Memory swizzling is the quiet tax that every hierarchical-memory accelerator pays. It is fundamental to how GPUs, TPUs, NPUs, ...