There is a very large market that needs more CPU performance. It's bigger than the one that needs more GPU performance.
If there were no CPU competition, Intel wouldn't have cycled through so many cores as fast as it has.
In the case of graphics, Intel's focus on the low end was such that neither Nvidia or ATI really tried to go there. It went exactly as far as it knew no one would challenge it, integrated graphics for its value segment.
It remained focused on what it does best, why push it?
Most of their CPUs don't go to gamers. The volumes x86 chips have are bigger.
Not all servers do all their talking on the internet.
A lightweight front-end server can have any number of application and data servers behind it. Niagra is somewhat better in some places than previously thought, but a lot of stuff happens behind what's visible to the web that takes serious work.
I don't think the system that runs Google's indexing service would do too well, considering there's a fair amount of heavy matrix math that goes on after a page has been parsed, and Niagra doesn't do multi-socket.
They opted away from heavy cores because they felt the gains weren't worth the costs. It wasn't until recently that was the case.
You're asking "why did Intel increase performance on its cores all those years ago when it was worthwhile to do so?"
It's not really that revolutionary. Heterogenous computing systems have been done before. They've been done with general purpose cores tied to smaller cores with high dsp or fp performance and local store.
It's just that it's all on one chip, and that's not really anything more than changing a physical location a couple inches.
What would the GPU do? If it's not being fed a command list from somewhere, what can it do?
It definitely sucks at that, so where's the gain of having 48 shader pipes and the ability to produce commands enough for less than one?
