Silicon photonics
Based on Wikipedia: Silicon photonics
The cable connecting your computer monitor probably can't move more than ten gigabits per second. The latest universal serial bus standard tops out at about the same rate. Yet inside data centers, silicon photonics now pushes one hundred gigabits per second through cables thinner than a pencil. We're witnessing a quiet revolution: light replacing electricity as the language computers use to talk to each other.
This isn't science fiction arriving decades hence. It's happening now, in server racks around the world.
Why Light?
Electrical signals face fundamental problems as they travel through copper wires. The faster you push data, the more energy you waste as heat. The farther the signal travels, the more it degrades. Crosstalk from neighboring wires creates interference. At some point, you hit a wall.
Light doesn't care about these limitations in the same way. Photons—particles of light—can carry staggering amounts of information without generating much heat. They zip through fiber optic cables for kilometers without significant degradation. Multiple streams of different wavelengths can travel through the same fiber simultaneously, like separate radio stations sharing the airwaves.
Telecommunications figured this out decades ago. The backbone of the internet runs on light. Fiber optic cables crisscross the ocean floors, carrying intercontinental traffic at the speed of light. But inside computers and between nearby servers? That's remained stubbornly electrical.
Until now.
The Silicon Advantage
Here's the clever part: silicon, the same material we've been using to make computer chips for half a century, happens to be transparent to infrared light. Not visible light—hold a silicon wafer up to a lamp and it looks opaque, like a gray mirror. But infrared light, with wavelengths around 1.55 micrometers (the same wavelength used by fiber optic telecommunications), passes right through.
This matters enormously. It means we can use the same manufacturing techniques that produce billions of transistors on a fingernail-sized chip to create microscopic optical components. The factories already exist. The expertise already exists. The supply chains already exist.
A typical silicon photonic device sits atop a layer of silicon dioxide—essentially glass. Engineers call this arrangement "silicon on insulator," borrowing terminology from the electronics world. The silicon layer might be only a few hundred nanometers thick. Light bounces off the boundary between silicon and glass, trapped inside the silicon like water in a pipe.
But these aren't ordinary pipes. They're waveguides with cross-sections hundreds of times thinner than a human hair.
How Light Gets Trapped
The physics depends on a property called the refractive index—a measure of how much a material slows down light. Silicon has an extraordinarily high refractive index of about 3.5. Glass is much lower, around 1.44. Air is nearly 1.
When light traveling through a high-index material hits a boundary with a lower-index material at a shallow angle, something remarkable happens: total internal reflection. The light bounces back entirely, like a mirror. This is the same principle that makes fiber optic cables work, and it's what keeps light confined inside silicon waveguides.
The tight confinement has profound consequences. When you squeeze light into such a small space, you concentrate its energy. Effects that would normally require powerful lasers become possible with modest light sources. The nonlinear behaviors of silicon—the ways that intense light can change how the material transmits other light—suddenly become accessible.
The Strange World of Nonlinear Optics
In everyday experience, light behaves predictably. Shine two flashlight beams through each other and they pass without interaction. Light doesn't affect other light.
Except that's not quite true. In certain materials, under certain conditions, light does interact with light. These nonlinear optical effects usually require enormous intensities—laser pulses powerful enough to damage the material. But in a silicon waveguide, the concentration of energy makes nonlinear effects visible at ordinary power levels.
The Kerr effect is one example. In materials exhibiting this behavior, the refractive index increases with light intensity. Brighter light travels through slower, in a sense. This creates all sorts of interesting possibilities. Light can modify its own path. Pulses can reshape themselves as they travel.
Four-wave mixing is another phenomenon enabled by the Kerr effect. When multiple light frequencies interact in a nonlinear medium, they can generate new frequencies—converting one wavelength of light into another. This enables wavelength conversion, where information encoded on one color of light gets transferred to a different color. It enables parametric amplification, where weak signals get boosted by borrowing energy from a stronger pump beam.
Most intriguingly, nonlinear effects enable solitons—pulses of light that maintain their shape over long distances. Normally, a short pulse spreads out as it travels because different wavelength components move at different speeds (a phenomenon called dispersion). But in a carefully designed waveguide with the right combination of dispersion and Kerr nonlinearity, these effects can balance perfectly. The pulse neither spreads nor compresses. It propagates indefinitely, like a wave that refuses to break.
The Two-Photon Problem
Silicon's nonlinear properties aren't all beneficial. Two-photon absorption creates significant headaches. In this process, two photons simultaneously excite an electron, knocking it free from its atomic bond. The photons vanish, their energy converted into an electron-hole pair rather than useful light transmission.
