While the prediction of surface smoothness has achieved a certain level of standardization and normalization, the rapid, high-frequency, and low-cost intelligent prediction of surface smoothness-specifically the International Roughness Index (IRI)-has gained widespread attention in the context of smart city and intelligent transportation infrastructure development. However, the prediction accuracy and performance of IRI based on different intelligent devices have yet to be thoroughly investigated. In this study, road driving experiments were first conducted using two types of intelligent prediction devices developed by the authors’research team: a smartphone App for road driving data collection and an intelligent terminal device for the same purpose. During these experiments, data such as driving vibration, speed, and location were collected. Next, a random forest model was employed to identify four vibration acceleration features that most significantly influence the prediction result of IRI. Finally, three neural networks, namely Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), and Long Short Term Memory Network (LSTM), were used to establish the prediction model for road roughness IRI. The prediction accuracy of different models and devices was then compared and analyzed. The results show that, LSTM model achieved the best robustness and highest prediction accuracy among three neural network models. The coefficients of determination of IRI prediction model for two devices were 0.864 and 0.789, respectively, with the intelligent terminal outperforming the smartphone in prediction accuracy. These research findings hold significant theoretical and practical value for enhancing the informatization level of surface smoothness prediction and monitoring in China, as well as for improving the scientific basis of road maintenance decision-making.