News
News

All use domestic chips! China's first 10000 card level full stack autonomous controllable intelligent computing cluster lights up

Time:2026-03-30Reading:620Second


According to media reports, Shenzhen's high-quality development of artificial intelligence has achieved a new breakthrough: the 11000P intelligent computing power cluster invested and built by Shenzhen has officially been lit up, and with the addition of 3000P that was lit up earlier last year, the cluster has fully built 14000P intelligent computing power.
06.png

This is the first self controlled intelligent computing cluster in China that uses advanced chips produced nationwide to build a 10000 card full stack.

The intelligent computing cluster that was lit up this time is also the first 10000 card Ascend 910C super node intelligent computing cluster in China, marking a breakthrough in Shenzhen's construction of an independent and controllable computing power base.

The project adopts domestically advanced chips and relies on full stack independent controllable technology to create an internationally competitive "Ascend CAAN" independent software and hardware ecosystem, effectively guiding innovation in underlying artificial intelligence technologies and promoting the localization process of key technologies.

The key technical indicators of the project have reached the international leading level. The failure rate has been significantly reduced, with an average daily failure rate of 0.3 ‰ for the first stage 3000P cluster tested intelligent computing equipment, significantly lower than the level of top enterprise clusters such as Meta.

At the same time, the linearity is significantly enhanced, achieving a training linearity of 93.12% for the Pangu-718B large model, with performance comparable to international top enterprise clusters such as Google; Green, intensive and efficient, adopting energy-saving solutions such as natural cooling sources, intelligent computing equipment full liquid cooling, and integrated power supply, comprehensively carrying out energy consumption optimization in the computer room, and continuously reducing operating costs.

Relying on high security facility deployment, low-cost training support, high-quality data corpus supply, and market-oriented ecological operation, this project will provide strong support for accelerating the construction of a national integrated computing power network and promoting the high-quality development of domestic large models and chips.