Meta Shares AI Networking Innovations at OCP Summit

What happened

Meta revealed networking advancements at the Open Compute Project (OCP) Global Summit, aimed at supporting its AI models and data center infrastructure. The company highlighted its latest projects, including Muse Wonder and Muse Spark, along with its Prometheus cluster, which operates at gigawatt scale.

Among the new technologies, Meta introduced IcePack and IceCube, which are liquid-cooled network racks. These systems use 102.4T switch blades and run on FBOSS, with Broadcom Tomahawk 6C and NVIDIA Spectrum-6 variants. The IceCube switch is part of the IcePack rack, which is designed with a hybrid cooling method.

Why it matters

These innovations are significant for the AI industry, as they address the growing demand for efficient and scalable networking solutions. By integrating open-source technologies and focusing on scalability, Meta is contributing to the broader goal of creating standardized, multi-vendor compatible systems for AI infrastructure.

Key details

The OSPA initiative, co-led by Meta, Microsoft, and Alibaba, aims to standardize the switch platform layer. The v1.0 specification for this initiative has been submitted to OCP. Meta also shared insights from talks by its network and hardware teams, with session materials now publicly available.

Meta's Scale-Across initiative includes BAG, a petabit-scale super-spine connecting buildings, campuses, and regions. Additionally, SRv6 regional backbones are being implemented using FBOSS and SAI. MetaRoCE, a RDMA transport for AI networks, has been made open source.