Looking for "fine tuning" advice on OC - i5 14600KF

[QcRMoD]

Hi everyone,

Would just like to look into what you think my next best move is.

The MB is ASRock Z790 Pro RS.

Totally out of my best expectations, it seems 5.7GHz on All P-Cores is "stable enough" to warrant tinkering around.

By "stable enough", I mean I'm running the OC in Adaptive Mode, so it's not purist override fixed Vcore/Ratio. That probably wouldn't go so well since I'm licking at the thermal throttling door as it goes, but keep well below it 99% of the time: HWInfo max registered temps of 92*C running OCCT CPU test and LinPack 2021 test for an hour each. No WHEA error, no problems otherwise encountered so far. The HWInfo average temps around 80*C, by eye "average higher temps" keep below 90*C.

Within these parameters/simplistic settings below, 5.9GHz doesn't boot properly and 5.8GHz gives a WHEA error within seconds with the LinPack test. Just to give the idea that I am near the limit and I do not intend to run a fixed Turbo Vcore ever anyway, and I believe there's little I can do to stabilize 5.8GHz that won't thermal throttle. Otherwise than the OC I have a fan profile that is noiseless below 80*C which I figure will be a rare occurrence in Real World use so I am also happy on that side and do not intend to make it more aggressive. More precise specs are in my System Specs if you need them.

BIOS Settings:

XMP Profile for RAM (and also a fixed frequency for 6000MHz since otherwise the OS would boot in 4800MHz even with the XMP profile loaded)
CPU Cooler Type: 120-140mm Liquid Cooler (the cooler is Air but it works there)
CPU Vcore Compensation Level 1 (not from me, comes with the XMP profile)
LLC Level 3 (by XMP it was level 4, and level 2 is too aggressive to use so close to thermal throttling in bench)
Long Duration Power Limit 140W (could probably lower that but is there a point?)
Short Duration Power Limit 170W (could probably increase to 180W but no further)
Short Duration maintained is on Auto (BIOS sets it to "56")
CPU Core/Cache Offset Mode -0.15 (gives just below 1.3V in LinPack and I guess I could do -0.10 as a last stabilization resort but not to help with P-Cores OC vs temps I believe)
PCore/ECore/Ring VF/L2 Offset -0.2 (probably useless but I liked the idea of as little voltage as possible in non Turbo clocks)
CPU PCore Ratio is set to All Cores @ 57
CPU ECore Ratio is set to All Cores @ 40 (I'm now testing with 41, stable so far, will continue up until something happens but will stop where it stops, no point tinkering around THAT much)
And obviously Undervolt Protection is Disabled.

That's it.

Of course - further testing will be done when I'll hit the ECore threshold with upping the Cache/Ring, and more stability tests will ensue to be fully sure of everything... that is obvious as well.

But do you think it's worth it to pursue a Per Core OC, or are there settings worth trying to perhaps stabilize 5.8GHz on All Cores and wherever I'll hit a wall with ECores and Cache/Ring?

RAM tinkering doesn't seem so obvious if I'm so close to thermal throttling and possibly further to be proved instability I might encounter and to which I'd rather keep a leeway to the 5.7GHz PCores.


Thanks for perusing this and give more experienced feedback with these balls-to-the-wall settings.

Quick update:

ECores won’t go over 42 doing a quick test right now with that.

Further update:

ECores at 4.2 ends up with multiple WHEA errors in Linpack.

Won’t even suck into the +100 MHz then. It’ll be 4GHz which has proven successful in multiple testings. I can’t care for ECores if they can run the Intel specs.

So much little overhead into making sure the 5.7GHz for PCores can be stabilized if ever I run into something day to day.

 

[Evelynn]

I recommend disabling the E-cores, overclocking the p-cores by themselves (at your target of 5.7 GHz), and seeing what the load voltage is for stability. For instance, my CPU needs 1.309v @ 5.5 GHz p-cores only.

The problem with re-enabling the e-cores is they create a vdroop, not so much the heat unless you're thermally limited. They pull a lot of power (8 e-cores is around 75W depending on the workload).

In my case, I have set:
5.5 GHz @ 1.46v (p-cores only) (vdroop to 1.309v)
5.5 GHz @ 1.46v with e-cores enabled causes a vdroop down to 1.267v (unstable)
5.5 GHz @ 1.49v with e-cores enabled causes a vdroop down to the correct 1.309v

Therefore I have to either add one tick of LLC (not recommended) or simply up the idle voltage to 1.49v to achieve 1.309v under load. This is the easiest and most efficient way to find stability on 12th/13th/14th gen, or so I've personally found.

I recommend using fixed voltage, a weak LLC setting (~150mV to ~180mV of a droop), and not pushing the ring past 48x as 13th & 14th generation uses a large die - causing the ring to lose stability more easily. You are not going to damage your system running 1.45 ~ 1.5v idle so long as your vdroop is large.

