Looks like they only used a single GPU for training though - you could trivially get that time down a lot if you used one of the cloud ML platforms (and were prepared to pay)