A Domain-Specific Supercomputer for Training Deep Neural Networks

Communications of the ACM 

Mixed precision training with 8-bit floating point.