Lightmatter Announces Passage L200 Optical Interconnect CPO and Passage M1000 Photonic 3D Superchip

Lightmatter has announced its first product, the Passage L200 3D co-packaged optics (CPO) optical interconnect, designed to integrate with the latest GPU and XPU designs and switches, designed to deliver significant increases in AI throughput across massive clusters of thousands of GPUs by eliminating interconnect bandwidth bottlenecks.

The L200 3D CPO family includes 32Tbps (L200) and 64Tbps (L200X) versions, delivering 5-10x the throughput of existing solutions. The L200 64Tbps enables multiple GPUs to be packaged together on a single chip, delivering over 200Tbps of total I/O bandwidth, enabling AI model training and inference to be accelerated by up to 8x.

Image source: Lightmatter

In conventional data centers, accelerators are connected to each other using an array of network switches, which form a sort of multi-level hierarchy. This architecture creates too much latency, since the signal must pass through several switches in order for one accelerator to communicate with another.

As Lightmatter CEO Nick Harris previously told SiliconANGLE, Passage solves the problem of bulky network connections by embedding its ultra-dense fiber optic technology into data center chips to improve throughput by 100 times over the best solutions in use today. Essentially, it bundles its fiber optic connections directly into the silicon chip package. “So instead of six or seven layers of switches, you have two, and each GPU can connect to thousands of others,” he explained.

Lightmatter calls its interconnect architecture “edgeless I/O,” and says it can scale bandwidth across the entire die on a GPU, whereas traditional dies can only connect to other dies at the shoreline. Passage 3D integration allows SerDes I/O to be placed anywhere on the die, rather than being limited to the edge, delivering the equivalent bandwidth of 40 pluggable optical transceivers. The modular 3D CPO solution reportedly leverages a standard UCIe die-to-die (D2D) compliant interface and simplifies the scalable chiplet-based architecture for seamless integration with XPUs and next-generation switches.

The company said the upcoming L200 CPO is designed for high-volume production, and it is working closely with semiconductor partners such as Global Foundries, ASE and Amkor, as well as leading CMOS manufacturers, to prepare it. Lightmatter L200 and L200X will enter volume production next year.

Lightmatter also announced the Passage M1000 reference platform, a 3D Photonic Superchip designed for next-generation XPUs and switches. Passage M1000 delivers a record-breaking 114 Tbps of total optical throughput for the most demanding AI infrastructure applications.

The 4,000mm²+ M1000 is a multi-mesh active photonic interposer that allows customers to create their own custom interconnects using silicon photonics, providing connectivity to multiple GPUs in a single 3D package.

The Passage M1000 is said to overcome the connectivity limitation at the edges by providing electro-optical I/O virtually anywhere on its surface to the die stack on top. The interposer’s pervasive connectivity is enabled by an extensive and reconfigurable waveguide network that carries high-bandwidth WDM optical signals throughout the M1000. With fully integrated connectivity supporting 256 fibers at 448 Gbps per fiber, the M1000 delivers an order of magnitude more bandwidth in a smaller package than conventional Co-Packaged Optics (CPO) and similar offerings. The Passage M1000 is scheduled to ship this summer.

Lightmatter’s investors include major tech companies like Alphabet and HPE. In its most recent funding round in October 2024, Lightmatter raised $400 million, bringing the company’s total raised to $850 million and giving it a market value of $4.4 billion.