Xeon Phi for password hashing

Likely more straightforward than defensive use of GPU
Should be able to efficiently run 8 bcrypt instances in SIMD vector elements per core (as limited by 32 KB L1 cache, leaving half the SIMD vector width unused), thus 400+ concurrent instances per chip at ~2.0 GHz
May be the bcrypt killer, especially if it is affordable
For defensive use, should run a hash function with sufficient parallelism at least to use a 512-bit vector
Probably can't hold a host-unreadable local parameter
Not intended as a security device, almost certainly allows DMA