CONCEPT Cited by 3 sources
Scale-up / scale-out / scale-across¶
Scale-up, scale-out, and scale-across are the three distance domains of AI cluster networking — a taxonomy that classifies an interconnect by how far apart the things it connects are, because distance dictates bandwidth, latency, and the transport/topology choices available.
| Domain | Connects | Distance | Classic tech |
|---|---|---|---|
| Scale-up | accelerator ↔ accelerator | within a rack / pod | NVLink, proprietary buses (now Ethernet via Tahan / ESUN) |
| Scale-out | GPU ↔ GPU | across a training cluster / building | RoCE / InfiniBand over Clos / multiplane |
| Scale-across | fabric ↔ fabric | across buildings / campuses / regions | Ethernet super-spine (BAG), SRv6 backbone |
Why the taxonomy matters¶
- Each domain is bandwidth-vs-distance-bound differently. Scale-up is the highest bandwidth / lowest latency (bytes travel centimeters); scale-across is the lowest bandwidth / highest latency (bytes travel kilometers). A design that works in one domain can break in another — e.g. lossless-Ethernet buffer headroom is trivial in-rack and a first-order problem across km.
- The industry push is "one transport across all three." Meta's OCP 2026 thesis is that Ethernet + lossless RoCE can serve every domain, unified under one NOS (FBOSS): Tahan (scale-up) reuses the scale-out RoCE transport; FPF (scale-out) and BAG (scale-across) are the other two layers.
- A single training job now spans all three. A model trains across accelerators in a rack (scale-up), GPUs across a building (scale-out), and buildings across a region (scale-across) — the collective communication must cross all three fabrics, which is why distance-awareness in routing and scheduling matters.
Seen in (wiki)¶
- Meta MetaRoCE (2026-08-24). Frames its transport around scale-up / scale-out AI Ethernet; topology-independent so it runs across the domains' varied fabrics. (Source: sources/2026-08-24-meta-metaroce-a-new-rdma-transport-built-for-ai-scale-ethernet)
- Meta OCP Summit 2026. The article's explicit organizing frame — "Powering AI with Open Innovations in Scale-Up, Scale-Out, and Scale-Across Networks." Maps each domain to a system: scale-up → Tahan/ESUN, scale-out → FPF, scale-across → BAG + SRv6 backbone, all on the systems/icepack|102.4T IceCube switch generation. (Source: sources/2026-10-12-meta-ocp-summit-2026-powering-ai-with-open-innovations-in-scale-up-scale-out-and-scale-across-networks)
- Meta OCP Summit 2026 — talk roundup. The "Validation Methodologies for the ESUN Ecosystem" talk (Meta + Keysight) widens the taxonomy from three to six AI scaling domains — scale in, up, out, across, edge, and above — situating scale-up within the full set rather than treating the three-domain cut as exhaustive. First canonical statement of the extended six-domain framing on the wiki. (Source: sources/2026-10-26-meta-networking-at-the-2026-ocp-global-summit-open-innovatio-ff1684df)
Related¶
- systems/tahan — scale-up (Ethernet).
- systems/flattened-planar-fabric — scale-out (multiplane Clos).
- systems/backend-aggregation-bag — scale-across (petabit super-spine).
- systems/roce-rdma-over-converged-ethernet / systems/metaroce — the transport reused across domains.
- systems/infiniband — the HPC-native alternative in scale-out.
- concepts/clos-topology / concepts/lossless-ethernet — the scale-out building blocks.