I didn't read anything in the paper that mentioned training time. For 100,000 iterations, approximately how long to train from scratch?