Estimating energy consumption of neural networks with joint Structure-Device encoding

Chaopeng Guo, Shiyu Wang, Ruolan Xie, Jie Song

Abstract: Highlights•We propose SDEnergy, a new method that combines Structure–Device encoding for rapid and accurate energy consumption prediction of neural networks.•SDEnergy leverages Graph Neural Networks for structural feature extraction and fully connected networks for capturing device-specific features.•SDEnergy employs a two-stage training approach for the training of Structure feature Encoder, as well as the integration training with Device features, thereby achieving high-precision energy consumption prediction.