Our paper, “WattLayer: Get Layers Right to Estimate Inference Energy of Neural Networks,” was accepted to SuRE'26, the 1st Workshop on Sustainability and Resource-Efficiency of Artificial Intelligence, held alongside IJCAI-ECAI 2026.
WattLayer is a method for estimating the inference energy of a neural network directly from its architecture, without running it. Instead of treating a model as a black box, it models the energy contribution of each layer type individually (convolutions, linear layers, attention, activations), then aggregates the per-layer estimates into a prediction for the whole architecture.
The method was evaluated on 295 neural architectures and over 100,000 energy measurements, across vision, text and audio tasks, and generalizes to large language models it was never trained on. An open-source Python implementation of the method is available.