Modern Intel CPUs can compute matrix operations more quickly through special instructions. This is done using the Math Kernel Library for Deep Neural Networks (MKL-DNN). Out-of-the-box TensorFlow does not exploit those instructions. Using them requires either compiling TensorFlow with the right options or installing a special build of TensorFlow called tensorflow-mkl.
Information on how to build TensorFlow with MKL-DNN is available at https://www.tensorflow.org/. Note that the toolkit currently only works on Linux.
