One thing I forgot to mention is that the next-generation Xeon Phi will be implemented in Intel's 14-nanometer process technology--which should shrink the die even though the whole package may stay the same size or even grow depending on how many memory die are added is alongside the processor.
Colin, something is not adding up for me...in teh title you say "customize", that menas to me many different versions for different applications, customers, etc...but then you add "highest integration" which implies one die that can do many things...kind of contrary to "customization"...I must be missing something here...Kris
I have to wonder who those key customers might be and how much of the market they would involve. I know that hard-core PC gamers like to overclock their systems, but this would potentially go well beyond that. Is the Intel architecture really modular enough to support a chinese menu approach to CPU design? What kind of volume would be required to qualify as a "key" customer? And finally, what are the implications for the software tool chain to the modifications?
Good questions all, but I don't think many answers will be available until next year when we get some examples. To hazard a guess, I'd say Google, Amazon and the like are buying Xeons in suffiencent quantity to be a "key" customer. Also some supercomputers use a lot of cores all by themselves, such as Milky Way which just won the Top500 with over three million cores.
Colin, any insights on how this change in business model shall impact ARM's server dreams as ARM claims one advantage that OEMs get by adopting ARM servers is that servers can be customized as per the requirements?