Maximus 12 series and i9 10900k overclocking guide and tech sheet

[Falkentyne]

Asus can't get the voltage requirements perfect without an actual stress test done in BIOS and you aren't going to see that. The problem is, besides leakage (one CPU that has the exact same VID and load vcore running cooler than a second CPU with the exact same VID and vcore, yet overclocking worse than the hotter chip, etc), /cooler score and VID, two CPU's that have the exact same identical VID can have a different guardband (basically, how much voltage leeway or grace voltage you have), so the CPU with a larger guardband will run at a lower load voltage stable, than a CPU with a smaller guardband. And there is absolutely no way to test for guardband without windows stress testing. But in general, the SP ratings give an average indication of the purity of the chip's silicon. My chip had very accurate avx disabled voltage points at 4.7 to 5 ghz but then things start getting wild at 5.1 ghz and higher.

[kvarq]

Just received my 10900KS, SP63 (probably most of the users have this), my aim was 5.2GHz daily / gaming.

My SP63 is really crappy.
#X019E439

5.2Ghz won't but at all below 1.30 V LLC6, 1.34V CinebenchR15 capable, so daily would be ~0.08V extra. No way.
5.0 GHz boots with 1.20V, 1.24 R15 capable, passed stress tests at 1.32V.
LLC6 also, probably for LLC4 will want some more juice.

I have 2x16GB 4000MHz C19 ( F4-4000C19D-32GTZKK), 17-17-17-37 pretty easy still @1.35 VRAM, would like to tune a bit less some numbers, maybe go even at CL15, but tRFC would be a must. I'll increase the RAM voltage if needed, although IO/SA voltages are still default, quite high though (1.45V, 1.52 V).

AiO Arctic II 360 / Asus XII Apex / 10900K / K|NGP|N 2080Ti / F4-4000C19D-32GTZKK / 970Evo 2TB / Seasonic TX850

[Falkentyne]

The voltages you mention are bios voltages or load voltages? Sometimes people just jumble them then it's impossible to know what they're referring to.
What bios voltage set do you need to pass 1 hour of Realbench 2.56 at 5.2 ghz, without CPU Cache L0 errors appearing in HWinfo64 (WHEA section) sensors?
And what are your temps like when you manage to pass that RB 2.56 test?

You should try 1.40v bios set + LLC6 at 5.2 ghz and see if you can pass RB 2.56 without L0 errors. Start with 1.36v Bios set LLC6 first.
And there's no problem with 1.40v bios set but I would stop there for daily. LLC6 is rather droopy still. 0.495 mOhms of LLC is healthy enough.

By comparison LLC "Turbo" on Gigabyte Z490 Aorus Master is 0.275 mOhms of LLC.

[geneo]

Hey

I assume in "5 ghz/4.7 cache @ 1000 khz" the 1000 KHz refers to the VRM switching frequency? Did you have the VRM at 1 MHz for all of these runs (at 5GHz, Realbench 2.56 my 10700k Hero XII VRM only hits 56C with 1 MHz switching frequency).

[Falkentyne]

Hey

I assume in "5 ghz/4.7 cache @ 1000 khz" the 1000 KHz refers to the VRM switching frequency? Did you have the VRM at 1 MHz for all of these runs (at 5GHz, Realbench 2.56 my 10700k Hero XII VRM only hits 56C with 1 MHz switching frequency).

I have it at 600khz now.
I noticed 'some' minor improvements messing around with FMA3 prime95 LLC8, going from 300 khz to 500khz (Like the system not insta-BSOD'ing at a low vcore), so I assume 600khz isn't worse. I have no idea if 1 mhz is worse or better than 500-600 khz. Doesn't seem to be at lower LLC levels. And the Minecraft issues have to do with some weird way that game hammers the cache hard. It's actually more stable the higher LLC you use (with the same load vcore target).

[kvarq]

The voltages you mention are bios voltages or load voltages? Sometimes people just jumble them then it's impossible to know what they're referring to.
What bios voltage set do you need to pass 1 hour of Realbench 2.56 at 5.2 ghz, without CPU Cache L0 errors appearing in HWinfo64 (WHEA section) sensors?
And what are your temps like when you manage to pass that RB 2.56 test?

You should try 1.40v bios set + LLC6 at 5.2 ghz and see if you can pass RB 2.56 without L0 errors. Start with 1.36v Bios set LLC6 first.
And there's no problem with 1.40v bios set but I would stop there for daily. LLC6 is rather droopy still. 0.495 mOhms of LLC is healthy enough.

By comparison LLC "Turbo" on Gigabyte Z490 Aorus Master is 0.275 mOhms of LLC.

I was referring to Bios voltages.

After some tests, I think I'll stuck at 5.0GHz for now.
I'm interested to be as steady as possible during game sessions.

5.2GHz boots just after 1.30V LLC6, then Cinebench15 capable only at 1.34V. From my experience, difference between CinebenchR15 and daily solid is around 0.08V-0.10V extra.
5.0GHz boots at 1.20V LLC6, CinebenchR15 capable 1.24V. Linpack requires 1.32V. Then, dropping timings from 19-19-19-39 to 16-16-16-36 would require 1.36V LLC6 (set in bios, 1.35V in Bios after restart or HWInfo) LLC6 for now.
During RealBench 2.56 max temps are around 87-89C (sometimes there are some 91-93C spikes on few cores) and the VCORE load voltage stays at 1.261V (though a minimum spike 1.252V reported by HWInfo). No errors.
Also I tried for 20mins small fft in prime95, temps are pretty much the same.

