1. Background and motivation
As mobile networks evolve toward 6G, artificial intelligence is moving from an optional optimisation layer to the core control mechanism of the network itself: AI-native architectures increasingly rely on models embedded directly in the control loop (e.g., forecasting traffic, managing topology, orchestrating resources, and making real-time decisions across base stations and edge sites). This project is situated within that shift toward AI-native network control, and asks a question that becomes unavoidable once AI moves from an occasional tool to a continuously running control component: what does it actually cost, in energy, to run these models as part of the network’s operation? While the wider AI industry debates the energy footprint of training large models, the models that actually govern network behaviour are queried constantly, at the timescale of scheduling intervals, which shifts the real energy burden toward inference rather than training. Understanding and measuring this properly is a precondition for making sound architectural choices in AI-native 6G networks.
The wider AI industry has started to reckon with training-time energy cost (large published training runs, GPU-hour audits), but inference-time energy cost is comparatively under-measured, despite the fact that a model queried thousands of times per second across a national network’s control loop will spend far more lifetime energy on inference than it ever spent on training. This is exactly backwards from where most public attention is: a great deal of visible discourse is about the cost of training frontier models, while use cases like: small-to-medium models running continuously and repeatedly inside network infrastructure, are dominated by inference cost. This project sets out to measure this properly, and to produce an energy profile that can inform every subsequent architectural choice.
2. Research questions
- RQ1 (Training cost): For representative network-intelligence model classes (small transformers/LSTMs for traffic forecasting, GNNs for topology-aware tasks, lightweight LLM fine-tunes for agentic control, federated-learning client updates), how does measured energy consumption during training scale with model size, dataset size, batch size, and training regime (full fine-tune vs. LoRA/parameter-efficient fine-tuning vs. training from scratch)?
- RQ2 (Inference cost): For the same model classes, how does per-inference energy consumption scale with model size, batch size, sequence/input length, hardware (CPU vs. GPU vs. edge accelerator), and serving configuration (single query vs. batched/streaming)? At what query rate does inference energy overtake training energy, given realistic network-operational query volumes (e.g. a per-cell traffic-prediction model queried every scheduling interval)?
- RQ3 (Precision/efficiency techniques): How much energy is actually saved by common efficiency techniques (quantisation to int8/int4, pruning, knowledge distillation, early-exit architectures, sparse Mixture-of-Experts activation) when applied to network-intelligence models specifically, versus the generic NLP/vision benchmarks these techniques are usually evaluated on?
- RQ4 (Federated/continual learning overhead): In a federated-learning setting where multiple base stations or edge sites each perform local updates, how does the aggregate energy cost (communication + local compute, summed across clients) compare to a centralised training run achieving comparable accuracy? Does continual/incremental learning (updating a model as new data arrives) cost less in total energy than periodic full retraining, once measured end-to-end rather than assumed?
- RQ5 (Reporting standard): What is a minimal, reusable measurement-and-reporting protocol (“energy card”) that can be attached to any trained and deployed model, analogous to a model card, so that energy cost becomes a first-class, comparable metric across different studies?