CONCEPT Cited by 2 sources

Clos topology

A Clos network is a multi-stage, non-blocking switching topology — named after Charles Clos (1953) — that builds a large, high-bisection-bandwidth fabric out of many small, identical switches instead of one impossibly large switch. In modern data centers it appears as the leaf-spine (2-tier) / spine-superspine (3-tier) fat-tree: every leaf connects to every spine, so any endpoint can reach any other endpoint across multiple equal-cost paths.

Why it dominates data-center and AI fabrics

• Non-blocking at scale. With enough spine capacity, the fabric offers full bisection bandwidth — every endpoint can talk to every other at line rate, the property AI all-to-all / all-reduce collectives need.
• Built from commodity switches. Scale comes from adding tiers and planes, not from a bigger chassis — the economic foundation of hyperscale networking.
• Many equal-cost paths → load spread via ECMP (or NIC-level spraying), and graceful degradation when a link/switch fails.

Tiers are a cost/latency dial

More tiers = more scale but more hops, latency, and switches. The number of tiers is a design knob:

• 2-tier Clos (leaf + spine) — fewer hops, lower latency, lower cost; caps the cluster size at what spine radix allows.
• 3-tier Clos (leaf + spine + super-spine/aggregator) — larger clusters, at the cost of an extra hop and tier of switches.

Seen in (wiki)

• Meta SIGCOMM 2024 RoCE fabric. Meta's AI training fabric is a two-stage Clos AI Zone (RTSW leaf + CTSW spine, with an optional ATSW aggregator tier for multi-Zone) — non-blocking inside a Zone, oversubscribed across Zones. (Source: sources/2024-08-05-meta-a-roce-network-for-distributed-ai-training-at-scale)
• Meta Flattened Planar Fabric (OCP 2026). FPF flattens the fabric from the 3-tier Clos of the NSF fabric to a 2-tier Clos, removing a switch tier to cut hops, latency, and cost — while a multiplane layout (8 planes per GPU) preserves non-blocking scale. A concrete instance of using the tier count as a cost/latency dial. (Source: sources/2026-10-12-meta-ocp-summit-2026-powering-ai-with-open-innovations-in-scale-up-scale-out-and-scale-across-networks; see systems/flattened-planar-fabric)

Related

• systems/ai-zone — Meta's two-stage Clos scale-out template.
• systems/flattened-planar-fabric — flattens 3-tier → 2-tier Clos with multiplane.
• systems/roce-rdma-over-converged-ethernet — the transport that rides the Clos fabric.
• concepts/lossless-ethernet — the loss/congestion discipline layered over the topology.
• concepts/scale-up-scale-out-scale-across — Clos is the scale-out building block.