Inside Intel's Secret Overclocking Lab: The Tools and Team Pushing CPUs to New Limits

[jimmysmitty]

This is where I'm at. Under 80 and less than 1.4v, for any system you care to have last beyond the next gen release.

Yes but that does also change depending on the uArch design and process node. Older AMD CPUs had a max temp of 62.5c if I remember.

 

[Karadjgne]

Those FX cpus were odd, the 62°C was 'in' core temp, which only got measured on a couple of engineering samples. Why AMD made the decision not to include any kind of thermal read on the production models is beyond my comprehension, I chalk it up to a serious bone-head move. But IHS and core temps aren't the same thing, nor the socket temps, so any reading from anywhere was inaccurate to say the least. And adding temp software poll times, it gets worse. I saw plenty of FX running in mid 70's (according to standard software) with no seeming adverse effects.

Overdrive thermal margins were derived partly from its algorithm reading vcore, and it wasn't an exact differential, you could loose 10° tm just moving a mouse cursor.

So glad Ryzen are not related to FX.

 

[jimmysmitty]

That and Phenom II. Deneb CPUs had a max of 62c. But again it all depends on the process node, uArch design etc what temperature they want you to look at and what voltage it can handle. My Q6600 had a max VID of 1.5V per Intels ARK. I wouldn't dare put 1.5V on any modern CPU.

 

[Karadjgne]

Yeah. For the longest time 3rd gen Intel limit was 1.4v, I searched multiple websites and forums looking for batch numbers that could get me 4.9-5.0GHz at under 1.4v. And then, of course, I spent a lot of time reading the many iterations of Computronix' pages and that safe level for 3rd gen is 1.3v. And yet if I look at my bios, 1.4v is still white, 1.41v-1.5v is yellow, 1.51v+ is red. Which I take as safe, slight damage over time, and serious damage period.

So I dunno, it's not an exact science. Mezoxin stated that you should be able to predict things, given certain stats and definitions, but when there's no physical correlation to specific, exacting numbers, even the math ends up being nothing more than variables. It's why Intel sets such a high stock vcore to begin with, just to cover all the variables of the cpus in that line, because even they can't predict what's needed exactly, just a general range.

 

[jimmysmitty]

Most of it will depend on the process design and what it can handle. The uArch might have some decision on the voltage but I would say uArch will determine temperatures more than voltage.

I would be that voltages will drop for Intels 10nm and then for their 7nm. Refinements may allow higher voltages but I would think they would lower voltages while allowing same or higher clock speeds, much like the Q6600 G0 that allowed for 3GHz on stock or less voltages for pretty much every sample or like the 9900KS that seems to be able to clock better while using less voltage and lowering temps and power.

Its a very different world I will say though. All the new tech in CPUs that control voltages, boosts and temps is well beyond what CPUs had when I started.

 

[Karadjgne]

Well yeah, the KS isn't anything but a K that's been tweaked some, been plenty of feedback from regular users on every kind of mobo made that'll handle it. Same architecture, same everything else. No different to what amd did with the 8350/9590, Intel just did it better.

And I started out on a Vic20, so totally get where the 'ancient designs' feeling comes from.

 

[Darkbreeze]

Yes but that does also change depending on the uArch design and process node. Older AMD CPUs had a max temp of 62.5c if I remember.

Well, we pretty clearly were talking about 14nm Intel. Obviously, older 22nm or AMD products are a different set of numbers.

 

[CompuTronix]

... For the longest time 3rd gen Intel limit was 1.4v ... I spent a lot of time reading the many iterations of Computronix' pages and that safe level for 3rd gen is 1.3v ...

As new information becomes available, many iterations are needed to keep pace with evolving technology. If you haven't seen more recent iterations, you might want to have a fresh look at Section 8 that shows the Vt Shift Degradation Curves, which for 22nm, does not support Vcore above 1.325. Moreover, without delidding and / or high-end cooling, 22nm flagship K's becomes exceedingly difficult to cool as Vcore approaches 1.3, which is also true for 14nm as Vcore approaches 1.4.

... why Intel sets such a high stock vcore to begin with, just to cover all the variables of the cpus in that line, because even they can't predict what's needed exactly, just a general range.

Although 3rd and 4th gen share the same 22nm uArch, they differ in that 4th gen has FIVR, which influences TDP. They also differ in VID, which at a glance, many users often misinterpret for Vcore.

Where 3rd gen max VID is 1.52, 4th Gen VID shows 1.82 on page 96. But if you follow the tables to page 98, VID ends at a spark-n-smoke value of 3.04. In addition to max VID, 1st gen datasheets include a table for "Processor Absolute Minimum and Maximum Ratings" which does not appear in subsequent datasheets.

Concerning 14nm, the datasheets for 8th & 9th gen show 1.52 on page 117. In contrast, the Degradation Curve for 14nm does not support Vcore above 1.425, which agrees with Intel's OC lab engineers. Yet as common sense prevails, we known that 1.52 is well beyond the realm of longevity for a daily driver. We also know that certain specs are worded somewhat ambiguously, which leaves them open to interpretation.

