Multi-physics ensemble simulations have emerged as a promising approach to ensemble hydrological simulations due to the advantages in process understanding and model development. As a multi-physics ensemble is constructed by perturbing the physics of multi-physics models, the ensemble members share a substantial portion of the same physics and hence are not independent of each other. It is unknown whether and to what extent the independence of the ensemble members affects the ensemble skill gain, especially compared with the multi-model ensemble approach. This study compares a multi-physics ensemble constructed from the Noah land surface model with multi-parameterization options (Noah-MP) with the North American Land Data Assimilation System (NLDAS) multi-model ensemble. The two ensembles are evaluated at 12 River Forecast Centers over the conterminous United States. The ensemble skill gain is measured by the difference between the performance of the ensemble mean and the average of the ensemble members’ performance, and the inter-member independence is measured by error correlations. The results show that the Noah-MP members outperform, on average, the NLDAS models, especially in the snow-dominated areas. In addition, the best-performing models among the two ensembles are mostly Noah-MP members. However, these two performance superiorities do not lead to the superiority of the ensemble mean. The Noah-MP multi-physics ensemble has a low ensemble skill gain, resulting from a high error correlation among the ensemble members. This study suggests that the methods of ensemble construction and optimization should be improved to also consider inter-member independence, especially for a multi-physics ensemble.