[CPO Deep Dive 4/6] Fiber-to-Chip: The Least Glamorous Step in CPO That Holds Everyone Back
Light leaks a little at every interface it crosses. CPO packs the laser, SiPh and EIC together, but eventually the light has to be "connected out," and that step, fiber-to-chip coupling, is where optical losses are most concentrated and where automation is hardest. It isn't cool and doesn't make headlines, yet it is the invisible ceiling on CPO yield and serviceability. This article breaks down the density race as V-groove pitch shrinks from 260 µm to 130 µm, the trade-offs between edge and grating coupling, the cost tug-of-war between passive and active alignment, and why "detachable" has become the one thing every player agrees on.
1. The Fate of Light: Every Interface Leaks
Start with a physical fact that underpins this entire article: for photons, the biggest challenge is paving them a continuous road. Every interface (laser to fiber, fiber to waveguide, waveguide to fiber) is a refractive-index discontinuity with a risk of leakage, and efficiency drops a notch each time.
Electrical signals barely care. Copper to copper, contact resistance is negligible. Light is different: insertion loss stacks up with every interface along the optical path, and every decibel of loss ultimately has to be made up by a more power-hungry laser. So coupling efficiency isn't a "nice-to-have" metric; it directly eats into the power budget you worked so hard to save in the light-source chapter.
Once CPO has pushed every earlier layer to its limit, the biggest source of loss on the optical path falls on the very last step: attaching the fiber to the chip. That step is coupling.
2. Everyone Watches the Chip, but Volume Production Hinges on "Aligning the Fiber and Still Being Able to Detach It"
First, clear up a misconception. CPO discussions always revolve around the PIC, EIC and laser, as if getting those right wins the game. But what actually holds back the production ramp is often this step that sounds like "manual labor": aligning a fiber (or a row of fibers) to the chip's optical interface with sub-micron precision, and reproducing that reliably at volume.
The difficulty lies on three levels. First, precision: misalign by a few hundred nanometers and coupling efficiency collapses. Second, speed: if every unit has to be tuned slowly by hand, volume production is impossible. Third, serviceability: when an optical engine fails or a fiber breaks, the data center has to be able to replace it rather than scrap the whole package.
Put those three together and you get every technology battle in the coupling layer. It isn't glamorous, but it is the toughest gate between CPO that "works in the lab" and CPO that "can be mass-produced on the line and serviced in the data hall."
3. The Density Race Is On: V-Groove Pitch Shrinks from 260 µm to 130 µm
The most visible progress in coupling is the shrinking pitch.
The V-groove is the classic way to hold fibers: V-shaped grooves are etched into silicon and fibers are seated in them one by one, relying on mechanical precision for alignment. It is mature, cheap and reliable. The most concrete spec progress of recent years is that fiber-array pitch is shrinking from 260 µm toward 130 µm.
The meaning is straightforward: halving the pitch fits nearly twice as many fiber channels into the same width. At the edge of a CPO package, where every millimeter counts, channel density doubles. This is real, quantifiable progress, and a hard metric for judging whether a coupling solution is "keeping up."
But the V-groove isn't the only answer; it mainly addresses how fibers are held in place. How light gets "in and out of the chip" is a separate technology battle.
4. Getting Light In and Out of the Chip: Edge vs. Grating Coupling, with No Single Right Answer
There are two mainstream ways to get light from a fiber into an on-chip waveguide, each with a cost:
Edge coupling: light enters and exits "horizontally" through the side of the chip. Pros: high coupling efficiency, broad spectrum, wavelength-agnostic. Cons: it only works at the chip edge, so channel count is limited by edge length; alignment tolerances are tight; and the chip must first be diced and its facet polished.
Grating coupling: a grating on the chip surface lets light enter and exit "vertically." Pros: it can be placed anywhere on the chip surface, alignment is relatively forgiving, and it enables wafer-level testing. Cons: efficiency is usually lower, bandwidth narrower, and it is more sensitive to wavelength and polarization.
Mainstream integration platforms currently "keep both": edge and surface/grating coupling coexist, because neither wins in every scenario. Which one you pick depends on whether your system needs a broad spectrum, how much alignment error it can tolerate, and whether you need wafer-level testing. It is the same logic as the three-way packaging battle in the packaging chapter: the goal isn't the optimal point solution, but system-level compatibility.
Beyond these two mainstream options, there is a more cutting-edge choice: Photonic Wire Bonding (PWB) uses an additive-manufacturing (3D-printing-like) approach to directly "write" a 3D optical waveguide connecting fiber and chip, bridging two misaligned interfaces with a "custom-made optical bridge" and sidestepping the traditional alignment problem. The potential is large because it can compensate for alignment errors during manufacturing, but maturity, write speed and production capacity remain open questions.
A practical rule of thumb for judging a coupling solution: don't just listen to a single "coupling efficiency" number. Ask three things: Is that efficiency reproducible under volume-production conditions, or cherry-picked in the lab? Is alignment passive or active (which determines throughput)? Can the interface be detached (which determines serviceability)? A solution with great coupling efficiency that requires active alignment and isn't detachable may lose, on volume-production economics, to one with ordinary efficiency that is passive and detachable.
5. Passive vs. Active Alignment, and the One Thing Everyone Agrees On: It Must Be Detachable
Continue reading in the subscriber section on vocus (in Chinese): https://vocus.cc/article/6a60277dfd897800018c3a98
Comments