(Yicai) Sept. 16 -- Chinese tech giant Huawei Technologies has launched the world's first three-dimensional data center architecture, aiming to enhance energy efficiency by vertically stacking functional spaces to meet artificial intelligence's growing demand for computational power.

The 3D data center vertically stacks different functional modules of traditional data centers, comprising cooling, information technology equipment, power supply, and backup power layers from bottom to top, Shenzhen-based Huawei said at the AIDC Industry Development Conference held in Wuhu yesterday. The framework can serve as a reference for building supercomputing clusters with over 100,000 cards, it pointed out.

Amid the explosion in AI demand for computing power, data center power consumption has drawn increasing attention, Peng Peng, secretary general of the Global Computing Consortium, told Yicai. The link between computing power, electricity, and cooling has become a critical challenge that must be addressed within data centers, Peng stressed.

Traditional flat data centers face various issues, including long power supply distances, energy transmission losses, and an inability to regulate temperature locally, Peng pointed out, noting that because the power of AI cabinets has increased from 20 to 30 kilowatts to 80 to 200 kW, these problems have become even more pronounced.

Through cross-floor distributed deployment, 3D data centers can effectively handle such pain points, Peter Zhou, director at Huawei and chief executive officer of the firm's cloud arm, noted at the conference. In addition, compared with traditional flat data centers, the 3D architecture can shorten construction cycles to keep pace with the accelerating iteration of AI chips, Zhou noted.

Huawei Cloud has already used the 3D data center architecture at its Wuhu base, reducing construction time to just nine months from the traditional 18 to 24 months, significantly speeding up the delivery of computing infrastructure, according to the company.

AI model parameters have increased from the hundreds of billions to the trillions and will likely expand to tens of trillions and beyond, Zhou said. This shows that AI infrastructure will enter a phase of rapid development over the next five to 10 years, he stressed.

Editors: Dou Shicong, Martin Kadiev