Huawei

AI Infrastructure · Semiconductors · Data Center Networking

Launch

Huawei launches Atlas 960E SuperPoD with first mass-produced NPO optics

Replacing tens of thousands of optical modules with a few thousand chip-adjacent engines cuts power draw and interconnect delay at once.

• Huawei unveiled the Atlas 960E SuperPoD at HUAWEI CONNECT 2026 in Shanghai on September 17, calling it the industry's first SuperPoD built on Near-Packaged Optics (NPO).
• A SuperPoD links thousands of computing nodes via high-speed interconnects and unified memory addressing so they function as one logical computer, aimed at training and inference for models approaching 10 trillion parameters.
• A single Atlas 960E supports up to 4,096 NPUs, delivers 8 EFLOPS at FP8 (16 EFLOPS at FP4), and holds up to 1 petabyte of high-bandwidth memory.
• The system's Hi-ONE optical engine delivers 7.2 Tbit/s of transmission and replaces roughly 48,000 conventional 800G optical modules with about 5,500 units, cutting power use by more than 550 kW while doubling fault-free operating time to reach 99.8% availability.
• Huawei says communications overhead can consume more than 40% of training time in conventional server-based clusters, the core bottleneck NPO and its UnifiedBus interconnect are designed to reduce.
• Huawei plans to link multiple Atlas 960E SuperPoDs into an Ascend SuperCluster scaling to up to 1 million NPUs via UnifiedBus and RoCE networking, with the SuperPoD itself scheduled for Q3 2027.
• As foundation models near 10 trillion parameters, chip-to-chip data movement rather than raw compute is becoming the limiting factor, pushing AI hardware competition from individual accelerators toward full system-level architecture.