Where Do the Joules Go? Diagnosing Inference Energy Consumption
Paper • 2601.22076 • Published
model stringlengths 9 38 | slug stringlengths 9 38 | model_id stringlengths 13 49 | params_b float64 8 671 | active_params_b float64 3 405 | precision stringclasses 2
values | gpu stringclasses 2
values | num_gpus int64 1 8 | max_num_seqs int64 128 4.1k | avg_running_batch float64 128 3.76k | wh_per_million_output_tokens float64 13 573 | joules_per_output_token float64 0.05 2.06 | median_itl_ms float64 33.9 272 | gpu_power_watts_total int64 589 7.53k | avg_output_tokens int64 351 1.46k | row_id stringlengths 20 49 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Llama 3.1 8B Instruct | llama-3-1-8b-instruct | meta-llama/Llama-3.1-8B-Instruct | 8 | 8 | BF16 | B200 | 1 | 2,048 | 2,042.9 | 12.97 | 0.046696 | 124.9 | 691 | 396 | llama-3-1-8b-instruct/B200/1/2048 |
Qwen 3 8B | qwen-3-8b | Qwen/Qwen3-8B | 8 | 8 | BF16 | B200 | 1 | 1,536 | 1,426.7 | 14.74 | 0.053058 | 91.2 | 801 | 649 | qwen-3-8b/B200/1/1536 |
Qwen 3 14B | qwen-3-14b | Qwen/Qwen3-14B | 14 | 14 | BF16 | B200 | 1 | 2,048 | 2,044.6 | 15.21 | 0.054771 | 112.6 | 869 | 613 | qwen-3-14b/B200/1/2048 |
Qwen 3 30B A3B Instruct | qwen-3-30b-a3b-instruct | Qwen/Qwen3-30B-A3B-Instruct-2507 | 30 | 3 | BF16 | B200 | 1 | 1,024 | 1,022.6 | 15.67 | 0.056408 | 88.7 | 601 | 928 | qwen-3-30b-a3b-instruct/B200/1/1024 |
Llama 3.1 8B Instruct | llama-3-1-8b-instruct | meta-llama/Llama-3.1-8B-Instruct | 8 | 8 | BF16 | H100 | 1 | 512 | 510.9 | 19.5 | 0.070195 | 47.4 | 688 | 397 | llama-3-1-8b-instruct/H100/1/512 |
Qwen 3 8B | qwen-3-8b | Qwen/Qwen3-8B | 8 | 8 | BF16 | H100 | 1 | 512 | 474.4 | 20.04 | 0.072129 | 47.2 | 616 | 645 | qwen-3-8b/H100/1/512 |
Qwen 3 30B A3B Instruct | qwen-3-30b-a3b-instruct | Qwen/Qwen3-30B-A3B-Instruct-2507 | 30 | 3 | BF16 | H100 | 2 | 1,024 | 1,018.8 | 24.69 | 0.088877 | 74.8 | 1,123 | 920 | qwen-3-30b-a3b-instruct/H100/2/1024 |
Gemma 3 12B | gemma-3-12b | google/gemma-3-12b-it | 12 | 12 | BF16 | B200 | 1 | 512 | 486.6 | 26.49 | 0.095356 | 43.5 | 928 | 723 | gemma-3-12b/B200/1/512 |
Qwen 3 14B | qwen-3-14b | Qwen/Qwen3-14B | 14 | 14 | BF16 | H100 | 1 | 256 | 253.4 | 31.09 | 0.111917 | 38.4 | 645 | 625 | qwen-3-14b/H100/1/256 |
Qwen 3 32B | qwen-3-32b | Qwen/Qwen3-32B | 32 | 32 | BF16 | B200 | 1 | 384 | 383.3 | 39.05 | 0.140581 | 50.2 | 972 | 635 | qwen-3-32b/B200/1/384 |
NVIDIA Nemotron Nano 12B V2 | nvidia-nemotron-nano-12b-v2 | nvidia/NVIDIA-Nemotron-Nano-12B-v2 | 12 | 12 | BF16 | B200 | 1 | 256 | 255.6 | 39.43 | 0.141962 | 33.9 | 885 | 1,464 | nvidia-nemotron-nano-12b-v2/B200/1/256 |
