One aspect Ive been examining recently is effects of having more CPUs within the same memory space (Larrabee has all the CPUs able to access a single main memory...) versus using multiple 'few-core' machines to scale up (where more network calls are required to transfer data).

Scaleability for the AI Im considering will STILL require multiple machine even when they are 'many-core' types but at least some processing can be grouped locally and gain efficiency from using memory-memory data access for inter object interactions.

I would assume that some basic processing efficiency is lost on these 'many-core' systems due to cache misses stressing the memory controller (when there are up to 32 CPU cores running heavy random access) but much of that may be offset by heavy use of intercommunications local to the machine and shared world representation data.

It may be that some processing types will have advantages being run on the 'many-core' machines (mass VM/FSM controlled simple behaviors), alongside
heavyweight AI (complex behavior planners - human like...) run on more conventional (quad?) CPU systems.