CEA-Leti, a French research institute, and NcodiN, a startup specializing in silicon photonics, have announced a partnership to industrialize 300mm silicon photonics technology. The collaboration aims to develop wafer-level optical interconnects to address the extreme bandwidth demands of next-generation AI computing.
The core challenge is the growing data bottleneck in AI systems, where electrical interconnects struggle with power consumption and bandwidth limitations as data volumes explode. Optical interconnects using light to transmit data offer a promising solution, providing higher bandwidth, lower latency, and reduced power consumption over longer distances.
This partnership focuses on scaling silicon photonics—integrating optical components directly onto silicon wafers—to a 300mm (12-inch) wafer platform. Moving to this larger, industry-standard format is critical for high-volume, cost-effective manufacturing, enabling the technology to move from lab prototypes to commercial production. The collaboration will leverage CEA-Leti’s expertise in advanced semiconductor processes and integration on its 300mm pilot line in Grenoble, combined with NcodiN’s design and system-level know-how in silicon photonics.
The joint effort targets creating a complete, scalable manufacturing process for wafer-level optical interconnects. These are envisioned as foundational elements for future AI hardware, potentially integrated into advanced packaging (like 2.5D/3D integration) to connect processors, memory, and other chiplets at unprecedented speeds and efficiency. The partners describe this as a "foundational step toward scalable, wafer-level optical interconnects for next-generation computing."
While no specific timeline or investment figures were disclosed, the announcement signals a significant push to mature a key enabling technology for AI infrastructure. Success could help alleviate one of the major bottlenecks in AI system performance, supporting the development of more powerful and efficient data centers and high-performance computing systems.