NVIDIA Nemotron Nano 9B V2 | nvidia-nemotron-nano-9b-v2 | nvidia/NVIDIA-Nemotron-Nano-9B-v2 | 9 | 9 | BF16 | B200 | 1 | 256 | 255.3 | 40.94 | 0.147368 | 40.7 | 800 | 1,242 | nvidia-nemotron-nano-9b-v2/B200/1/256 |
NVIDIA Nemotron Nano 9B V2 | nvidia-nemotron-nano-9b-v2 | nvidia/NVIDIA-Nemotron-Nano-9B-v2 | 9 | 9 | BF16 | H100 | 1 | 256 | 255.3 | 43.11 | 0.155207 | 43.4 | 693 | 510 | nvidia-nemotron-nano-9b-v2/H100/1/256 |
Gemma 3 12B | gemma-3-12b | google/gemma-3-12b-it | 12 | 12 | BF16 | H100 | 1 | 128 | 127.8 | 46.14 | 0.166113 | 34.7 | 589 | 723 | gemma-3-12b/H100/1/128 |
Gemma 3 27B | gemma-3-27b | google/gemma-3-27b-it | 27 | 27 | BF16 | B200 | 1 | 256 | 247.3 | 49.1 | 0.176746 | 39.6 | 974 | 717 | gemma-3-27b/B200/1/256 |
NVIDIA Nemotron Nano 12B V2 | nvidia-nemotron-nano-12b-v2 | nvidia/NVIDIA-Nemotron-Nano-12B-v2 | 12 | 12 | BF16 | H100 | 1 | 256 | 241.2 | 50.08 | 0.180303 | 46 | 694 | 462 | nvidia-nemotron-nano-12b-v2/H100/1/256 |
Llama 3.1 70B Instruct | llama-3-1-70b-instruct | meta-llama/Llama-3.1-70B-Instruct | 70 | 70 | BF16 | B200 | 2 | 2,048 | 2,041.9 | 52.41 | 0.188675 | 140.5 | 1,907 | 351 | llama-3-1-70b-instruct/B200/2/2048 |
Llama 3.3 70B Instruct | llama-3-3-70b-instruct | meta-llama/Llama-3.3-70B-Instruct | 70 | 70 | BF16 | B200 | 2 | 1,024 | 1,021.6 | 60.72 | 0.218586 | 105.6 | 1,935 | 409 | llama-3-3-70b-instruct/B200/2/1024 |
Qwen 3 32B | qwen-3-32b | Qwen/Qwen3-32B | 32 | 32 | BF16 | H100 | 2 | 256 | 255.6 | 63.47 | 0.228481 | 44.5 | 1,178 | 634 | qwen-3-32b/H100/2/256 |
Qwen 3 235B A22B Instruct | qwen-3-235b-a22b-instruct | Qwen/Qwen3-235B-A22B-Instruct-2507 | 235 | 22 | BF16 | B200 | 4 | 4,096 | 3,764.6 | 66.73 | 0.240245 | 67.8 | 3,196 | 874 | qwen-3-235b-a22b-instruct/B200/4/4096 |
Gemma 3 27B | gemma-3-27b | google/gemma-3-27b-it | 27 | 27 | BF16 | H100 | 2 | 192 | 187 | 80.42 | 0.28952 | 39.8 | 1,210 | 714 | gemma-3-27b/H100/2/192 |
Llama 4 Scout 17B 16E Instruct | llama-4-scout-17b-16e-instruct | meta-llama/Llama-4-Scout-17B-16E-Instruct | 109 | 17 | BF16 | H100 | 4 | 512 | 472.8 | 93.39 | 0.336198 | 69.1 | 1,881 | 402 | llama-4-scout-17b-16e-instruct/H100/4/512 |
Llama 3.1 70B Instruct | llama-3-1-70b-instruct | meta-llama/Llama-3.1-70B-Instruct | 70 | 70 | BF16 | H100 | 4 | 1,024 | 898.5 | 103.65 | 0.373142 | 105.6 | 2,548 | 351 | llama-3-1-70b-instruct/H100/4/1024 |
Llama 3.3 70B Instruct | llama-3-3-70b-instruct | meta-llama/Llama-3.3-70B-Instruct | 70 | 70 | BF16 | H100 | 4 | 768 | 724.2 | 117.53 | 0.423094 | 97.5 | 2,546 | 413 | llama-3-3-70b-instruct/H100/4/768 |
