The number of training examples used in one iteration of model training. It determines how often the model's weights are updated. Larger batches arguably provide a more stable gradient estimate but require more VRAM.
Fundamental training parameter.
The first thing you lower when you get an OOM error.