Doesn't Vega have tons of HPC dedicated transistors and features not present with Polaris? That would account for the less efficient die usage factor. And when you say "doesn't clock as high as Polaris," which iteration of Polaris are you referring? Clearly we see that clocking up Polaris with the GTX 580 had pronounced negative effects on perf/w. But what if Vega is 25% more efficient that RX 480 (which is more efficient than the "tuned up" RX 580)? That'll translate into still being about 25% less efficient than GTX 1080 TI, which means that at the same performance the average power consumption will be ~290 watts. With spikes and peaks, power consumption would be well north of 300 watts and would be spec-breaking.
People on here or rant and complain about other people making a big deal about efficiency fail to see what happens when dies and core clocks are scaled up. Unless AMD wants something hotter and louder than GTX 480 while simultaneously breaking the PCIe spec for a single card, they need to play within the confines of their efficiency limitations.
I too have been maintaining that Vega will be less power efficient than Pascal due to having a combined design for gaming and HPC.
That being said, Titan Xpp is not a 300W TDP design, (even though some sites have found it pulling 300W in a few instances), HBM also uses less power than a 384bit GDDR5X memory bus. Those two facts mean Vega can probably have another 70W or so for the core over Pascal while staying at or below 300W TDP. (I know AMD and NV don't be measure power the same way but I'm going 1st order approximation here)
Pascal having a smallish die, less than max board power, and a nice but reasonable cooling solution leaves performance on the table but makes for a lower cost to manufacture.
Now that could be used to provide a great cost to consumers or some pretty impressive profits to NV. As NV has released three versions of the chip ranging from $700-$1200 I'll leave it up to the reader to decide which way NV went.
Since NV has left power and price on the table, that carves out margin for AMD to compete on performance and performance/$ with a larger hotter, more expensive to manufacture GPU at the cost of less profit. Since profit does not directly effect FPS that's a metric I'm willing to ignore. The GPU just needs to be profitable. We already know from Fiji that a 600mm^2, 275W TBP, HBM, water cooled GPU is profitable at $649.
So I think you are underestimating the margins AMD has to play in to make a marketable performance competitive GPU with HPC capabilities.
(Now I could be wrong. AMD could pull another 2900XT but I think under current management that is highly unlikely)