英伟达在 SEMICON West 2026 公布下一代 OOI（optics-on-interposer）CPO 原型及 DWDM 技术，仍处研究阶段

Nvidia has unveiled next-generation co-packaged optics (CPO) technology that places optical engines directly on the interposer carrying computing chips rather than on the package substrate.
By eliminating the organic substrate from the electrical signal path, the technology can increase bandwidth density and reduce power consumption. Nvidia said it has developed a prototype.
Nikola Nedovic, a principal researcher at Nvidia, presented the research at SEMICON West 2026 in San Francisco on Oct. 13.
The technology, called optics-on-interposer (OOI), moves optical engines from the organic package substrate onto the interposer that carries computing chips. Electrical signals travel directly through the interposer’s short, fine-pitch interconnects, similar to the architecture connecting graphics processing units (GPUs) and high-bandwidth memory (HBM). While existing mass-produced CPO brings optical engines closer to chips, OOI places them even closer. Current designs route signals from the chip through the interposer and organic substrate to the optical engine. OOI eliminates the organic substrate from that path.
Nvidia also unveiled dense wavelength-division multiplexing (DWDM) technology designed to increase the amount of data carried by a single optical fiber. The technology transmits separate data streams using multiple wavelengths of light, which are separated at the receiver. Nvidia demonstrated eight wavelengths operating at 32Gb/s each, delivering 256Gb/s per fiber.
Both technologies remain in the research stage. “We are proposing and evaluating technologies for future products while addressing the challenges that prevent commercialization,” Nedovic said.
CPO integrates optical engines and switch application-specific integrated circuits (ASICs) within the same package. Conventional networking equipment sends electrical signals to pluggable optical transceivers mounted on the front panel of a switch, where they are converted into optical signals. CPO moves optical engines containing components such as optical modulators and photodetectors next to the ASIC. Lasers are separated into external light source modules, enabling high-density connectivity while reducing power consumption and heat generation.
Nvidia has already deployed CPO technology in mass-produced products, including Quantum-X Photonics and Spectrum-X Photonics. Quantum-X mounts fiber-connected optical engines on the substrate using sockets, while Spectrum-X uses soldered optical engines and detachable fiber connectors. Quantum-X is used in InfiniBand networks connecting Blackwell Ultra-based GB300 NVL72 racks, while Spectrum-X is deployed in Ethernet switches for the Vera Rubin platform.
Nvidia has established a CPO manufacturing ecosystem with multiple partners. TSMC provides its COUPE silicon photonics packaging technology, which integrates electronic integrated circuits (EICs) and photonic integrated circuits (PICs). TSMC also handles contract manufacturing alongside Taiwan-based semiconductor assembly and testing company SPIL. Coherent, Lumentum and Sumitomo Electric manufacture external laser source modules, while Corning, Senko, Browave and TFC supply optical fibers and related components. Foxconn and Fabrinet handle cooling and final testing.
“An entire ecosystem is needed to make CPO technology suitable for mass production,” Nedovic said. “Key partners must provide pre- and post-assembly testing, fiber alignment, temperature control, cooling and maintenance technologies to ensure optical engines can withstand package assembly processes.”