I can be a chip-head sometimes when it suits me, but most of this discussion is way over my level of expertise. You can read the 16-page writeup over at Anandtech if you want, but you better be ready to deal with terms like “16-wide Vector ALU” and “cache coherency.”
See what I mean? From what I understand, Intel is looking into creating a sort of chameleon CPU with a ton of parallel, non-specialized cores that can be organized in software to fill the role of a GPU or, conceivably, a physics processor or what have you. We’re talking 24, 32, 64 cores here, being organized according to whatever is needed by an application or instruction set; DirectX could reserve half, drivers could include special pipelines for shadow, physics, or AI calculations, or at least that’s how it would work in my fantasy world. There’s a lot more info at the link.
It’s not a full volley at Nvidia; doubtless by the time Larrabee is a reality, Nvidia will have made its counter-move (ATI’s been working on a mutant hybrid CPU/GPU for a while), but it’s an intriguing approach and could yield some interesting results down the line.