Noticed few posts of yours where you recommend LLC4 - required voltage would be higher though, probably 1.40-1.42 V to be set in bios.
When 1.40V LLC4 set in bios, after restart in bios it shows 1.376V, also in HWInfo.
During RealBench 2.56V, the load VCORE varies between 1.234-1.225V (also min spike 1.217V reported by HWInfo). Temps during test are around 85-88C, with spikes of 90-91C (overall 1-2C less than LLC6).
But with LLC4, drops seem higher in prime 95, almost instant crash in prime95.

[geneo]

The V-F curve is useful for more than just fixing idle issues. The default v-f for my 10700k had 5.1 GHz point way too high and the 4.6 GHz point was anomalous, Below are the default (top) and my corrected (bottom) v-f curves. I verified the lower 4.6 GHz point was stable. This curve is stable at LLC5.

If I have some time, I may go through and optimize each point to its lowest value. I have only verified 5.0, 5.1 are lowest.

This is a very nice and useful feature Shamino, thanks!

@Shamino

There is one oddity with v-f . When I try to OC 5.2 GHz, point 8 is for 5.2 GHz and there is a gap then between 5.0 and 5.2 GHz. In order to get 5.2 stable, the interpolated voltage for 5.1 is higher than necessary, I would have thought that the point 8 would still be for 5.1, with 5.2 being taken care of by the adaptive "additional turbo voltage" which kicks in at 5,2. Is this a bug?

[geneo]

@Shamino

To add to the above. On my 10700k, when I select a multiplier of 52 in the BIOS. point 8 gets applied to the 52 multiplier (which is above turbo threshold) instead of 51. In the AISuite vf editor, it appears as applied to 51, when in fact it is 52. I think the BIOS VF point 8 should always be applied to the maximum turbo multiplier (in this case 51), since above maximum turbo the additional core voltage can be controlled by other means.

[Jane Doe]

I got 10700k 2 days ago and now testing... My SP is only 51 and using Noctua cooler.
My prediction for x45 AVX is 1.355v it is too big value for only 45 So, I start testing cpu with
svid disable, and manual voltage without HT.Difernce with or wihout HT is about 0.030 by prediction.
Now I.m on 1.150 in bios and 1.134v in windows...in prime AVX2 after 10min testing, temps are 70C and
voltage drop to 1.083v.
So I will continue with drop manual voltage by 0.010 until prime errors. My question is, is it posible
that prediction is wrong??? I tested couple cpus and pridiction was 99,9% acurate, but in this case it
is something new for me, bcs diference between prediction and test are 0.2v. I will continue with 4.5Ghz
and after that try 46,47 ....My limit is Noctua and I cant go max than 1.25v to be safe.

[pianobench]

I got 10700k 2 days ago and now testing... My SP is only 51 and using Noctua cooler.
My prediction for x45 AVX is 1.355v it is too big value for only 45 So, I start testing cpu with
svid disable, and manual voltage without HT.Difernce with or wihout HT is about 0.030 by prediction.
Now I.m on 1.150 in bios and 1.134v in windows...in prime AVX2 after 10min testing, temps are 70C and
voltage drop to 1.083v.
So I will continue with drop manual voltage by 0.010 until prime errors. My question is, is it posible
that prediction is wrong??? I tested couple cpus and pridiction was 99,9% acurate, but in this case it
is something new for me, bcs diference between prediction and test are 0.2v. I will continue with 4.5Ghz
and after that try 46,47 ....My limit is Noctua and I cant go max than 1.25v to be safe.

I don't think the prediction is entirely accurate. But it is very close. My prediction varies by temperature sometime but generally it floats between 1.339v and 1.334v for 5.0GHz@LLC4.

I set via AC/DC Load Line to 0.4 mOhms to generate a bios set of 1.314v (as displayed in the top right panel). This is stable for me after a 1-hour prime 95 small fft (no avx/avx2) test. It is also stable after 1-hour CB R20 loop test.

Try to find the setting under the advanced tab --> CPU Config --> CPU - power management and find the boot performance mode. Set it to turbo so that you can see voltage at your highest frequency within bios if it is not already set this way.

Now set AC/DC Load Line to 0.01 save & reboot into bios and this should display what your Vmin is for the CPU VID. This is the lowest VID for that particular frequency. Once you know the floor you can maybe judge what next to do.

0.01 mOhm should be the lowest VID value for your target frequency with increasing values of AC/DC load line providing higher VID values.

I am using this technique based upon Shamino's suggestion of trimming the AC/DC Load Line to get the VID that you want. A higher AC/DC Load Line of 0.4mOhm or 0.5mOhm is what worked for me.

Asus's own SVID behavior suggests that 0.6mOhm is the "typical" behavior of the VID. So most people should fall into that range. I don't know if this is helpful information but I thought to add as it could be an indicator of your voltage performance relative to others or Asus's data.

edit: I also have an i7 10700K