Why bigger batch sizes don't always mean faster training.

A larger batch size processes more data per step, which sounds like it should mean training finishes sooner.

But larger batches also demand more memory, may require adjusting the learning rate to keep training stable, and, past a certain point, stop improving hardware utilization.

Bigger isn't automatically faster. Past the point where hardware is fully utilized, a bigger batch is just a bigger number, not a shorter timeline.