Steering member Broadcom's 800G AI Ethernet NIC: what it means for NoLimit
This month steering member Broadcom introduced an 800G AI Ethernet NIC that is fully compliant with the NoLimit architecture. Built for the next generation of large-scale AI clusters, it can connect hundreds of thousands of XPUs while preserving the openness and interoperability that define the foundation's work.
The NIC brings advanced RDMA features designed for AI-scale networking: packet-level multipathing, out-of-order delivery directly into XPU memory, selective retransmit, and programmable congestion control. It supports PCIe Gen6 x16, 100G or 200G PAM4 SerDes, line-rate encryption and decryption with PSP offload, and secure boot with device attestation. It ships in both PCIe CEM and OCP 3.0 form factors.
The launch is an important step for the ecosystem. It shows how open Ethernet standards can deliver performance and scalability for AI workloads while giving operators flexibility across XPUs, optics, and switches rather than locking them into a single vendor's stack.
It is also a proof point for the architecture itself. Features like multipath spraying, out-of-order direct data placement, and programmable congestion control were designed on paper in the working groups over the past year. Seeing them in shipping silicon, at 800G, is exactly the kind of implementation the foundation set out to enable.
As AI infrastructure grows in scale and complexity, products like this one highlight the value of open, standards-based networking. If you are interested in shaping the future of open Ethernet for AI and HPC, join the NoLimit Foundation and get involved in building the standards that will define tomorrow's data infrastructure.