China’s supercomputer chips get 10 times more powerful than Nvidia, claims study

China’s supercomputer chips get 10 times more powerful than Nvidia, claims study

Could this be an unintended consequences of Washington’s escalating tech sanctions?

Chinese computer researchers have achieved a near-tenfold boost in performance over Nvidia-powered US supercomputers using domestically made graphics processors, according to a peer-reviewed study.

This accomplishment challenges the longstanding dominance of American-made chips in advanced scientific research while also potentially highlighting the unintended consequences of Washington’s escalating tech sanctions. 

The researchers credit innovative software optimization techniques for enhancing computer efficiency powered by Chinese-designed graphics processing units (GPUs). These optimizations enabled their system to outperform traditional US supercomputers in specific scientific computations. 

Escaping the ‘Chokepoint’ 

According to South China Morning Post (SCMP), some experts caution that software tweaks alone cannot close the hardware gap indefinitely. The development highlights Beijing’s broader strategy to mitigate “chokepoint” risks—its heavy reliance on Western chip technologies amid persistent US sanctions.

The stakes are particularly high in fields that depend on extensive computational resources. Scientists frequently rely on large-scale, high-resolution simulations for real-world applications such as flood defense planning and urban waterlogging analysis. 

These simulations require significant processing power and time, often limiting their broader application. For Chinese researchers, the challenge is compounded by the fact that the production of advanced GPUs like Nvidia’s A100 and H100 is dominated by foreign manufacturers and the export restrictions imposed by the US. 

Furthermore, the restriction of Nvidia’s CUDA software ecosystem from running on third-party hardware has hindered the development of independent algorithms.

The software-enabled solution 

Seeking a breakthrough, Professor Nan Tongchao from the State Key Laboratory of Hydrology-Water Resources and Hydraulic Engineering at Hohai University in Nanjing spearheaded research into a “multi-node, multi-GPU” parallel computing approach. His team focused on leveraging domestic CPUs and GPUs to build a more efficient supercomputing model. 

Their findings demonstrate that efficient data transfer and task coordination between multiple nodes are key to minimizing performance losses in parallel computing.

In 2021, Oak Ridge National Laboratory researchers introduced a “multi-node, multi-GPU” flood forecasting model known as TRITON using the Summit supercomputer. Despite deploying 64 nodes, TRITON only achieved a processing speed increase of about six times.

In contrast, Nan’s innovative architecture combined multiple GPUs into a single node to counterbalance the performance limitations of domestic hardware. By refining data exchanges between nodes at the software level, his model drastically reduced communication overhead.

Implemented on a domestic general-purpose x86 computing platform, with Hygon processors (model 7185, featuring 32 cores, 64 threads, and a 2.5 GHz clock speed) and domestic GPUs supported by 128GB of memory and a network bandwidth of 200 Gb/s, the new model achieved a speedup of six using just seven nodes, an 89 percent reduction in node usage compared to TRITON.

Nan’s team simulated the flood evolution process to validate the model at the Zhuangli Reservoir in Zaozhuang, Shandong province. Using 200 computational nodes and 800 GPUs, the simulation was completed in just three minutes, reaching a speedup of over 160 times. 

Nan said,

Simulating floods at a river basin scale in just minutes means real-time simulations of flood evolution and various rainfall-run-off scenarios can now be conducted more quickly and in greater detail.

“This can enhance flood control and disaster prevention efforts, improve real-time reservoir management, and ultimately reduce loss of life and property,”

The research code is available on an open-source platform, and Nan noted that the findings could extend beyond flood modeling to simulations in hydrometeorology, sedimentation, and surface-water-groundwater interactions.

He added:

Future work will expand its applications and further test its stability in engineering practices.

READ the latest news shaping the AI Chips market at AI Chips News

China’s supercomputer chips get 10 times more powerful than Nvidia, claims study, source