This might be a bit of an incoherent rant, but I'm mostly putting this down because I've had this (and others here too) on my mind ever since Vega got announced and I'm still trying to make sense of the whole situation.
It's been a while so I don't remember the exact details, but I think Vega first got mentioned in an interview on AT with Raja(?). "We have Polaris 10 and 11, 10 being the bigger one, and we also have Vega 10 and 11" It was then speculated that Vega 11 would be the bigger of the two.
Time and speculations went on and the seemingly logical guesstimate would be that since P10 being of similar size to Pitcairn, that it was going to be its successor and competitor to GP106. As the mid to low range was established, Vega could only be two higher end chips. My thought, and other people's, was that perhaps small Vega would end up being in the range of Tahiti (=>350mm²) to go up against GP104 and big Vega being somewhere inbetween Hawaii and Fiji (>450mm²) as a GP102 competitor.
Then we had that 64CU rumor floating around since forever and you'd think that one would be small Vega. You take Fiji, shrink it to 14nm and throw a bunch compute (if Fiji even had that much) out because this would be primarily a gaming chip. I know I'm oversimplifying process shrinks and not all parts of a GPU scale that easily, if at all. To me it seemed plausible that this could end up in that aforementioned die size range. Nowadays we're starting to see Fiji stretch its legs in same games and it actually manages to keep up with GP104. So far so good, I guess.
Big Vega on the other hand still seemed like a mystery to me. Are they going to throw more shaders at it? I think I saw some people say like 80CUs or something. I guess it could work given HMB2 doesn't need a memory controller as wide as Fiji's if they use two stacks and you still have a smaller process.
Now here's the confusing part. We've been shown a Vega chip being about >500mm² with two stacks of HBM2. They called it RX Vega which pretty much means one chip (with its derived SKUs). That basically confirms it being the big one, unless AMD are insane enough to go even beyond that. After all this talk of NCU and all the other tweaks they did, it seems a given that all this stuff will eat up extra die space. Perhaps 64CU does make sense. Again, Fiji without its bottlenecks and a decent bump in frequency might be able to compete with GP102. Then they've also shown that Doom demo just beating a 1080, which, as mentioned plenty of times, could be down to an ES running at lower clocks, immature drivers, etc etc.
So now the question is, what happened to small Vega? Did they decide to can it because they couldn't feed it properly? HBM2 on a GPU that should slot into the GP104 range might not be economically viable. GDDR5 could be an option, but since Polaris barely manages to get by with its 256-bit bus, they'd probably have to go 384-bit wide. I don't think it'd look good from a marketing POV that a cheaper/slower card has 6/12GB memory while RX Vega only has 8GB. That kind of memory controller would also eat into their power budget too. I guess GDDR5X + 256-bit is an option and I think I read Polaris supports it, so who knows.
I guess you could go into tinfoil hat territory and say the demos they showed were actually running on small Vega while the package they showed was the big one to throw people off, but that sounds too ridiculous.
I could be totally off on any of this, so feel free to point it out.
TL;DR: I don't know anymore.