Onur Mutlu’s course at ETH. The reason it’s on the list: most of my work sits on top of GPUs and accelerators that I treat as black boxes. The point of going through this course is to stop doing that - to understand what “memory bandwidth” actually means and why a particular kernel is slow.
Pairs well with the Programming Massively Parallel Processors book on the bookshelf. Plan to start once I’m further along on the deep-learning course.