On August 21, 2026, I gave an invited talk at the Member Day of the 33rd Annual Conference on Information Theory of the Chinese Institute of Electronics (CIEIT 2026), held in Kunming, Yunnan and hosted by Kunming University of Science and Technology.
The talk, “Network Coding Theory in Collective Communications” (网络编码理论与集群通信前沿进展), addresses the communication bottleneck in large AI clusters: with tens of thousands of GPUs training large models, collectives such as Broadcast, Reduce, All-Gather and All-Reduce can dominate the runtime — up to 6× on 1024 GPUs and 12× on 2048 GPUs relative to compute. It introduces our recent results on floating-point sum reduction via network coding and on linear network coding for sum All-Reduce over ring networks, based on joint work with Yi Chen and Zhuoqi Tu.