cuDNN 7.2 has simplified the usage of Tensor Cores for convolutions and RNNs, and explicit casts to float16 can be avoided: https://devblogs.nvidia.com/tensor-ops-made-easier-in-cudnn/

Are there plans to expose this in MXNet already? If not, lets discuss a design in this issue.

@szha