@szha, @DickJC123: TensorCore support for RNNs is [currently 
disabled](https://github.com/apache/incubator-mxnet/blob/225f71f744ac5e7bd29868b6d3ba0e4fe2527c43/src/operator/cudnn_rnn-inl.h#L47).
 We would really like to make use of this for training large RNN models.

Are there any plans to add support for it soon? 

Also: what about a flag `cudnn_use_tensorcores` (similar to cudnn flags in 
convolution). When `true`, it will use tensorcores even with `float32` (and 
handle internal downcasting or [use 
new](https://devblogs.nvidia.com/tensor-ops-made-easier-in-cudnn/) 
`cudnnSetRNNMatrixMathType(cudnnRnnDesc, 
CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION)`). If its `None`, it will use standard 
heuristics (environment variable + check for `float16`).

[ Full content available at: 
https://github.com/apache/incubator-mxnet/issues/9543 ]
This message was relayed via gitbox.apache.org for [email protected]

Reply via email to