I assume you are talking about Tensorflow with MKL-DNN optimizations.
I would not be able to comment on the package sizes. That would be a question for the TensorFlow forums. However, this section on their website answers your other questions: TensorFlow Documentation
According to their documentation, Intel MKL-DNN optimizations are only available on Linux for now.
But accroding to https://software.intel.com/en-us/articles/intel-optimization-for-tensorflow-installation-guide , there is an installation for windows for MKL. I need to install it as offline. But could not be sure if I am correct conda packages as file size are low in space.
Thank you for pointing that out. Intel supports Linux builds because the Windows and Mac builds are less reliable. If you really need to use them, you can, but the reliability may not be guaranteed. "tensorflow-mkl" is the correct package but it is not managed by Intel. It is managed by the anaconda community themselves. So I cannot answer your question about package sizes as an Intel engineer, however, I will try to do so as any other engineer/developer.
Visiting the Intel website that you linked and the anaconda repository made me think that may be tensorflow-mkl is a selector package which still requires main tensorflow package to be installed. To explain what I mean:
https://anaconda.org/anaconda/tensorflow-mkl/files shows the tensorflow-mkl package, and the description says that it is a "Metapackage for selecting a TensorFlow variant." This implies that it needs the package "tensorflow" to be installed as well.
Also to clarify what a metapackage is, refer to this: https://askubuntu.com/questions/66257/what-is-the-difference-between-a-meta-package-and-a-package
I hope this helps answer your questions. Thank you.