Sending model weight updates between devices and servers during distributed training, a major bottleneck on bandwidth-limited networks.