below 80C.

But then, there were reasons folks took to de-lidding those IB's.

 

[MongGrel]

4.3 on my Westmere X5650 myself, works well.

 

[BonzaiDuck]

4.3 on my Westmere X5650 myself, works well.

I'd never thought nor attempted to use a Xeon. But since they were meant for either workstation or server applications, I don't see why not.

So . . . Socket-1366?

 

[PhIlLy ChEeSe]

BonzaiDuck,
I miss me 2500K and my maximus iv gene, here's a link

http://hwbot.org/submission/2234142_philly_cheese_steak_cpu_frequency_core_i5_2500k_5544_mhz

I had a 2600K but it didn't clock as high, I have had a 3570K, a 3770K, 4790K. None came close to that 2500K, if it was the board or what but it screamed. I think it still would have went higher, but i sold it and the board just to up grade ugh!:| PS ive had your board, its a great OC board!

A list of my CPU'S

My mother boards, not all of them but close

 

[Blitzvogel]

I'm the balls to the wall type, but only after I've spent time looking at how others have done with the same hardware. Sure, not every chip will have the exact same characteristics within the same bin, but I figure it's a starting point.

It's been a while since I played with any OCing, but just today I set my Phenom II x4 to 3.5 GHz @ about 1.43v. It's been stable (I think) so far while I've been playing the original Crysis the past couple hours. I don't do the crazy multi-hour burn in tests others do. I just keep an eye on the temps.

 

[MongGrel]

I'd never thought nor attempted to use a Xeon. But since they were meant for either workstation or server applications, I don't see why not.

So . . . Socket-1366?

Yeah, I've been hanging out on the L5639 thread with Burpo and friends a long time now.

http://forums.anandtech.com/showthread.php?t=2335636&highlight=l5639&page=40