Coupling these two designs
enables us to train large models efficiently and effectively: we accelerate training (by 3x or more) and improve accuracy.