Grain growth is one of the most fundamental phenomena affecting the microstructure of polycrystalline materials. In experimental studies, three-dimensional (3D) grain growth is usually investigated by examining two-dimensional (2D) cross sections. However, the extent to which the 3D microstructural characteristics can be obtained from cross-sectional observations remains unclear. Additionally, there is some disagreement as to whether a cross-sectional view of 3D grain growth can be fully approximated by 2D growth. In this study, by employing the multi-phase-field method and parallel graphics processing unit computing on a supercomputer, we perform large-scale simulations of 3D and 2D ideal grain growth with approximately three million initial grains. This computational scale supports the detailed comparison of 3D, cross-sectional, and 2D grain structures with good statistical reliability. Our simulations reveal that grain growth behavior in a cross section is very different from those in 3D and fully 2D spaces, in terms of the average and distribution of the grain sizes, as well as the growth kinetics of individual grains. On the other hand, we find that the average grain size in 3D can be estimated as being around 1.2 times that observed in a cross section, which is in good agreement with classical theory in the stereology. Furthermore, based on the Saltykov-Schwartz method, we propose a predictive model that can estimate the 3D grain size distribution from the cross-sectional size distribution.