Kaveri would be impressive if 20nm this year, not 28nm.
It's difficult to add embedded ram without blowing up cost (they already lose money on every cpu sale, that's how you rack up 1.18B in losses when the gpu side barely breaks even - you're being killed on cpu sales), where Intel offsets some with 22nm saving die cost. AMD will have a tough time with Broadwell which again will concentrate on gpu perf and leave us CPU people at 10-15% if lucky. I'd expect broadwell to double that cache to 256mb or something else that will dramatically up perf. They can afford it unlike AMD. At 14nm they will have plenty of room to play with gpu. If they released iris to the desktop A10-6800 wouldn't make a cent. They will do this move next time though, as Arm will be in their cpu space on all fronts by then (notebook, server, desktop) so they will be able to legally kill AMD with FCC coming to get them.
You could argue they could pull this now, but I think they'll just wait until they KNOW the FCC won't come for them at 14nm. By then Arm will be selling desktops, notebooks etc with chrome, firefoxos, etc and of course servers with all the Arm chips coming there too (Project Boulder etc, cavium etc, everyone planning arm server chips). If there are 6-7 cpu competitors, Intel can kill AMD and get it over with.