Also Parallella is MIMD whereas the Phi (according to WP) and GPUs are SIMD, so Parallella is more general (surprisingly).
I've always thought that was more future proof, really. I'm no engineer, but it makes sense to sacrifice chip complexity for scalability. Plus, it's probably easier to write/compile against. No special SIMD instructions or significant architectural changes from chip to chip (theoretically). Just simple cores at different amounts.