Info PSA: Don't Undervolt Ryzen 3000 If Using Auto Clock Multiplier

[rigg42]

Optimum Tech posted a video yesterday showing incredible results from undervolting a 3900x all the way down to 1v and leaving clocks on auto. It's very confusing because when you do this the clocks will go up in monitoring software while temps and power usage plummet and the CPU remains stable. I observed this behavior in my own testing and was extremely excited about until running performance benchmarks and seeing performance plummet as well. Myself and others brought this to his attention in the comment section of the video. He pulled the video and has since posted a follow up.

View: https://www.youtube.com/watch?v=2wM3obN2pAE




Possible explanation from the comments section of the video:

Respect for addressing and correcting the mistakes. The Stilt over at Overclock.net pointed me into the direction of the true explanation, which - according to him - has to do with "Clock Stretching", or in laymans terms, a proprietary AMD failsafe mechanism intended to deal with unexpected Vdroop (ah sounds familiar). In short, the board detects a lower supply voltage from the board (as we set in the BIOS), tries to compensate voltage by requesting higher-than-usual voltages for the individual cores, and makes sure each core still gets 1.3'ish volts. Once you undervolt too much, however, you run out of headroom and the CPU is actually short for voltage, and treats this as if it's experiencing "Vdroops". It fixes this by lowering the clockspeed on a nanosecond scale, but also does this a shitload of times because it's not just a Vdroop but a constant low voltage. All software still reports high frequencies, while under the hood this failsafe is throttling the CPU on a nanosecond scale in order to keep data integrity. So your effective clockspeed is lower without monitoring tools noticing.

 

[CountMike]

It's tuned for best performance, looks like XFR is fighting itself and PBO.

 

[rigg42]

So it appears with some CPU's a slight negative offset can be beneficial. I'm skeptical that this will carry across all workloads, and would be difficult to test effectively, but it's always nice to have more testing data.

View: https://www.youtube.com/watch?v=JscRwIH3OAY