Japan’s optical-chiplet project tackles the production line

A Japanese research project announced on 20 October will build and evaluate a pilot production line for optical chiplets used in AI computing. Its target is more than a tenfold improvement in manufacturing productivity over current research-stage processes. That is a production goal, with the comparison tied to laboratory processes—not a measured increase in computing speed.[1]

NTT Innovative Devices leads the four-company project selected by NEDO. Japan Semiconductor contributes silicon-photonics manufacturing; Furukawa Electric develops compound-semiconductor optical devices; Shinko Electric Industries handles packaging processes and inspection. NTT participates separately as an adviser.[1]

Optical chiplets combine photonic and electronic circuits in a package, moving data between processing chips using light. Making them in quantity requires a repeatable route from fabricated components to tested, assembled devices.[1]

The light must be tested as well as produced

An earlier imec engineering account explains why testing can become a production problem. Manufacturers check photonic devices while they are still on a wafer so they can avoid spending money packaging defective chips. Unlike a purely electrical test, an optical test must also couple light reliably into the device.[2]

Imec’s work with FormFactor used motorised fibre positioning and automatic alignment. A misaligned fibre can change the measured light loss; changes in polarisation, connector losses or contamination can also disturb the measurement. The station therefore needs calibration as well as movement between devices. The published account dates to 2018 and describes that test system, not equipment specified for the new Japanese line.[2]

This makes measurement consistency an economic issue. If a device appears poor because the test connection moved, discarding it wastes a good component. If a genuinely poor device passes, expensive assembly can add value to something that will later fail. Faster measurements help only when they remain meaningful.

The production-line target consequently has a different purpose from a record optical data rate. A fast link answers how much information a device can carry. Manufacturing productivity asks how effectively a factory can make usable devices. Testing and packaging are part of the route by which an impressive component becomes a repeatable product.

Report an error or suggest a correction