THE structure of convection in the mantle is still the subject of considerable debate. The now standard modelling of the convective flow as driven in a viscous mantle by density anomalies derived from seismic tomography has successfully explained the longest-wavelength (degree 2 to 8) geoid anomalies and provided important information concerning the viscosity structure of the mantle(1-8). With this approach, however, the predicted response of surface topography to convective stresses (the ’dynamic topography’) has a typical magnitude of several kilometres, which does not conform with observations(9-11). A possible source of this discrepancy lies in the severe underestimation, by tomography, of density anomalies due to deflections of the boundary between the upper and lower mantle, at 660 km depth. Here we model the mantle flow implied by seismically derived density heterogeneities, using an empirical method to account for the 660-km boundary topography. The predicted dynamic (surface) topography thus obtained is significantly reduced, to values that conform with the observations; in addition, the 660-km boundary topography appears to have a strong influence on the computed mantle circulation, inducing local layering of the convective flow.