Numerical simulations of stationary thermal convection in a differentially heated enclosure corresponding to the AFRODITE solidification benchmark experiment [1-5] are presented. The cavity of relative dimensions 10:6:1 (length:height:width) is characterized by a small transverse width. The Prandtl number Pr is varied within the range [0.0045, 0.03], typical of liquid metals, whereas the Grashof number, defined as Gr = g beta(Delta theta/L)H-4/v(2), is varied within the range [1.3 x 10(6), 1.6 x 10(7)]. As shown by the reference 3D simulations, the temperature field in these situations is 2D (independent of the transverse direction); 2D simulations are, however, not able to catch the physics of the flow and the resulting temperature results are also erroneous. To improve these 2D simulation results while keeping reasonable computational times, a 2D1/2 model is developed, which will take into account the no-slip condition at the side walls. This model is obtained by averaging the governing equations over the width of the domain, with a transverse profile for the velocity featuring a uniform central part and two boundary layers of size delta (delta is fixed for the whole domain). The relative deviation of the temperature field between the 2D1/2 and 3D computations is investigated as a function of the Prandtl number, the Grashof number and the chosen boundary layer thickness. It is shown that an optimum value exists for delta, which gives a mean deviation in the middle plane of less than 2%, whereas the choice of a more usual parabolic profile would lead to a twice larger deviation. Good comparisons are also obtained with the original experimental results reported at the end of the paper. The 2D1/2 model is thus able to give results which compare well with fully 3D results. It can then be used for extensive parametric studies at a reasonable cost.