Abstract: Flow direction algorithms are fundamental for slope hydrology and hydrodynamic simulations. Implementing a parallel D8 flow accumulation algorithm under the CUDA architecture can effectively accelerate simulation speeds, making the parallelization strategy a key research focus for addressing computational access and conflict handling. This paper optimizes the parallel strategy of the D8 algorithm using the atomicAdd function within the CUDA architecture. We selected watersheds at different spatial scales within the Ganjiang River basin (upper reaches, upper-middle reaches, and full basin) as study areas to assess the accuracy of extracted drainage networks, parallel acceleration performance, and scale effects at each spatial scale. The findings indicate:The drainage network extraction results using the parallel D8 strategy closely align with those from the classical algorithm, with relative errors in stream length, basin area, and drainage density all below 0.3%.Computation times under the CUDA architecture for the parallel D8 algorithm were less than those of the serial ArcGIS implementation, which were in turn less than the serial Matlab algorithm. The speedup ratio was proportional to the number of thread blocks and grids.When the number of threads was ≤128, the optimal speedup occurred with grid numbers below 1024; when thread counts exceeded 128, optimal speedup was achieved with grid numbers above 65,536.The speedup ratio exhibited a decreasing effect as spatial scale increased. Speedup ratios relative to ArcGIS decreased by over 20% for the upper-middle reaches and the full basin compared to the upper reaches.The parallelization strategy for the D8 algorithm may serve as a theoretical reference for parallel computing in hydrological and hydrodynamic models.