Gradient Tape is much slower than keras fit method

Question

I have trained a simple Dense Layer on gradient tap and keras api both, but gradient tape has stuck on local minima but fit() method decresed loss till global minima, both h