Optical Switches: What They Are and How They Work

 

In telecommunications, data centers, and increasingly in quantum computing research, optical switches represent both a mature technology with decades of deployment history and an active frontier where researchers still chase performance improvements that seemed impossible five years ago.

The gap between "conceptually straightforward" and "actually building one" is where things get expensive and interesting.

 

Why Bother with Light?

 

The case for optical switching comes down to a single, frustrating bottleneck: the O-E-O conversion. Every time an optical signal hits a conventional electronic switch, it must be converted to an electrical signal, processed, then converted back to photons for the next fiber segment. This isn't just inefficient-it's becoming untenable.

Modern data center traffic has a nasty habit of doubling every few years. Electronic switches are hitting a wall. Power consumption scales poorly. SerDes (serializer/deserializer) circuits generate heat that requires aggressive cooling. And there's the latency-each O-E-O hop adds processing delay that accumulates across a multi-tier network architecture.

An optical switch sidesteps all of this. Light goes in, light gets redirected, light goes out. No conversion. No packet inspection. No buffering. The speed-of-light latency is essentially the propagation delay through the switch fabric itself, which for most practical purposes might as well be zero.

Sounds perfect. So why isn't everything optical?

 

The Switching Zoo

 

Here's where it gets complicated. There is no single "optical switch" technology. There's a whole taxonomy of approaches, each with different trade-offs that make sense for different applications. The major categories:

 

Mechanical switches physically move optical elements-mirrors, prisms, fiber ends-to redirect light. Crude? Maybe. But they've been deployed for decades and they work. Polatis (now part of Huber+Suhner) has built a business on 3D beam-steering switches using piezoelectric actuators. These things are slow by data center standards-switching times measured in milliseconds-but they're reliable. I've heard stories of accumulated actuator life exceeding a billion hours across field-deployed units without failures. That's not a typo.

MEMS switches (micro-electro-mechanical systems) take the mechanical concept and shrink it dramatically. Tiny mirrors fabricated on silicon or glass substrates using photolithography can tilt to redirect beams. The switching speeds improve to microseconds. Port counts can reach into the hundreds. But MEMS fabrication is finicky, and the devices remain sensitive to shock and vibration in ways that make deployment outside controlled environments tricky.

Thermo-optic switches exploit the temperature dependence of refractive index in silicon waveguides. Heat a section of waveguide with a thin-film resistor, change the refractive index, shift the phase relationship in a Mach-Zehnder interferometer, redirect the output. Silicon has a strong thermo-optic coefficient-about 1.8×10⁻⁴ K⁻¹-which makes this approach practical. Switching times land in the microsecond-to-millisecond range. Power consumption is the catch: those heaters need continuous current to maintain state.

Electro-optic switches can theoretically switch in nanoseconds. Silicon doesn't have useful linear electro-optic effects, so you're either using carrier injection (which adds loss) or looking at exotic materials like lithium niobate. LiNbO₃ modulators have been around since before I was born-Pockels cells, Mach-Zehnder modulators, the whole catalog. Thin-film lithium niobate on insulator is having a moment right now, with half-wave voltages dropping and integration density improving. But CMOS compatibility remains elusive.

 

And then there are the more exotic approaches: liquid crystal, acousto-optic, semiconductor optical amplifiers as gates, photonic crystals. Each has niche applications. None has become the universal solution.

 

 

MEMS: The Technology That Keeps Almost Arriving

 

Silicon photonic MEMS deserves its own discussion because it represents what might be the most promising path toward large-scale optical switching, and also one of the most frustrating.

The pitch is compelling: fabricate optical switches using the same CMOS-compatible processes that churn out billions of transistors. Leverage existing foundry infrastructure. Achieve the cost reductions that come with semiconductor manufacturing scale.

UC Berkeley researchers demonstrated a few years back that you could build photonic MEMS switches on standard 200mm SOI wafers using regular photolithographic and dry-etching processes in commercial foundries. No exotic fabrication steps. The switches worked: 7.7 dB fiber-to-fiber loss, 30nm optical bandwidth around 1550nm, 50 microsecond switching times.

