Interesting post. It would be much better if the author included a few code snippets to show how to identify the failing GPU during training.