Comment by ptasker

I would imagine using a GPU compute instance would bring this time down significantly.