Research map / Large language models / Agents, code and math
LLMCarbon: Modeling the End-to-End Carbon Footprint of Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LLMCarbon, an end-to-end model that predicts operational and embodied carbon footprints of dense and MoE LLMs before training, validated against published numbers for T5, GPT-3, GShard, Switch, XLM and Meta/Noor data.
Problem
Existing tools (mlco2) can predict ML carbon footprint only for CNNs and GPUs, ignore architectural parameters and MoE models, and cannot model embodied carbon, so LLM carbon cannot be estimated before training.
Approach
An analytical projection model combining a parameter model, Chinchilla neural scaling law for test loss, FLOP model, parallelism-dependent hardware efficiency model, PUE and carbon intensity for operational carbon, and chip area times CPA amortized over hardware lifetime for embodied carbon, extended to MoE LLMs.
Key results
- LLMCarbon predicts training operational carbon within 8.2% of actual values for five LLMs (T5 -2.22%, GPT-3 +0.32%, GShard +3.8%, Switch +8.2%, XLM -3.54%), versus 69% to 132% overestimation by mlco2.
- Embodied carbon for Meta XLM hardware predicted at 0.64 tCO2eq vs actual 0.66 (-3.05%).
- Embodied carbon accounts for about 24% to 35% of total life-cycle footprint of LLMs and 92% to 95% in a 97% renewable Meta data center.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LLMCarbon instead of mlco2 for operational carbon projection | LLM training · T5, GPT-3, GShard, Switch, and XLM training | absolute disparity from actual operational carbon: mlco2: +91.3%, +73%, +95.3%, +132%, +69%; LLMCarbon: -2.22%, +0.32%, +3.8%, +8.2%, -3.54% | Helped |
| Use optimal parallelism settings | LLMCarbon · training operational carbon footprint | operational carbon footprint: 16%∼39% | Helped |
| Use H100 GPUs instead of V100 GPUs | LLM processing · training, inference, and experimentation | operational carbon footprint: 71% | Helped |
| Use TPUv4 instead of V100 GPUs | LLM processing · training, inference, and experimentation | operational carbon footprint: 41% | Helped |
| Use a sparse MoE architecture rather than a dense architecture | LLM designs · training carbon footprint and test loss | test loss at the same training carbon footprint | Helped |
Surprising findings
- Embodied carbon can dominate the life-cycle footprint of LLMs when data centers run on renewable energy (92% to 95% in a 97% renewable Meta data center).
- mlco2 overestimates training carbon footprints by 69% to 132% because it assumes peak throughput and peak power.
Limitations
- MoE LLM predictions have larger errors due to architectural complexity and extra host-device memory swaps (about 80% of dense base efficiency).
- Several validation inputs (e.g., Noor architecture) are unavailable, and parameter model deviates for Gopher and Chinchilla due to unaccounted positional encoding weights.
- Validation relies on a small number of published carbon numbers and one embodied-carbon data point.
Experimental setup
Datasets: published LLM training carbon footprints (T5, GPT-3, GShard, Switch, XLM) published Meta XLM embodied carbon data Noor storage data published data center PUE and carbon intensity
Benchmarks: published operational carbon footprints of T5, GPT-3, GShard, Switch, XLM GPT-3 inference latency and energy Noor storage energy Meta XLM embodied carbon
Metrics: percent difference between predicted and actual tCO2eq latency error energy error
Baselines: mlco2
Builds on: mlco2 Chinchilla scaling law Megatron-LM parallelism
Closest papers by meaning
- CO2-Meter: A Comprehensive Carbon Footprint Estimator for LLMs on Edge Devices (AAAI 2026 · similarity 0.86)
- Holistically Evaluating the Environmental Impact of Creating Language Models (ICLR 2025 · similarity 0.86)
- Towards Green AI in Fine-tuning Large Language Models via Adaptive Backpropagation (ICLR 2024 · similarity 0.85)
- What can large language models do for sustainable food? (ICML 2025 · similarity 0.84)
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.84)
- LLMCBench: Benchmarking Large Language Model Compression for Efficient Deployment (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.