Measured Energy Consumption for AI-Based Precoding

This paper presents the first measurement-based study on the energy consumption of artificial intelligence (AI) algorithms for radio access networks (RAN), aiming to investigate whether the integration of AI leads to higher energy consumption than traditional non-AI algorithms. We use the multiuser precoding problem as a representative example, where AI is widely recognized for its potential in supporting the growing deployment of antennas for 6G systems. Our assessment examines three AI algorithms using the convolutional neural network (CNN), edge graph neural network (EGNN), and communication model-based graph neural network (MGNN), in comparison with two non-AI algorithms, the weighted minimum mean square error (WMMSE) and regularized zero-forcing (RZF) algorithms. Through extensive measurements of energy and power consumption, we examine how efficient architecture designs of deep neural networks (DNNs) affect energy consumptions, and provide a quantitative comparison between measured results and estimates from existing power estimation models. The results indicate that the MGNN, which is the most efficient architecture among the three DNNs, consumes substantially less energy than the non-AI algorithms for large-scale systems, while the CNN consumes significantly more energy for both training and inference compared to the MGNN and EGNN. Moreover, GPUs do not always consume more power than CPUs; the power consumption of CPUs should be taken into account when GPUs are used, and DRAM contributes little to total power consumption. We demonstrate the shortcomings of existing utilization-based and FLOPs-based models in estimating energy consumption, highlighting the need for more accurate models to support green AI designs for RAN.