I'm sure if they threw enough money at it they could given the engineering talent they have. Unfortunately they're stuck in the mindset of avoiding fixed function hardware and they're also not willing to give up programming flexibility in order to get more performance. Nvidia and AMD are able to get so many execution units (and thus gflops) on their GPUs because the execution units are primitive and use a simple instruction set. Intel is trying to use x86 cores, and because of the greater complexity of the x86 instruction set they aren't able get the same performance with the same die size, although it does allow for greater flexibility.