This improves GPU memory-efficiency and allows you to train much larger models on fewer GPUs.