Shanghai AI Laboratory achieves efficient integration of long-distance cross-domain heterogeneous computing power for the first time globally, solving industry bottlenecks.

The Shanghai Artificial Intelligence Laboratory released a milestone achievement yesterday, their research and development of the DeepLink ultra-large-scale cross-domain mixed training technology solution, successfully applied to the China Unicom network. This technology "combines" two heterogeneous smart computing centers separated by 1500 kilometers into a "super node", and completes the training of a large model with billions of parameters. This is the first time in the world to achieve efficient integration of long-distance cross-domain heterogeneous intelligence computing power, which not only can solve the problem of uneven distribution and low utilization of national computing resources but also reduce the dependence of the AI industry on specific chips. In the event of supply chain fluctuations, it will provide important backup computing power support for the AI industry and prevent being bottlenecked. According to the Shanghai AI Laboratory, in February of this year, they, together with more than ten partners, built a prototype of an ultra-large-scale cross-domain mixed training cluster in Shanghai, achieving uninterrupted training of a large model with billions of parameters for 20 days. Building on this foundation, they integrated the China Unicom AINET computing power network, spanning 1500 kilometers, connecting the smart computing centers between Shanghai and Jinan, completing the mixed training of a large model with billions of parameters.