If there is randomness in some part of the training process (and most surely there will be), then it makes sense. Different trainings may end up in different local minimums/maximums, so there is value in running multiple trainings simultaneously in different machines.