The news says "silicon photonics". What is it, and why send data with light instead of electricity?

The news often mentions "silicon photonics" and "CPO", and says AI data centers are "replacing copper with light". Silicon photonics means using chipmaking methods to build tiny channels that guide light on silicon, plus small parts that turn electrical signals into light and light back into electrical signals. The problem it tackles: data moving between chips and between racks isn't fast enough, and moving it takes too much power. To see why, start with how data moves today.

How does data move today?

Inside chips and between them, data is a long string of 0s and 1s, mostly sent as electricity: a high or low voltage stands for 1 or 0, running along copper wires.

An AI server has several chips computing together; a data center has rows of tall cabinets full of servers, called racks. Data keeps flowing back and forth among all of them. Why AI has to move so much data is covered in the article on "HBM" in the news; this one is about trips longer than the one from a chip to its memory.

Why is copper struggling?

The faster you send, the more times per second the voltage has to swing up and down. The faster it swings, the more of the signal is lost along copper: the current crowds to the surface of the wire, which works like higher resistance, and the insulation around the wire soaks some up too. The longer the wire, the weaker and blurrier the signal at the other end.

So that the far end can still tell 0 from 1, circuits at both ends boost and clean up the signal, which takes power and makes heat. That's why the faster the speed, the shorter copper can reach. Take the public Ethernet standards: in the generation where one lane sends ten billion 0s and 1s per second, copper cables are specified for about 7 meters; at a hundred billion per second, it's about 2 meters, roughly enough to stay inside one rack.

What's better about light?

An optical fiber is a very thin strand of glass, and light travels along it, trapped inside.

• Little loss: in fiber made for long distances, light loses only a few percent per kilometer. That's why links between racks and between cities have used fiber for a long time.
• Lots at once: one fiber can carry several colors of light at the same time, each with its own data, like one road divided into several lanes.

The price is converting at both ends. Chips compute only with electricity, so before data goes out, the electrical signal has to become flashing light; at the far end, the light becomes electricity again. That's done by small boxes called optical transceivers, plugged into the front of a switch (the machine that forwards data to the right place).

So what is silicon photonics?

The parts that do this conversion were mostly made separately and then assembled. The idea of silicon photonics: chipmaking is very good at making things small and making lots of them, so use the same methods to build the optical parts on silicon too. Three are common:

1. Waveguides (paths for light): a thin line of silicon on the chip, about 0.5 micrometers wide and around 0.2 thick (a micrometer is a thousandth of a millimeter), surrounded by silicon dioxide, the main ingredient of glass. Silicon photonics uses infrared light the human eye can't see (wavelengths of about 1.3 or 1.55 micrometers), and silicon is almost transparent to it. Light in silicon travels at only about 40% of its speed in glass; with that big a difference between the two materials, light stays trapped in the silicon line and can even go around tight bends.
2. Modulators (electricity to light): a laser's light is always on, and the modulator rapidly makes it brighter and dimmer to follow the electrical signal. The silicon way: add or remove some mobile electric charge in the light path, which slightly changes how fast light travels through the silicon; then split the light into two paths and bring them back together, or send it around a tiny ring, and that small change becomes bright and dim.
3. Photodetectors (light to electricity): parts that make a current when light hits them. Silicon is almost transparent to this light, so it can't catch it; a small piece of germanium is often grown on the silicon instead. Germanium is another semiconductor, and it does absorb this light.

All of these are made the way a wafer fab makes things. The starting point is often a "silicon, glass, silicon" sandwich wafer: a very thin layer of silicon on top, with a layer of silicon dioxide underneath. Litho prints the pattern of the light paths, etch cuts them in, and then come films, implant and metal wiring, on the same kinds of tools. How wide and how thick a light path is decides how light travels in it, so line width has to be watched here too.

One thing to keep straight: the silicon photonics in the news is mostly about sending data, not computing with light. The computing inside chips is still done by transistors.

What is co-packaged optics (CPO)?

The common setup today: optical transceivers plug into the front of the switch, while the switch chip sits in the middle of the machine. An electrical signal leaving the chip first travels a few centimeters to a few tens of centimeters across the circuit board before it reaches a module and becomes light. The faster the speed, the harder that stretch gets.

Co-packaged optics (CPO) moves the electricity-to-light step into the same package as the switch chip, right up against it. Electricity travels only a few millimeters before it becomes light, and fiber connects straight to the package. With a shorter path, the circuits that clean up the signal can be much simpler, which saves power, and more data paths fit around the edges of the package. That's the general picture; some also want to put it next to AI chips.

It's the same direction as the article on "advanced packaging" in the news: put chips that talk constantly closer together. Silicon photonics makes the optical parts small enough to fit inside a package, which is why the two terms show up together; one is a way to make the parts, the other a way to arrange them.

Why isn't it everywhere yet? What's hard?

• Silicon is bad at giving off light: silicon struggles to turn electrical energy into light, so it doesn't make a useful laser. The light source is often made separately from another kind of semiconductor (such as indium phosphide) and then attached or soldered on, or kept outside the package altogether, with fiber bringing the light in.
• Fibers must line up very precisely: the light-carrying core of a fiber is about 10 micrometers across, while a light path on silicon is only a few tenths of a micrometer, a difference of tens of times. Off by just a little, and much of the light is lost.
• Heat: lasers and some optical parts are sensitive to temperature; when it changes, the color and strength of the light drift. And the switch chip itself runs very hot. That's one reason some designs keep the light source outside.
• Hard to replace: a front-panel module that fails can be pulled out and swapped; one built into the package may mean replacing the whole thing. So parts are tested before they go in, and they have to last.

Next time you see it in the news

• "Silicon photonics": using chipmaking methods to build optical parts on silicon, making the electricity-to-light and light-to-electricity steps small.
• "CPO", "co-packaged optics": putting the optical parts in the same package as a switch chip (or an AI chip), so electricity travels only a few millimeters before becoming light.
• "Replacing copper with light": usually means the links that keep getting faster and longer move to light; inside chips, and between chips sitting very close together, it's still mostly electricity.
• "Optical transceivers", "optical modules": small boxes plugged into a switch that convert between electricity and light.

In fab24hr

There is no silicon photonics in the game, and no chip that sends data with light. The game's chip, Glowmere, is a fictional hearing-aid chip; at the test-and-assembly house it goes into a small module and is connected with fine metal wires (wire bonding), carrying electricity.

The closest thing is the litho area in the production fab: light there prints circuit patterns onto wafers, and the light paths in silicon photonics are printed and then etched the same way. In "A day in the fab", on the "Your role today" card, tap "Production fab", swipe to the process engineer, then tap "Litho" under the name, and you're the litho process engineer. In the afternoon shift, for example, two lots from the coater/developer track TRK-01 come out with line width wide on one side and narrow on the other, and you pick which wafers to rework.

The companies, chips and people in the game are fictional; what silicon photonics is and why it's hard come from public, general sources. For other jargon, see Chip fab jargon in plain English.