it seems he should rather focus on 3800 Mhz CL14 or CL15?
This ^ (imho), I don't think there will be any benefit from 2000 fclk overall (bandwidth raised, latency worsened, and add to it close to 100% chance of getting a whea errors shitload), that's the way things are now. Hope they resolve it in not so distant future.
-5 on best, -10 on 2nd best, and -15 on others
Heyahh!
Why don't you try -18-20 on "bad" cores?
I spent half a day testing various PBO configs and think now I'm close to being finalize "release candidate".
At the end of the day I came to conclusion, that I was mostly wasting time looking for appropriate settings with boost over 350. I have basically two configurations left to test from the original pile of data.
1. +200 PBO (your case) and +350 PBO - the one I lean towards.
First, the most I regret is wasted time when trying to get around that damned CPU resets (reboots) suddenly occured when benchmark just started (and what is remarkable - NEVER during the test itself). It turned out that it was caused by Vcore +10mv offset, that I used everywhere. Just here in this very thread I saw someone suggested to aplly it to compensate Vdroops from weak LLC.
All in all, after I set it back to default and instead tried modest scalar for the same purpose (usually 3 or 5) - all nightmare has gone. Now if that even happened, it's much easier to track down what was wrong.
So, back to the settings, I would ask you to try those below and tell me, if you'll see any improvement (if at all):
PBO Boost: +200, Limits: ppt 115W and tdc/edc 70/140A, scalar: 3x, Curve: 8 and 10 for the best two cores, and 20 for the rest.
I see it, like the following steps can make the tuning process easier to understand and accomplish (the order is important):
1. Define your boost override value (if you're still capped at 200, that's the value)
2. Pick up values for Curve. A couple of tips: establishing values for best cores are kinda tricky, since OS prefer them for light loads, both cores need to get as close as possible to the boost frequency you have set previously and at the same time, be high enough not to crash themselves. For this purpose CB R20 ST suits well (of course, better to avoid stretching, so watch for the effective clocks, not the performance).
3. For the rest of the cores you may set basically the same value, i.e. the lowest possible, which won't crash the weakest one.
4. And finally, depending on cooling capability, you may want to limit your CPU power appetite setting PPT to a couple percent below default PBO value. For the TDC and EDC I'd keep them just the way they won't ever be reached in most workloads.
So what I see as the main idea behind all of it, is to find the point of intersection of the descending (complementary to the aggressiveness of PBO tuning) curve of heavy-load performance and ascending one for light loads. A compromise determined based on the intended type of workloads.