Question 13900K, KF, and KS Undervolting wizards... What is your methodology?

[Hulk]

I have a 13900K and have noticed that many others with the same CPU are getting much better power/thermal results than I am at equivalent frequency/performance. I realize that there is a difference in the quality of the silicon of each CPU but that being said I'd like to learn more about how you go about undervolting? The reason I have specifically mentioned the 13900K series is because unless you have some type of extreme cooling most people are generally performance limited by heat/temperature. But some around here are getting much better results than others, like myself. Anyway, onto the specifics of my question.

Someone hands you a 13900K, KF, or KS, in a step-by-step analysis how do you go about optimizing for a given cooling solution? Let's say it's Noctua U12A air cooling in a case than can effectively deal with about 200-225W of heat.

 

[Just Benching]

It's very hard to test undervolting stability if you can't properly cool the CPU to sustain it's maximum clocckspeeds.

 

[Hulk]

It's very hard to test undervolting stability if you can't properly cool the CPU to sustain it's maximum clocckspeeds.

So how do you optimize for a given cooling solution?

 

[Kocicak]

It's very hard to test undervolting stability if you can't properly cool the CPU to sustain it's maximum clocckspeeds.

It is very easy to disable parts of the chip to let the remaining part run at maximal frequency and test its stability.

 

[Just Benching]

So how do you optimize for a given cooling solution?

Well, personally I would just set Lite Load to 6 (instead of the default intel spec which is 9) and then just...test it in day to day and see if it's stable or not. Download benchmate and run all of the cinebench tests, 11.5 - 15 - 20 --23. Even though R23 is the heaviest of all - I've noticed that R15 and R11.5 were giving me errors when R23 would easily pass. After you pass all of them with max priority, try ycruncher. If one of them crushes with an error, you are close to stability, if you bsod you are probably far away. If you pass everything without crushing, either try a 30 minute stability test of something like real bench, or any test of your choice.

 

[Carfax83]

From my experience, Raptor Lake seems to be limited by power draw more than by voltage, hence I tend to try and keep PL1 and PL2 as high as possible to hit and maintain my desired frequency, while reducing voltage as much as possible. Lowering the voltage also reduces power draw in light to moderate loads and has a massive effect on temps.

When I was tweaking my CPU, I decided that I wanted it to consume no more than 235w due to the fact that I'm a 4K gamer (therefore predominantly GPU bottlenecked) and I occasionally do some encoding/transcoding. Unlimited power draw was a no no for obvious reasons. After setting it to 235, I ran tests to see how high my CPU would boost with all P cores active at 100% and maintain the clock speeds, and it turned out to be 5.3ghz. Efficiency cores were also able to boost to their maximum clock speed, being 4.3ghz.

So after seeing that, I started lowering the voltage. I guess my CPU has some high quality silicon, because I've had my CPU undervolted by as much as -165mV and it was perfectly stable under load. Only when idling did it become unstable. My final settings were -135mV and I haven't had any crashes since. My CPU has exceptional performance per watt and power efficiency for gaming workloads and is easily cooled by my stock Noctua NH-U12A when running full bore at 5.3/4.3ghz. I think eventually I will get some of those fans that you have on yours and see if I can squeeze some more blood out of the stone

If I can drop another 4-5c off the temps, I could bump the clock speed to 5.4ghz without increase PL or voltage.......or so I think.

 

[JoeRambo]

There are multiple ways to undervolt 12-13th generation chips. I think at least 3 or 4. Intel is providing incredible amount of options to dial in the tunings both in BIOS and from XTU app in Windows.

The complexity is a problem tho and amount of information to absorb, you need to watch like 1-2 hours of material to understand the concepts and then there's a question of how/if to translate those concepts to the BIOS of motherboard cause of different vendor than in videos.
But the toolbox is there and it sure is working.

I'd recommend ScatterBencher on Youtube, he has some great stuff, including platform voltages, where and how they are regulated.

So after seeing that, I started lowering the voltage. I guess my CPU has some high quality silicon, because I've had my CPU undervolted by as much as -165mV and it was perfectly stable under load. Only when idling did it become unstable. My final settings were -135mV and I haven't had any crashes since.

