Q1. Why is there regression in performance/mm2 compared to Polaris?
A1. There's 45MB of SRAM (think of it like Intel's separate Crystal Well, or Xbox1's EDRAM) on Vega, an unprecedented amount for a GPU to have on-die. Some of this contributes to the front-end changes which lifts Vega's performance at lower resolution & CPU bound scenarios. The rest, make up the bulk of the HBCC to enable it to function. This feature is basically useless for gaming graphics. It's purely intended for big-data acceleration, something even NVIDIA's P100 Tesla is very weak at, once datasets gets beyond 32GB, it stalls and performance drops off severely. A big chunk of die-space is devoted to a feature that do not contribute in a meaningful way to gaming graphics performance is the correct answer here. In combination, there's some advanced Vega features which RTG's driver team have trouble with either due to a lack of time/manpower or talent. DSBR has an optional deferred rasterization step, it is not active, and efficiency drops because it's brute forcing some steps.
Q2. What can we expect for Vega 12nm refresh?
A2. FP64 big-Vega for scientific HPC markets. Maybe Vega 11 finally show itself for gamers. AMD has a better chance now with their software stack. Though there's a good chance AMD may simply abandon these and push straight to Navi, ASAP.
Q3. Why not make multi-uarch or chip for gaming & professional markets?
A3. Expensive. Return on investment is volatile and uncertain. It really only works once you have marketshare and mindshare dominance to ensure gamers buy your gaming-optimized chip, and professionals buy your pro-optimized chip. It's something NVIDIA can do and they will likely do it soon, a tic-toc (Volta -> Ampere) approach to their architectures since they already have a strong lead. If you are in AMD's position, where both markets are uncertain, increasing expenditure could be seen as wasting resources which they sorely need to ensure Zen succeeds. Now, AMD also needs to remain focused on Zen+, Zen 2, and future architectures to have a strong core business.
AMD's future is solely in multi-chip, cheap to mass produce, cheap to link up either via an interposer or cheaper & more flexible if they use Intel's EMIB (cross-licensing no issues). If they become profitable enough, a small chip is much cheaper on R&D and so they could potentially have multi-market optimized small chips as well. Though this applies to Intel & NV in the longer term as well, so to me, rather than say this approach is AMD going on top, no, this is AMD being competitive. They are still the underdog for the foreseeable future.