This can't be the case... I have 8 nitro 64s on a rig pulling 1112w at the wall - maybe ~35w for system, so that's <135w per GPU on average, each doing 18.5 kh/s (850/1107/818/L18+18.)  I also have a ref 64 and a flashed ref 56 (850/1107/818/L22+22) doing the same on another rig.  These were all bought over a period of many months during the vega stock shortages, so they certainly weren't all binned together.

Granted I'm on platinum PSUs and a 240v line, so your power needs could be slightly higher, but prob not more than 5%-10% max.  I'm also on linux, but my tests on windows (18.12.x drivers iirc) showed similar results.

Really, memory manufacturer is the only thing I see impacting performance - mine are all samsung, and i expect based on widespread reports, and my experience w/ polaris, that hynix would underperform.  Drivers are the other possibility - though I have yet to see any versions since blockchain significantly impact performance for any algo (other than on xmr-stack derived miners.)