Abstract: Traditional hydrodynamic models are often constrained by low computational efficiency, making it difficult to meet the demands of large-scale, high-accuracy flood forecasting and warning. To enhance the efficiency of urban flood simulation, this study develops a high-performance hydrodynamic model based on multi-GPU parallel acceleration. The model integrates three core computational modules: a 2D surface runoff module, a surface-sewer flow interaction module, and an underground sewer flow module. To optimize the most computationally intensive 2D surface runoff module, Metis graph partitioning was employed for spatial domain decomposition and load balancing, while MPI-OpenACC technology was implemented to achieve multi-GPU parallel acceleration. The model was applied to the Zhongshundawei (Zhongshan-Shunde Embankment) area in Guangdong Province to simulate the flooding process during the "2024.05" extreme rainstorm event using nowcasting rainfall data. The results demonstrate that the developed model achieves high computational accuracy, with a coefficient of determination (R²) of 0.91 between the measured and simulated water depths at major waterlogging points. The multi-GPU parallel acceleration technology substantially improved the model's efficiency, achieving a 26.38-fold speedup with 8 GPUs compared to a 64-core CPU setup. Notably, the model simulated a 6-hour flood process for an 811 km² urban area (50.68 million computational meshes) at a 4-meter spatial resolution within 10 minutes. This study demonstrates the significant potential of the multi-GPU parallel hydrodynamic model for large-scale, high-resolution flood forecasting and warning, and the findings provide robust technical support for urban flood management.