You did not list what the load voltage is for 5.7 GHz on your CPU, so I recommend finding out what the load voltage for stability is.
Setting your vCore to 1.45 ~ 1.50 fixed alongside a weak LLC.
And not pushing more than 1.32v through the chip under load. Any higher causes thermal hotspots due to the density of these chips.

The reason why a weak LLC is better is due to the VRM being given more voltage space to adjust. More voltage = more time for the VRM to compensate = better stability. Couple that with a fixed voltage and transients are far less of an issue for your system.

As for WHEA errors, I recommend disabling the following:
Link State Power Management under Edit Power Plan (Windows) > advanced settings > PCI Express
and: Native ACPI PCI-e support in your BIOS

If you google: ACPI error z690 or Z790
you'll receive a lot of forum threads discussing WHEA errors related to the ACPI functionality. It's recommended to disable it.

You can also set the Core PLL voltage within the BIOS to 1.10v and the CPU Input Voltage to 1.88v. Those both assist with achieving stability at the wall of your CPU's clocks.

Now, speaking personally - I tend to only aim for Cinebench R23 stability for three runs. After I'm stable at that point, I set a wattage cap that is 20% lower than what R23 requires. So for example, if R23 consumes 250 watts, I set a PL2 of 200W. This all but guarantees stability in everything I personally use, and if I encounter a crash, I simply up the voltage by a measly 0.010 mV.

For your efficiency cores, if you want to achieve 4.2 ~ 4.4 GHz, you'll need to adjust the L2 Cache Voltage. Do not go above 1.30v on the L2 Cache Voltage. Try to aim for 1.25v. You can either do adaptive or fixed: I recommend fixed voltage since every BIOS handles the V/F scaling on the cache differently.

 

[QcRMoD]

I recommend disabling the E-cores, overclocking the p-cores by themselves (at your target of 5.7 GHz), and seeing what the load voltage is for stability. For instance, my CPU needs 1.309v @ 5.5 GHz p-cores only.

The problem with re-enabling the e-cores is they create a vdroop, not so much the heat unless you're thermally limited. They pull a lot of power (8 e-cores is around 75W depending on the workload).

In my case, I have set:
5.5 GHz @ 1.46v (p-cores only) (vdroop to 1.309v)
5.5 GHz @ 1.46v with e-cores enabled causes a vdroop down to 1.267v (unstable)
5.5 GHz @ 1.49v with e-cores enabled causes a vdroop down to the correct 1.309v

Therefore I have to either add one tick of LLC (not recommended) or simply up the idle voltage to 1.49v to achieve 1.309v under load. This is the easiest and most efficient way to find stability on 12th/13th/14th gen, or so I've personally found.

I recommend using fixed voltage, a weak LLC setting (~150mV to ~180mV of a droop), and not pushing the ring past 48x as 13th & 14th generation uses a large die - causing the ring to lose stability more easily. You are not going to damage your system running 1.45 ~ 1.5v idle so long as your vdroop is large.

You did not list what the load voltage is for 5.7 GHz on your CPU, so I recommend finding out what the load voltage for stability is.
Setting your vCore to 1.45 ~ 1.50 fixed alongside a weak LLC.
And not pushing more than 1.32v through the chip under load. Any higher causes thermal hotspots due to the density of these chips.

The reason why a weak LLC is better is due to the VRM being given more voltage space to adjust. More voltage = more time for the VRM to compensate = better stability. Couple that with a fixed voltage and transients are far less of an issue for your system.

As for WHEA errors, I recommend disabling the following:
Link State Power Management under Edit Power Plan (Windows) > advanced settings > PCI Express
and: Native ACPI PCI-e support in your BIOS

If you google: ACPI error z690 or Z790
you'll receive a lot of forum threads discussing WHEA errors related to the ACPI functionality. It's recommended to disable it.

You can also set the Core PLL voltage within the BIOS to 1.10v and the CPU Input Voltage to 1.88v. Those both assist with achieving stability at the wall of your CPU's clocks.

Now, speaking personally - I tend to only aim for Cinebench R23 stability for three runs. After I'm stable at that point, I set a wattage cap that is 20% lower than what R23 requires. So for example, if R23 consumes 250 watts, I set a PL2 of 200W. This all but guarantees stability in everything I personally use, and if I encounter a crash, I simply up the voltage by a measly 0.010 mV.

For your efficiency cores, if you want to achieve 4.2 ~ 4.4 GHz, you'll need to adjust the L2 Cache Voltage. Do not go above 1.30v on the L2 Cache Voltage. Try to aim for 1.25v. You can either do adaptive or fixed: I recommend fixed voltage since every BIOS handles the V/F scaling on the cache differently.

