Right, they can fit so many cores on a chip because they are very optimized for one type of work (streaming vector math) and don't have things like branch-predictors that would make them good for general purpose work.