Перейти к содержимому
Новости

NVIDIA Shows Prototype Optics-on-Interposer CPO Architecture and 256Gb/s Single-Fibre DWDM Demo

Редакция ЦОД:Главное

TechNews 科技新報 — иллюстрация к материалу
Источник изображения: TechNews 科技新報

NVIDIA has presented Optics-on-Interposer (OOI), a prototype co-packaged optics architecture that places the optical engine on the compute chip interposer rather than the organic substrate. The company also demonstrated DWDM transmission of 256Gb/s over a single fibre using eight 32Gb/s wavelengths.

Главное
  • OOI is a prototype CPO architecture, not a confirmed commercial deployment.
  • The design places the optical engine on the compute-chip interposer rather than an organic substrate.
  • NVIDIA demonstrated eight 32Gb/s DWDM wavelengths, totaling 256Gb/s on one fibre.
  • The reported production use of Quantum-X Photonics and Spectrum-X Photonics concerns related CPO solutions, not OOI itself.
  • Commercial timelines, product plans and manufacturing volumes for OOI were not disclosed.

At Semicon West 2026 in San Francisco, NVIDIA senior researcher Nikola Nedovic presented development results for Optics-on-Interposer, or OOI, a new co-packaged optics (CPO) approach.

The architecture moves the optical engine from an organic substrate onto the interposer of the compute chip. By shortening the electrical path between compute silicon and the optical engine, the approach is intended to address power and density constraints associated with high-bandwidth AI-cluster interconnects.

NVIDIA describes OOI as a prototype rather than a commercial product. The company is evaluating potential integration into future products while addressing technical barriers to commercialization. No timeline, product designation or deployment volume for OOI was disclosed.

Separately, NVIDIA demonstrated dense wavelength-division multiplexing using eight simultaneous wavelengths at 32Gb/s each, for aggregate transmission of 256Gb/s on a single optical fibre. The result is a demonstration metric and does not establish volume deployment in data centres.

The source also reports that related CPO technologies are already used in production Quantum-X Photonics systems for GB300 NVL72 rack interconnects and in Spectrum-X Photonics Ethernet switches for the Vera Rubin platform. NVIDIA is also building a CPO manufacturing ecosystem spanning wafer production, advanced packaging, external laser modules and fibre optics.

For operators designing AI-scale networks, the development points to continued efforts to remove electrical-interconnect bottlenecks in power and bandwidth density. If OOI reaches commercialization, it could increase demand for advanced packaging, optical engines, laser modules and high-density fibre infrastructure. That transition remains subject to unresolved technical and manufacturing risks.

Почему это важно

For AI data-centre networks, moving the optical engine closer to the compute silicon could reduce the electrical path to optics, potentially improving bandwidth density and lowering interconnect power consumption. OOI remains a prototype, however, and its effect on deployed AI infrastructure has not yet been demonstrated.

Источники

Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.