I was wondering if there are any problems with using different gpus with DataParallel, for example, 1080ti and titian Xp? Or does it just work at the rate of the slowest gpu?
It will split the work equally to all GPUs. So that means that yes it will run twice as fast as the slower. And you won’t be able to use more memory than the smallest have.
It should still be almost twice as fast as using only one of them in this case. Given that the perfs of the two cards are similar.