Summary
This article provides an in-depth explanation of why silicon photonics, CPO (Co-Packaged Optics), and optical I/O technologies are necessary in the semiconductor industry, and particularly why light plays a central role in information transmission. Starting from fundamental principles of information transmission, it compares the properties of copper and light, and covers the historical background of light gradually expanding from long-distance to short-distance communication. It explains in detail how the AI era has exponentially increased data transmission demands, accelerating the need for light-based technologies, and describes the three stages of technological development bringing optics closer to chips (pluggable transceivers, CPO, optical I/O). Finally, it presents the direction of technological development through silicon photonics' working principles and the pJ/bit metric.
1. The Most Efficient Way to Transmit Information: Light
You've probably seen terms like Silicon Photonics, CPO (Co-Packaged Optics), and Optical I/O in semiconductor news. What does it mean when NVIDIA declares it will "connect chips with light"?
To transport goods you need trucks, to transmit power you need wires, to transmit force you need gears or pulleys. The common thread: something with mass must physically move. But information transmission is fundamentally different. The key is that transmitting information doesn't require physical matter to move — only a 'pattern' (signal) agreed upon by sender and receiver. And the lighter, faster, and more energy-efficient that signal carrier is, the better.
Something with no mass, incredible speed, and that doesn't interfere with itself? That's light.
2. Copper vs. Light: Two Pipes for Data Transmission
2.1. The Copper Pipe and Electrons
Steel balls shot through a metal pipe lose energy from colliding with pipe walls (resistance), converting to heat. Faster balls mean more collisions and heat. Adjacent pipes create crosstalk (interference). This is exactly how electrons move through copper wire — they have mass and charge, constantly colliding with conductor atoms and electromagnetically interfering with neighboring conductors.
2.2. The Laser Pipe and Light
A glass tube with mirror-coated insides: laser light, having no mass, loses no energy bouncing off mirror walls. Adjacent tubes cause zero interference since light creates no electric field. Light also passes through itself — different wavelengths can travel through the same space without interfering. This technology is called Wavelength Division Multiplexing (WDM).
The performance difference is staggering. Today's technology allows a single copper wire to transmit about 100-200 Gbps. But a single optical fiber with WDM can transmit 64-96 Tbps — hundreds of times more than copper.
Physics overwhelmingly favors light. So: why do we still use copper?
3. Light's Victory and Copper's Resistance
Light already defeated copper — starting from the longest distances. Since the first transatlantic fiber-optic cable TAT-8 in 1988, light has steadily expanded inward: intercontinental (1980s), inter-city (1990s), inter-datacenter (2000s), intra-datacenter (2010s).
But copper still holds its short-distance fortress. At a few meters and hundreds of gigabits, copper is cheap and easy to install with massive existing infrastructure. Optics requires expensive transmitter/receiver modules for electrical-to-optical-to-electrical conversion.
3.1. AI Changes Everything
AI didn't just bring "more data" — it brought the requirement that tens of thousands of GPUs operate as one giant brain. As AI models grow exponentially, physical distances between GPUs extend from centimeters to tens of meters across hundreds of server racks, and they must exchange results every few milliseconds. At every speed doubling, copper's viable distance shrinks 30-50%.
The current frontline between optics and copper is at server-to-server distance (meters). The next battleground will be chip-to-chip distance (centimeters). NVIDIA has shown confidence in converting NVLink itself to optical by around 2028.
4. Silicon Photonics: Silicon's Clever Partnership
Silicon excels at making transistors, memory, and processors but is terrible at making light due to its indirect bandgap. The solution: partner with III-V compound semiconductors (InP, GaAs) that create light efficiently, while silicon guides and manipulates that light through its high refractive index waveguides — all manufacturable in existing CMOS fabs.
Light-creating materials (InP, GaAs) generate the light, and silicon guides and manipulates it. It's a partnership.
5. pJ/bit: The Key Metric for Optical Technology Progress
The single most important metric is pJ/bit (picojoules per bit) — the energy needed to transmit one bit. Lower pJ/bit means better technology. Three generations of technology bring light closer to chips:
5.1. Generation 1: Pluggable Transceivers
Easily swappable modules on server front panels. But electrical signals must travel tens of centimeters, requiring power-hungry DSP restoration. An 800G transceiver consumes ~15W; a 51.2Tbps switch needs 64 such modules (~960W just for transceivers).
5.2. Generation 2: CPO (Co-Packaged Optics)
Optical engine packaged alongside the switch ASIC, reducing distance from tens of centimeters to millimeters. Power drops up to 70%. But the thermal paradox — lasers need cool stability while ASICs generate hundreds of watts — poses a challenge.
5.3. Generation 3: Optical I/O
Photonic engine becomes a chiplet inside the same package as GPU/CPU chiplets. Electrical distance shrinks to hundreds of micrometers — hair-width. The UCIe standard enables GPU, memory, and optical I/O chiplets to communicate in a common language.
Conclusion
Information can be transmitted without mass, and light — having no mass, no charge, and no self-interference — holds overwhelming physical superiority. Light conquered from the farthest distances inward, and the AI era is enormously accelerating this advance. Current technology evolves through three generations to bring light closer to chips: pluggable transceivers hit power walls, CPO faces thermal paradoxes, and optical I/O is just crossing the threshold. Beyond these lies photonic computing — using light itself for computation.