---
title: "From Tokens to Watt-hours: Analytical Energy Estimation for LLM Inference on Modern GPUs"
description: "Large language model inference is a growing part of the environmental footprint of AI systems, yet measuring its energy directly usually needs hardware…"
url: https://greencode.ai/publications/from-tokens-to-watt-hours-analytical-energy-estimation-for-llm-inference-on-mode/
---

# From Tokens to Watt-hours: Analytical Energy Estimation for LLM Inference on Modern GPUs

- **Authors:** [Tina Vartziotis](https://orcid.org/0009-0009-5300-2572 "ORCID profile of Tina Vartziotis") , Rodopi Kosteli, Elli Danae Vartziotis, [George Dasoulas](https://orcid.org/0000-0002-0562-5136 "ORCID profile of George Dasoulas") , [Michael Keckeisen](https://orcid.org/0009-0000-2214-9445 "ORCID profile of Michael Keckeisen") , [Konstantinos Skianis](https://orcid.org/0009-0002-8804-6320 "ORCID profile of Konstantinos Skianis") , [Sotirios Kotsopoulos](https://orcid.org/0000-0003-2709-752X "ORCID profile of Sotirios Kotsopoulos") , [Francesca Dominici](https://orcid.org/0000-0002-9382-0141 "ORCID profile of Francesca Dominici")
- **Contributing partner:** [TWT Science & Innovation](https://greencode.ai/partners/twt-gmbh-science-innovation/ "TWT GmbH Science & Innovation in GreenCode")
- **Published in:** arXiv preprint arXiv:2607.26571 (2026); accepted for oral presentation at GREEN-AI: Grounding Research on Energy- and resource-Efficient Networks for sustainable AI, workshop co-located with ECML-PKDD 2026, Naples, Italy, 7 September 2026
- **Published:** Jul 2026
- **Type:** Paper
- **DOI:** [10.48550/arXiv.2607.26571](https://doi.org/10.48550/arXiv.2607.26571)

### Synopsis

Large language model inference is a growing part of the environmental footprint of AI systems, yet measuring its energy directly usually needs hardware telemetry, power instruments or infrastructure-specific monitoring. That rules direct measurement out for many comparative studies, early design decisions and sustainability reports. This report sets out a way to estimate inference energy on NVIDIA H100-class GPUs without measuring at run time. The estimator combines transformer arithmetic counts scaled by model size, calibrated memory-traffic factors and hardware-specific energy figures for tensor-core computation and high-bandwidth memory access. It treats prompt processing separately from token-by-token generation, so it can estimate energy for input tokens, output tokens and whole requests. It also splits total energy into computation, parameter access, key-value cache writes and attention reads, showing how energy scales with model size, context length and the number of tokens generated. The estimates are not meant to replace physical power measurement; they are transparent, reproducible approximations with explicit assumptions, suited to comparing models, green coding analysis and design-time evaluation of inference workloads.

[View at the publisher](https://doi.org/10.48550/arXiv.2607.26571) [All publications](https://greencode.ai/publications/)Cite this work

Vartziotis, T., Kosteli, R., Vartziotis, E., Dasoulas, G., Keckeisen, M., Skianis, K., Kotsopoulos, S., &amp; Dominici, F. (2026). <i>From Tokens to Watt-hours: Analytical Energy Estimation for LLM Inference on Modern GPUs</i> (Version 1). arXiv. <https://doi.org/10.48550/ARXIV.2607.26571>
