High-precision precipitation simulation over complex terrain in arid regions remains uncertain. This study compares two Weather Research and Forecasting (WRF) model experiments, a convection-parameterized simulation with 12 km resolution (CRM) and a convection-permitting simulation with 4 km resolution (CPM), evaluated for summer precipitation over Northwest China (NWC) during 2009-2011. Observational and model data were analyzed on a unified 4 km grid and in local time to conduct statistical harmonic analysis and thermodynamic diagnostics. The results show that the CPM substantially improves the simulation of precipitation amount, intensity, and frequency relative to the CRM, especially in mountainous regions. For total precipitation, the CPM yields a regional mean of 81.53 mm, closest to observations and outperforming ERA5 and CRM. Although the CPM slightly overestimates precipitation intensity, it exhibits higher spatial consistency and better captures the frequency of moderate to heavy rainfall. For wet-day frequency, spatial correlation reaches 0.79 with a Root Mean Square Error of 0.11%, and the regional mean is also closer to observations. The CPM reproduces the observed dual-peak diurnal cycle with reduced phase errors and amplitude bias. The CRM, constrained by cumulus parameterization, responds weakly to orographic forcing, limiting convective initiation and reducing precipitation intensity. In contrast, the CPM captures moisture convergence and vertical motion along windward slopes, producing stronger updrafts and more favorable thermodynamic conditions for convection. Overall, the CPM achieves simulations most consistent with observations, providing a more realistic representation of precipitation processes over the arid complex-terrain region of NWC and demonstrating its potential to advance high-resolution climate modeling.