A new natural convection ＇benchmark problem＇ for validating CFD codes is defined. In the subject problem, a cubical air-filled cavity, tilted at 0, 45 degrees, or 90 degrees, has one pair of opposing faces at different temperatures, T-h and T-c, respectively, the remaining faces having a linear variation from T-c to T-h. In contrast to some other benchmark problems, this problem is physically-realizable. Experimental techniques to establish the thermal boundary conditions and to measure the Nusselt number to 1% accuracy are reported. Measured Nusselt numbers at Rayleigh number equal to 4 x 10(4) are shown to agree with CFD predictions to within +/- 0.3%.