Skip to content

Latest commit

 

History

History
64 lines (46 loc) · 2.69 KB

File metadata and controls

64 lines (46 loc) · 2.69 KB

Carbon Emissions Report

This document reports the carbon footprint of the LLM inference runs conducted during the ASOS-CRM evaluation. These measurements were omitted from the DaWaK 2026 paper due to space constraints and are provided here in the interest of reproducibility and environmental transparency.

All emissions were estimated using EcoLogits, a framework for measuring the environmental impact of LLM inference requests.


Measurement Context

Parameter Value
Tracking tool EcoLogits
Emission unit gCO₂eq (grams of CO₂ equivalent)
Electricity mix zone France (FRA)
LLM requests (per model) 22
Scope Full ASOS-CRM evaluation across all four datasets

Results

Llama 4 Scout

Parameter Value
Backend Ollama
Provider huggingface_hub
Model llama4:scout
LLM requests 22
Output tokens 17,658
Request latency 4,589.80 s
Quantization bits 4
Estimated energy 0.019445 kWh
Est. CO₂ emissions 4.9872 gCO₂eq

GPT-4o

Parameter Value
Provider OpenAI
Model gpt-4o
LLM requests 22
Output tokens 53,699
Request latency 1,133.03 s
Estimated energy 0.107612 kWh
Est. CO₂ emissions 43.8989 gCO₂eq

Summary

Model LLM requests Output tokens Energy (kWh) CO₂ (gCO₂eq)
Llama 4 Scout 22 17,658 0.019445 4.99
GPT-4o 22 53,699 0.107612 43.90

Discussion

GPT-4o produces approximately nine times more CO₂ emissions than Llama 4 Scout (43.90 vs. 4.99 gCO₂eq) despite completing the same 22-request workload. This disparity is partly explained by output token volume, GPT-4o generated 53,699 tokens compared to 17,658 for Llama 4 Scout, a factor of approximately three, but the ratio in emissions (×9) substantially exceeds the ratio in output tokens (×3), reflecting the additional carbon overhead of large-scale cloud data centre infrastructure as modelled by EcoLogits.

These figures should be read in conjunction with annotation quality. ASOS-CRM with Llama 4 Scout achieves 85.3% precision and 73.2% recall at a footprint of approximately 5 gCO₂eq, roughly one ninth of the 44 gCO₂eq required by GPT-4o for a gain of approximately 6 percentage points in precision. For institutions operating under data governance, local deployment with an open-weight model therefore represents the environmentally preferable configuration, provided that the modest reduction in annotation quality is acceptable for the target use case.