The widespread deployment of Deep Neural Network (DNN) models on heterogeneous devices requires accurate performance evaluation and prediction. Key metrics like inference latency and energy consumption are critical for optimizing DNN deployments, but these metrics are often difficult to obtain. This paper introduces a unified system, which provides comprehensive and accurate on-device performance profile and a novel end-to-end Graph Neural Network (GNN) based predictor for inference latency. Unlike prior kernel level prediction methods that incur significant memory and latency overhead, our approach directly predicts the end-to-end DNN inference latency by leveraging its computational graph’s features. This results in a compact and efficient predictor. Experiment results show that our GNN predictor reduces memory footprint by 17 times and prediction overhead by at least two orders of magnitude compared to kernel level method. It also achieves over 96% ACC10 on various DNN models and devices. We believe this system can provide valuable insights for serving more efficient deployment and evaluation in edge computing scenarios.