Huawei Ascend 950 SuperPoD — largest commercial AI supernode wins global award
At the World Internet Conference in Wuzhen on November 3, 2026, Huawei's Ascend 950 SuperPoD received the event's Pioneering Science and Technology Award, recognition that China's largest commercial AI supernode has earned global authoritative standing.
A Supernode at Scale
The Ascend 950 SuperPoD is the largest commercial supernode shipped today. Built on Huawei's self-developed Lingqu interconnect protocol, it logically operates as a single giant computer and is configurable with 128 or 1024 NPUs. A full configuration packs 1,024 NPUs alongside more than 1,300 interconnect chips and over 4,000 optical modules, purpose-built for trillion- and ten-trillion-parameter pretraining and inference.
Efficiency and Latency
The system reaches 45 percent training compute utilization — about 1.8 times the industry average — thanks to unified memory addressing that lets thousands of chips pull data without repeated hops. Inference latency can drop to 5 milliseconds, well suited to real-time agent applications. More than 40 large models have already been natively pretrained on Ascend hardware.
Deployment Footprint
Ascend supernodes have been deployed in over 1,000 units across internet, telecom, finance, education, and energy sectors. From 430-plus submissions spanning 43 countries and regions, only 20 projects won this year, underscoring the competitiveness of the system-architecture innovation.
Key Data Points
- NPUs in full configuration: 1,024
- Training compute utilization: 45 percent (about 1.8 times industry average)
- Lowest inference latency: 5 milliseconds
- Models natively pretrained on Ascend: 40 plus
- Supernodes deployed: 1,000 plus
Sources: C114, ifeng, AIBase, World Internet Conference, threerouter
Top comments (0)