Llama 4 Maverick 17B 128E Instruct FP8 | llama-4-maverick-17b-128e-instruct-fp8 | meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8 | 400 | 17 | FP8 | H100 | 8 | 384 | 383.3 | 134.5 | 0.484193 | 47.9 | 3,462 | 408 | llama-4-maverick-17b-128e-instruct-fp8/H100/8/384 |
Qwen 3 235B A22B Instruct FP8 | qwen-3-235b-a22b-instruct-fp8 | Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 | 235 | 22 | FP8 | B200 | 2 | 256 | 255.7 | 139.08 | 0.500694 | 64.5 | 1,891 | 876 | qwen-3-235b-a22b-instruct-fp8/B200/2/256 |
DeepSeek V3.1 | deepseek-v3-1 | deepseek-ai/DeepSeek-V3.1 | 671 | 37 | FP8 | B200 | 8 | 3,072 | 3,007.5 | 168.43 | 0.606344 | 185.5 | 5,470 | 842 | deepseek-v3-1/B200/8/3072 |
Qwen 3 235B A22B Instruct FP8 | qwen-3-235b-a22b-instruct-fp8 | Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 | 235 | 22 | FP8 | H100 | 8 | 512 | 511.2 | 205.96 | 0.74147 | 75.8 | 3,961 | 881 | qwen-3-235b-a22b-instruct-fp8/H100/8/512 |
Llama 3.1 405B Instruct | llama-3-1-405b-instruct | meta-llama/Llama-3.1-405B-Instruct | 405 | 405 | BF16 | B200 | 8 | 4,096 | 3,588.7 | 281.96 | 1.015042 | 272.3 | 7,530 | 370 | llama-3-1-405b-instruct/B200/8/4096 |
Qwen 3 235B A22B Instruct | qwen-3-235b-a22b-instruct | Qwen/Qwen3-235B-A22B-Instruct-2507 | 235 | 22 | BF16 | H100 | 8 | 256 | 228.5 | 303.53 | 1.09271 | 55.5 | 4,038 | 874 | qwen-3-235b-a22b-instruct/H100/8/256 |
Llama 3.1 405B Instruct FP8 | llama-3-1-405b-instruct-fp8 | meta-llama/Llama-3.1-405B-Instruct-FP8 | 405 | 405 | FP8 | B200 | 4 | 512 | 510.5 | 318.53 | 1.146695 | 98.2 | 3,883 | 380 | llama-3-1-405b-instruct-fp8/B200/4/512 |
Llama 3.1 405B Instruct FP8 | llama-3-1-405b-instruct-fp8 | meta-llama/Llama-3.1-405B-Instruct-FP8 | 405 | 405 | FP8 | H100 | 8 | 256 | 255.2 | 572.94 | 2.062586 | 66.9 | 4,875 | 367 | llama-3-1-405b-instruct-fp8/H100/8/256 |
Green Tokens: GPU energy per million output tokens
Lead Green is a public index of LLM inference energy per token: the GPU energy it takes to serve a million output tokens (watt-hours per million output tokens), by model, GPU and configuration, with grid carbon intensity to price it anywhere, published as open data with a JSON API and an MCP server under CC BY 4.0. Every figure comes from a named public source with a retrieval date; it is a Katonic project, run and paid for by Katonic, and takes no money from hardware, cloud or model vendors.
This dataset is the 2026-10-19 release of the Lead Green JSON API, file for file, with a flat CSV of the index for the viewer. Figures change only when a source publishes a dated release; a release listed as published never changes. Cite a release, not the live files.
