To evaluate the capability of dynamical climate downscaling for future climate projection, the next-generation mesoscale weather prediction system Weather Research and Forecasting model (WRF) has been used to simulate the regional climate over the North America during 2000-2004 with fine resolution (36 km). The initial and lateral boundary conditions of WRF are provided by the fully coupled atmosphere-ocean-land-ice NCAR Community Climate Model version 3 (CCSM3) at T85 (~155 km) resolution. Two integration approaches are studied: 1) A 5-year continuous long-term integration. 2) Consecutive short-term weekly reinitialized integrations. The simulations given by these two approaches are compared with the NCEP North American Regional Reanalysis (NARR) data, to evaluate the model performance and uncertainty. Result shows that the WRF regional climate downscaling can significantly improve the CCSM3 climate projection. The precipitation and surface temperature are reasonably simulated, especially over the mountainous region, in terms of capturing the regional variations and seasonal evolutions. Comparing the performance between the two integration methods, the continuous simulation has 2 to 3 degree cold biases of surface temperatures over the US continent. The reinitialized integrations significantly improve those systematic biases; however, it results in under prediction of precipitation in the summer.