I've yet to see a Vega56 within 10% of the VII. Any sources for that?
If you want a huge sample size, 3D Mark's online results database.
It may be a synthetic test with no relevance to real-world game results, but in terms of A/B testing two identical architectures that differ only in bandwidth and clockspeed, it's perfectly valid.
There are almost 100 Vega56 results scoring north of 3800 points on Time Spy Extreme from a variety of different users, ranging from 1600-1700Mhz core clocks.
Radeon VII scores about 4200 on the same test at stock speeds.
(I picked 3800 points because that's 90% of the Radeon VII's 4200 score)
I've seen videos of modded, waterblocked Vega56 at 5-10% faster than Vega64 stock in a battery of gaming tests, so that correlates because the VII is typically 15% faster than a stock Vega64. I think it was Steve over at GamersNexus but it's been a while since he was messing with the Vega56. It's old news now
I'm not sure why you're so surprised - Radeon VII is a die-shrink, not a new architecture. if you clock a Vega56 close to a Radeon VII, the only real difference is 56 compute units vs 60 and the extra memory bandwidth on the VII is largely unneeded, because Vega's HBM2 is already an insane amount of bandwidth. As Krogoth implies it's not bandwidth that's the issue when the VII pulls ahead, just those extreme examples where the 8GB cards (he cited the 2080) actually run out of VRAM.
The whole discussion is getting a bit silly though, because for the price of a single VII you can almost buy three Vega56 cards, and Crossfire caveats aside, that's a lot
of GPU power.