Correct, this is not he modding "undervolting".
Despite CPUs being devices that work with digital signals, the physical device still deals with analog processes. For something to be a 1 or a 0 the analog signal must be below or above a certain threshold. During the transition from low to high the signal raises over time based on the resistance, current and voltage. The higher the field strength (the voltage) that drives that change, the steeper (closer to vertical) the signal will move in time domain, the sooner (higher clock speed) you can sample the signal and determine if it is above or below the threshold needed to say whether it is a zero or 1. So, higher voltage - higher clocks. There are tables in the CPU that tell the VRMs what voltage the CPU needs to run at a certain speed. These tables have a healthy margin overprovisioned for stability. The better the silicon etching (less resistance/leakage), the actual margin can be even higher.
The "undervolting" by the modding community is offsetting the demanded voltage stored in the CPU clock/voltage tables by some amount (decreasing the stability margin), so you supply less, hoping that the silicon is good enough. that the signal still raises fast enough so when sampled at the time needed it does not produce an ambiguous or a false reading (should be 1 but reads 0, or should be zero but reads indeterminate). As established before P=V.I so less voltage equals less power (less heat). If you are running against any of the thermal points where the CPU will start suppressing boost, then the lower temperature will help to keep the higher clocks for longer, but again, this as the cost of your stability safety margin. If you got a well etched CPU on a wafer that didn't have inclusions - it works. If you got a not so well etched CPU on a wafer that wasn't so perfect - the system will crash. AMD has something called "clock stretching" that tries to ensure stability even if you run into situations where the signal isn't clear what it is - clocks apparently stay high, but actual performance is worse as the internals need to wait more than what they should at the given speed to be able to sample the signal reliably.
If you are just playing and learning - good for you, but if you are making money by using these CPUs just make them stable and go on with your business. Time wasted monkeying around with PBO is time you are not billing a customer for.
If you know you have a not so good CPU sample, or only care about temperatures (e.g. live in a particularly hot climate - 45C ambient in some part of the day without AC, no space for a bigger cooler, etc.) - just go into PBO and change the thermal limit only and call it a day. I have a particularly bad 3600 sample that needs 1.38 to run at its regular boost clocks, does not take any negative voltage offset without crashing (practically non-existing margin), and as a result is very hot. I just put 68C as thermal limit in PBO and kept the rest of the values at default.
