Numerical Differences and Computational Performance Across Floating-Point Precision Configurations in a GPU-Accelerated 1D–2D Urban Flood Model

Coupled sewer–surface flood simulations are computationally demanding. We evaluated twelve surface-flow and exchange-discharge precision combinations in a graphics processing unit (GPU)-accelerated urban flood model for Jinju, Korea, using rain-on-grid (RoG) and rain-on-subcatchment (RoS) inputs. In 6 h hydraulic tests with 64-bit double-precision (FP64) exchange calculations, reduced-precision surface settings gave final water-depth root mean square errors (RMSEs) of 0.172–1.187 mm for RoG and 0.746–2.867 mm for RoS relative to the full-FP64 reference. Using 32-bit single precision (FP32) in selected exchange operations gave a maximum within-setting RMSE of 11.347 mm relative to FP64 exchange in the 48 h test. In a separate performance experiment, ten repeated 1 h runs per condition kept coupling data in GPU memory. Using FP32 for state storage and all targeted surface-flow operations, with FP64 exchange, achieved median speed-ups relative to the full-FP64 configuration of 1.883 (95% interval from repeated sampling: 1.876–1.887) for RoG and 1.205 (1.199–1.212) for RoS. The results identified this combination as an efficient candidate for the tested catchment. Local water-depth and exchange-volume differences should be assessed when applying it.