CPO Germanium Photodetectors Explained: Silicon Is Blind to Its Own Light, So TSMC Grows a Germanium Eye

Bottom line: Co-packaged optics moves the optical engine inside the switch package, but the receiver side hides an awkward bit of physics: silicon is essentially transparent at telecom wavelengths. Light travels meters of fiber to reach the chip, and the chip can't see it. The fix is to grow a small block of germanium on top of the silicon waveguide and let it act as the eye, soaking up photons and handing a current to the amplifier that turns it back into ones and zeros.

This piece expands a post from our Facebook page: CPO's awkward secret: the silicon chip can't see light

What happened

• NVIDIA says its Spectrum-X silicon photonics switch built on TSMC's COUPE (Compact Universal Photonic Engine) is in volume production, and TSMC lists its COUPE co-packaged solution for 2026 production.
• COUPE uses SoIC to bond an electronic die face-to-face onto a 65 nm silicon photonics die. Micro-ring modulators handle transmit; germanium photodetectors (GePDs) handle receive.
• Public COUPE spec slides list the GePD at 110 GHz bandwidth and 4.5 nA dark current, and an industry summary cites responsivity around 1.0 A/W.
• On Nov 7 we covered the transmit side: silicon can't emit light, so CPO uses an external laser. Today is the other half: silicon can't detect it either.

The physics in one line: a semiconductor only absorbs photons with more energy than its bandgap. Silicon's gap is 1.12 eV, which cuts off near 1.1 µm. A 1310 nm photon carries about 0.95 eV, not enough. Germanium's direct gap is about 0.80 eV (around 1550 nm), so it covers both the O-band and the C-band.

Engineering point 1: the light rides a 220 nm silicon wire

After the fiber couples in, light travels in a silicon waveguide roughly 220 nm thick and 500 nm wide on the SOI top layer. Here silicon's blindness is a feature: because it doesn't absorb 1310 nm, it can deliver the light to the detector with very little loss.

Engineering point 2: the germanium trap

In the detector section, germanium is grown epitaxially on top of the waveguide. Germanium's refractive index (about 4.2) is higher than silicon's (about 3.5), so the optical mode leans upward and leaks into the germanium through evanescent coupling, getting absorbed as it goes. Our rough estimate (not official): germanium's absorption coefficient at 1310 nm is on the order of 7,000 to 9,000 per cm, an absorption length of about 1 to 1.5 µm. A germanium pad a few tens of microns long can swallow nearly all the light while staying small and low in capacitance.

Engineering point 3: speed versus dark current

Each absorbed photon makes an electron-hole pair. A reverse bias across P and N regions sweeps them apart into a small photocurrent. Our math: the ideal responsivity ceiling is R = λ(µm)/1.24, or about 1.06 A/W at 1310 nm, so a measured 1.0 A/W means roughly 95% quantum efficiency. The catch is the roughly 4% lattice mismatch between germanium and silicon, which seeds threading dislocations that leak current. Dark current sets the noise floor, so pairing 110 GHz with 4.5 nA says both the depletion region and the defect density are under tight control.

Engineering point 4: SoIC puts the TIA right on top

The photocurrent is in the microamp range, and a transimpedance amplifier (TIA) converts it to voltage. Longer wires add parasitic capacitance, the TIA loses sensitivity, and the laser has to run brighter to compensate. TSMC's ECTC 2024 paper says the low-capacitance SoIC bond in COUPE improves TIA sensitivity enough to cut laser power by more than 40%. That closes the loop with the external laser story: the more sensitive the receiver, the less light the laser has to make.

Open questions

• TSMC hasn't published production GePD responsivity, dark-current distribution or yield; the specs above come from public slides and industry summaries.
• The absorption length and 95% efficiency figures are our textbook-based estimates, not measurements.
• Fiber alignment, thermal control of micro-rings and serviceability remain the real tests after ramp.

Related videos

Silicon Photonics explained

TSMC CoWoS S, R and L explained

Silicon can't emit light and can't see it, yet it's still the star of CPO. What do you think breaks first: fiber alignment, micro-ring tuning or yield? Tell us in the comments.

Sources: TSMC COUPE public specs (Swissphotonics 2025 CPO talk); IEEE ECTC 2024 COUPE paper; SemiWiki, "How TSMC Is Wiring the AI Era With Light"; CommonWealth Magazine, Aug 24 2026; original Facebook post. Figures marked as our estimate or math are not official.
Videos by @TheEngineeringCore-v. Published on The Automation Core.
Original post: https://www.facebook.com/1147074221177116/posts/1151079287443276
This one sparked a lot of discussion on our Facebook page. Come join the conversation: PLC on Facebook