Also, you keep throwing around the "5-6X is not enough" but don't go into WHY. WHY won't it be enough?
It comes down to what you expect for graphics next generation. Just consider that moving from this gen's 720p/slight sub-HD to 1080p will require more than double the pixel processing and rendering bandwidth. Now factor in that current gen consoles are stuck with 32-bit per pixel formats. FP16 (64bpp) is more than just "HDR" - the precision for shader effects is just that much more important once you start piling on the post-processing and what not.
So even before we can begin to drool about higher quality graphics, we're already worrying about a rather significant jump in resource requirements just to maintain the current gen quality of shading/effects etc.
And since we are talking about just pixel processing, that number is fairly simple to derive, even if it is rather naive and simplistic. TBH I hate these multipliers being thrown around recently just because they have no real context. But here: 1080p is 2.25x 720p (bigger if you want to consider sub-720p current gen titles), and FP16 simply requires double the speed (or higher cost in transistors for full speed vs 32-bpp formats) and double the memory and memory bandwidth (who knows what the edram or main memory situation will be, that's not the point).
In short: if people want 1080p and high precision pixel processing as starting points, then we already require at least 4.5x the processing power of current gen. Then you'll obviously need much more if you're going to expect leaps in lighting and shadowing quality.
Of course, ALUs are fairly inexpensive. We can never get enough Z/Stencil performance as well as memory for those more complex shadow buffers (we are soooo far from having decent shadows in the general case).
Anyways. That's all moot if devs end up targeting 720p, which really isn't an awful idea since they can also more easily accommodate MSAA, and it's not like the GPU hardware scalers have been stale for the last 7 years. Devs will just naturally be able to do a shit ton more with less than half the shading or fillrate requirements.
----------------
FWIW, if we take a look at the geometry side of things, the base performance considerations haven't really seen any significant leap. And by that, I mean that the setup rates are still 1 triangle per clock, and so we're ultimately limited by the clock speed of the GPU. And I think I've heard enough about people trying to cut down on power consumption of a high end GPU by gimping clock rates. There's always a trade-off somewhere, and it's perhaps not so simple to gimp the GPU in such a manner if we are to expect more from the geometry side. Certainly, I don't expect to see integer multiples of Xenos' clock speed, but it's just something to ponder (shadows, particles, decals, multipass rendering and other things will eat up the setup rate).
The multiple geometry engines being added to GPUs now are only useful when tessellating, and we're quite a ways away from it being adopted universally in the artist asset pipelines. Tessellating just puts an extra load on the setup and shading anyway.
Fortunately, ALUs are cheap, so at least we'll have considerably more vertex/geometry shading power than this gen to actually do some more complex manipulations of vertices.
.
