Really, really depends on the PoW algo. Cryptonight, for example, has vega blowing past Titans, and equihash is posting very competitive results. Dagger-Hashimoto (Ethereum’s PoW) is actually memory bandwidth-limited, not FP32. That’s not to say a lopsided setup will magically get better, but we’ve typically seen any small improvement in memory conveying a relatively proportional advantage on almost every card.
(The 1080 and other GDDR5X cards are outliers because none of the cuda mining authors and optimizers can figure out how the hell G5X works, this isn’t the case with HBM)
Also note that geth --mine is not what the majority of folks are hashing with, and isn’t representative of performance for optimized setups. Big miners don’t use publicly available software for the most part.
What’s also puzzling is the quoted performance by AiB partners on their internal tests. I don’t have direct lines to them, but Steve Burke and others have independently validated the high hash rates. Some of my contacts in the mining business have intimated that vega 56 should be about 40-55 MH/s on a large scale rig according to their FE hacking and testing, and 64 should be around 55-65 without going into it in more detail.
My guess would be that the differences mainly come down to how far they can push the HBM on each chip.
tl;dr is that we’re gonna need more time to see how it shakes out.