This study contrasts Reynolds-Averaged Navier-Stokes (RANS) and (Numerical) Large-Eddy Simulation (NLES) for their use in predicting conjugate heat transfer for a low Reynolds number flow over a surface roughness element. The (N)LES predictions are in good agreement with experimental data, the heat transfer estimate is within 7.7% error. The linear RANS model shows larger errors up to 40%, especially in modelling the turbulent stresses. Using a one-equation LES turbulence model slightly increases the heat transfer prediction compared to a numerical LES. This indicates that more advanced turbulence models might not give more accurate heat transfer predictions.Additionally, this study investigates the time dependant development of the flow and temperature fields and how long data needs to be collected for statistically stationary results. The flow needs to develop for approximately 1800 through-flow times, T L , and 150 T L is required for collection of statistics. A significant range in turbulence length scale prediction between (N)LES and RANS was found. (N)LES scales appeared physically reasonable and should inform mesh resolution in future studies. Two additional cases were run with different cube heights showing changes in the time dependant development of the temperature field and greater sensitivity to turbulence modelling.