How does Apache MxNet used by AWS compare with keras, theano, tensorlfow

I was looking at this video from Amazon AWS where they are saying that they are building their stack on top of Apache MxNet. I was wondering how does that compare with with the frameworks that are being used in this course.

Has noone used or heard about it ever? AWS picked a library which no one uses?

MXNet can be used as a Keras backend (instead of Theano or TensorFlow):

MXNet is great. I don’t think there’s any major reason to use it over what we use in the course, but there’s no reason to avoid it either…