I think scale will have an impact on these rates.
PowerGPU are a system builder, in the short time they have been building Ryzen 5000 PC's how many have they sold? a few hundred? so they order several CPU batches of 50 units, the first two batches they may have had one dead CPU, so "0.5%" they order another two batches and there are 3 dead CPU's "3%" and another two batch has 4 dead in one 50X batch and two in the other "6%"...
The thing here is if you're only shifting a few 100 units you don't really have a sample size that's big enough for a conclusion, if you're shifting tens of thousands you probably do and they ain't necessarily going to agree with eachother.
We do actually have enough data to start doing this now. Am a statistician by trade so did a 99% confidence interval for the PowerGPU 5950X and 5900X on the back of a napkin and it gives a range 3.6% to 20.4% rejection rate for their CPUs. However combining PowerGPU with Mindfactory.de's dataset you get a much narrower failure rate:
a 99% chance it's between 0.44% to 0.67%.
The trouble is we also need PC world's data, given they suggest a large manufacturer has a rejection rate of 2.9%.
Looking at it overall I'd suggest there's
both low volumes for some outfits leading to unusual findings and also different failure rates for different batches (e.g. whomever PC world's large 2.9% source is, vs mindfactory). On top of that there are obviously differing definitions of failure.
My sense is that the true 'fail rate' is about of 0.44% to ~3.2% most likely, but I really need all the raw data.