Chinese researchers have achieved a groundbreaking performance boost in supercomputing, using domestically developed GPUs. Their new computing model delivered a nearly tenfold speed improvement compared to Nvidia-powered US supercomputers. This achievement challenges America’s dominance in high-performance computing and shows China’s growing self-reliance in advanced technology.
The supercomputing system was built using a domestic x86 computing platform. It used Hygon 7185 CPUs, which feature 32 cores and 64 threads, running at 2.5 GHz. The GPUs, developed in China, were equipped with 128GB of memory. The system also utilized a 200 Gb/s network bandwidth to enable efficient communication between nodes.
With this breakthrough, China is not only catching up but also surpassing US supercomputers in certain areas. This marks a significant shift in the global tech race, as China continues to strengthen its position in high-performance computing.
China’s response to US Tech sanctions
The United States has placed strict sanctions on China, restricting access to Nvidia’s advanced GPUs like the A100 and H100. Additionally, Nvidia’s CUDA software is not compatible with non-Nvidia hardware, making independent algorithm development difficult. In response, China has focused on developing its own hardware and software solutions to bypass these technological barriers.
What is the innovation in Parallel Computing
Led by Professor Nan Tongchao from Hohai University in Nanjing, the research team created an innovative “multi-node, multi-GPU” parallel computing model. This approach optimizes data transfer and task coordination, reducing performance losses and improving efficiency. Unlike the US TRITON system, which required 64 nodes to achieve a 6x speedup, the Chinese model accomplished the same with only 7 nodes. This represents an 89% reduction in node usage, making the system more efficient and cost-effective.The supercomputing system was built using a domestic x86 computing platform. It used Hygon 7185 CPUs, which feature 32 cores and 64 threads, running at 2.5 GHz. The GPUs, developed in China, were equipped with 128GB of memory. The system also utilized a 200 Gb/s network bandwidth to enable efficient communication between nodes.
Flood simulation success
The research team tested their supercomputer by simulating a flood at the Zhuangli Reservoir in Shandong, China. Using 200 nodes and 800 GPUs, the system completed the simulation in just three minutes. This marked a 160x speed improvement over traditional methods. Such advancements make real-time flood simulations possible, improving disaster response, flood control, and reservoir management.What are the future applications?
The study, published in the Chinese Journal of Hydraulic Engineering on January 3, 2025, has made its research code open-source. This allows other researchers to build upon the findings. Potential applications include hydrometeorology, sedimentation modeling, and surface-water interactions. Future work will focus on expanding the model’s applications and testing its stability in real-world engineering projects.US-China rivalry in supercomputing
China’s success in supercomputing highlights the unintended effects of US sanctions. By restricting access to Nvidia technology, the US aimed to slow down China’s progress. However, the sanctions have instead pushed China to develop its own powerful computing solutions.With this breakthrough, China is not only catching up but also surpassing US supercomputers in certain areas. This marks a significant shift in the global tech race, as China continues to strengthen its position in high-performance computing.