Model A
01-ai/Yi-1.5-9B
23.001 Wh · GPU electricity
Carbon not estimated
Published baseline: 23.001 Wh / 1,000 tasks.
Model evidence
Compare model energy tests
Choose two models to compare their GPU electricity use in the same published text-generation test. These results exclude the rest of the computer and data centre.
Browse 27 open models tested in February 2025. It does not cover every model or measure the service handling your requests.
The test conditions and original data
AI Energy Score · February 2025 text-generation cohort. NVIDIA H100 80 GB, one GPU. Class A describes a model-size grouping, not an RTX 4090 measurement.
Published method: 1,000 prompts sampled from WikiText, OSCAR and UltraChat; batch size 1; text generation in FP16; mean of ten runs.
Reported GPU energy for preprocessing, prefill and decode. Excludes CPU, RAM, facility cooling, networks, training and hardware manufacture.
The imported summary does not supply per-row output-token counts, exact weight revision hashes, run-level spread or complete software configurations. The separately pinned method describes the protocol, not a verified historical configuration for each row.
Published benchmark summary; AI Footprint Lab extraction not independently reviewed. Energy alone does not establish answer quality, usefulness or overall environmental performance.
Selection, units and original data
All February 2025 Class A text-generation records with one GPU explicitly listed in the parameter table. Alphabetical order; no selection by energy result.
CSV total_gpu_energy is kWh per 1,000 queries. Multiply by 1,000 for Wh per 1,000 queries; divide that by 1,000 for an allocated mean per query.
Pinned energy table · Hardware-count table · Published protocol
Source snapshot checked 2026-09-05. Download original energy CSV · Download parameter CSV · Attribution and extraction notes
Explore two models
This compares electricity only. Answer quality is not scored here, so lower energy use does not necessarily mean a model is a better fit for your task.
Model B
meta-llama/Meta-Llama-3-8B
17.333 Wh · GPU electricity
Carbon not estimated
Published baseline: 17.333 Wh / 1,000 tasks.
Linear scaling of a published benchmark mean, not a new measurement or a prediction of a commercial service. Water is not estimated.
These choices belong to this comparison only and reset when you leave the page. Download to preserve them; the main scenario report does not include this comparison.
Browse the collection
27 of 27 records
| Model identifier | GPU Wh / 1,000 tasks | Compare |
|---|---|---|
| 01-ai/Yi-1.5-9B | 23.001 | |
| allenai/OLMo-1.7-7B-hf | 16.831 | |
| allenai/OLMo-1B-hf | 5.594 | |
| berkeley-nest/Starling-LM-7B-alpha | 19.011 | |
| CohereForAI/aya-23-8B | 17.817 | |
| distilbert/distilgpt2 | 1.313 | |
| facebook/opt-125m | 1.944 | |
| facebook/opt-2.7b | 8.06 | |
| google/gemma-2b | 8.68 | |
| google/gemma-7b | 18.9 | |
| HuggingFaceTB/SmolLM-1.7B | 7.645 | |
| HuggingFaceTB/SmolLM-135M | 7.28 | |
| HuggingFaceTB/SmolLM-360M | 8.095 | |
| meta-llama/Llama-2-13b-hf | 26.102 | |
| meta-llama/Llama-2-7b-hf | 16.441 | |
| meta-llama/Meta-Llama-3-8B | 17.333 | |
| meta-llama/Meta-Llama-3.1-8B | 17.376 | |
| microsoft/phi-1 | 6.393 | |
| microsoft/phi-1_5 | 6.29 | |
| microsoft/phi-2 | 9.864 | |
| mistralai/Mistral-7B-v0.1 | 19.13 | |
| openai-community/gpt2 | 2.146 | |
| openai-community/gpt2-large | 6.121 | |
| openai-community/gpt2-medium | 3.956 | |
| openai-community/gpt2-xl | 8.921 | |
| openai-community/openai-gpt | 2.704 | |
| tiiuae/falcon-7b | 15.875 |
Benchmark evidence and review dates
○ Study or modeled estimate
Reported GPU energy for preprocessing, prefill and decode. Excludes CPU, RAM, facility cooling, networks, training and hardware manufacture.
Water in a real place · historical case study
The Dalles, Oregon
Facility disclosure: Google’s April 2023 profile reports almost 275 million gallons consumed at its The Dalles data centres in 2021. This covers the site’s services collectively; it is not an AI-only total or a current-year reading. Read the company’s original disclosure.
Water-supply context: the city’s 2025 drinking-water report identifies a municipal watershed fed by Dog River, Alder Creek, Crow Creek and the South Fork of Mill Creek, with wells drawing from The Dalles Pool aquifer. A location beside the Columbia River does not by itself identify the water delivered to a facility. Read the city’s water report.
These documents describe different periods and boundaries. The city’s supply description does not establish the facility’s exact 2021 source mix. We do not turn either document into litres per AI request or a watershed-impact score.
What would complete the investigation?
- Matching facility withdrawal, discharge and consumption records, split by source and month.
- Workload allocation and water used to generate the facility’s electricity.
- Seasonal streamflow, groundwater levels, competing demand and ecological needs.
Use the Oregon Water Resources Department’s city water-use records to investigate supply over time. Company disclosure, utility records and ecological observations answer different questions. This case is selected for its accessible primary documents, not as a ranking of providers.
Water case evidence and review dates
○ Study or modeled estimate
Historical facility disclosure and separately dated municipal supply context.
Historical teaching case; no live monitoring or scarcity classification.