Intel's paper on Larrabee in the ACM's Transactions on Graphics, published to coincide with Siggraph next week, is now up. Although aimed at 'visual computing', Intel stresses in the final pages that it could be used for a lot of other things. The obvious one is finance, as trading systems already use FPGA and GPU accelerators to speed up calculations. Less obvious is this one:
"Larrabee's highly threaded x86 architecture benefits traditional enterprise throughput computing applications, such as text indexing [Intel's emphasis]."
I'm not sure when text indexing became a "traditional enterprise application" but it's precisely what search-engine operators like to do a lot of. Maybe it's no coincidence that one of the researchers who wrote on asymmetric multiprocessing in IEEE Computer this month is now at Google.
As far as 3D graphics goes, it looks as though Intel has taken a similar approach to ARM to try to reduce the amount of memory bandwidth issue in 3D graphics. It is using tile-based algorithms to split up the processing rather than the immediate mode favoured by high-end GPUs. The caches can also be loaded by software - their cache behaviour can be overridden, which should help in 3D work.