Hey thanks for replying...

I'm quite further down the rabbit hole at this point. The 5.7GHz was a dud. Oh it worked but with choking up power to keep thermals in check, it was a glorified 4.8-5.1GHz UNDERclock in real load still pushing way too much voltage.

See the "problem"...

I have better Cinebench score on optimized stock clocks in Offset mode and a very low 1.13V average.

By now I have stabilized a "real" 5.5/5.4 GHz around 1.18V average in Offset (Vcore). Had to open up the top slots covers of my case for thermals.

Currently playing with using FIVR instead which is a much more interesting approach. I've stabilized the above 5.5/5.4 above in FIVR with 4.8 Ring and 4.1 E-Cores.

Now experimenting with a silent fan profile OC of 4.4GHz on all P-cores in FIVR. After 2 weeks of "whirring" fans when under load or testing I decided it was fun to reach the limits of my CPU cooler, but I won't have any of it for day to day operations. I want to close the top of my case with the dampening plates again and have something inaudible or about. Of course that minimal OC is stable since a good while, but it's not what is interesting. FIVR and a closer to operational Core VID and Uncore VID vs. Vcore is what is much more interesting. Pushing less voltage and power to the chip than ever there for the same OC in regular Offset. Minimal gains in thermals at that low level, but exporting the idea to someone that has a cooler that can withstand higher clocks volts and power, the gains would multiply. Still pretty sure my electricity bills will thank me running the same OC needing 5W less every second under load.

Between my original post and now... water under the bridge.

Come on, fixed Vcore is soooo Sandy Bridge... I said I wouldn't use that already.

Too easy, and not what is optimal for these CPUs. Even Offset is outdated methodology. Still too easy. Still not enough control.

FIVR is much more fun.

 

[chrcoluk]

What kind of god tier CPU is that, probably the best vcore I have ever seen reported anywhere.

 

[QcRMoD]

What kind of god tier CPU is that, probably the best vcore I have ever seen reported anywhere.

I didn't fully understand how these "things" are supposed to work... I mean, you know the so called "better performance cores" (in my case XTU was expecting the same VID for all cores, whatever that was) or what is called SP I believe?

Here's what my motherboard thinks of the quality of the chip in the so called CPU Indicator...

In my mind that "looked" like one of the best yeah... whatever it's worth.

TESTING is the only wise way to find your "most behaved cores". Which is starting from the very lowest voltage this CPU can run stock clocks (like any minor adjustment below yields WHEA/freezes) and pushing each core one by one +100MHz and see who does it, see who crashes Cinebench.

In my case, P0 and P3 are good, P1 is a bit above that with lower temps/power, and P2-P4-P5 are basically rednecks. If anything with some more attuned testing, P2 is the best out of the rednecks, but I wouldn't expect it to be of a real consequence in OC. Just like the fact that P1 is a bit above "good" didn't mean I could push it further than the other good ones. But obviously my 5.5/5.4GHz OC was using 5.5 on the good ones and 5.4 on the rednecks.

From there, a "Fixed" voltage is not a wrong idea per see. What I mean is, it doesn't work well with these CPUs to try and push the better cores too far ahead and use significantly different voltages for them and for the lesser cores than run lower. Like, I tried 5.6/5.4 and 5.7/5.3 and I also tried that with lowering E-Cores to 3500. It doesn't do much good. You'd likely get a lot of WHEA CPU TLB errors. They just don't like so much a significant difference of voltage unless I guess you try to fix it up by using "Selection Mode" in FIVR and trying to perfect a curve for both groups. A lot of work. Could have been worth it, you know, if my CPU cooler could sustain that kind of clocks on the good cores, but it can't. By testing again, I go (ALL CORES) from 1.13V @ 5.3GHz to 1.15V @ 5.4GHz to beyond 1.2V @ 5.5GHz. And that's the very LIMIT of my cooler. It's even too much except for bragging. This kind of shit needs 195W+ until better tuned and that just won't go well with that cooler or my case and fans. I guess an open bench would about be doable with the cooler.

Yadda yadda yadda, back to the "Fixed" voltage not being the wrong idea, guess why Intel doesn't work like that. What they do with the crazy CPUs out there is regroup a number of cores that will operate at the highest boost clock, then more cores at an intermediate boost clock, and then all of them at the lowest boost clock. These things want to see a regular voltage for a specific number of cores under load, not an array of different voltages distinctive for each group as it goes. FIVR surely doesn't help with THAT. It's not the point of using it with an Adaptive offset, nor the point of using it with Selection. The truth of what it does great is elsewhere.

 

[chrcoluk]

82 seems pretty good, I am tempted to grab the latest bios for my board as ASRock added the CPU indicator feature to it. Your screenie seems to suggest you in the top 1% for p-core so maybe I wasnt crazy on the golden sample comment.

