CPUs forte is serialized tasks. GPUs is parallelized tasks. Add memory for frizzante taste. You follow me?
You integrate all in the same die, thus avoiding many bottlenecks like MB copper lanes, difference in speed and latency between buses/parts.
Add to that a compiler capable of efficiently handling and tasking your code to the part it's more cool to it, and coders/programmers capable of further optimizing and tasking that code, and you see gains all around.
It's the future, boys/girls. Don't go crazy because you see AMD as the first releasing such computational paradigm change, and because first implementation is not a good performer. FIRST IMPLEMENTATION is the tag. They just have to keep working on it. Intel and Nvidia are about to turn the corner on integration also...Nvidia's Maxwell is supposed to have ARM(CPU) cores in it. Tim Sweeney doesn't have any problem with the technology...in fact, he's saying he loves it. The problem for him is DirectX/OpenGL, which have to be changed in order to be more in consonance with this new computational path, to see something more than those marginal gains...but as he states, changes to the main graphic libraries come slow, so the ball is on Microsoft/Khronos Group's roof now.
OpenCL 2.0 and C++ AMP, as programming languages capable of doing such flavour based programming(serial/parallel) automatically are in its infancy, so there goes another "problem".Edited by Nonehxc - 8/27/13 at 3:14pm