Modeling the dynamics of a free floating surface vehicle is known to be challenging due to the complicated vehicle-fluid interaction and inherent nonlinearity in the model. Datadriven machine-learning technologies can be applied to model the vehicle dynamics and predict its motion over a particular time horizon given specific control inputs. However, the learned model typically is not directly interpretable and is susceptible to out-of-distribution data, which could result in significant modeling errors. To overcome this limitation of learning-based models, we propose using a Gaussian process (GP) to model the dynamics of a surface vehicle, enabling the prediction of the motion with uncertainty. However, a naive implementation of GP algorithms is computationally very intensive. Therefore, efficient state-of-the-art techniques are employed to ease the computational complexity of the traditional GP. The performance of several algorithm variants was compared using actual experimental data, which demonstrated the capability of the proposed GP-based hydrodynamics modeling.