This matters because the generated free carriers—electrons and holes wandering through the silicon—cause additional problems. They absorb more light. They change the refractive index unpredictably. The effect compounds at high intensities and builds up over time.
Engineers have developed clever countermeasures. One approach embeds the waveguide in a PIN diode—a semiconductor device that can sweep free carriers away before they cause trouble. Apply a reverse voltage, and the electric field pulls electrons one direction and holes the other, clearing the optical path. An even more sophisticated version recovers energy from this process, partially compensating for the two-photon losses.
Another approach uses slot waveguides. Instead of confining light entirely within silicon, these devices trap light in a narrow slot between two silicon ridges. The slot can be filled with a different material—often a specially designed polymer—that exhibits strong Kerr nonlinearity without the problematic two-photon absorption.
Modulators: Encoding Information onto Light
A photonic system needs to do more than just transmit light. It needs to encode information. In electronics, this is straightforward: voltage levels represent ones and zeros. In photonics, information typically gets encoded as variations in light intensity or phase.
Silicon modulators exploit a quirk of the material. When you inject electrons and holes into silicon, its refractive index changes. The effect is modest but measurable, and more importantly, it can be controlled electrically. Run current through a silicon waveguide, and you alter how light travels through it.
The simplest modulators use this principle to vary light intensity. A common design, the Mach-Zehnder interferometer, splits incoming light into two paths, changes the refractive index in one path, then recombines the beams. If the beams arrive in phase, they reinforce each other—bright output. If they arrive out of phase, they cancel—dark output. Toggle the refractive index at billions of cycles per second, and you've got a modulator that can stamp data onto a light beam.
Mach-Zehnder modulators work well but take up considerable space—typically millimeters. For applications demanding extreme miniaturization, ring resonators offer an alternative. These circular waveguides, only tens of micrometers across, resonate at specific wavelengths. Tuning the resonance by changing the refractive index modulates light passing through.
In 2013, researchers demonstrated ring modulators that could be manufactured using standard semiconductor fabrication processes—the same processes that produce computer processors. This wasn't just a technical achievement. It was an economic one, proving that silicon photonics could piggyback on the enormous infrastructure of the semiconductor industry.
Detectors: Converting Light Back to Electricity
At the receiving end of an optical link, light must become electrical signals again. Silicon itself can't do this job at telecommunications wavelengths—its electrons don't respond to 1.55-micrometer photons. The light passes through without generating current.
Germanium solves this problem. This element, silicon's neighbor on the periodic table, has a narrower band gap. When a 1.55-micrometer photon hits germanium, it can kick an electron free, generating detectable current. Better still, germanium grows compatibly on silicon substrates, allowing integration into standard manufacturing processes.
Modern germanium photodetectors achieve remarkable speeds. Researchers have demonstrated devices operating at 40 gigabits per second—fast enough to download a high-definition movie in less than a second. Avalanche photodiodes, which use internal amplification to boost weak signals, push performance even further.
Graphene represents an intriguing alternative still in development. This single-atom-thick carbon sheet absorbs light across an enormous range of wavelengths. Graphene photodetectors could potentially receive multiple data streams encoded on different wavelengths simultaneously. They operate without applied voltage, reducing energy consumption. And they might integrate more simply onto silicon chips.
The catch: graphene absorbs light weakly. Current devices remain about ten times less sensitive than germanium detectors. But progress has been rapid, and pairing graphene with silicon waveguides—routing light to maximize interaction with the graphene sheet—has improved performance substantially since the first demonstrations in 2011.
The Laser Question
Every optical system needs a light source. Here, silicon photonics faces a genuine challenge. Silicon is terrible at generating light. Its electronic structure prevents efficient photon emission. An electron dropping to a lower energy state in silicon typically transfers energy as heat rather than light.
The debate over how to address this has persisted for years. Should the laser live on the same chip as the photonic circuitry, or should it remain a separate component?
On-chip lasers offer elegance. No external connections to worry about. No alignment problems. Everything integrated into a single package. Hybrid silicon lasers achieve this by bonding a different semiconductor—indium phosphide, for example, which does emit light efficiently—to the silicon chip. The indium phosphide generates photons; the silicon routes them.
All-silicon approaches exist too. Raman lasers use the Raman effect—light transferring energy to atomic vibrations—to generate new wavelengths from an external pump beam. Brillouin lasers exploit a similar phenomenon involving sound waves in the material. These avoid the complexity of bonding dissimilar materials but still require an external light source to get started.
