Blockage effect refers to the influence of wind farms on the flow upstream, affecting the performance of the first row of wind turbines and potentially the turbines downwind. Blockage effects affect the overall wind farm production, being very difficult to characterize and quantify. The global blockage effect is not only due to the wind turbine arrangements; it also depends on the interaction between the wind farm and the atmospheric boundary layer (ABL). The presence of gravity waves could potentially amplify blockage. In the present work, we study the different physical mechanisms affecting blockage in wind farms consisting of thousands of wind turbines under neutral, nocturnal and diurnal stability conditions. We quantify blockage by comparing the wind turbines power output in the wind farm's front row against the power output of an isolated wind turbine under the same ABL conditions. The simulations are carried out with Reynolds Average Navier Stokes (RANS) equations, using k-e and k- Wilcox 2006 models, with modifications for atmospheric flow problems. The simulations include Coriolis forces without using a maximum mixing length limitation model, which artificially affects the wake length. The simulations are also solved using an LES model with Deardoff closure. The wind turbines are modelled using uniformly loaded actuator discs. The URANS equations are solved using the highly parallelized finite element code Alya, developed at the Barcelona Supercomputing Center (BSC). Being solved steady (RANS) when simulating neutral conditions and as transient (URANS) when the ABL is thermally stratified. The LES equations are solved using the spectral finite element code SOD2D developed at BSC. Using a Galerkin formulation stabilized with the entropy viscosity model, the LES model has very low numerical dissipation. SOD2D code runs in graphical processors. We use RANS and LES models to simulate the same wind farms under the same thermal stratification conditions. By comparison of the simulation results, we assess the ability of the RANS models to reproduce different physical mechanisms that generate blockage. We want to clarify in which situations URANS is a reliable model. The CPU-time spent by the two codes to run RANS and LES wind farm simulations is discussed.