Lightelligence Connects Chiplets

Lightelligence’s ONoc interposer connects chiplets optically through silicon waveguides on a monolithic multireticle substrate. The company launched a machine-learning application as a demonstrator.

Bryon Moyer

As more companies adopt chiplets, the technology for connecting those pieces is evolving. Electrical connections built on silicon or organic interposers dominate, although links must be short to manage power, complicating networks in large chiplet assemblies. Lightelligence’s ONoc enables long-distance interchiplet signals on a monolithic silicon interposer. The company targets hyperscalers and other chip builders requiring extremely low latency.

ONoc acts as a passive 2D interposer between chiplets, with separate lasers providing signal power. Analog electrical signals on the chiplet drive the ONoc to modulate laser light. At the end of the photonic link, a photodetector converts the signal back to the electrical domain for delivery to another chiplet. Total electrical-to-electrical latency is approximately 1.0 ns. The company’s Hummingbird PCIe card, available now, demonstrates the technology but applies it differently, using it to connect computing cores on a single die.

Lightelligence isn’t the first to announce an optical interposer; Lightmatter did so last year. But ONoc is a simpler product and is the first to be available for design starts, with Lightelligence having held off its announcement pending readiness of its demonstrator.

Since its founding in 2018, Lightelligence has raised more than $230 million in eight funding rounds, the most recent adding $20 million in September of last year. The company’s primary product is its OMac AI inference accelerator. It also offers an ONet product, which helps move data from chips onto fiber for delivery to a different location. ONoc is similar but serves for intrapackage data movement.

Each ONoc engagement will involve a custom interposer design, adapting for both chiplet footprint and fabric (mesh, torus, etc.). Experience will guide the possible future development of productivity tools to automate parts of the design process.