What is RoCEv2?

RoCEv2 (RDMA over Converged Ethernet, version 2) is a networking standard that brings RDMA to routed Ethernet networks, enabling low-latency, low-CPU-overhead data transfer on the infrastructure most industrial and edge AI deployments already use.

RDMA originally required a dedicated fabric such as InfiniBand. RoCE made it possible to run the same RDMA transport over Ethernet, and RoCEv2, published by the InfiniBand Trade Association in 2014, extended it to routed IP networks, which is what most industrial and OT environments actually are.

Underneath, RoCEv2 keeps the InfiniBand transport layer that handles RDMA reads, writes and sends, and wraps it in UDP and IP headers instead of InfiniBand network headers. Packets use UDP destination port 4791, so switches and routers handle them like any other IP traffic.

Compared with RoCEv1, RoCEv2 made three changes that matter for industrial deployment.

RDMA was designed for a fabric that almost never drops packets. Standard Ethernet drops packets when a switch buffer fills, and early RoCE hardware recovered from loss slowly. RoCEv2 networks therefore use a handful of switch features to keep loss low and react to congestion early.

Newer network cards cope much better with occasional loss, and some deployments run RoCEv2 on lossy Ethernet with ECN alone. Even then, consistent traffic classes and congestion settings across the path matter more than raw link speed.

InfiniBand is a dedicated fabric with its own adapters and switches, built for lossless RDMA from the start, and it remains common in large AI training clusters; our InfiniBand FAQ covers it. iWARP carries RDMA over TCP, so it tolerates loss without special switch settings but adds TCP processing in the network card.

RoCEv2 sits between them. It uses ordinary Ethernet switches and cabling, so it fits networks that already run Ethernet everywhere, but it rewards careful switch configuration. Because it rides on UDP and IP, our guide to the TCP/IP model is a useful reminder of where it sits in the stack, and our Layer 2 vs Layer 3 switch guide explains the routed networks RoCEv2 was built for.

For edge AI, RoCEv2 is the practical way to get the CPU-efficiency and latency benefits of RDMA without redesigning the physical network. The main prerequisite is that the network cards, switches and operating system all support it cleanly, a point worth checking during platform selection rather than after procurement.

Most RoCEv2 networks are in data centres, but the same approach suits edge AI systems with several GPUs or nodes in one cabinet, on a vehicle or across a site. At that scale the network is small enough to configure consistently: often one or two managed switches between a handful of edge servers, a camera aggregation computer and network storage.

Keep the RDMA traffic on its own VLAN and traffic class, apart from cameras, PLCs and office traffic, so a burst of video cannot cause the congestion RoCEv2 is most sensitive to. Test the whole path under realistic load before deployment, because a single switch with a missing setting can quietly undo the benefit.

We build edge AI computers and servers as standard, checking RoCEv2 support across the network card, switch and operating system and fitting the cards and drivers before the system ships.