The study I posted above indicates some are quite good when compared with the "gold standard" of lab-grade PSG. In particular the Whoop 3.0 is dead on. Caveat that the Garmin data isn't from the newest generation of Garmin HRV. These graphs are showing the deviation from PSG, so a horizontal line would be perfect agreement.

Edit: Important context with the graph that the Whoop 3.0 was the only device that provided raw R-R intervals. For the rest the study had to use the output of proprietary black box systems. So it's not apples-to-apples, and the underlying sensors may not be primary source of error. E.g. the authors speculate that the Garmin may have done relatively poorly because it was mixing in waking HRV with sleeping.


In separate sections of the same study, they're also fairly good at estimating broad sleep metrics. Less good at identifying specific phases of sleep.