I like big batches ... Spent a morning trying to optimize GPU usage. Turns out my batch size was too small .... GPU stayed cold. Cranked up the batch size from 10k → 100k rows. GPU memory barely budged, but utilization jumped. Anything smaller and you’re paying the kernel launch cost for nothing. Moral: GPU is like a freight train. Fill it or don’t bother. [[ML]] [[Serendipity]]