Switching light physically — steering a whole fibre path with mirrors rather than forwarding packets electronically.
What an OCS Does Not Do
A packet switch receives light, converts it to electrical, recovers the bits, looks up a destination, re-serialises and converts back. An optical circuit switch does none of that. It steers the light itself, usually by tilting micro-mirrors, so that an input fibre is mapped to an output fibre and everything on it passes through untouched.
The consequences follow directly. There is no rate limit imposed by the switch, no protocol awareness, no packet buffer, and no forward error correction — only insertion loss. Published systems sit around 2 dB worst case across any permutation of a 136 × 136 array, and draw roughly a hundred watts for the whole chassis rather than the kilowatts an electrical switch of comparable capacity would.
The Trade-Off Is Time
Mirrors move in milliseconds — commercially quoted around 10 to 20 ms — which is four to five orders of magnitude slower than a per-packet forwarding decision. An OCS therefore cannot carry flows in the sense a leaf switch does. What it can do is decide the topology.
For AI clusters that is genuinely valuable, because a training job’s communication pattern is known before it starts and stable for hours. Wiring the fabric to suit the job, then rewiring it for the next one, replaces a permanent over-provisioned spine tier with a much smaller one plus a patching layer.
What It Demands of the Fibre Plant
This is where an OCS stops being an abstract idea and becomes a cabling specification. The path is single-mode and it is APC end to end, because a mirror-steered free-space hop is unforgiving of reflections. Every connector between the transceiver and the mirror array counts against the same budget as the 2 dB the switch itself consumes.
Deployed designs also put an optical circulator at each end so one fibre carries both directions, which halves both the port count and the cable count — a considerable saving when the alternative is two fibres per link across a hall. Budget the whole path deliberately with the link budget calculator rather than assuming the transceiver specification leaves room.
What Goes Wrong
Contamination, overwhelmingly. A dirty or scratched end face costs a fraction of a decibel on an ordinary link and can push an OCS path past its margin, because the switch has already spent most of the budget and there is no regeneration anywhere to recover it. Inspection discipline is not optional in this plant.
The second issue is planning: since the OCS is transparent, nothing in it reports a link problem. Faults appear only at the endpoints, on transceivers that may be hundreds of metres away in either direction.
Optical circuit switch reference
| Switching element | typically 2D MEMS mirror arrays |
|---|---|
| Published radix example | 136 × 136 ports |
| Insertion loss | on the order of 2 dB worst case, port to port |
| Return loss | APC-grade, better than −38 dB specified |
| Reconfiguration time | milliseconds — commercially quoted at 10–20 ms |
| Power | of the order of 100 W for a whole chassis |
| Connectors | APC throughout, to hold return loss |
| Bidirectional trick | circulators halve the port and fibre count |
| Fibre type | single-mode only |
Related Terms
rail-optimised topology, leaf-spine, APC and UPC, insertion loss, OS2 single-mode fibre.