HappySqurriel on 05 December 2008
The next stage will (possibly) be to return to a CISC mindset and increase the instruction set of the cores using more complicated instructions in order to increase performance in specific situations ...
Consider 4x4 Matrix multiplication which is used heavily in videogames, in order to multiply two matricies together it takes 64 floating point multiplications and 48 floating point additions; it is easy to reduce this to take the time of 4 floating point multiplications and 3 floating point additions through parallel execution on a single core (which can be improved further using other methods) but this would (dramatically) increase the size of a core.
