RDMA

Overview

Remote Direct Memory Access (RDMA) is a networking paradigm enabling kernel-bypass, zero-copy, hardware-offloaded data transfer between hosts — foundational to AI training cluster interconnects. The dominant implementation, RoCEv2 (RDMA over Converged Ethernet v2), was based on the InfiniBand specification standardized in 2000 and carries design choices optimized for an era when transistors were expensive: go-back-N error recovery, no multipathing, no integrated security solution, and challenging connection-state scalability. As AI clusters have grown, these design choices have become structural limitations: RoCEv2 requires lossless (PFC-based) Ethernet, causing deadlock/livelock risks; ECMP hash collisions cause underutilization; and go-back-N retransmits excessively on any packet loss. The industry was already diverging with proprietary alternatives (AWS SRD, Google Falcon) before the Ultra Ethernet Consortium standardized a modern replacement.