Molecular dynamics (MD) simulations of metal solidification require both high interatomic accuracy and scalability to millions of atoms to capture physically relevant grain and microstructure evolution. Classical potentials for metal simulations, such as EAM and MEAM, are limited in accuracy and often fitted to a few similar elements due to their empirical nature. Recent machine-learning interatomic potentials (MLIPs) approach the accuracy of first-principles calculations. However, their applicability to large-scale, multi-component systems remains an open challenge due to computational cost and scaling limitations. In this work, we train equivariant graph neural network potentials (GNNPs) for pure aluminum and systematically compare them to existing classical and machine-learning potentials. We assess the model performance by predicting key material properties and conducting a million-atom solidification simulation. We find that MEAM potentials perform comparably to MLIPs on many equilibrium solid-state properties, whereas EAM potentials exhibit significant deviations from the reference data. At high-energy configurations and in the liquid state, the MLIPs yield better agreement with reference data, highlighting the limitations of classical potentials outside their fitted regimes. Due to their equivariant and graph-based formulation, the GNNPs can be readily retrained for multiple species while maintaining favorable scalability with respect to the system size, unlike previous approaches. This provides a pathway for feasible simulations with near-first-principles-level accuracy at experimentally relevant system sizes, benefiting multiscale materials design.