Skip to content
← FrontpageMeta hallmark logoMetaAnalysis

Meta open-sources IcePack liquid cooling and IceCube 102.4T networking blueprints for gigawatt AI clusters

Topographic elevation contour card for Meta open-sources IcePack liquid cooling and IceCube 102.4T networking blueprints for gigawatt AI clusters
VISUAL ARTWORK:Original editorial artwork. Third-party brand marks used under nominative fair use.

What happened.Meta released open hardware and network specifications at the Open Compute Project Global Summit for liquid-cooled server racks and 102.4-terabit switch architectures engineered for gigawatt training clusters.

Why it matters.Hyperscale AI training infrastructure has historically remained behind proprietary vendor moats. Open-sourcing 102.4-terabit liquid-cooled switch designs and RDMA transport layers allows independent data center operators and private cloud builders to construct multi-gigawatt facilities without paying custom silicon markups.

The technical specifications:

  • Hardware architecture: The IcePack liquid-cooled rack integrates IceCube 102.4-terabit switches hosting 128 non-blocking 800-gigabit ports across Broadcom Tomahawk 6C and NVIDIA Spectrum-6 silicon variants.
  • Fabric topology: Meta transitioned backend clusters to a 2-tier Flattened Planar Fabric with network-interface packet spraying to connect accelerators directly across eight distinct planes.
  • Transport protocols: The release open-sources MetaRoCE, a specialized Remote Direct Memory Access transport protocol designed to eliminate tail latency during multi-tenant distributed training.