Huawei Ascend 950 SuperNode Wins 2026 World Internet Conference Leading Science and Technology Award
At the 2026 World Internet Conference held in Wuzhen, Huawei's Ascend 950 SuperNode won the Leading Science and Technology Award. In its fully configured form, the SuperNode can hold 1,024 NPUs, paired with more than 1,300 interconnect chips and more than 4,000 optical modules, achieving a computing power utilization rate of 45% and a minimum inference latency of 5 milliseconds. More than 40 large models have completed native pre-training on it, and the Ascend series SuperNodes have been deployed in more than 1,000 sets cumulatively. This year's Leading Science and Technology Award received more than 430 applications from 43 countries and regions, and ultimately selected 20 award-winning achievements.
Why read this
AI score 64
The original text provides the fully configured specifications, computing power utilization rate, and deployment scale of the Ascend 950 SuperNode, which can be used to understand the current form of domestic SuperNode computing power systems.
Topics
Body · AI translation
Powering trillion-parameter large models! Huawei Ascend 950 SuperNode wins a top global award
Fast Technology Download the app for exclusive early access November 03, 2026 07:17:11 From Henan Updated November 03, 2026 07:26:04
Fast Technology news, November 3: At the 2026 World Internet Conference held in Wuzhen, the Huawei Ascend 950 SuperNode received the Leading Science and Technology Award. This is also a top-tier international science and technology honor in the global internet field, representing that this domestic computing power system has gained authoritative industry recognition.
The Ascend 950 SuperNode is currently the SuperNode device with the largest commercial scale. Relying on the self-developed Lingqu interconnect protocol, it fuses a large number of computing power chips together and logically operates them as one super large computer.
The fully configured version can be equipped with 1,024 NPUs, paired with more than 1,300 interconnect chips and more than 4,000 optical modules, and is specifically used to run ultra-large AI models with trillion and ten-trillion parameters.
Its computing power utilization rate can reach 45%, roughly 1.8 times the industry average, meaning the same hardware can produce more actual computing power.
The minimum inference latency can be reduced to 5 milliseconds, making it very suitable for running AI applications such as agents that have high speed requirements. Currently, more than 40 large models have completed native pre-training directly on this Ascend hardware.
In traditional clusters, communication between chips incurs considerable overhead. The Ascend 950 SuperNode uses unified memory addressing, so more than a thousand cards can retrieve data from one another without repeated relaying, reducing wasted waiting.
To date, more than 1,000 Ascend series SuperNodes have been deployed, and they are used in many industries including the internet, telecom operators, finance, education, and energy. As large models grow larger, demand for computing power has surged, and SuperNodes are gradually becoming the mainstream choice for AI infrastructure.
This Leading Science and Technology Award solicits entries globally. This year, it received more than 430 applications from 43 different countries and regions, and in the end only 20 winning achievements were selected. That the Ascend 950 SuperNode stood out from them shows that the system architecture innovation of the domestic computing power system has been recognized.
"Special statement: The content of the above works (including videos, pictures, or audio if any) is uploaded and published by users of "Dafeng Hao," a self-media platform under Phoenix Net. This platform only provides information storage space services.
Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."