Meta Publishes OCP Networking Talks: 102.4 Tb/s Liquid-Cooled Switches, Scale-Across Fabrics and Open MetaRoCE

Meta publishes its OCP networking talks: liquid-cooled 102.4 Tb/s switches and an open RDMA transport

On October 26, Meta's engineering blog posted an index of its networking talks from the 2026 Open Compute Project (OCP) Global Summit. Meta says the work supports its latest AI models, Muse Wonder and Muse Spark, and its gigawatt-scale Prometheus cluster. The recordings and slide decks are now public. The post is a catalogue of talk abstracts, not a measurement report. Every specification below is Meta's own description of its designs, and the post gives no throughput, power or failure-rate results. It is a different item from the earlier piece in this series about NVIDIA's rack designs at the same summit.

Scale-out: a liquid-cooled switch rack and a flatter fabric

Meta's headline hardware is IcePack, which it calls its first liquid-cooled network rack, and the IceCube switch that goes into it. The switch blades deliver 102.4 Tb/s, built as 128 ports of 800 Gb/s each, in variants using either Broadcom's Tomahawk 6C chip or NVIDIA's Spectrum-6, and they run FBOSS, Meta's own network operating system. A talk describes IceCube as a design with hybrid liquid and air cooling and no separate PHY chips. For the network between racks, Meta describes a Flattened Planar Fabric. In it, each GPU connects to eight network planes, a three-tier Clos topology becomes a two-tier one, the network cards spray packets across paths, and one control plane combines host and switch. Meta says it built this with NVIDIA and shared it with cloud partners.

Scale-across: joining buildings, campuses and regions

Meta uses "scale-across" for links between data-center buildings, campuses and regions. Its piece for this is BAG, a petabit-scale super-spine that joins buildings and campuses inside the Prometheus cluster. A talk on regional backbones says Meta chose SRv6 routing for a new layer, running on FBOSS and the SAI switch interface, and moved from concept to deployment in months. Another talk covers lessons from BAG deployment, including distance management, reliability and capacity planning. A panel brings in speakers from Google Cloud, Microsoft Azure, Broadcom and Arista to compare approaches. These are session summaries, so the engineering trade-offs sit in the slides and videos, not in this post.

Scale-up and open standards

Inside a rack, Meta presents Tahan, an Ethernet-based scale-up switch for its MTIA 300 accelerator system, built on Broadcom's Tomahawk5. It also previews a design with 72 accelerators per rack on 102.4 Tb/s silicon. ESUN, the open Ethernet-for-scale-up effort, moves from specification 1.0 toward 1.1. Meta also says it co-leads a new OCP initiative called OSPA with Microsoft and Alibaba, which aims to standardize the switch platform layer, and that a 1.0 specification has been submitted. One talk says a copper scale-up domain with thousands of connections per rack would have roughly 10 bad links even at 99.9% yield, and treats this as a statistics problem. That figure is an illustration from a talk abstract, not a field measurement.

An RDMA transport going open

The post also says Meta is releasing software versions of MetaRoCE, its RDMA transport for AI networks. It begins with a user-space version on GitHub, and a version based on DPDK is described as coming soon. Remote direct memory access (RDMA) lets one machine write into another's memory with little processor involvement, and AI training depends on it. Readers who want to use the code should check the repository's license, state and supported hardware, because the post is a pointer, not documentation. The post tells you what Meta presented and where the material lives. It does not show whether these designs beat alternatives.