Abstract Simulation of urban airflow and ventilation potential is desirable for building design, however the complex and transient nature of flows in urban environments makes this a challenging task. This study aims to evaluate the capability of a lattice-Boltzmann method (LBM) code deployed on a graphical processing unit (GPU) using a large eddy sub-grid turbulence model for cross-flow ventilation of an idealised cubical building at wind-tunnel scale. ANSYS Fluent is used as a numerical comparison. Facade pressure and ventilation of the cube are investigated for parallel and perpendicular wind directions with the building in isolation and regular array format. Pressures, velocities and ventilation rates are compared to experimental data from wind tunnel and full-scale experiments of the Silsoe cube. Simulations compare favourably with experimental values and between each other. When the cube was surrounded by other cubes, simulations suggest that vortex shedding from up-wind buildings provides pulsating ventilation, improving airflow ingress in the parallel wind cases. A parametric study showed that doubling surrounding building height had a small negative effect on ventilation but was mitigated by high levels of downdraft and flow fluctuations in the vertical plane. Comparatively, doubling the central building height had a net positive effect but caused high internal airspeeds for both angles. The LBM code running on one GPU was several orders of magnitude faster than Fluent with similar accuracy. Simulation time using the LBM approach was several orders of magnitude lower than Fluent.