Urban atmospheric flows over rough surfaces exhibit strong turbulent interactions that remain challenging to reproduce numerically with sufficient fidelity. In this work, we evaluate the capability of the SOD2D high-order GPU-accelerated finite element solver to replicate controlled experiments conducted in the Boundary Layer Wind Tunnel (BLWT) at the University of Bristol. The experimental configuration consists of a developing turbulent boundary layer generated by vortex-generator plates followed by an array of cubic roughness elements representative of an urban canopy. Experiments were performed at inflow velocities of 5--10 m/s, resulting in Reynolds numbers on the order of $10^5$, with measurements of mean velocity, turbulence intensities, and turbulent kinetic energy at multiple downstream stations. The numerical simulations implementing SOD2D in its incompressible LES mode using a high-order spectral finite-element discretization, low-dissipation operator splitting, and entropy-based stabilization optimized for GPU architectures. The computational domain reproduces the BLWT geometry and roughness layout, while the inflow velocity profile is prescribed based on the wind-tunnel measurements. A sponge layer is applied near the outlet to ensure non-reflecting boundary treatment. Time-averaged quantities are compared against the experimental data at corresponding streamwise locations. The simulations successfully capture the rapid boundary-layer growth induced by the vortex generator, the wake dynamics behind the cube array, and the evolution of mean velocity and turbulent kinetic energy along the tunnel. Good agreement with the BLWT measurements demonstrates that SOD2D can reproduce rough-wall turbulent flows with high accuracy and computational efficiency. These results establish SOD2D as a reliable numerical counterpart to boundary-layer wind-tunnel experiments for urban-flow research.