The technical results were solid. What remains challenging is everything else.

MEMS actuators need relatively high drive voltages-tens of volts-which complicates the control electronics. The mechanical structures must be released from the underlying oxide layer using HF vapor etching, which adds process complexity. Packaging becomes painful when you're dealing with hundreds of optical ports that need precision alignment to fiber arrays. And then there's the control plane: how do you coordinate switching across a 64×64 matrix without creating scheduling bottlenecks?

A group recently published work on split waveguide crossings-essentially MEMS-actuated couplers where the switch operates by physically separating or joining two halves of a waveguide crossing. They demonstrated a 64×64 Benes switch array with remarkably low crosstalk and ran it through a billion switching cycles without performance degradation. Impressive. Still not in production.

 

The Crosstalk Problem Nobody Wants to Talk About

 

Here's something that tends to get glossed over in the marketing materials: crosstalk accumulates.

In a small switch-2×2, 4×4-crosstalk might be -30dB or better. Acceptable. But large-scale switch fabrics cascade many elementary switching elements. A 64×64 fabric might have light traversing dozens of individual switches and waveguide crossings. Each one contributes a bit of stray light to the wrong output port.

The worst-case scenario isn't one aggressor signal leaking into your victim channel. It's N-1 aggressors all contributing coherent or incoherent crosstalk simultaneously. Testing for this is a nightmare-you'd need to illuminate all input ports except one and measure what shows up where it shouldn't. Most published results report single-path crosstalk, which is... optimistic.

Researchers at IBM and elsewhere have been working on ultra-low-crosstalk designs, pushing extinction ratios to -60dB or better in individual switching cells. Whether those numbers survive scaling to large fabrics with real manufacturing variations is another question.

 

Thermo-Optic: The Workhorse Nobody Loves

 

Thermo-optic MZI switches don't get the glamour. They're slow compared to electro-optic. They burn power compared to MEMS. But they work, they integrate cleanly with silicon photonics platforms, and they've been demonstrated at scale.

A 32×32 thermo-optic switch fabric was packaged and characterized several years ago with something like 1,560 electrical I/O ports handled through ceramic BGA wire-bonding. That's a lot of wires. The thermal management involved CuW substrates and thermoelectric coolers. Not elegant, but functional.

The power consumption comes from those resistive heaters needing continuous current. Each phase shifter might draw milliwatts. Multiply by hundreds or thousands of elements in a large fabric and the thermal budget becomes a real constraint. Some groups have explored suspended waveguide structures to improve thermal isolation-less heat leaking into the substrate means faster response and lower power-but at the cost of mechanical fragility.

For applications that can tolerate microsecond switching times and can handle the thermal load, thermo-optic remains the pragmatic choice. Data center reconfiguration, wavelength routing, test-and-measurement-nobody needs nanosecond switching for these.

 

 

The Electro-Optic Promise

 

Nanosecond switching unlocks use cases that slower technologies simply cannot address. Packet-by-packet optical switching. Burst-mode operation. Dynamic bandwidth allocation that tracks application demand in real time.

Silicon doesn't help here. Its electro-optic effects are too weak. You need either carrier-injection PIN diodes (which work but add loss and have limited speed) or materials with real Pockels coefficients.

Lithium niobate has been the go-to for decades. The electro-optic coefficients are substantial-r₃₃ around 31 pm/V. Commercial LiNbO₃ modulators from Thorlabs and others operate to 40GHz or beyond. The problem has always been integration density. Bulk lithium niobate devices are centimeter-scale. Waveguide widths are micron-scale in silicon; they're much larger in diffused LiNbO₃.

Thin-film LiNbO₃ on insulator changes the calculus. Researchers are now demonstrating Mach-Zehnder modulators with bandwidths exceeding 100GHz and half-wave voltages under 2V. The footprints are shrinking toward what silicon photonics achieves. Nature papers are appearing with regularity.

