rahul003 commented on issue #11842: Strange results from fp16 benchmarks on 
Alexnet on CIFAR10
URL: 
https://github.com/apache/incubator-mxnet/issues/11842#issuecomment-409763085
 
 
   For CIFAR I had observed similar performance. 
   
   Here's a related thread with my observations and logs of nvprof. 
   https://github.com/apache/incubator-mxnet/issues/9774#issuecomment-366835890
   I was told here that the dimensions of convolutions for CIFAR are not 
suitable for tensor cores. 
https://github.com/apache/incubator-mxnet/issues/9774#issuecomment-367057821
   
   I'm not sure if we can improve fp16 support for CIFAR. @OneRaynyDay 
mentioned that he's transforming the image size to 224x224, in which case I 
don't think why he doesn't see any speedup. You can verify with synthetic data 
of 224x224 size to see that it does improve the speed. Need more profiling to 
understand what's going on here. 
   
   I request a committer to add Performance label to the issue. 

----------------------------------------------------------------
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