What is in it
index.csvandindex.json: 32 rows, one per model and GPU, each the configuration with the lowest energy per output token whose median inter-token latency stays under 500 ms (workload: text-conversation).models/: every measured configuration for each of the 18 models (concurrency, parallelism, latency, power), with the best row per GPU.grid.json: grid emissions intensity in gCO2e per kWh for 194 countries and regions plus the world figure (2024 annual).hardware.json: GPU reference data (7 parts, 24 MIG profiles) transcribed from manufacturer documents.per-query.json: 22 rows of energy per query from the AI Energy Score (December 2025).measurements.json: rows contributed under the measurement protocol (0 at this release); never merged into the index.
Reading a row
The unit is Wh per million output tokens: the GPU energy a server drew while generating one million output tokens, at steady state, in the configuration named by the row. Selection rule, verbatim from the API: per model and GPU, the configuration with the lowest energy per token whose median inter-token latency does not exceed itl_deadline_ms (the leaderboard's default ceiling; no measured configuration reaches it).
At this release the index runs from 13.0 Wh/M (Llama 3.1 8B Instruct, 1 x B200) to 572.9 Wh/M (Llama 3.1 405B Instruct FP8, 8 x H100), a span of 44x. The reference row, used on the site for every comparison, is Qwen 3 32B, 2 x H100 at 63.5 Wh/M.
Boundary
GPU energy only, measured at steady state on a busy server (the most energy-efficient configuration under a latency limit). Host, network, cooling and idle time are not counted. Every figure is a floor; a lightly used server sits well above it. The boundary field on every document reads gpu_only.
Carbon
Multiply watt-hours per million output tokens by the grid's gCO2e per kWh and divide by one thousand to get kilograms of CO2e per billion output tokens. grid.json carries the intensity for each area; the world average at this release is 473 gCO2e per kWh. The site's per-country pages apply exactly this arithmetic.
At the reference row (Qwen 3 32B, 2 x H100, 63.5 Wh/M), a billion output tokens emits the following on 2024 grid intensity (Ember, lifecycle, location-based, annual average). All 194 areas are in grid.json; the site has a page per country at https://leadgreen.ai/grids/.
| Grid | gCO2e per kWh (2024) | kg CO2e per billion output tokens |
|---|---|---|
| World | 473 | 30.0 |
| United States | 383.8 | 24.4 |
| China | 559 | 35.5 |
| India | 707.1 | 44.9 |
| Japan | 484.3 | 30.7 |
| South Korea | 410.1 | 26.0 |
| Germany | 339.9 | 21.6 |
| United Kingdom | 216.5 | 13.7 |
| France | 40.3 | 2.6 |
| Poland | 608.5 | 38.6 |
| Serbia | 668.4 | 42.4 |
| Norway | 29.6 | 1.9 |
| Brazil | 101.8 | 6.5 |
| Australia | 549.5 | 34.9 |
| Saudi Arabia | 689.1 | 43.7 |
| United Arab Emirates | 477.3 | 30.3 |
| Qatar | 581.7 | 36.9 |
| Singapore | 498.7 | 31.7 |
| Malaysia | 604.3 | 38.4 |
| Indonesia | 710.4 | 45.1 |
| Viet Nam | 477 | 30.3 |
| South Africa | 717.8 | 45.6 |
These are GPU-only floors. A facility figure multiplies by the server's system-to-GPU power ratio and the site's PUE first.
H100 against B200
14 models have an index row on both cards. The B200 row uses less GPU energy per output token on 14 of 14; the H100 row's energy over the B200 row's is 1.63x at the median (a 39 percent reduction), from 4.55x down to 1.05x. The two rows of a pair are each the energy-optimal configuration on their own card, so they can differ in card count and running batch; every configuration is in models/ so any pair can be re-cut at a matched cap. The leaderboard team's own analysis of the same measurements agrees: across three latency limits the B200 wins 63 of 72 comparisons (88 percent) with a median 35 percent energy reduction (Chung et al. (The ML.ENERGY Initiative, University of Michigan), Where Do the Joules Go? Diagnosing Inference Energy Consumption, arXiv:2601.22076, v2, January 2026).
