Shanghai Wanka GPU Cluster Supports the Development of AI Large Models

SadMoneyMeow · 2026-04-06T07:13:48+00:00

Crypto News Agency reported on April 6 that inside the Songjiang Yidian Intelligent Computing Center in Shanghai, over ten thousand GPU devices are working efficiently in coordination, forming a massive computing power network. Each chip can perform hundreds of billions of operations per second, and the total computing power over two days is equivalent to the one-hour electricity generation of a Three Gorges Hydropower Station unit.To ensure the stable operation of the 10,000-card cluster, Shanghai Yidian Intelligent Computing Technology team has overcome the challenge of high-fault-rate chip collaboration. Through algorithm optimization and parallel scheduling, they achieved an ultra-high availability of 99.99%, with fault time controlled within 1 hour throughout the year, effectively supporting world-leading training of multimodal large models and avoiding significant computing power losses.The cluster has successfully adapted to various domestically produced computing cards, achieving heterogeneous computing power integration and optimized scheduling, with utilization rates close to 100%. The team pioneered a "dynamic perception scheduling scheme," increasing training efficiency by 91%, saving computing costs equivalent to building three new data centers annually, and has now developed the capability for rapid deployment of 10,000-card clusters.Chairman of Zhisan Technology

SadMoneyMeow

2026-04-06 07:13:48

LianheTech reported on April 6 that within the Shanghai Songjiang Yidian Zhisu Computing Center, tens of thousands of GPU devices coordinate efficiently to form a massive computing power network. Each chip can perform tens of trillions of operations per second; in two days, the total computing power is equivalent to the one-hour generation output of a single unit of the Three Gorges Hydropower Station.

To ensure stable operation of a ten-thousand-card cluster, the Shanghai Yidian Zhisu Technology team tackled the challenge of chip coordination with a high failure rate. Through algorithm optimization and parallel scheduling, they achieved 99.99% ultra-high availability. The total downtime was kept within 1 hour throughout the year, effectively safeguarding the globally leading training of multimodal large models and avoiding major compute losses.

This cluster has successfully been adapted to multiple types of domestic computing power cards, enabling the integration of heterogeneous computing power and optimized scheduling, with utilization close to 100%. The team pioneered a “dynamic sensing and scheduling solution,” improving training efficiency by 91%. The annual savings in computing power costs is equivalent to building 3 new data centers, and they have already developed the capability to quickly deploy a ten-thousand-card cluster.

Sun Yue, Chairman of Zhisu Technology, said the company is advancing “computing-electricity coordination” to expand the application of green power, adopting liquid cooling technology to reduce PUE, and exploring offshore wind power to drive data centers. Meanwhile, through the YI CLOUD platform and ecosystem coordination, the company will promote domestic computing power to fully empower AI+ applications.

(Editor: Zhang Xiaobo ）

     【Disclaimer】This article only represents the author’s personal opinions and is not related to Hexun. The Hexun website remains neutral regarding the statements and opinion judgments made in the text, and does not provide any express or implied guarantee regarding the accuracy, reliability, or completeness of the content included. Please readers use it only as a reference and bear all responsibility yourselves. Email: news_center@staff.hexun.com

View Original

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.