That's the problem with offset undervolting, undervolt applies to all operating point VIDs, idle/load and on some motherboards even uncore VIDs are offset by that amount and it might get unstable when idle ( but on cold winter night only ).
Alternative is so called V/F point offset voltage tuning, so You only apply offset at upper clock range V/F points. So one can go ham on them and dial in them properly. ( like set 48x to 0.0V, -0.150V on 52x and 55x each and then run your tests and go to each multiplier from 48-55x to see if they have enough stability at resulting interpolated voltage ).

 

[Hitman928]

I don't have an Intel 12th/13th gen CPU to play with but I will add a general statement if you don't mind. I have found through the years of tinkering hardware that my standards for stability and full load temps leave me unable to replicate/match what others online report they accomplish in terms of over (or under) clocking and cooling. Of course, the silicon lottery comes into play here, but based on many years of building and testing systems, I've come to the conclusion that it's more than that. I hope that someone with specific experience with the platform can help you eek out a little more perf or perf/w with your system, but ultimately I've learned to just be happy with what I can do with my system, especially when you already have a top of the line computer like you do. Best of luck

 

[Rigg]

Do you have a contact frame installed? If not it could be part of the reason for the discrepancy. The VRM on your motherboard isn't very good either. It's pretty under built for a 13900k IMO. It's possible it could be throttling under heavy load.

I'm firmly on the temp limit side of the temp vs power limit debate. I run my 13700k with an 85c limit and no power limits. As far as undervolting goes, i think @JoeRambo covered it pretty well. There are several ways to manipulate voltages on these CPU's. The best recipe for pure heavy multi-thread performance is arguably going to be LLC set low with a fixed voltage and clock multiplier that can be adequately cooled under heavy loads.

We should be discussing clock speeds and vcore under heavy load and not how big of offsets are being achieved. The size of the offset is kind of meaningless because you can manipulate the SVID behavior. The default SVID behavior is possibly highly variable from board to board. If i set my board to "best case scenario" SVID behavior I'm not even stable at stock speeds without increasing load line. If I leave it default I can offset -100 mV. You can also manually offset the SVID instead of the vcore but the smallest offset is 100 mV.

 

[Hulk]

You can see my rig in the signature. Temp limit set to 85C. If I reduce LLC from 3 to 2 it'll restart now and then under heavy load. LLC=3 completely stable. I guess I could limit frequency and possible run LLC 2 for a little undervolt but I'd like the ability to ramp up to 5.5GHz for my bursty loads and I like that.

I always get a mediocre to bad chip as well. I remember back in the days of the Celeron 300. Every one of those could get to 450, mine as well. But some could get to 504 or whatever that specific number was. I had a few but none could ever get to 504. I remember putting one in the freezer for a while and then putting it back in the computer to try and get it there for a quick run. Oh to be young and dumb.

 

[Carfax83]

That's the problem with offset undervolting, undervolt applies to all operating point VIDs, idle/load and on some motherboards even uncore VIDs are offset by that amount and it might get unstable when idle ( but on cold winter night only ).

Yep, that's exactly what happened.

Alternative is so called V/F point offset voltage tuning, so You only apply offset at upper clock range V/F points. So one can go ham on them and dial in them properly. ( like set 48x to 0.0V, -0.150V on 52x and 55x each and then run your tests and go to each multiplier from 48-55x to see if they have enough stability at resulting interpolated voltage ).

Man, that seems like a lot of work. Properly tuning a PC can take hours and hours, or even days depending on how in depth your methodology is. I've always been one of those "get it over and done" types, so I usually take the path of least resistance.

The only exception being memory performance. I will take more time to tweak the memory than I will the CPU, and unfortunately, tuning memory is a lot more technical and time consuming.

 

[Carfax83]

I'm firmly on the temp limit side of the temp vs power limit debate. I run my 13700k with an 85c limit and no power limits. As far as undervolting goes, i think @JoeRambo covered it pretty well. There are several ways to manipulate voltages on these CPU's. The best recipe for pure heavy multi-thread performance is arguably going to be LLC set low with a fixed voltage and clock multiplier that can be adequately cooled under heavy loads.

