What is the relationship between CNN and OMP parallel compute?

[Litchy]

In Intel mkldnn Tensorflow , it recommended CNN should use KMP_BLOCKTIME=0

And it also say that if OpenMP parallel compute is heavy, KMP_BLOCKTIME should increase in order to block other threads.

Normally, CNN is heavily using parallel computing?  (I suppose). So why CNN is recommended to use a smaller KMP_BLOCKTIME? 

[AthiraM_Intel]

Hi,

We are forwarding your case to Subject matter Experts.

Thanks

[Litchy]

Any ideas?