The practical answer, for now, often keeps the laser off-chip. Computer processors run hot, and laser efficiency drops with temperature. Quantum effects matter: the probability of a laser emitting a photon declines as the device warms up. Separating the laser from the processor lets each operate at its optimal temperature.
The first microprocessor with optical input and output, demonstrated in 2015, used an off-chip laser. Light entered and left the processor through integrated waveguides and modulators, but the light source itself sat elsewhere. This "fiber-to-the-processor" approach delivered data at 2.5 gigabits per second in each direction—modest by modern standards, but a genuine breakthrough in integration.
Co-Packaged Optics and the Datacenter Revolution
The architecture that's reshaping datacenters is called co-packaged optics. Rather than treating processors, memory, storage, and networking as separate boxes connected by cables, this approach integrates optical transceivers directly into switch packages. Light replaces copper for chip-to-chip communication.
The advantages compound. Optical connections consume less power per bit than electrical ones at high speeds. They take up less space. They don't generate electromagnetic interference. They can span longer distances without repeaters.
Intel's 2010 demonstration of 50 gigabit-per-second silicon photonic links was an early milestone. By 2013, the company announced technology pushing 100 gigabits per second through cables about five millimeters in diameter. For comparison, conventional cables of the era topped out around eight gigabits per second.
IBM reached another milestone in 2012, achieving optical components at the 90-nanometer scale using standard manufacturing techniques. The significance wasn't raw performance—90 nanometers was already old technology in the processor world. The significance was compatibility. These components could be manufactured in existing factories, integrated into conventional chips.
Beyond Communications: Sensing and Computing
Data transmission dominates commercial interest in silicon photonics, but researchers explore other applications. Biosensors exploit the sensitivity of optical waveguides to surface conditions. A molecule binding to the waveguide surface changes the effective refractive index, shifting resonance frequencies in measurable ways. Lab-on-a-chip devices could detect diseases from a drop of blood using integrated photonic circuits.
Augmented reality presents another frontier. Magic Leap, a startup that generated enormous hype in the 2010s, worked on light-field displays using silicon photonics. The goal: glasses that overlay digital images onto the real world, with the photonic circuitry responsible for directing light to create the illusion of depth.
Perhaps most surprisingly, silicon photonics has found applications in artificial intelligence. Neural networks perform vast numbers of multiplications, and photonic implementations can execute these operations with remarkable energy efficiency. Light passing through a Mach-Zehnder interferometer undergoes transformations analogous to the matrix multiplications at the heart of deep learning. Adjusting the interferometer—by physically bending it with nanoscale mechanical actuators—changes the transformation.
These photonic AI accelerators remain experimental, but they promise dramatic reductions in power consumption. Training large language models currently requires megawatts of electricity. If photonic alternatives mature, they might slash those requirements by orders of magnitude.
The Dispersion Engineering Trick
One of the most elegant aspects of silicon photonics involves controlling dispersion—the phenomenon that makes short pulses spread out as they travel. In bulk silicon, longer wavelengths travel faster than shorter ones (called "normal" dispersion). This causes problems for ultrashort pulse applications, where pulse spreading ruins the signal.
But dispersion depends on geometry as well as material. By carefully designing waveguide dimensions, engineers can flip the sign of dispersion, making shorter wavelengths travel faster ("anomalous" dispersion). They can even flatten dispersion across a range of wavelengths, keeping all colors traveling together.
Anomalous dispersion isn't just a curiosity. It's essential for soliton propagation and modulational instability—nonlinear phenomena that enable some of the most sophisticated photonic applications. The ability to engineer dispersion gives designers an additional lever to pull, making silicon waveguides far more versatile than the raw material properties would suggest.
The Road Ahead
In 2006, Intel's Pat Gelsinger—then a senior vice president, later the company's chief executive—made a bold prediction: "Today, optics is a niche technology. Tomorrow, it's the mainstream of every chip that we build."
We're not there yet, but the trajectory is clear. Moore's Law—the observation that transistor density doubles roughly every two years—faces physical limits. You can only shrink transistors so far before quantum effects take over. But the data demands keep growing. Artificial intelligence, in particular, creates insatiable appetite for bandwidth.
Silicon photonics offers a path forward. Not by replacing electronics, but by complementing it. Electrons handle computation inside chips. Photons handle communication between them. Each does what it does best.
The technology that carries internet traffic across oceans is migrating inward, from transoceanic cables to intercontinental links to datacenter racks to the spaces between chips on a circuit board. Eventually, perhaps, to the millimeter-scale gaps between processor cores.
Light is coming home to silicon.