Перейти к содержимому
Новости

Huawei Announces Ascend 960 SuperPoD and Updated UnifiedBus Interconnect

Редакция ЦОД:Главное

TechNews — иллюстрация к материалу
Источник изображения: TechNews

Huawei has introduced the Ascend 960 SuperPoD and an updated UnifiedBus interconnect architecture incorporating Hi-ONE optics. The company says the architecture can connect up to 4,000 processors within one SuperPoD and has outlined planned releases for the Atlas 950 and Atlas 960 SuperPoD systems.

Главное
  • Huawei says UnifiedBus with Hi-ONE optics can connect up to 4,000 processors in one SuperPoD.
  • The company has introduced near-packaged optics into the SuperPoD architecture for the first time.
  • Atlas 950 is planned for the fourth quarter of 2026, while Atlas 960 is planned by the end of 2027.
  • Huawei states that Atlas 960 can scale to 15,488 Ascend cards in 220 racks across 2,200 square metres.
  • No power, energy-use, PUE, CAPEX or grid-infrastructure parameters were disclosed.

Huawei unveiled its Ascend 960 SuperPoD and an updated UnifiedBus interconnect technology at Huawei Connect 2026. The company said UnifiedBus, combined with its Hi-ONE optical system, can connect up to 4,000 processors in a single SuperPoD.

Huawei said it is using near-packaged optics (NPO) in the SuperPoD architecture for the first time. The company also plans mass production of an NPO product with an integrated light source, although no delivery date was provided for that product.

The roadmap includes the Atlas 950 SuperPoD, planned for release in the fourth quarter of 2026, followed by the Atlas 960 SuperPoD by the end of 2027. Huawei said the Atlas 950 is designed to use 8,192 Ascend NPUs. For Atlas 960, it stated a maximum configuration of 15,488 Ascend cards across 220 racks and 2,200 square metres.

For data-centre infrastructure, the relevant shift is not solely accelerator count but the architecture needed to operate large AI systems as a coherent cluster. Huawei’s NPO and optical-interconnect approach could increase the importance of scalable optical fabrics within AI deployments. The stated Atlas 960 configuration also suggests that rack deployment, cooling, power delivery and intra-data-centre networking will need to be engineered together for high-density AI systems.

Separately, one source reported that Huawei researchers have described a deployed unified AI computing system at the 256K-node scale. The available material does not specify its location, configuration or operational status.

The available material is based on a single source, and Huawei’s stated cluster scale, system configurations, roadmap and other claims have not been independently confirmed.

Почему это важно

The announcement points to high-speed optical interconnects moving deeper into AI compute-system design as cluster scale increases. If Huawei’s stated specifications are realised, data-centre operators and AI-infrastructure suppliers will need to treat interconnects, optical infrastructure, rack density, cooling, power delivery and internal networking as an integrated design problem.

Источники

Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.