You selected carefully the facts which support the GPU in the CPU approach omitting the other
facts: the memory bandwidth of the RAM in the GPU is in fact far higher than main memory's
bandwidth and the quantity of RAM present in a GPU is bigger than CPUs caches.
So if your textures fit in the GPU's memory, in many case the standalone GPU will be faster
that the GPU-in-CPU..
IMHO, the GPU-in-CPU will be interesting for the 'low end' of the gaming market which is also
the one with the biggest number of users (maybe also for some scientific computations this
remain to be seen).