Chinese Research Team Develops Technology with 10x Performance of US Supercomputer

Hong Kong Media Reports
"Using Domestically Produced GPUs"

A Chinese research team has developed technology using domestically produced graphics processing units (GPUs) that delivers performance approximately 10 times greater than that of American supercomputers equipped with advanced Nvidia chips.

The photo is unrelated to the article content.

According to the South China Morning Post (SCMP) in Hong Kong on the 12th, a research team led by Professor Nan Tongchao of the State Key Laboratory of Hydrology-Water Resources and Hydraulic Engineering at Hehai University in Nanjing, Jiangsu Province, published a peer-reviewed paper on this topic in the Chinese Journal of Hydraulic Engineering on the 3rd of last month.

Exports of Nvidia's advanced AI chips A100 and H100 to China are prohibited. Additionally, from a software perspective, Nvidia restricts its AI development platform CUDA from running on third-party hardware.

Under these circumstances, Professor Nan's research team focused on enhancing the efficiency of computers equipped with Chinese GPUs through innovations in software optimization technology. To overcome the performance limitations of Chinese central processing units (CPUs) and GPUs, they introduced a new architecture that combines multiple GPUs into a single node.

The team also improved data exchange between nodes on the software side to reduce inter-node overhead (indirect and additional computing resources required to perform specific functions). This outperformed the flood prediction model "TRITON" developed using the American Oak Ridge National Laboratory's high-performance supercomputer "Summit" by about ten times.

Professor Nan's team explained that when applying this model to the Jiangli Reservoir in Zaozhuang, Shandong Province, eastern China, they simulated the flood progression process in just three minutes using 200 computing nodes and 800 GPUs.

The research team stated, "The results of this study can be applied not only to flood modeling but also to simulations of complex systems such as hydrometeorology (the study of water presence in the atmosphere), sediment, and surface water-groundwater interactions."