Gamasutra was paid to do an interview with Intel's Mike Burrows, who is Senior Graphics Software Architect Manager for the Larrabee creator. Paid content is usually not too interesting, but this interview contains some interesting snips. You can check it out over here.
It's questions like, "How do you map an algorithm to a more constrained set of computing resources, while also being significantly more flexible in terms of the width of the data?" That's SIMD: Single Instruction, Multiple Data -- maximizing the effective compute power per instruction. That's something that's been on the CPU side for a reasonable amount of time already.
We have a lot of vector computes on the CPU, but the CPU, particularly from the game side, is busy trying to deal with a lot of the complexities of game-specific problems.
Longer term, yes, we probably want to be able to load-balance those two sets of discrete compute units on the graphics card and the CPU, but having more flexible systems long-term is just as important. The buzzword in the industry today is heterogeneous computing. It's a matter of finding the most appropriate set of computing technology to run a given algorithm. That is very exciting to me at the moment.