Turns out the motherboard Vcore reading is pretty much useless. It's the same wide "stepped" readings all the time! The SVI2 seems to better reflect what the CPU is getting.
As opposed to 4.0/1.3V/Auto I'm currently testing 4.0/1.25V/Turbo. Auto droops to 1.237V (-0.063), current setting droops to 1.225V (-0.025). There may be room to drop Vcore further, I don't know. I'll save the higher clocks for when my U9S arrives this week, but so far the LLC work seems to be paying off. It's actually running cooler with these settings and cooling normally, as opposed to fans full blast, Auto LLC and side panel off.
Okay, this is crazy. I'm down to 4.0GHz / 1.2V, with 16 thread P95 Smallest sustaining an effective 1.175V. Two repetitions completed, no errors anywhere. My CPU temps are now below 77c with auto fan curves, and my VRMs are down to 68c. I may have been wrong about Ryzen 3000. This is absolutely delightful.
One thing I'm not wrong about, though, is the horrible firmware. I mean, my chip spends most of its time at 4.0GHz anyway when I leave everything on auto, since it can't last under higher boost. How is it "normal behaviour" to be pumping 1.4v of full load into a chip that doesn't even need 1.2V?? And they said it was "safe". I call bollocks on that AMD statement.
EDIT: found the tentative limit. With Turbo LLC, at 4.0GHz the floor is at 1.181V drooping down to 1.162V. For some reason, the lower you go, the more the BIOS wants to round up; if that's not enough, some of the Vcore settings share the same practical results (exact same max/min on SVI2 and Vcore). The result is a gaping hole between 1.181V and 1.16875V that I can't bridge, so this is the limit for now.
