The rise in interest in deep learning chips for training and inference has reignited interest in how reduced precision compute can cut down on energy, bandwidth, and other constraints inherent to ...