Huawei has introduced the Atlas 960E SuperPoD, a new large-scale AI computing system designed for training and inference workloads involving increasingly large AI models.
The company announced the system on September 17 at HUAWEI CONNECT 2026 in Shanghai.
Huawei says the Atlas 960E is the industry’s first SuperPoD built around Near-Packaged Optics, or NPO, and is designed to support AI models approaching 10 trillion parameters.
Built for Large-Scale AI
The Atlas 960E combines Huawei’s UnifiedBus technology, also known as Lingqu, with the company’s Hi-ONE optical system.
The architecture uses unified memory addressing, allowing computing resources across the system to operate more like a single computing environment rather than separate machines.
A single Atlas 960E SuperPoD can support up to 4,096 Ascend AI processors.
Huawei says the system can deliver up to 8 EFLOPS of FP8 performance and 16 EFLOPS of FP4 performance.
The Atlas 960E also uses a fully liquid-cooled design and an orthogonal system architecture.
Hi-ONE Reduces Optical Hardware
A major part of the system is Huawei’s Hi-ONE NPO optical engine.
Huawei says each Hi-ONE unit supports up to 7.2 terabits per second of transmission capacity and includes an integrated light source.
The Atlas 960E uses around 5,500 Hi-ONE units, replacing roughly 48,000 conventional 800G optical modules.
According to Huawei, the design can reduce power consumption by more than 550kW.
The company also claims the system can double fault-free operating time and achieve 99.8% system availability.
Scaling to 1 Million AI Processors
Huawei plans to extend the SuperPoD architecture beyond AI accelerators.
Its upgraded TaiShan 950 SuperPoD supports up to 4,096 nodes and provides a 256TB unified memory pool.
The OceanStor M900, meanwhile, provides petabyte-scale key-value caching for AI workloads.
Huawei also plans to connect multiple Atlas 960E SuperPoDs into much larger AI clusters using UnifiedBus and RoCE networking.
The company says the architecture can connect up to 512,000 AI processors.
Using a multi-rail topology, Huawei says an Ascend SuperCluster can scale to as many as 1 million processors.
Launch Timeline
The Atlas 960E SuperPoD is scheduled to launch in Q3 2027.
Huawei is also preparing several new generations of Ascend AI processors.
The Ascend 960 DT is scheduled for Q1 2027, while the Ascend 960 PR is expected in Q3 2027.
Huawei then plans to introduce the Ascend 970 in 2028, followed by the Ascend 980 in 2029.
Huawei Atlas 960E SuperPoD Specifications
| Feature | Huawei Atlas 960E SuperPoD |
|---|---|
| Launch | Q3 2027 |
| AI Model Target | Up to 10 trillion parameters |
| NPO Optical Engine | Hi-ONE |
| Transmission Capacity | 7.2 Tbit/s |
| Maximum Ascend Cards | 4,096 |
| FP8 Performance | 8 EFLOPS |
| FP4 Performance | 16 EFLOPS |
| Optical Modules | 5,500 Hi-ONE |
| Replaced 800G Modules | Around 48,000 |
| Claimed Power Reduction | Over 550kW |
| System Availability | 99.8% |
| Interconnect | UnifiedBus / Lingqu |
| Cooling | Fully liquid-cooled |
| Maximum SuperCluster Scale | Up to 1 million Ascend processors |
The post Huawei Unveils World’s First NPO-Based AI SuperPoD With 4,096 Processors appeared first on ProPakistani.
