pengzhao-intel commented on issue #14173: [WIP] MXNet AMP (automatic mixed 
precision)
URL: https://github.com/apache/incubator-mxnet/pull/14173#issuecomment-476056958
 
 
   Thanks @ptrendx @eric-haibin-lin this is no FP16 GEMM from Intel MKL/MKL-DNN 
till now so I think we have to fall back to FP32 GEMM or simulated FP16 GEMM 
(maybe from `mshadow` or other BLAS library).
   
   The current design is good for APEX.  I am thinking about a general solution 
for mixed precision training, such as INT8, FP16, BF16, FP32. The different HW 
supports for the different data type, like INT8 in CPU and small GPU chip and 
BF16 on CPU and FP16 on GPU. So we need more flexibility for data type change 
or other library integration.
   
   We are going to support BF16 and the ideal situation is the user can switch 
FP16 and BF16 based on the devices transparently.
   
   We are looking into the changes and will give back more details later. 
   

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to 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