I believe there's huge difference between CPU-offloading for common tasks, not requiring superior speed, and "I need to have the fastest possible" type of computations.

Larrabee might be superior solution for first type of tasks. Like eg. 2D acceleration under Linux. Currently it's done in complex way - first transforming 2D to 3D, then going through OpenGL transforming 3D back to 2D. With general programmability of Larrabee you could just load huge part of conventional 2D XWindows on Larrabee itself, without all that transformations, and with keeping lots of tried-and-true code. Same with sound or printing system, computational parts of these can be put on Larrabee. Or any other code that is non-performance critical, but it's beneficial to offload from the CPU.

But when you aspire to have the fastest computations possible, then "optimized for parallelism by default" models of OpenCL might be much better choice. Converting conventional code may not even work, as it containst a lot of performance stoppers, like shared data (locks, static class members) or liberal calling functions from hordes of libraries, that are not ported to GPGPU, or doing ocassional system calls interleaved with computations.