[rdavl]

Hmm, I always saw each of them as an answer to completely different problem, so I have hard time contemplating this scenario.

How exactly would CPU be more like GPU? What would happen to branching and out of order execution pipelines? Do you see today's complex superscalar cores so cheap and small in the future that you could have thousands of them on single die? What other way is there get vector computations to work as well as scalar computations on same hardware? Would you rather have few superscalar cores and a lot of SIMD's around it? Or do you envision a completely new architecture?

I'm just trying to get a picture of how this would be done. Any reading material would be welcome.