Urban microclimate analysis is often associated with computational fluid dynamics simulations of urban aerodynamics and thermal conditions, air quality, heat island, humidity, and rain at a building level. Conventional microclimate analyses are often limited by the size of the simulation domain, turbulence models, and sparse digital urban data available. This study introduces CityFFD with its supporting web portal of digital cities for large-scale urban simulation problems. A few novel algorithms, including backward and forward sweep interpolation schemes, time-adaptive technique, and the 2nd-order temporal scheme, were implemented to enhance the computing speed and accuracy of the model. Parallel computing on graphics processing units (GPU) speeds up the simulation. CityFFD is equipped with a large eddy simulation method to capture the turbulence behavior in the urban roughness sublayer. This paper reports the structure of CityFFD, and a group of benchmark cases, including airflow around buildings, natural ventilation, and thermal stratifications around building blocks in an actual urban area, and CityFFD is shown to predict acceptable results of the urban scale wind speeds, temperatures, and flow patterns. This study demonstrates a new approach to modeling urban microclimate by CFD simulations on GPUs.