Traffic speed prediction is the task of forecasting the average moving speed on certain roads and highways. Since transportation is an essential part of our daily life, traffic prediction helps reduce wasting time caused by traffic congestion and the negative environmental impacts of idling vehicles during traffic jams. Recently, graph neural network has achieved a certain level of success in forecasting traffic, thanks to the ability to capture both spatio-temporal features of traffic data. Although many publications propose and benchmark different traffic prediction methods, those models’ performance and other profiles are hardly mentioned and evaluated, such as the running time and memories taken. This research aims to evaluate the performance of five mainstream GNN models in traffic prediction: DCRNN, Graph Wavenet, MTGNN, STGCN, and T-GCN. Training time, inference time, and other profiles of mentioned models are also further investigated and reported. During the experiments, we record various behaviors/factors from models, such as the memory allocation. Different models are believed to be better in specific scenarios in traffic forecasting based on their profiles. The finding shows that some prediction methods can only perform well if the training time is significant, such as T-GCN. The performances on single/multiple step forecasting vary with different models, where MTGNN outperforms other models. Additionally, the memory usage and inferencing time of GNN models are different, potentially impacting the selection of models for some limited memory machines and other scenarios.