My comment was based on how low your vcore is, the level of voltage needed for a given clock stability will be lower on better silicon. From what I have seen so far 12/13/14 gen has a really wide variance between the best and worst.

I can tell you my stock VID for 5.3ghz is 1.34v. 900mv higher than yours. But I dont know if ASRock shenanigans have manipulated that. It might not be the one programmed into my chip.

I also was at one point tempted to try different things on different cores as these chips now allow, but probably wont ever touch that now.

 

[dgianstefani]

82 seems pretty good, I am tempted to grab the latest bios for my board as ASRock added the CPU indicator feature to it.

My comment was based on how low your vcore is, the level of voltage needed for a given clock stability will be lower on better silicon.

True to an extent.

But leaky silicon can actually overclock better.

Efficient silicon may be just that, but may have a lower clock ceiling.

The KS seem to be both efficient voltage curves and have clock headroom, so best of both. If you set both K and KS to 5.5 GHz for instance, the KS will draw less power, but the chips also have a higher clock ceiling in general.

Regardless, if you're running into cooler limitations, you're probably better off with an undervolt offset on the stock voltage curve, instead of attempting a manual overclock.

 

[QcRMoD]

For those that "would have some doubts" here's the 30 minutes Cinebench results at optimized stock clocks.

By the RPM for the monitored CPU/CHFans I was still running my crazy whirring most advanced cooling profile for my situation. I'm not anymore with the 5.4GHz OC. I'll show the progression of the latter soon, just need to fine tune a couple more things and test.

True to an extent.

But leaky silicon can actually overclock better.

Efficient silicon may be just that, but may have a lower clock ceiling.

The KS seem to be both efficient voltage curves and have clock headroom, so best of both. If you set both K and KS to 5.5 GHz for instance, the KS will draw less power, but the chips also have a higher clock ceiling in general.

Regardless, if you're running into cooler limitations, you're probably better off with an undervolt offset on the stock voltage curve, instead of attempting a manual overclock.

Just to be sure, you meant "KF"?

And yeah I'm not saying I'm presenting anything scientific here regarding silicon quality. As I said, the "so called" CPU Indicator and so called "better performance cores".

With a relaxed fan profile and the 5.4GHz 1.15V average I'm seeing 84-86C Maximum and around 79-81 average on the package. Since I'm not rendering all day long, my temps ceiling is just nothing in Red so I have room, especially at low V/W.

More knowledgeable than me people seem to agree that beyond temps/voltage, CPU degradation is mostly a question of the insane power they are made to run "within specs".

And I mean, they do put i9 in laptops. I've seen a couple of threads about some of these. OTB they're running the 95*C/100*C tier downclocking to base boost clock all the times. Folks smartly going "stupid" on the best approach to undervolt and downclock these without losing too much out of their unconsidered investment.

Not talking about this thread here with the big dispute that an 14600K was going to cook an ITX config. I mean... try to beat an i9 in a laptop.

 

[chrcoluk]

No doubts here, your screenshot does seem to confirm is a golden sample, people can routinely get 10s of mv below programmed VID, so your VID is so low at its starting point anyway, then stable under 1.2v is believable.

 

[QcRMoD]

No doubts here, your screenshot does seem to confirm is a golden sample, people can routinely get 100s of mv below programmed VID, so your VID is so low at its starting point anyway, then stable under 1.2v is believable.

In fact, my motherboard doesn't allow more than -100mV for an offset in Voltage menu, which is what I ran these tests with as proven by the fact HWi doesn't report an Offset at all, which it indeed doesn't whenever it's not in FIVR. So I couldn't ever even crash Cinebench or get WHEA errors upon which to fine tune the LOWEST voltage. From there, the "best core" test to 5.4GHz per each core had even the P2 Core pass and only the P4-P5 freeze.

I had to start with the basic 5.4GHz I did with a Voltage Offset of -80mV to then push each at 5.5 and discover the P2 wouldn't pass so close to limits.

Running a LLC5 on the same optimized stock settings yielded 1.1V and below and that got crashed.


Therefore, my average Vcore in the tests shown above had some minimal amount of headroom.

 

[chrcoluk]

I typo'd I meant 10s of mv up to around 150 in the most aggressive setups. But still you within that

 

[QcRMoD]

Are there still folks venturing into a manual delid with a razor blade nowadays?

T'was easy back in the Haswell and Kaby Lake days.

Now either folks are pissing their pants, or there's some difficulties I cannot foresee about it.

 

[phanbuey]

it's annoying to heat the chip up and then delid it. Higher risk of killing it

 

[QcRMoD]

Heat it up?

Are these still glued like back in the days?

Never had to heat up anything. Just cut, wiggle, remove.

Ok then I guess I’d have to find out on my own.