yes, it is said to have a cluster computation code already built-in. It is the recommended approach if you are using a Tensorflow. It is mentioned in the course module from Google (https://www.coursera.org/learn/intro-tensorflow/home/week/2)
If a premade estimator is used it basically does parallel computing to run it on various different hyperparameters and gets the best result.
Even for a custom estimator, a certain range can be specified to test the parallel.