have a look here:
I cannot seem to get reliable performance out of TF.
I am making sure the model will fit into allocatable memory before running Tensorflow.
But with the allow growth option turned on it will overallocate and segfault.
Do you have any advice for me or the Tensorflow team?