A GitHub commit to AMD's ROCm GPU software suggests AMD will support the BFloat16 numeric data format for deep learning.

[Only registered and activated users can see links. ]