| Model | H100 row (cards, running batch) | Wh/M | B200 row (cards, running batch) | Wh/M | H100 / B200 |
|---|---|---|---|---|---|
| Qwen 3 235B A22B Instruct | 8 x H100, 228 | 303.5 | 4 x B200, 3,765 | 66.7 | 4.55x |
| Qwen 3 14B | 1 x H100, 253 | 31.1 | 1 x B200, 2,045 | 15.2 | 2.04x |
| Llama 3.1 70B Instruct | 4 x H100, 898 | 103.7 | 2 x B200, 2,042 | 52.4 | 1.98x |
| Llama 3.3 70B Instruct | 4 x H100, 724 | 117.5 | 2 x B200, 1,022 | 60.7 | 1.94x |
| Llama 3.1 405B Instruct FP8 | 8 x H100, 255 | 572.9 | 4 x B200, 510 | 318.5 | 1.80x |
| Gemma 3 12B | 1 x H100, 128 | 46.1 | 1 x B200, 487 | 26.5 | 1.74x |
| Gemma 3 27B | 2 x H100, 187 | 80.4 | 1 x B200, 247 | 49.1 | 1.64x |
| Qwen 3 32B | 2 x H100, 256 | 63.5 | 1 x B200, 383 | 39.0 | 1.63x |
| Qwen 3 30B A3B Instruct | 2 x H100, 1,019 | 24.7 | 1 x B200, 1,023 | 15.7 | 1.58x |
| Llama 3.1 8B Instruct | 1 x H100, 511 | 19.5 | 1 x B200, 2,043 | 13.0 | 1.50x |
| Qwen 3 235B A22B Instruct FP8 | 8 x H100, 511 | 206.0 | 2 x B200, 256 | 139.1 | 1.48x |
| Qwen 3 8B | 1 x H100, 474 | 20.0 | 1 x B200, 1,427 | 14.7 | 1.36x |
| NVIDIA Nemotron Nano 12B V2 | 1 x H100, 241 | 50.1 | 1 x B200, 256 | 39.4 | 1.27x |
| NVIDIA Nemotron Nano 9B V2 | 1 x H100, 255 | 43.1 | 1 x B200, 255 | 40.9 | 1.05x |
Concurrency
Energy per token depends on how many requests the server decodes at once. Qwen 3 32B on 2 x H100 reads 452.0 Wh/M at a running batch of 8 and 63.5 Wh/M at 256: 7.1x on unchanged hardware, while board power rises from 850 W to 1,178 W. Every row on the index is the busy end of such a ladder; a lightly used server sits well above it. Single-request benchmarks therefore read many times higher than the index for the same model and GPU, and both are correct measurements of different things. The full ladder for every model and GPU is in models/.
Per query
One answer on the reference row costs about 0.040 Wh of GPU energy (63.5 Wh/M times 634 average output tokens). The three published whole-stack per-prompt figures are 0.24 to 0.34 Wh: Google 0.24 Wh per median Gemini Apps text prompt (August 2025; accelerators 58 percent of it, narrow accelerator-only figure 0.10 Wh), Epoch AI 0.3 Wh per typical 500-token ChatGPT reply (February 2025) and OpenAI 0.34 Wh per average query (June 2025). The gap is boundary (whole estate against GPUs only), load and answer length. per-query.json holds the AI Energy Score December 2025 round, 0.0042 to 1.27 Wh per query across 22 models at batch size 1 with a fixed ten-token answer: a probe, not a serving load, so it is not converted to Wh/M. The conversion with each term stated is at https://leadgreen.ai/per-query/.
