Korean version will be posted soon]

In the last post
View Tweet, we established that copper is struggling. At the data speeds that modern AI systems demand (hundreds of gigabits to terabits per second, across distances of meters and beyond), copper wires heat up, lose signal, and require so much circuitry just to recover that signal that the power bill becomes unsustainable.
To be more precise, there are two distinct bottlenecks. The first is between a GPU and its memory devices. This is being (partially) addressed with technologies like High Bandwidth Memory (HBM), which stacks memory dies directly on top of the GPU to dramatically shorten the electrical path. The second bottleneck that we are more interested in is longer-distance communication, such as GPU-to-GPU links. As AI clusters grow larger, GPUs spread across multiple servers and racks must constantly exchange data over distances that copper simply cannot handle at the required speeds. This is where the industry is turning to light.
But that answer immediately raises a more practical question: how? Light travels through optical fiber between cities, sure. But how do you get it onto a chip? How do you route it, control it, split it, and combine it, all within a device the size of your fingernail?
That is what this post is about. The technology is called a Photonic Integrated Circuit (PIC)β a circuit where light partially replaces electric current (the word "partially" indicates that a PIC still requires electronic integrated circuits to operate it).
# The Electronic Integrated Circuit (IC) Analogy
Let's start with something more familiar: the electronic integrated circuit (IC), what people usually just call a chip. But letβs call this EIC (electronic IC) to distinguish it from photonic integrated circuit (PIC).
An EIC is a collection of electronic components, such as transistors, resistors, capacitors, metal interconnects, which are all fabricated together on a single piece of silicon. Before the EIC existed (pre-1960s), engineers built electronic circuits from individual discrete components: a transistor here, a capacitor there, all wired together by hand. The EIC changed everything by bringing all of those components onto a single substrate, dramatically reducing size, power consumption, and cost, which makes modern computing possible.
A Photonic Integrated Circuit (PIC) does the same thing, but for light.
# What Does a Photonic Integrated Circuit (PIC) Look Like?

A PIC integrates optical components: lasers (which generate the light in the first place), waveguides (which is a path where light can travel like copper wires in EICs), modulators (which convert electrical signals into optical signals), photodetectors (which convert the optical signals back into electrical signals), and/or so on. Note that you do not need to know what each component does yet; it will be explained in later posts.
The most common application of PICs today is the optical transceiver. As mentioned earlier, longer-distance communication, such as GPU-to-GPU links, requires light. But GPUs speak in electrical signals. So, the electrical signal first needs to be converted into an optical signal, sent through optical fiber to the vicinity of another GPU, and then converted back into an electrical signal that the GPU can read. In short, an optical transceiver is a device that converts electrical signals from electronic hardware into optical signals or does the reverse.

For example, as shown in the image above, GPU1sends electrical signals to Optical Transceiver 1 located relatively close to the GPU. Optical Transceiver 1 converts the electrical signal into an optical signal and sends it through an optical fiber to Optical Transceiver 2 that is farther away. Finally, the received optical signal is converted back into electrical signal by Optical Transceiver 2 and sent to GPU 2.
This illustrates the fundamental concept of how light can address the data transmission bottleneck in the AI era. As the distance between devices that need to communicate each other (e.g., optical transceivers in this case) increases, light becomes significantly more advantageous than copper traces. This simple image intuitively highlights why light is so important.
However, several challenges remain. (1) Each optical component must exhibit high performance, (2) these components must be integrated onto a single chip, or at least placed close together, and (3) the entire system must be manufacturable at low cost and high volume. These are the main challenges of PIC development.
Before diving into the components of a PIC and how they work, letβs first cover some background.
# The Waveguide: The Highway for Light
This is the highway through which light travels, which is the equivalent of a copper trace on a circuit board. But instead of electrons flowing through metal, photons travel through a carefully engineered structure.
Generated by Thread Navigator
Press β + S to quick-export
