Tensorflow: Difference performance between keras.Dense and tf.matmul

Question

I have an input with shape 16x23x43, where 16 is the batch dimension.

I tried to implement some kind of MLP (only important parts are listed here). First