IMO the likely cause was some kind of cache bug due to manufacturing. Threadripper processors were binned for tighter cache timings (they had the same timings

IMO the likely cause was some kind of cache bug due to manufacturing. Threadripper processors were binned for tighter cache timings (they had the same timings as the 2000 series desktop chips) and thus didn't suffer from it. EPYC was on a different stepping entirely and didn't share dies with the consumer or HEDT processors. I suspect they likely had the tighter cache timings although I'm not sure on that. Regardless, it's also possible that they were just binned out, or nobody ever encountered it. EPYC was not a very high-volume product, most people did test installations and said "yeah, we'll wait for Rome". Quad-numa on a package and lower-than-Intel IPC on a pretty poor node was not a winning formula. Once the problem was realized, desktop Ryzen processors started getting binned for it as well. A few slipped through here and there, so it wasn't a manufacturing change, I think binning is the most likely explanation.