Files
| File | What it holds |
|---|---|
manifest.json |
Sources, versions, retrieval dates, licences and this endpoint list. |
index.json |
The index: 32 rows, one per model and GPU, energy-optimal under the latency limit, sorted ascending. |
index.csv |
The index rows as CSV, for a spreadsheet. |
models.json |
Every model on the index with its slug, Hugging Face id, parameter counts and measured GPUs. |
grid.json |
Grid emissions intensity, gCO2e per kWh, 2024 annual average, 193 countries plus World. |
hardware.json |
Manufacturer power ceilings, memory, MIG slice profiles and one embodied-carbon figure. |
per-query.json |
AI Energy Score text-generation rows, December 2025 round: Wh per 1,000 queries at batch size 1 on H100 80GB GPUs (one card for classes A and B, several for C). |
measurements.json |
Contributed measurements accepted under the protocol, with contributor, affiliation, date, versions and MIG profile; never merged into index.json. |
changes.json |
Every change to the published data or its description: date, summary, documents, reason and source version. |
openapi.json |
OpenAPI 3.1 description of these endpoints. |
models/<slug>.json (18 files) |
One model: every measured configuration (concurrency, parallelism, latency, power) and the best row per GPU. |
schema/measurement.schema.json |
The JSON schema a contributed measurement must satisfy (protocol 0.2). |
NOTICE.md |
Per-source provenance and terms for every data snapshot the index is computed from. |
Sources and licences
| Source | Version | Retrieved | Terms |
|---|---|---|---|
| ML.ENERGY Leaderboard | 3.0 (released 2025-12-01); data compiled 2026-02-16 | 2026-09-26 | Apache-2.0 per the ml-energy/benchmark-v3 dataset card |
| Ember, Yearly electricity data | October 2026 release (file last modified 2026-10-06) | 2026-10-17 | CC-BY-4.0 |
| NVIDIA Multi-Instance GPU User Guide, Supported MIG Profiles | last updated 2026-09-11 | 2026-09-26 | manufacturer specifications transcribed as facts; NVIDIA retains its rights in the source documents |
| NVIDIA product specifications: A100, H100, H200, DGX B200 | as published on the retrieval date | 2026-09-26 | manufacturer specifications transcribed as facts; NVIDIA retains its rights in the source documents |
| NVIDIA, Product Carbon Footprint Summary for NVIDIA HGX B200 | July 2025 | 2026-09-26 | manufacturer specifications transcribed as facts; NVIDIA retains its rights in the source documents |
| AI Energy Score leaderboard | December 2025 round | 2026-09-26 | Apache-2.0 as tagged on the Space; the project states no separate licence for the results file, which is reproduced as facts with attribution |
| OpenAI, What are tokens and how to count them | as published on the retrieval date | 2026-09-26 | cited, not redistributed |
Compilation and site text: CC BY 4.0. Code: Apache 2.0. Third-party figures remain subject to their sources' terms, listed per source.
Citation
Attribution line to copy (fill the Changes field): Green Tokens index, Lead Green, a Katonic project, https://leadgreen.ai/api/v1/, CC BY 4.0. Contains data from the ML.ENERGY Leaderboard v3.0 (The ML.ENERGY Initiative; data compiled 2026-02-16, retrieved 2026-09-26) and Ember Yearly Electricity Data (CC BY 4.0). Changes: [none | describe].
The site root carries a CITATION.cff; GitHub renders it as "Cite this repository". The concept DOI https://doi.org/10.5281/zenodo.23478909 always resolves to the latest release.
Latest change
2026-10-19: Launch. Three blog posts: how much energy an AI query uses, what watt-hours per million tokens means, and why an idle GPU wastes energy. The site now serves from GitHub Pages at leadgreen.ai. No change to the data or its description. The full record is in changes.json.
Elsewhere
- Website and documentation: https://leadgreen.ai/
- JSON API (GET only, CORS open): https://leadgreen.ai/api/v1/ (https://leadgreen.ai/api/v1/manifest.json lists every endpoint)
- MCP server for assistants: https://leadgreen.ai/mcp/
- Dated releases, each frozen once published: https://leadgreen.ai/api/v1/releases/index.json
- Source, build script and issue tracker: https://github.com/katonic-dev/leadgreen-ai
- Concept DOI (always resolves to the latest release): https://doi.org/10.5281/zenodo.23478909
- Wikidata item: https://www.wikidata.org/wiki/Q141778099
Publisher and funder
Lead Green is run and paid for by Katonic, which sells a sovereign AI platform and benefits when buyers ask the questions on this site. It takes no money from hardware, cloud or model vendors.
Publisher: Katonic AI Technology Ltd, registration CL9944, Dubai International Financial Centre (DIFC), United Arab Emirates, trading this project as Lead Green. Corrections and contributions: hello@leadgreen.ai, answered within 10 working days. Contributed measurements are made under the protocol described at https://leadgreen.ai/contribute/.
- Downloads last month
- 226