Power Prediction Modelling for CPU–GPU Heterogeneous Platforms Based on Heterogeneous Feature Sequence Modelling
TL;DR
A Transformer, a convolutional neural network, and a long short-term memory (LSTM) architecture (Transformer-CNN-LSTM) for system-level power prediction, indicating its potential for runtime power estimation and energy-aware platform management.
Abstract
Central processing unit-graphics processing unit (CPU-GPU) heterogeneous platforms are widely deployed for AI-related, data-processing, and compute-intensive workloads. Their practical deployment is constrained by fluctuating node power, cross-device coordination overhead, and uneven energy efficiency. This paper develops a Transformer, a convolutional neural network, and a long short-term memory (LSTM) architecture (Transformer-CNN-LSTM) for system-level power prediction. Runtime features are cleaned, normalised, and reorganised into sliding-window sequences, allowing the model to learn global feature interactions, local workload fluctuations, and temporal evolution within a single prediction pipeline. Experiments on the collected CPU-GPU dataset demonstrate that this method achieves better results than CNN and LSTM, CNN-LSTM, performance monitoring counter statistical (PMC-Stat), and GPUWattch-Inspired baselines. The model achieves mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and R2 results of 50.31, 7.09, 4.79, and 0.98, respectively, indicating its potential for runtime power estimation and energy-aware platform management.