Sustainability AI
Abstract
How much does it cost the planet to run a neural network? This project combines YOLOv5 object detection with Carbontracker to measure the real carbon footprint of ML inference — making the invisible environmental cost of AI visible and measurable.
Inference Carbon Footprint
per 1K images
~2.3g
CO₂ per 1K inferences
YOLOv5s on consumer GPU
0.8 kWh
Energy per training run
Tracked via Carbontracker
73%
mAP accuracy
On accessibility dataset
Methodology
Language
Model
Deep Learning
Emissions Tracking
Background
Sustainability AI is a research project that combines YOLOv5 object detection with Carbontracker to measure the carbon footprint of running machine learning inference. The goal: make the environmental cost of AI visible, not invisible.
The project tracks GPU energy consumption during YOLOv5 detection runs and reports estimated CO₂ emissions per inference batch. It was built as an experiment in carbon-aware computing — understanding not just what AI can do, but what it costs the planet to do it.