Perhaps the form factor will allow for far more dense clusters? and that its on dram boards, the cost for these is certainly low-ish.

AFAIK GPUs are floating-point centric, meaning you'd have to reimplement some functionality on top of FP. Plus it seems that each 'cell' can be it's own state machine, I also believe that GPUs are less fine-grained than that (you allocate clusters of units to compute a programmable shader).

Actually the throughput of integer instructions isn't that much lower than floating point operations, at least for arithmetic and bitwise operations.

http://docs.nvidia.com/cuda/cuda-c-programming-guide/#maximi...