AsianFin -- Chinese researchers published a peer-reviewed study in January, revealing that domestic Chinese GPUs outperformed Nvidia chips by nearly 10 times in supercomputing simulations.

The study, based on the development of a "multi-node, multi-GPU" large-scale distributed heterogeneous parallel computing hydrodynamic model, was used for efficient parallel solving of two-dimensional shallow water equations at a watershed scale.

The research, titled "Heterogeneous Parallel Computing of Hydrodynamic Models Based on Domestic CPU/GPU Platforms," has been publicly released on the academic platform CNKI.

The study highlights that the model is built on China's standard x86 computing platform, utilizing the domestically-produced Haiguang 7185 processor with 32 cores and 64 threads, operating at 2.5 GHz.

The GPU used is also domestically produced, with 128GB of memory and 200 Gb/s network bandwidth. This new model achieved a sixfold acceleration with just seven nodes, requiring 89% fewer nodes compared to the TRITON flood forecasting model used by Oak Ridge National Laboratory's Summit supercomputer.