As digitalisation accelerates and artificial intelligence (AI) integrates into every facet of our lives, the information and communication technology (ICT) sector faces a critical challenge: a steady rise in energy consumption and associated carbon emissions. To build a sustainable 6G future, we must first deeply understand exactly where, when, and how power is consumed in current 5G infrastructures.
At the recent EUCNC & 6G Summit (June 2-5, 2026), EXIGENCE partner TNO presented innovative research addressing this exact challenge. Their work introduces a novel measurement and attribution framework designed to track power consumption down to the network function and individual user level.
By creating this granular visibility, operators can dramatically optimise service power consumption or even incentivise end-users to reduce their digital carbon footprint.
The Three-Pronged Approach
The TNO research team divided their methodology into three interconnected pillars spanning across the network architecture:
- The Radio Access Network (RAN): Measuring real-world power consumption of physical RAN nodes.
- The Core Network: Attributing power consumption to Virtualised Network Functions (VNFs) like the User Plane Function (UPF) running in complex cloud environments.
- AI-Driven Optimisation: Using reinforcement learning and modeling to dynamically minimize end-to-end service power consumption.
1. Inside the RAN Setup: Physical Node Measurement
To capture accurate physical power data, TNO deployed a hardware-in-the-loop testbed:
- The Hardware: One or two Quectel 5G Test Kits were used as User Equipment (UE) to trigger data sessions. They established standalone data streams ranging from 2 to 30 Mbps, as well as joint data calls.
- The Network: An Amarisoft Callbox gNB acted as the base station.
- The Measurement: A Shelly Plug S smart meter monitored the total energy consumption of the Amarisoft Callbox every second across multiple 10-minute experimental iterations.
RAN Key Findings & Formula
The baseline idle traffic condition (zero load) of the Amarisoft CallBox sat at 79 W. To distribute power equitably among active users, TNO used a direct user attribution formula:
Where P(ue) is the attributed power per user, T(ue) is user throughput, and T(total) is total throughput.
- Linear Scaling: The results showed a stark, clean, linear relationship between effective downlink throughput and increased power consumption.
- Zero Multi-Stream Overhead: Crucially, data revealed no power consumption overhead from serving two individual traffic streams simultaneously compared to a single aggregated stream.
2. Core Network Setup: The Complexity of Cloud Attribution
While measuring physical hardware is straightforward, cloudified environments hide infrastructure power behind layers of virtualisation. TNO’s primary goal here was developing an attribution method capable of parsing power consumption inside an OpenStack cloud environment.
Using a T-Rex traffic generator to create highly variable and high-throughput data streams , TNO combined measurements across three distinct domains: the compute node, the Virtual Machine (VM), and the Open5GS UPF container.
The Core Attribution Formula
To isolate the exact power drawn by a specific virtualised User Plane Function (UPF), TNO used Running Average Power Limit (RAPL) metrics from the physical machine alongside CPU cycles:
Where CPU-Sec tracks the isolated processing seconds, R(Active) represents active ratios, and P(VM) represents total virtual machine power.
Core Network Results
- Cross-Domain Necessity: To get an accurate power attribution for a virtualised network function, cross-domain information from the physical compute node is non-negotiable.
- Traffic Alignment: Power consumption of the UPF scales proportionally to both traffic load and CPU load.
- Bounded Overhead: Much like the RAN, the power consumption overhead of serving multiple traffic streams simultaneously remains strictly limited.
3. Minimising Service Power Consumption via AI
Having established reliable metrics for the RAN and Core, TNO moved from measurement to mitigation.
They deployed a Double Deep Q-Network (Double DQN) with NoisyNet(s) to intelligently distribute user network flows across a simulated 50-node edge compute infrastructure. The algorithm was explicitly tasked with minimising the combined idle and active load power consumption:
The Outcome
The AI-driven policy successfully learned to consolidate traffic and manage workloads efficiently. The Learned Policy improved overall power consumption by 2% to 5% over a standard random masked baseline policy. To push these savings even further, TNO is currently exploring advanced architectures like Pointer Networks and REINFORCE algorithms.
Key Takeaways for the 6G Horizon
The TNO presentation highlights four critical insights for the telecom industry as we design next-generation networks:
- Load Proportionality: Network power consumption scales predictably and proportionally to traffic and CPU load.
- Virtualisation Complexity: Measuring power in cloudified environments is highly complex and mandates transparent data sharing between cloud providers and cloud tenants.
- User Incentivisation: Granular attribution models mean operators can now fairly allocate power metrics to users, opening up new avenues for carbon-incentivisation tools.
- A Foundation for Green AI: Precise measurement and attribution frameworks are the absolute prerequisite for training effective AI models to minimise network energy waste.
This research was co-funded by the European Union under the HORIZON-JU-SNS-2023 Grant Agreement No. 101139120 as part of the EXIGENCE Project.
References
Reference: https://zenodo.org/records/20645561
Authors
This work is a collaborative effort by researchers from TNO; Belma Turkovic, Ganesh Aditya, Ljupco Jorguseski, Toni Dimitrovski and Sarah N. Lim Choi Keung, contributing to the broader EXIGENCE research ecosystem.
TNO
Sarah Lim Choi Keung is a project manager at TNO, focusing on ICT research projects in the areas of next-generation networking technologies, including 5G/6G, cloud-based, radio and immersive networking. Her interests extend to technology for verticals, sustainability and technology policy. With a background in Computer Science, Sarah has research, teaching and project management experience in health informatics while working at the University of Birmingham and University of Warwick in the UK.