Optical Switching Fabrics

Co-packaged optics pushed the light right up to the die — but at every hop through the network, the bits still surrender: they convert back to electrons so a packet switch can buffer them, read their headers, and fire them out the right port. Each conversion costs the full E/O toll, twice. The tantalising question: why not switch the light? Steer the beam itself from an input fibre to an output fibre and a bit could cross a whole datacenter without ever becoming an electron.

The answer is that we can — and do, at serious commercial scale — but only on optics' terms. An electronic packet switch makes a fresh decision for every packet, and it can do that because it can store packets while it decides. Optics has no store. There is no practical optical RAM, no register that holds a photon; light either moves or is gone. So optical switching is circuit switching: configure a path, let torrents of data flow through it unexamined, and reconfigure only occasionally. This lesson is about the hardware that does the steering, and the topologies — descended from a 1953 telephone-exchange paper — that let small switches compose into big ones.

Three ways to steer light

Every optical switch element is a device you already know, repurposed. The MEMS mirror physically tilts a micro-machined mirror to bounce a free-space beam from any input collimator to any output — slow (milliseconds, it is moving matter) but superbly indifferent to what the light carries. The MZI mesh uses Module 3's workhorse: a Mach–Zehnder with its phase shifter set to \varphi = 0 routes light to the cross port, at \varphi = \pi to the bar port — a 2×2 switch with no moving parts, microseconds thermally or nanoseconds electro-optically. And wavelength routing needs no moving parts and no active fabric: an arrayed-waveguide grating router (AWGR) passively delivers each wavelength from each input to a fixed, distinct output — so a transmitter chooses its destination simply by tuning its laser's colour, in nanoseconds.

Technology	Mechanism	Reconfiguration	Port count	Insertion loss
3-D MEMS mirrors	tilting micro-mirrors, free space	~10 ms	100s–1000s	~1–3 dB, rate-agnostic
MZI mesh (thermal)	interference, heater phase shift	~10 µs	10s	grows with stage count
MZI mesh (electro-optic)	interference, junction phase shift	~ns	10s	grows with stage count
AWGR + tunable laser	passive wavelength routing	~ns (laser tune)	10s–100s	moderate, fixed

Note the recurring trade: the fast switches are small and lossy-per-stage; the enormous one is slow. No entry in the table does what an electronic switch does — thousands of ports and nanosecond decisions and per-packet flexibility. Choosing an optical fabric means choosing which column you can live without.

Composing switches: crossbar versus Beneš

A 2×2 element is not a network. To connect N inputs to N outputs you must compose elements into a topology, and the choice is the same one you met in electronic interconnection networks — but with a photonic twist: light accumulates insertion loss and crosstalk at every element it crosses, so the number of stages on a path is now a link-budget line item, not just a latency. The brute-force crossbar spends N^2 crosspoints; the elegant Beneš network reaches every permutation with only

\frac{N}{2}\,(2\log_2 N - 1)\ \text{elements, in}\ 2\log_2 N - 1\ \text{stages}.

Work the accounting for N = 64. A crossbar needs 64^2 = 4096 crosspoints; a Beneš needs 32 \times 11 = 352 elements — a twelvefold saving — but every path now threads 11 stages. At, say, 0.3 dB of insertion loss and −25 dB of crosstalk per element, that is 3.3 dB of loss and eleven whispering neighbours added to your link budget. The saving has a second price: the Beneš is only rearrangeably non-blocking — any permutation is realisable, but adding one new connection may force re-routing (hence briefly interrupting) existing ones, where a crossbar is strictly non-blocking. Circuit-switched fabrics can usually stomach that; it is one more reason they suit scheduled, long-lived flows.

The deepest fact in this lesson hides in plain sight: there is no optical memory. An electronic switch resolves contention — two packets wanting one output — by parking one packet in SRAM for a few hundred nanoseconds. Light cannot be parked. The only "optical buffer" ever fielded is a delay line: a spool of fibre that postpones a signal by its flight time, about 5 µs per kilometre — so buffering a mere millisecond means 200 kilometres of glass, per port, with no random access and no ability to hold a bit until a condition is met. (You can slow light in exotic media; the storage times and capacities remain uselessly small.) This is not an engineering gap awaiting a clever fix — storing a signal means confining energy in a resonant or material state, and doing that cheaply, densely and re-readably is precisely what electrons in capacitors do and photons do not. Every serious optical network design therefore designs the contention out: circuits configured in advance, wavelengths pre-assigned, schedules agreed — decide with electrons, then let photons fly.

The Beneš topology is not a photonics invention — it comes from Bell Labs' Charles Clos (1953) and Václav Beneš (1962), who were minimising a very different scarce resource: electromechanical crossbar relays in telephone exchanges. A phone call is a circuit — set up once, held for minutes, torn down — and the Clos/Beneš construction answered "how few relays can connect any caller to any callee?". When packet switching conquered data networking, the old circuit mathematics looked like history. Then it re-conquered from two directions at once: the fat-tree fabrics inside modern datacenters are folded Clos networks (the topology, re-used for packets), and optical circuit switches are Clos's problem statement made literal again — long-held connections, expensive crosspoints, no buffering at the crosspoints. The final lesson completes the irony: Google's spine-replacing optical switches bring telephone-exchange thinking back to the very heart of the packet-switched internet.

Where this goes next

So optics can carry (Module 8 so far) and steer (this lesson) — always wholesale, never per-packet. The natural next question is how far inward the idea can push: if light can cross a datacenter, why not a chip? The next lesson takes optical networking to its most ambitious proposed venue — the network-on-chip — and performs an honest autopsy on why that beautiful idea has not shipped.