与传统概念性水文模型相比，二维水动力模型可提供更丰富的流域地表水力要素信息，但是计算耗时太长的问题限制其推广应用，提升二维水动力模型的计算效率成为当前数字孪生流域建设工作中的关键技术难题之一。采用基于Godunov格式的有限体积法离散完整二维浅水方程组建立模型，通过消息传递接口(message passing interface, MPI)与统一计算设备架构(compute unified device architecture, CUDA)相结合的技术实现了基于多图形处理器(graphics processing unit, GPU)的高性能加速计算，采用理想算例和真实流域算例验证模型具有较好的数值计算精度，其中，理想算例中洪峰的相对误差为0.011%，真实流域算例中洪峰的相对误差为2.98%。选取宝盖寺流域为研究对象，分析不同单元分辨率下模型的加速效果，结果表明：在5、2、1 m分辨率下，使用8张GPU卡计算获得的加速比分别为1.58、3.92、5.77，单元分辨率越高，即单元数越多，多GPU卡的加速效果越明显。基于多GPU的水动力模型加速潜力巨大，可为数字孪生流域建设提供有力技术支撑。

Compared to the traditional conceptual hydrological models, two-dimensional hydrodynamic models can provide more comprehensive hydraulic information of watershed surfaces, but the issue of long computational time restricts its widespread application. Exploring ways to improve the computational efficiency of two-dimensional hydraulic models has become one of the hot points and key technological challenges in current digital twin watershed development. The rapid advancement of GPU hardware technology has enabled the utilization of two-dimensional hydraulic models for the purpose of simulating watershed flood processes in real time. The potential applications of this technology in the construction of digital twin watersheds appear bright.The model is established based on a structured grid and adopts the Godunov scheme based on the finite volume method to discretize the complete two-dimensional shallow water equations. A high-performance accelerated calculation based on multiple GPUs is realized by combining MPI and CUDA computing architecture to meet the requirements of large-scale parallel computing tasks and realize the simultaneous work of multiple GPUs. MPI implements message passing between parallel processes based on the distributed storage model. Each process has a unique process rank at runtime and controls a GPU device. When using multiple GPUs for computation, the computational domain needs to be divided into multiple subdomains, and each GPU is assigned to compute a specific subdomain. Each subdomain is surrounded by an additional layer of grid cells that is used to communicate with adjacent subdomains. This outer layer of grid cells receives data from the adjacent subdomains to perform updates. Once the communication is completed, the computation continues within each subdomain.The model's numerical accuracy has been verified using ideal and real watershed cases, with a relative error of 0.011% for the peak discharge in the ideal case and 2.98% for the peak discharge in the real watershed case. The acceleration effect of the model under different cell resolutions was analyzed in the Baogaisi watershed. The results showed that when the total number of grids reaches a certain scale, the multi-GPU acceleration technology can obtain a satisfactory acceleration effect. When the grid resolutions of the watershed are 5 m, 2 m, and 1 m, the corresponding grid units are 861,605, 5384,807, and 21,539,061. The speedup ratios obtained by 8 Tesla V100 GPUs are 1.58, 3.92, and 5.77, respectively. Higher cell resolutions lead to more significant acceleration effects with multiple GPUs. The hydrodynamic model based on multi-GPU has great potential for acceleration and can provide strong technical support for the construction of digital twin river basins.