Research map / Large language models / Reasoning and chain of thought

CO2-Meter: A Comprehensive Carbon Footprint Estimator for LLMs on Edge Devices

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Efficient LLM serving · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces CO2-Meter, a framework estimating operational (peripheral equation models plus a GNN predictor for prefill/decode LLM energy) and embodied (unit-level SoC) carbon for LLM inference on edge devices.

Problem

Existing carbon estimators for LLMs ignore peripheral energy on edge devices, do not separate prefill and decode phases, and model embodied carbon only at chip-level, missing SoC unit-level bottlenecks.

Approach

Equation-based energy models for peripherals (WiFi, Bluetooth, camera, microphone, video output, speaker, display), a GNN over transformer-layer kernel graphs that predicts phase-specific (prefill and total) LLM inference energy, and a unit-level embodied carbon model summing per-unit area times carbon-per-area for SoCs.

Key results

• CO2-Meter achieves 10.3% average MAPE on rk3588 and 18% on AGX Orin, versus 33.1% and 28.1% for NNLQP.
• CO2-Meter reaches 70.3% (rk3588) and 48.6% (AGX Orin) average accuracy within 10% error bounds, outperforming RF and NNLQP.
• On unseen Q2 inference, CO2-Meter improves 10% EB accuracy by 129% on rk3568 and 75% on AGX Orin relative to NNLQP.

Surprising findings

• The TFT LCD display dominates operational energy (>55%), and LLM inference accounts for only 2–12% of total energy in the virtual reality assistant case.
• AGX Orin has about 3.44x the embodied carbon of rk3588 yet uses 14–20% less operational energy for the same task.

Limitations

• LLM energy predictor trained on only three LLMs and two SoCs; generalization tested on one unseen LLM and two devices.
• Embodied carbon model shows 10-20% deviation from reported values and relies on assumed per-area carbon factors.
• Batch size 1 only and token length distributions assumed from cloud traces.

Experimental setup

Datasets: Azure LLM inference traces self-collected LLM inference energy measurements on rk3588 and AGX Orin peripheral energy measurements

Benchmarks: self-collected LLM inference energy dataset (rk3588, AGX Orin) unseen Q2 configuration on rk3568 and Orin NX

Metrics: MAPE 10% error bound accuracy mean absolute error

Baselines: RF NNLQP 2P

Builds on: NNLQP GraphSAGE graph attention networks graph isomorphism networks Azure LLM inference traces

Closest papers by meaning

• LLMCarbon: Modeling the End-to-End Carbon Footprint of Large Language Models (ICLR 2024 · similarity 0.86)
• CATransformers: Carbon Aware Transformers Through Joint Model-Hardware Optimization (NeurIPS 2025 · similarity 0.84)
• Carbon Footprint Reduction for Sustainable Data Centers in Real-Time (AAAI 2024 · similarity 0.81)
• CE-NAS: An End-to-End Carbon-Efficient Neural Architecture Search Framework (NeurIPS 2024 · similarity 0.81)
• Holistically Evaluating the Environmental Impact of Creating Language Models (ICLR 2025 · similarity 0.80)
• TokenPowerBench: Benchmarking the Power Consumption of LLM Inference (AAAI 2026 · similarity 0.80)