Ultra Ethernet

Ultra Ethernet is an open networking standard developed by the Ultra Ethernet Consortium to connect GPUs and accelerators inside AI training and inference clusters, positioned as a higher-performance, more scalable alternative to proprietary fabrics like NVIDIA's InfiniBand. Built on standard Ethernet physical layers running at 200-800 Gbps per port, it adds features such as packet spraying, selective retransmission and improved congestion control to cut the tail latency that hurts distributed AI workloads. Backers including AMD, Broadcom, Cisco, HPE, Intel, Meta and Microsoft published the first Ultra Ethernet Specification 1.0 in 2025, aiming to let hyperscalers mix commodity Ethernet switches from multiple vendors rather than lock into a single-vendor interconnect stack. For data center operators, wider adoption could ease the design of AI-ready halls by relaxing dependence on InfiniBand-specific cabling, switch and NIC ecosystems, though most large GPU clusters deployed through 2026 still rely primarily on InfiniBand or vendor-proprietary NVLink-class fabrics for the tightest coupling.