Just because developers want something doesn't mean it is the best implimentation of hardware. Mind you developers thought faster serial processors was the best path as well as beefier single cores.
Just because, in theory, a single processor would produce best/easiest results, doesn't mean that is the best use of the hardware.
e.g. Repi mentions a OOOe being a dot in a sea. Maybe this same concept can be achieved through a very fast link between a GPU and CPU.
Anyhow, my problem is first that while it sounds great in theory, we aren't seeing any benefits of a single CPU/GPU now and we DON'T know how it stands up against a current CPU GPU. And the 2nd issue is that a single chip has fewer execution units for the same $.
So you pay the same $ for lower peak performance and no one has shown that the advantage of a single die can overcome this issue to be at least on par. That is a big challenge--and there is always the counter arguement that why cannot a 2 die solution evolve (e.g. optical link) to get the same concept without the sacrifices of a single die.
Just because, in theory, a single processor would produce best/easiest results, doesn't mean that is the best use of the hardware.
e.g. Repi mentions a OOOe being a dot in a sea. Maybe this same concept can be achieved through a very fast link between a GPU and CPU.
Anyhow, my problem is first that while it sounds great in theory, we aren't seeing any benefits of a single CPU/GPU now and we DON'T know how it stands up against a current CPU GPU. And the 2nd issue is that a single chip has fewer execution units for the same $.
So you pay the same $ for lower peak performance and no one has shown that the advantage of a single die can overcome this issue to be at least on par. That is a big challenge--and there is always the counter arguement that why cannot a 2 die solution evolve (e.g. optical link) to get the same concept without the sacrifices of a single die.