Industry giants unite to define open co-packaged optics for next-gen AI infrastructure

Lightmatter has formally launched the Open Silicon Photonics for AI Systems initiative, an Open Compute Project workstream backed by Celestica, Dell, Keysight, and Qualcomm, to define open co-packaged optics specifications for AI clusters of 72 to over 1,000 nodes. The initiative's 294-page white paper outlines an architecture that supports existing OCP ORv3 and modular hardware system platforms, aiming to reduce power consumption and future-proof interconnects. Lightmatter CEO Nick Harris said, 'Interconnect has become as fundamental to AI infrastructure performance as compute.'

Some of the industry’s top interconnect designers have come together to form a coalition developing open specifications for co-packaged optics CPO for AI infrastructure. Dubbed the Open Silicon Photonics for AI Systems initiative, the effort was first teased back in March https://lightmatter.co/press-release/lightmatter-announces-reference-architecture-initiative-with-industry-leaders-in-the-open-compute-project-for-co-packaged-optics/ and has now formally launched with California-based Lightmatter leading the charge. The likes of Celestica, Dell, Keysight, and Qualcomm have pledged support for a project in which vendors will create a shared blueprint for CPOs capable of supporting AI clusters ranging from 72 to more than 1,000 nodes. The initiative, now an official workstream within the Open Compute Project OCP , kicked off with a white paper outlining its “architectural vision.” The 294-page opus https://www.opencompute.org/documents/vision-opensiphforaisystems-final-pdf details design goals including open co-packaged interconnects that can extend to existing OCP ORv3 or modular hardware system MHS platforms – meaning engineers wouldn’t have to alter liquid-cooling manifolds or tray form factors to deploy photonic interconnect offerings. Lightmatter and co want agnosticism at the platform level, with passive optical infrastructure like fiber routing connectors able to work with coherent optics and future interconnect options like micro-LEDs. Desired blueprints have further futureproofing desires, with the white paper suggesting passive components should largely remain unchanged even as serializer-deserializer SerDes data rates rise, with only the optical engines at the tray level being the hardware needing routine updates. The OCP-backed effort aims to create truly open standards, with the prospective architecture able to scale-up protocols like NVLink and UALink as well as scale-out protocols. The white paper outlines that multiple independent fabric planes “may coexist on the same passive infrastructure,” and that deployments should be able to support distinct protocols across different networking domains. “Interconnect has become as fundamental to AI infrastructure performance as compute,” said Nick Harris, Lightmatter’s founder and CEO. “This white paper lays the foundation upon which this growing OCP community will develop a shared MHS-based CPO rack architecture that fits into hyperscalers’ build flows. The work begins now.” The formation of the Lightmatter-led initiative comes amid an ever-growing interest in co-packaged optics. While not a new concept – Nvidia, for example, has been hyping up CPOs https://www.sdxcentral.com/news/nvidia-unveils-silicon-photonics-networking-switches-to-reduce-data-center-power-consumption/ for some time – the idea of placing optical engines directing on the same package as a switch or ASIC Application-Specific Integrated Circuit aims to dramatically reduce hardware power consumption while providing faster speeds for interconnects. Lightmatter’s own work to bring silicon photonics to life includes the unveiling of a liquid-cooled laser network interface card https://www.sdxcentral.com/news/lightmatter-unveils-guide-dr-laser-nic-to-support-cpo-scale-up/ in May that supports high-bandwidth optical interconnects. Meanwhile, last August, it boasted test results that saw a bidirectional dense wavelength division multiplexing DWDM link capable of carrying 16 independent wavelengths https://www.sdxcentral.com/news/lightmatter-claims-ai-friendly-eight-fold-leap-in-wavelength-density/ over a single strand of standard single-mode fiber SMF , an apparent industry first that offers a glimpse at possible future bandwidth densities for interconnected chips. “As scaling demands accelerate, optimizing energy efficiency and data transfer speeds across AI systems is now just as critical as raw processing power,” said Jeff Janukowicz, VP and global lead for compute infrastructure and technologies at IDC. “The physical limitations of copper are increasingly constraining system scalability and accelerator utilization in large-scale AI clusters, and that gap widens with each new generation of AI accelerators.” “The Open Silicon Photonics for AI Systems initiative establishes a vendor-neutral framework that can accelerate ecosystem collaboration and help move co-packaged optics from early deployments toward mainstream AI infrastructure,” Janukowicz added.