Loss of model performance when reducing numerical precision (e.g., from 16-bit to 8-bit), caused by accumulated rounding noise.