emolero/open-telco-energy-logs

Open-Telco Energy Logs

Per-model GPU energy measurements for 67 open-weight LLMs evaluated on the Open Telco benchmark suite. This is the raw data backing the energy and AMEI (AI Model Energy Index) tables in the AMEI paper.

Three measurement campaigns are included:

Campaign	Models	Hardware	Runtime	Suite	Samples/model	Schema
single_h100/	24	1×H100-80GB SXM	vLLM 0.8.5 (TP=1)	open_telco_5_benchmarks	1,400	v1.0
4x_h100/	35	4×H100-80GB SXM	vLLM 0.15.1 (TP=4)	open_telco_7_benchmarks	4,200	v1.2
ollama/	8	1×H100-80GB SXM	Ollama	orange_business_resolver	2–120	v1.1

What this dataset is: aggregated per-model energy, throughput, and token counts plus per-task breakdowns. What it is not: model weights, prompts, completions, trajectories, or raw NVML power traces. For benchmark code and prompts see the AMEI paper repository.

Quick start

from datasets import load_dataset

ds = load_dataset("emolero/open-telco-energy-logs", data_files="all_runs.jsonl", split="train")
print(len(ds))                                            # 67
print(set(ds["campaign"]))                                # {'single_h100', '4x_h100', 'ollama'}

df = ds.to_pandas()
df.sort_values("wh_per_1k_queries").head(10)              # most efficient models

To read a single run's full per-task breakdown:

from huggingface_hub import hf_hub_download
import json

path = hf_hub_download(
    repo_id="emolero/open-telco-energy-logs",
    filename="4x_h100/qwen2.5-7b/energy_results.json",
    repo_type="dataset",
)
run = json.load(open(path))
for task in run["per_task"]:
    print(task["task"], task["energy_j"], task["scores"])

Layout

emolero/open-telco-energy-logs/
├── README.md
├── SCHEMA.md                     # field-by-field reference for v1.0/v1.1/v1.2
├── all_runs.jsonl                # 67 rows, one per (campaign, model), schema-normalized
├── single_h100/
│   └── <model>/energy_results.json   # 24 files (v1.0)
├── 4x_h100/
│   └── <model>/energy_results.json   # 35 files (v1.2)
└── ollama/
    ├── run_summary.json              # campaign-level outcome log
    └── <model>/energy_results.json   # 8 files (v1.1)

Schema

See SCHEMA.md for the full field reference. Headline numbers in every run:

Field	Unit	Meaning
metrics.total_energy_wh	Wh	Total GPU energy for the run.
metrics.wh_per_1k_queries	Wh per 1K queries	Energy intensity. AMEI input. Lower is better.
metrics.j_per_token	J per output token	Token-normalized energy. Lower is better.
metrics.total_tokens_out	tokens	Total generated tokens.
metrics.total_samples	int	Samples processed (called n_samples in v1.1).
metrics.mean_score	float	Cohort-mean benchmark score (v1.2 only).

Measurement protocol

• Tool: Zeus wrapping NVML's nvmlDeviceGetTotalEnergyConsumption() — a hardware cumulative energy counter, no polling required.
• Accuracy: ±5% on H100 per the SC24 GPU energy measurement paper.
• Why not CodeCarbon: CodeCarbon polls every 15 s, which misses sub-second inferences entirely. NVML's cumulative counter captures every joule.
• Warmup: 10 samples discarded per model for thermal equilibrium and CUDA kernel caching.
• Scope: Full request lifetime (including GPU idle during agent tool calls on the Ollama campaign). This matches deployment cost: you pay for GPU time whether it's computing or waiting.
• Carbon intensity: When present, carbon_gco2e uses a constant 400 gCO₂e/kWh (US grid average). It's a linear transform of energy and does not change rankings on the same hardware.

Caveats

• Cross-campaign comparison is apples-to-oranges. Different hardware (1× vs 4× H100), different vLLM versions (0.8.5 vs 0.15.1), and Ollama uses different concurrency than vLLM. For absolute energy comparisons, stay within one campaign. Fourteen models overlap between single_h100/ and 4x_h100/; see Section 4 of the paper for the ratios.
• DeepSeek-R1-Distill-Qwen-32B is absent from single_h100/ — it OOM'd and triggered worker instability. The 7B, 8B, and 14B distillations completed; their chain-of-thought behavior inflates token counts substantially.
• Reasoning models look efficient per-token but expensive per-query. DeepSeek-R1 variants have low j_per_token (fast generation) but very high wh_per_1k_queries (because they generate many tokens per query).
• Ollama runs use a different benchmark (orange_business_resolver, an agentic ReAct workload) and run on smaller sample counts. Don't merge these with the vLLM campaigns without thinking about it.
• Single GPU host. All measurements were taken on a single OCI bare-metal H100 server. We have not characterized inter-host variance.
• Two partial runs in single_h100/. olmo2-7b (10 samples) and olmo2-13b (8 samples) did not complete the full 1,400-sample sweep. They are preserved for transparency but should be filtered out for any cohort statistic — check n_samples == 1400 before averaging.

Contributing a new run

If you add a model to one of the campaigns, follow this checklist:

1. Drop the new file at <campaign>/<model-slug>/energy_results.json. Model slugs are lowercase with hyphens (e.g. qwen2.5-7b, not Qwen2.5-7B).
2. Match the schema version of the campaign. If you legitimately need new fields, bump the minor version in SCHEMA.md — do not rename or remove existing fields.
3. Re-run the publish script with --dry-run to confirm all_runs.jsonl regenerates cleanly:

   HF_TOKEN=hf_... python scripts/publish_energy_logs_to_hf.py --dry-run
   

   The output should show one more row and the new model under the right campaign.
4. Update the campaign table at the top of this file if the count changes.
5. If the new run measures something materially different (new benchmark, different hardware), open an issue first — it likely belongs in a new campaign directory.

Reproducing the measurements

The measurement harness lives in the paper repo:

• energy/run_suite.py — top-level driver
• energy/runners/ — per-provider runners (vLLM, Ollama)
• experiments/energy_consumption_*/README.md — campaign-specific protocols

You need: an H100-80GB host, Zeus, vLLM, and the Inspect AI benchmark code from the paper repo.

Citation

@article{molero2026amei,
  title   = {AMEI: An AI Model Energy Index for LLM Inference in Telecommunications},
  author  = {Molero, Enrique and others},
  year    = {2026},
  note    = {Paper repo: https://github.com/eaguaida/AMEI_paper},
}

License

CC-BY-4.0. Free to use with attribution.

Contact

Issues, corrections, or new runs: https://github.com/eaguaida/AMEI_paper/issues.