After a couple of silicon spins, the Larrabee hardware actually came out looking pretty good. For a radical new architecture (shared cache, ring-based, many-core etc) I think that's quite impressive. The problem was that the million or more lines of software needed to make it behave like a DirectX graphics part just weren't good enough. And that's what it would have taken to make this successful in the market. Fixing those problems was a long way off, so it made sense to cut our losses, at least for now, and move on.
-- Intel blog on Larrabee cancellation