Lightmatter, a US startup developing light-based computing hardware, has set out a roadmap for connecting artificial intelligence chips with optical links as copper wiring reaches its physical limits.
As AI systems grow to span multiple server racks, the copper connections that carry data between chips can only reach a few dozen centimetres before signals degrade, according to the trade publication DigiTimes.
That limit is tightening as data transfer speeds climb towards 448 gigabits per second, pushing the industry towards silicon photonics, which use light rather than electrical current to move information.
The approach is often paired with co-packaged optics, or CPO, in which the optical components sit alongside the main processor rather than at the edge of the circuit board.
Lightmatter's plan covers three stages of optical integration.
These run from near-package optics, through co-packaged optics, to a photonic interposer, a base layer that routes light between stacked chips.
Its most advanced product, the Passage M1000, is a photonic interposer larger than 4,000 square millimetres that can move 114 terabits per second of data across 256 optical fibres.
The M1000 is built on the Fotonix platform from GlobalFoundries, the contract chipmaker, with packaging by Amkor.
Lightmatter has also joined NVLink Fusion, the ecosystem Nvidia has opened up to let other companies connect their hardware to its chips.
The wider effort, run through the Open Compute Project, adds to signs that optics are becoming a necessity rather than an option for the industry.
Notably, Lightmatter is working with several manufacturers at once, including GlobalFoundries as its first partner, Tower Semiconductor and TSMC.
