CPO Intelligence

Three optical engines, taken apart

Where the photonic IC sits, how the electronics reach it and how the fibre gets out. Each engine is reconstructed from public disclosures, and every part is tagged as sourced or inferred.

Loading reconstruction…
Reconstruction · not vendor CAD

Drag to orbit. Hold Ctrl and scroll, or use + and −, to zoom. Models are exported from our editable Blender reconstructions. Explosion spacing is a viewing choice, not a physical clearance.

Analysis

Co-packaged optics: three ways to stack an engine

Published
Sources assessed through
Evidence labels
F vendor disclosure, attributed to its source   I our analysis or inference   U undisclosed

Co-packaged optics moves the optical engine from the faceplate onto the switch or processor package. The three engines here make different choices about one question: which chip goes on top. That choice sets how the light leaves the package, which sets what can be serviced in the field.

Broadcom's Bailly flips the photonic IC over a fan-out package and couples from the die edge. The NVIDIA Spectrum-X engine, built on TSMC's COUPE platform, bonds the electronic IC directly onto the photonic IC and sends light up through a lensed carrier to a detachable connector. The Ayar Labs and Alchip reference design also stacks the two dies, with a detachable fibre array.

Vendor disclosures are not independent certification. The reconstructions are not dimensional teardowns: footprints, bond counts and most dimensions are illustrative. No energy, cost or reliability winner is established, because no public test uses equal reach, payload, FEC and temperature across vendors.

Architecture comparison

Figures

Reading the energy numbers

Energy per bit is power divided by throughput. State whether the divisor is transmit capacity, receive capacity or their sum. A comparison that uses one-way capacity for one vendor and a bidirectional sum for another already carries a factor-of-two bias. Bailly's 7.01 pJ/bit is our calculation from 359 W of engines and lasers against 51.2 Tb/s of one-way capacity. It does not transfer to newer generations.

What remains undisclosed

  • Die areas, bond pitches and production package drawings for all three engines.
  • A complete, comparable power budget for the NVIDIA engine including lasers.
  • Fleet reliability data and the field repair unit for any of the three.
  • Performance figures for the Ayar Labs / Alchip stacked reference.

Primary sources

    Vendor and product names identify publicly disclosed architectures. Photonic Intelligence is not affiliated with these companies.