<p>Demand-side management (DSM) is a promising mechanism for reducing peak demand and improving the operational efficiency of smart grids without requiring additional generation or network expansion. However, the trade-off between fully distributed and semi-distributed DSM architectures remains insufficiently clarified, particularly when comparing appliance-level and household-level scheduling. This paper presents a comparative game-theoretic framework for residential DSM under dynamic electricity pricing, where the scheduling problem is formulated as a non-cooperative potential game and solved through best-response dynamics. Two architectures are studied: a fully distributed single-appliance model and a semi-distributed multi-appliance model. The proposed framework is evaluated using real residential consumption data under four start-time flexibility scenarios. The results show that both methods reduce peak demand compared with the baseline case; however, the semi-distributed model achieves better overall performance, with an average peak-demand reduction of 4.1%, compared with 3.72% for the distributed model. In addition, the semi-distributed strategy reduces the average energy cost by 1.16% relative to the distributed benchmark. The results also indicate that medium flexibility provides a favorable balance between load smoothing and scheduling stability, whereas excessive flexibility may introduce unnecessary demand oscillations in some periods. These findings demonstrate that the proposed semi-distributed game-theoretic DSM framework offers a practical and computationally effective solution for residential load scheduling in smart grids.</p>