Business · AI Infrastructure

Huawei launches the Atlas 960E SuperPoD at HUAWEI CONNECT 2026, supporting AI models with up to 10 trillion parameters

Scales up to 4,096 NPUs per cluster, delivering 8 EFLOPS (FP8) using the Hi-ONE Optical Engine running at 7.2 Tbit/s

By Techhaus ThailandFriday, September 25, 20262 min read
Share
Huawei launches the Atlas 960E SuperPoD at HUAWEI CONNECT 2026, supporting AI models with up to 10 trillion parameters

HUAWEI CONNECT 2026 opened in Shanghai, with David Wang, Huawei's Deputy Chairman of the Board and Rotating Chairman, delivering a keynote titled "Advancing the Agentic World, Building a Solid Silicon Foundation," reflecting the company's AI infrastructure vision. Huawei's strategy centers on computing power through SuperPoD and SuperCluster systems, alongside promoting open systems and open source.

As Foundation Models approach the 10-trillion-parameter scale, using a SuperPoD becomes increasingly important. Huawei's design connects every component within a cluster using a single protocol via peer-to-peer connectivity, supporting cross-server memory access, combining short-reach copper with optical communication for long distances. Huawei launched Hi-ONE, the world's first mass-produced NPO-based optical engine, transferring data at up to 7.2 Tbit/s, and the only NPO in the industry with a built-in light source.

The Atlas 960E SuperPoD, powered by UnifiedBus and Hi-ONE, uses an Orthogonal architecture and fully liquid-cooled design, with unified memory addressing via UnifiedBus allowing it to scale up to 4,096 NPUs per unit, delivering up to 8 EFLOPS at FP8 and 16 EFLOPS at FP4. Integrating 5,500 units of Hi-ONE into the system reduces the need for 48,000 units of 800G optical modules that would normally be required to connect all the NPUs, significantly cutting power use.

Source: Techhaus Thailand