Integration with the rest of a photonic circuit remains the issue. LiNbO₃ doesn't grow on silicon. Heterogeneous integration involves bonding, which adds cost and complexity. The supply chain for thin-film LiNbO₃ wafers is nascent compared to silicon photonics.

Still. If you need speed, this is where the physics points.

 

What Data Centers Actually Want

 

The hyperscalers have specific requirements that don't always align with what academic researchers find interesting.

They want cost per port around $10. They want insertion loss under 10dB for cascaded switch architectures. They want reconfiguration speeds fast enough to track traffic matrices that shift unpredictably. They want power efficiency measured in picojoules per bit or better. They want reliability numbers that let them deploy at scale without dedicated maintenance staff babysitting each switch.

MEMS-based optical circuit switches from companies like Polatis have penetrated some data center applications. The switching times-milliseconds-are slow, but for persistent "elephant" flows that dominate inter-cluster bandwidth, millisecond reconfiguration is fine. You're not trying to switch packet-by-packet; you're trying to avoid the O-E-O conversion overhead for bulk data movement.

The dream of sub-microsecond optical packet switching remains mostly that-a dream. The control plane problem alone is daunting. Without optical buffers (which don't practically exist), you can't absorb contention the way electronic switches do. Scheduling must be perfect. Synchronization across potentially thousands of servers must be tight. Some research groups have demonstrated 40-nanosecond switching-and-control systems, but productization is another matter.

 

Acousto-Optics: A Detour

 

I should mention acousto-optic switches because they keep appearing in research contexts, and because the physics is genuinely interesting even if the applications remain limited.

An acousto-optic modulator uses acoustic waves-typically surface acoustic waves launched by interdigital transducers-to create a periodic refractive index grating in a material. Light diffracts off this grating. Control the acoustic wave, control the light.

Lithium niobate again: strong piezoelectric coupling for efficient acoustic generation, decent photoelastic coefficients for interaction with light. Researchers have demonstrated AO modulators with VπL products (the figure of merit for modulation efficiency) below 0.1 V·cm on thin-film platforms.

The switching speeds are limited by acoustic propagation-microseconds, not nanoseconds. The applications tend toward RF photonics, frequency shifting, and laser Q-switching rather than telecom routing. But for completeness, the technology exists.

 

The Integration Question

 

Here's what keeps coming up in every serious discussion about optical switching: how does it fit with everything else?

A switch by itself is useless. You need transceivers, wavelength multiplexers, amplifiers, monitors, control electronics. The more of these you can integrate onto a single chip or into a single package, the better the system economics become.

Silicon photonics has a head start. Foundries like GlobalFoundries, TSMC, and imec offer process design kits. Modulators, photodetectors, wavelength filters, and passive routing all coexist on the same platform. Adding MEMS actuation to this stack-as several research groups are now doing-could enable switches that integrate seamlessly with the rest of the photonic circuitry.

Lithium niobate takes a different path. The material can host electro-optic modulators, acousto-optic devices, nonlinear optical elements, and low-loss waveguides all on one substrate. The toolbox is arguably richer than silicon. But the manufacturing ecosystem is less mature.

III-V semiconductors (InP, GaAs) enable semiconductor optical amplifiers and lasers that silicon can't match. Heterogeneous integration-bonding different materials together-might combine the best of each. Or it might just combine the fabrication challenges of each.

Nobody has figured out the winning formula yet.

 

The Honest Assessment

 

Optical switching is real technology deployed in real networks. It's also technology that has been "five years away" from transforming everything for at least twenty years.

The physics works. The engineering is advancing. The economics are improving. For certain applications-protection switching, wavelength cross-connects, reconfigurable add-drop multiplexing, test automation-optical switches have established themselves as the right solution.

For the grander vision of optical packet switching eliminating electronic routers entirely? The challenges remain formidable. Control plane complexity. Lack of optical buffering. Manufacturing costs at scale. Standardization across vendors.

Progress continues. Research papers appear weekly. Startups get funded. Big companies acquire small ones. The underlying need-moving more data with less energy-isn't going away.

Maybe this time, the next five years really will be different.