# Optical Fabric **Domain:** Networking / High-Performance Computing **Doc Type:** Canonical Technical Concept **Maturity:** Developed ## Definition An **optical fabric** is a high-bandwidth network in which fiber links, optical transceivers, switches, topology, transport and control operate as an integral part of distributed computation. At cluster scale the network becomes part of the processor: accelerators waiting for parameters, gradients or collective results are consuming capital and energy without advancing the calculation. ## Historical Lineage Data-center networks evolved from north-south service traffic toward dense east-west communication. Distributed training increases synchronized collective operations and sensitivity to congestion, tail latency and path failure. ## Governing Variable The fabric governs usable communication: bandwidth, latency, congestion, route diversity, collective progress and failure recovery. ## Material and Computational Manifestation Fiber, transceivers, leaf-spine or rail-aware topologies, InfiniBand or Ethernet-derived transports, RDMA, multipathing, telemetry and collective libraries compose the fabric. [[wiki/Ultra Ethernet|Ultra Ethernet]] standardizes an interoperable Ethernet-based scale-out stack; [[wiki/UALink|UALink]] addresses accelerator scale-up inside a pod. ## Relation to Governance Fabric topology shapes which resources can behave as one domain. Schedulers therefore operate on an access geometry: a nominally available accelerator may be practically unusable if communication locality or path health makes the job inefficient. ## Constitutional Question and Failure Modes Open protocols can broaden participation, while proprietary management or congestion behavior can reintroduce lock-in. Failure modes include oversubscription, incast, correlated optics failure, hidden shared paths and control decisions that optimize average throughput while stranding large synchronized jobs. ## Related Ontology [[wiki/Fiber-Optic Network|Fiber-Optic Network]] · [[wiki/Rack-Scale Computing|Rack-Scale Computing]] · [[wiki/UALink|UALink]] · [[wiki/Ultra Ethernet|Ultra Ethernet]] · [[wiki/AI Factory|AI Factory]]