RDMA over Converged Ethernet

<p>RoCE (pronounced "rocky") enables Remote Direct Memory Access over Ethernet networks, allowing one machine to directly read and write the memory of another without involving the remote CPU or operating system. This dramatically reduces latency and CPU overhead for data-intensive workloads.</p><p>RoCEv2 (the current standard) runs over standard Ethernet using UDP, requiring priority flow control (PFC) or DCQCN congestion management to prevent packet loss — which would cause RDMA retransmissions and performance collapse. RoCE over 400GbE is now a common alternative to InfiniBand for AI cluster fabric, particularly where Ethernet-native switching is preferred. NVIDIA supports both InfiniBand and RoCE on its GPU clusters; hyperscalers including Meta and Microsoft have standardised on RoCE at scale.</p>