NeuralPower

NeuralPower provide a systematic modeling and prediction method for power, runtime, and energy consumptions in running deep neural networks (DNNs) on target devices, including GPUs. It helps the DNN designers to get a detailed breakdown of the runtime, energy consumptions for a given DNN on a target device. The main contributions are:

NeuralPower is the first to apply learning-based models to predict runtime, power, and energy consumption for CNNs;
NeuralPower outperforms state-of-the-art analytical models, by achieving an improvement in accuracy up to 68.5% compared to the best previously published work;
NeuralPower also provides the detailed breakdown of runtime and power across different components (at each layer) of the whole network.

Work Flow

The main processes of NeuralPower can be divided by the following steps:

Collect real data for model training and testing
Build models
Test against real-time data.

Data Collection

The data are collected real time through various DNN applications running on target devices. We adopted and modified the code framework developed by Qi et al. to support both network-level and layer-level real-time data collection. The original code can be found at Paleo. We acknoledge the helps from Qi et al. throughout the Paleo modification process. The modified version is in the directory neuralpower_paleo.

Model training

In this work, we mainly use polynomial models to learn the layer level power and runtime values. The sample code can be found inthe directory model_training, while the file parser_raw_data.py can be used to parse the produced data from the data collection process.

Model testing

With the trained model, we can use it directly in the runtime, power predictions. The file predict_runtime_power.py provide such functionality for this process.

Reference Paper

E Cai, DC Juan, D Stamoulis, D Marculescu. NeuralPower: Predict and Deploy Energy-Efficient Convolutional Neural Networks. In Asian Conference on Machine Learning 2017 Nov 11 (pp. 622-637).

@article{cai2017neuralpower,
  title={Neuralpower: Predict and deploy energy-efficient convolutional neural networks},
  author={Cai, Ermao and Juan, Da-Cheng and Stamoulis, Dimitrios and Marculescu, Diana},
  journal={arXiv preprint arXiv:1710.05420},
  year={2017}
}