I think I'm going to try the temp limit method tonight and see how far I get. So do you mess with the voltages at all for a temp limit?

 

[Rigg]

I think I'm going to try the temp limit method tonight and see how far I get. So do you mess with the voltages at all for a temp limit?

Yes. This is an alternative way of inducing throttling. It's not a replacement for undervolting.

 

[alcoholbob]

The lite load method works, or you can just do manual undervolting. Most people report anywhere from -120mv to -150mv is stable with Raptor Lake by offsetting the turbo frequency curve only. That way it won't affect your idle voltages.

 

[coercitiv]

Man, that seems like a lot of work.

It's not more work than a static undervolt, at least not in the easy version. Take your static undervolt and apply it only in the high frequency region. The rest of the frequency spectrum will be using stock.

I'll use semi-random numbers to illustrate the principle. Say you have a -100mV static undervolt and your target frequency is 5Ghz+ all-P-core, meaning that your power target is setup in such a way that the CPU hits 5GHz+ under MT loads. This means you only really need to undervolt for 5GHz+ speeds, since the rest will not really be used during real work.

The 13900K has a V/F point at 4.3Ghz and the next is 5.1GHz. You can set the 4.3Ghz offset to 0 and the 5.1Ghz to -100mV. The firmware will use interpolation to calculate voltages between these two points. The rest of the V/F points (5.4, 5.7, 5.8) can all be set to -100mV, though you may even have the option of reducing the offset on the OC ratio if you want to go for more than 5.8Ghz.

In the end all you're setting up is a static offset with a high-pass filter - everything above 4.3Ghz receives the undervolt (with interpolation until 5.1Ghz), everything bellow stays stock and fully stable. If your system is tested stable at 5.1Ghz /w -100mV and 5.8Ghz /w -100mv then you're pretty much done, as every other frequency is safe as well by the nature of the applied voltage curve.

 

[JoeRambo]

It's not more work than a static undervolt, at least not in the easy version. Take your static undervolt and apply it only in the high frequency region.

The only extra know-how piece is that VID curve has to be increasing and resulting value of next VF point has to be same or above the previous one.

So in that example with 4.3 and 5.1, if one was to set -0.2V on 5.1ghz and leave 4.3 at stock value, the resulting 5.1 voltage might not apply due to resulting VID being lower than what is at 4.3.

There is also an advanced topic of uncore VID and clocks. It might or might not need offset as well to achieve desired clocks. For example Your undervolt at 5.1 might drop VID for cores so hard, that uncore can no longer clock above 4Ghz in "default" state. So two things will happen, either Uncore is reduced in clocks killing perf or resulting VID is increased to that of what Uncore demands and results in more voltage than expected.

 

[Hulk]

I've noticed this when trying to undervolt. I'll set LLC at 2, it'll run CB for 10 minutes no problem. Then I'll stop it and when I try to restart it again instant computer restart.

With the additional insight I've gained from this thread I'm thinking that the CPU is not still hot so when I restart the off-idle voltage isn't high enough as the CPU begins to ramp up in frequency and the extra voltage doesn't come fast enough so the restart occurs.

I'm going to try what coercitiv wrote. Keep LLC at 3 but offset voltage down at higher clocks.

If I set offset at "0" at say 4.5GHz and then -0.1at 5.8GHz does that mean I'd get my "normal" voltages due to LLC=3 alone until 4.5GHz then -0.1V interpolation down from the LLC=3 from 4.5 to 5.8?

Just want to be sure so I can approach this methodically and not chase my tail (too much).

 

[Carfax83]

OK I tried the thermal limiting approach instead of the usual power limit approach and I must say, I quite like it. Seems a lot easier than twiddling with the PL values. Did a single test CB23 run to see how much higher a score I could get with unlimited PL and thermal limit set to 90c, and I got 41,055. CPU maxed the P cores at 5.5ghz and the E cores at 4.3ghz for the first with no throttling, which generated the 41,055 score, the highest I've ever had. Then I did a multiple back to back runs to test how much the CPU would throttle when put under prolonged load, and it honestly didn't throttle that much. It hit a max of 91c then dropped down to 5.4ghz on some of the P cores, but was still hitting 5.5ghz on a few of them so the algorithm is very good for thermal limiting. The highest the power draw was during the CB23 run was 263w or so.

