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.

 

[Carfax83]

Most likely it starts to throttle cause your front fans dont provide enough fresh air to the t30. Try removing the front panel and test again

I honestly don't care about whether it throttles in CBR23 because I don't do any rendering at all. Even when it did throttle though, it was only to 5.4ghz on a few cores. Most of the cores stayed at 5.5ghz. I tested 4K60 encoding and it didn't throttle at all, and that's more important because it's actually a workload that I do on a fairly regular basis.

 

[Carfax83]

Drop AC towards 1 (keep DC at stock), something like AC 3 + slight VCore offset might be more stable, though (looking at you, Folding@home). Increase CPU LLC to middle value ("High" in my Gigabyte BIOS, aka step 4 out of 7). If you do per core OC then use TVB to limit boosting above 60-65°C.

And most importantly: Use a power limit slightly above what CB23 uses, no realworld load will ever hit that limit! I am currently using 245W PL1/2.

Don't use silly memory overclocks. 5600 MT CR 1T is enough for almost all realworld applications and runs at 1.15v IMC and dimm voltages for me.

What cooling are you using? Looks like you might be throttling, because your CPU is overclocked and mine isn't, but your CBR23 score is less than mine. I scored 41,200 at 5.5/4.3.

Or it could possibly be due to your lower memory bandwidth.

 

[Timur Born]

That's within measuring tolerances. 40-41k is normal for 55x/43x settings. My CPU only uses 56x below 65°C and downbins to 55x via TVB above that.

 

[Timur Born]

Take note how I wrote "towards AC 1/0.010", not "to". The lower you go the less of a Vcore offset you need, but at the same time low-core load increases in voltage. This can be used for higher per-core OC (like 2x 6 Ghz), but may be more than you need/want. So finding the right balance between AC and Vcore offset is the tricky part.

Personally I am currently working on "flattening the AC vs. offset curve" to lower low core Vcore a bit, because my stable settings currently see it peak at 1.36v for CB23 single-core at 6 GHz, which I hope to decrease. Doing the flattening part is easy, but keeping all other load scenarios stable alongside is the real task at hand.

 

[Timur Born]

It seems that 1.35v (1.335 VID) is as low as I can get for single-core CB23 at 6 GHz without sacrificing in other areas. That's around 40W package power for single-core while maintaining about 236W for multi-core. I can drop to 1.33v at 5.9 GHz for single-core, which is below 40W, but wouldn't improve in multi-core and still needs to maintain stability in other workloads.

 

[Timur Born]

Currently I am testing the last few steps of AC vs. offset. AC 6 with -0.025v offset was not overnight Folding@home stable, so I will try AC 5 with -0.020v now. AC 3 with -0.010v offset was what I tested stable so far, but the lower AC means higher single/low core voltage, which currently I am trying to shave off the last 0.01v from. This is more about "let's try to squeeze the last 1%" just to learn about the behavior.

F@h is a good stability test in that it causes serious instabilities. But it's also a bad stability test in that every "work unit" is different, causing different load at considerably different power usage.

 

[Timur Born]

Unfortunately AC 3 - 0.010v turned out to not be as stable as I thought. So back to the drawing board. Folding@home is killing me. :-D

 

[Rigg]

I've been playing around with my OC a bit trying to find daily driver settings. I've found that capping clocks on individual p cores can help with stability. I have my favored cores capped at 58, my 2 hottest cores at 56, and the rest at 57. My core usage OC is 58/58/57/57/56/56/55/55. I was occasionally getting instability on load transitions when the CPU would try to momentarily turbo up. I found when doing extended CBR23 multi tests CB would crash when transitioning to a new render even though it was stable under load. Capping my cores seems to have fixed that problem.

 

[Timur Born]

You can also (ab)use per core maximum clocks to make the Windows/CPU scheduler prioritize cores. It will always use highest maximum cores first, even if they are set to a higher maximum than your allowed Turbo ratios. Here is what I use:

Turbo Ratio Limits (P-cores) - IA/SSE: 60x (1-2c), 59x (3-4c), 57x (5-6c), 56x (7-8c)
Turbo Ratio Limits (E-cores): 43x (1-16c)

56, 57, 60, 61, 57, 56, 59, 59, 43, 43, 43, 43, 43, 43, 43, 43, 45, 45, 45, 45, 44, 44, 44, 44