sbodenstein commented on issue #12463: cuDNN 7.2 Tensor Core support
URL: 
https://github.com/apache/incubator-mxnet/issues/12463#issuecomment-419409686
 
 
   @szha: the most obvious option is to have a new flag for convolution + rnn 
ops, something like `use_tensor_cores`, which both calls 
`cudnnSetConvolutionMathType(cudnnConvDesc, 
CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION)` and selects a compatible algorithm 
(probably overriding `cudnn_tune`):
   
   > One limitation is the use of ALGO_1 (IMPLICIT_PRECOMP_GEMM for forward). 
No other convolution ALGOs in cuDNN make use of tensor ops yet.
   
   It would be great to get feedback from some NVIDIA folks (@DickJC123 
@ptrendx) as to how they envision frameworks using this stuff (and it will be 
even more important soon, as I believe that even consumer-grade Turing GPUs 
will have Tensor Cores).

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
[email protected]


With regards,
Apache Git Services

Reply via email to