Overclocking, Vcore and Core temperatures are highly controversial and hotly debated topics. Regardless of theoretical science, practical implementations and empirical insights, there are always those who will cite unknown variables, rare exceptions and wishful speculation to support a position where sensible recommendations are either too extreme or too conservative.

 

[Beetlebox]

At home, the lab engineers consider a load temperature above 80C to be a red alert, meaning that's the no-fly zone, but temps that remain steady in the mid-70’s are considered safe.

Please would you clarify if this is core or package or Intel standard center of IHS specification temperature.

Also

It's the person that sets their system up at elevated voltages and just leaves it there 24/7 [static overclock], that's the person that is going to burn that system out faster than someone who uses the normal turbo algorithms to do their overclocking so that when the system is idle your frequency drops and your voltage drops with it.

Since halt states (C1) stop the core clock (frequency drops to zero) and C6 power gates the core off while C0 races to idle with static voltage and max clock when active then how does this burn out the system faster unless it's something else such as LLC failing when using fixed voltage?

 

[Karadjgne]

Static OC means c-states are disabled. You get a processor that even in windows idle environment runs the same clocks as under loads.

 

[Beetlebox]

Static OC means c-states are disabled.

I didn't know that. I'd always taken it to mean highest p-state with fixed voltage. So people are also disabling use of core state C1, which needs to be done via the OS as there is no way of doing it via BIOS? I can understand some of the other C-States not wanting to be used but other than for testing purposes not C1.

 

[Darkbreeze]

To the best of my knowledge, a "static overclock" is all "available" C states disabled in the BIOS, and the Windows power plan set to performance with 100/100 min/max CPU power state configuration, as well as Intel speed step, Intel speed shift, or AMD cool N quiet, disabled. If any of those conditions are not met, then it is not a static overclock by any definition I've ever known.

 

[Karadjgne]

You can disable all, C1, C1e, in my bios, also can disable sleep in windows power settings, turn cpu usage to min 100% etc.

Not that I'd ever advise such, newer gen cpus aren't as affected by speed stepping as the older gens are. My OC is a turbo max OC, so c-states 1,1e, 2 and 3 are still active, hibernation is gone, hiberfil.sys deleted/disabled totally.

 

[Darkbreeze]

Right. While we can define a "static overclock", it doesn't mean that most, or any, of us are recommending that anybody should configure that way for their daily drivers. For competitive or lab use machines, do what you want. For the rig in your bedroom, it's probably not wise to configure for static behavior. I always recommend leaving the C-states, Speed step, Cool N Quiet, and an 8-10% min CPU power state in the advanced power plan options when overclocking your regular use machine.

 

[Beetlebox]

You can disable all, C1, C1e, in my bios

How things have changed. Newest system I've used is Broadwell so behind the times. Typically package C-States such as C1E could be disabled by hardware register but core states are a different matter, especially C1 as it is the halt (hlt) x86 instruction. The Core C-States therefore require announcement through ACPI tables unless that's changed too. For instance if I disable C-States in my MSI Z97 BIOS then Windows will not use them except for C0 and CC1 however Fedora 30 will ignore the ACPI table and still use them unless the default boot configuration is modified. Do you mind me asking which BIOS you are using, just curious how it goes about it?

 

[Karadjgne]

I'm using a MSI Z77 mpower Big Bang edition. Basically the Msi version of the Asus Z77 ROG Hero. So the bios is quite extensive, it's got stuff in it I've left at default because I still haven't figured out its purpose, the manual isn't very explicit and online help next to useless. But it works, so I leave it. There's also a software that came with it that has even more switches, that when applied will change bios settings either on the fly or permanently, so get adopted on reboot.

 

[Beetlebox]

Thanks for that, I downloaded 17.12 from https://us.msi.com/Motherboard/support/Z77_MPOWER.html

I see the following options

For Package C-States
Limits for C0,C1,C3,C6,No Limit,Auto (MSR configured)
C1E (MSR Configured)

For Core C-States
C3, C6, C7 (Report only via ACPI), note also that ACPI will refer to these as something else such as C1, C2 which can be confusing.

As far as Intel C-State C1 is concerned Core C-State C1 (CC1) doesn't have an option to be disabled by that BIOS, package C-State C1 (PC1) does by limiting however only C1E actually does anything.

An excerpt from 3rd gen datasheet that might be helpful

For a bit of trivia, if using Ivy-Bridge one can monitor the transitions to/from active/idle states which go through the LFM (Lowest Frequency Multiplier/Mode) by setting the MPO (Maximum Performance Override/Limit) to 8x IIRC (my memory is terrible and getting worse). The LFM is 16x and one can see the core momentarily going to 16x before being forced to 8x.

 

[Beetlebox]

Anyone knows how to made contact with them? cause I fould a big bug on 10th corex chip about adaptive mode overclocking

You can post on Intel Communities under the processor section but need patience as progress can take a while.