CPU usage far too high and training inefficient

okay, I see. But do you have any clue why the code is running so much slower when running on 80 CPUs as opposed to when running on one CPU?