I also lowered my undervolt from 135mV to 115mV. Not sure if I'll keep these settings though, because I don't really gain much performance at all due to running at 4K. Then again, gaming power draw is so minimal that it probably won't matter as I never get anywhere near the 90c limit. I'll have to try some encoding tests.

 

[Hulk]

OK I tried the thermal limiting approach instead of the usual power limit approach and I must say, I quite like it. Seems a lot easier than twiddling with the PL values. Did a single test CB23 run to see how much higher a score I could get with unlimited PL and thermal limit set to 90c, and I got 41,055. CPU maxed the P cores at 5.5ghz and the E cores at 4.3ghz for the first with no throttling, which generated the 41,055 score, the highest I've ever had. Then I did a multiple back to back runs to test how much the CPU would throttle when put under prolonged load, and it honestly didn't throttle that much. It hit a max of 91c then dropped down to 5.4ghz on some of the P cores, but was still hitting 5.5ghz on a few of them so the algorithm is very good for thermal limiting. The highest the power draw was during the CB23 run was 263w or so.

I also lowered my undervolt from 135mV to 115mV. Not sure if I'll keep these settings though, because I don't really gain much performance at all due to running at 4K. Then again, gaming power draw is so minimal that it probably won't matter as I never get anywhere near the 90c limit. I'll have to try some encoding tests.

Wow! You've got a sweet setup. When running CB 5.5/4.3 what are your VID and Vcore average values?

 

[Carfax83]

Wow! You've got a sweet setup. When running CB 5.5/4.3 what are your VID and Vcore average values?

At 5.5ghz, the average was 1.274v. I can only fully sustain 5.5ghz for about 2 runs before it starts to throttle though, then it will mostly be at 5.4 and 5.3ghz.

Not really a problem though, because CB23 isn't a workload that I ever do

 

[Carfax83]

Correction, the Vcore during the run was 1.2v. HWinfo doesn't report accurate Vcore values from what I've seen. CPU-Z is more accurate, because when I change the voltage settings in the UEFI, CPU-Z registers the change unlike HWinfo.

 

[Hulk]

Correction, the Vcore during the run was 1.2v. HWinfo doesn't report accurate Vcore values from what I've seen. CPU-Z is more accurate, because when I change the voltage settings in the UEFI, CPU-Z registers the change unlike HWinfo.

What was the reported VID when the Vcore was 1.2? Crazy that your chip can do 5.5 Vcore 1.2.
During those two runs where it was holding 5.5/4.3 what was the package power?

 

[Carfax83]

What was the reported VID when the Vcore was 1.2? Crazy that your chip can do 5.5 Vcore 1.2.
During those two runs where it was holding 5.5/4.3 what was the package power?

Sorry, the terminology had me a bit confused. After doing some googling, VID is what the CPU is requesting correct? And Vcore is what is being delivered?

I took a screenshot and I think my VID was 1.278v, while the actual Vcore was 1.163v. Package power was in the 270w range.

 

[Just Benching]

Just a tip, instead of lowering llc, try to play with lite load instead, leave LLC to the default option. Lowering LLC drops the voltage for heavy workloads, but not for lighter ones.

The truth is in some games the 13900k draws a LOT of juice and you need to play with lite loads to fix that. In cyberpunk I've managed to drop the power to 140w but still its far away from my 12900k which was just drawing 90 to 100 watts. Yes it is slower but the power draw difference is huge

 

[Hulk]

Sorry, the terminology had me a bit confused. After doing some googling, VID is what the CPU is requesting correct? And Vcore is what is being delivered?

I took a screenshot and I think my VID was 1.278v, while the actual Vcore was 1.163v. Package power was in the 270w range.

Jeez, those are some great numbers. I'll post mine for comparison purposes. I don't see Vcore listed in your HWinfo? I hate to keep bothering you but what was that reading as compared to CPUz?