On November 3, at the 2026 World Internet Conference held in Wuzhen, the Huawei Ascend 950 SuperPoD officially won the Leading Technology Award. As a top international technology honor in the global internet field, the award solicits entries worldwide. This year, it received more than 430 applications from 43 countries and regions, and only 20 winning achievements were ultimately selected. This marks that this domestic computing power system has gained highly authoritative industry recognition.

Core Architecture: Self-developed Lingqu Interconnect Breaks Through the Trillion-Parameter Bottleneck

As the largest commercially deployed SuperPoD device currently available, the core advantage of the Ascend 950 SuperPoD lies in its self-developed Lingqu interconnect protocol, which tightly interweaves a large number of computing chips and logically enables them to operate together as a single supercomputer.

In terms of hardware scale, its fully configured version can house up to 1,024 NPUs, along with more than 1,300 interconnect chips and over 4,000 optical modules. It is purpose-built to support pre-training and inference for ultra-large AI models with trillion- to ten-trillion-parameter scales.

image.png

Performance Breakthrough: Leading Compute Utilization, Latency Reduced to 5 ms

To address common pain points in traditional large-model training, such as high inter-chip communication overhead and long waiting times, the Ascend 950 SuperPoD innovatively adopts unified memory addressing technology. This allows thousands of accelerator cards to retrieve data from one another without repeated transfers, effectively reducing wasted waiting.

In actual operation, the system’s compute utilization can reach 45%, about 1.8 times the industry average. This means the same hardware configuration can unlock more real compute. At the same time, its minimum inference latency can be compressed to 5 ms, perfectly matching AI application scenarios with extremely high real-time response requirements, such as AI agents. Currently, more than 40 large models have completed native pre-training directly on this hardware platform.

Industrial Deployment and Future Iteration: Accelerating the Building of a Domestic Computing Power Foundation

As large models continue to scale up and demand for computing power keeps rising, the SuperPoD architecture is gradually becoming a mainstream choice for new AI infrastructure. Official data shows that Ascend series SuperPoDs have now been deployed in more than 1,000 sets nationwide, widely covering core industries such as the internet, telecom operators, finance, education, and energy.

Huawei officials said that in the future it will continue to iterate and upgrade along the technology path of “SuperPoD plus super cluster,” building a solid, independent and controllable computing power foundation for the rapid development of China’s AI industry.