Blockbuster release | AsiaInfo Data fills the gap in domestic technology, and has made a major breakthrough in the field of GPU heterogeneous adaptation in the field of localization

Recently, AsiaInfo Data's first artificial intelligence generic technology development support and achievement transformation platform supported by domestic computing power has been put into operation in Zhongguancun No. 1 Park. This means that Chinese enterprises have made a major breakthrough in the field of GPU heterogeneous adaptation, and the platform adopts the "Tongxin" AI heterogeneous adaptation all-in-one machine and supporting software independently developed by AsiaInfo Data, which realizes the localization of the AI computing platform and greatly improves the operation efficiency, breaks the "controlled by others" situation of relying on imported network equipment to improve the performance of AI computing power clusters in the past, builds China's own GPU software and hardware ecosystem, and fills a number of domestic technology gaps.

 

 

As the No. 1 project in Haidian District, the platform is a key scientific and technological innovation project of the Beijing Municipal Science and Technology Commission, and is built by AsiaInfo Data, which adopts GPU localization heterogeneous compatibility and adaptation technology with completely independent intellectual property rights, GPU computing power cluster performance optimization technology, as well as self-developed "Tongxin" AI heterogeneous adaptation all-in-one machine and supporting software, realizing the unified scheduling management and adaptation framework of heterogeneous computing resources, and providing low-cost and high-performance artificial intelligence large model training and inference services for 100,000 high-tech enterprises in the capital.

 

GPU is the core of computing power in the development of artificial intelligence, and its key field technology is difficult to develop and has been monopolized by Western countries for a long time. In recent years, China has adhered to innovation-driven and made significant progress in localization, and some products can be compared with international high-end chips. However, due to different design ideas, there is a certain degree of incompatibility and incompatibility of software and hardware between GPUs of different domestic manufacturers, imported GPUs and domestic GPUs, which affects the maximum play of the integrity of GPU clusters, resulting in low computing efficiency and increased operating costs of artificial intelligence large models, which seriously restricts the development and growth of China's artificial intelligence industry.

 

"In response to the pain points of the domestic GPU industry, we have fully integrated the three core technologies of artificial intelligence, chips, and cloud computing, concentrated on technical research, and realized the comprehensive adaptation of localized GPU hardware resources and the optimal scheduling of AI computing resources; By optimizing the cluster network, storage, and computing performance, the cost performance of domestic AI clusters surpasses that of similar international first-line products. Shen Danting, chairman of AsiaInfo Data, said that after preliminary testing, the use of GPU computing power cards designed by AsiaInfo Data on mainstream artificial intelligence models can improve computing power performance by more than 30%, and the operating cost has been significantly reduced, realizing the localization of domestic intelligent computing centers and greatly improving the operation efficiency.