I also think that NVidia compute apps generally run a lot faster in CUDA mode than OpenCL.
I'm not sure if the app developers have just spent more time optimizing CUDA, or if it's something NVidia hasn't done much of yet. But I don't think you can just say that the AMD 79xx > NVidia at compute.
http://geizhals.at/783302 with 2666mhz
even with the same die size you get more MHZ on the cpu and gpu APU because of 22nm but they will be bigger dies as well.
also 256bit ram interface is on the AMD roadmap for the desktop in 2013 they can't go with 128bit interface forever.
so what? then the "APU" do have 100% more speed than now will you still claim that its not faster?
It's also worth noting that GCN is a completely new architecture which they really haven't tuned all of the performance out of yet so tests from 3 months ago fail to give the whole story. So while Nvidia has probably gotten most of the performance out of their 680 because it's an evolutionary and a simplification step off of Fermi, AMD actually has some real play left in terms of optimization.
Yes AMD is still working on problems with the new design. In theory it has got a hardware h264 ENCODER like kepler but there is still no software that can access it.