Summary: NVIDIA has revealed its next-generation co-packaged optics (CPO) research at SEMICON West 2026, introducing Optics-on-Interposer (OOI) and an eight-wavelength DWDM solution capable of delivering 256Gbps per optical fiber. Prototype samples have been completed, but both technologies remain in the research stage with no confirmed commercialization timeline.

As AI clusters continue to expand, data transmission between chips is emerging as a major system bottleneck. NVIDIA's newly disclosed OOI architecture addresses this challenge by removing the organic package substrate and placing optical engines directly onto the silicon interposer that carries the GPU. This allows electrical signals to reach the optical engine through shorter interconnections without passing through the organic substrate.

Existing mass-produced co-packaged optics (CPO) solutions typically position optical engines on the organic package substrate alongside switch chips. Although this approach significantly reduces electrical transmission distances compared with traditional pluggable optical architectures, signals must still travel through both the silicon interposer and organic substrate before optical conversion.

NVIDIA presented a comparison showing that electrical signal loss falls from approximately 22dB in traditional pluggable architectures to around 4dB in existing CPO designs. OOI takes this integration a step further by eliminating the organic substrate from the electrical signal path. Its shorter, finer-pitch interconnections on the silicon interposer could reduce power consumption and increase bandwidth density.

The architecture resembles the 2.5D packaging approach used to connect GPUs with high-bandwidth memory (HBM) through silicon interposers. However, NVIDIA has not disclosed specific signal-loss measurements for OOI.

Earlier research by NVIDIA Chief Research Scientist Nikola Nedovic explored how high-density interposer routing could support wider, lower-speed electrical interfaces, potentially improving interconnect efficiency. However, bringing optical engines closer to high-power computing chips also introduces more demanding thermal-management requirements.

Alongside OOI, NVIDIA demonstrated a dense wavelength-division multiplexing (DWDM) solution that transmits eight wavelengths simultaneously through a single optical fiber. Each wavelength supports up to 32Gbps, delivering a combined bandwidth of 256Gbps per fiber.

NMDIACPO's new technology OOI revealed: the optical engine is directly mounted on the silicon interposer, and paired with DWDM, a single optical fiber can reach speeds of 256Gb/s-1.

DWDM increases the capacity of existing optical links without requiring a proportional increase in fiber count. Combined with OOI's high-density electrical interfaces, the technology could further expand bandwidth capabilities within advanced semiconductor packages.

Despite the completed prototypes, both technologies remain at the research stage. Nedovic said the team is evaluating potential technology pathways for future products while addressing engineering challenges that could prevent commercialization.

Key obstacles include thermal management, manufacturing yield, and serviceability. Integrating temperature-sensitive optical components alongside high-power GPUs requires careful thermal control, while advanced packaging processes must achieve sufficient reliability and production consistency for large-scale deployment.

Overall, OOI represents a potential next step in the evolution of CPO, moving optical engines closer to computing chips by integrating them directly onto silicon interposers. Meanwhile, DWDM increases optical transmission capacity through wavelength multiplexing. Although prototype development marks technical progress, further engineering work will be necessary before either technology reaches commercial production.