InfiniBand provides RDMA (remote direct memory access) networking that bypasses the CPU for data transfers between nodes, achieving latencies under a microsecond and bandwidths of 400+ Gb/s per port. For distributed training with thousands of GPUs, the all-reduce communication pattern requires moving gigabytes of gradient data every step, making InfiniBand's consistent low latency critical. Ethernet alternatives are closing the gap but InfiniBand remains dominant in frontier training clusters.