It looks like AMD mainly plans to use integrated GPUs for mobile processors, to save power.
On top of that, mobile is far less demanding than a pimped-out high-end GPU. This is the main way that DAAMIT will address the memory bandwidth disparity by lopping huge amounts of bandwidth use out of the equation, low end embedded parts need low end embedded bandwidth. These parts are, duh, better to embed.
The diagrams shown had the GPU with a separate cache, but may not make it into the real world. Intel internally seems to think this is where they will get a big bang out of the integration of graphics in Gesher, cache, cache and more cache.
Also added to the mix is AMD's process technology. ATI can go from a process made for a generalist outsourced fab to a more tightly tuned product. This can add up to big gains for the graphics side, basically a free process node gain.