Estimate energy use, cost, and CO₂ emissions of large-language-model inference in real time.
Generative AI is amazing—but it comes with an energy bill. This tool makes the environmental and economic impact of LLM inference transparent & actionable for engineers, researchers, and product teams.
- 📊 Instant feedback – tweak model, precision, tokens, location, electricity price & more
- ⚙️ Research-backed formulae – FLOP analysis + region-specific grid intensity
- 🌐 Global perspective – compare 15+ grid regions from Iceland to Australia
- 💰 Cost breakdown – see dollars and kWh for any workload size
- 🍃 Carbon insights – quantify CO₂ per token and total session
"What gets measured gets managed." – Peter Drucker
pnpm install
pnpm dev # then visit http://localhost:3000- Total FLOPs (F = 2\times N_{params}\times T)
- Energy (kWh) (E = \dfrac{F}{\eta}\div 3.6\times 10^{6})
- PUE overhead (E_{total} = E \times \text{PUE})
- Carbon (CO₂ = E_{total} \times I_{grid})
Key assumptions:
- NVIDIA H100 efficiency: 6.59×10¹¹ FLOPs/J (conservative)
- Precision multipliers: FP32 (1×), FP16 (2×), FP8 (4×)
- Region-specific carbon intensity from latest open-data sets
Sources: Hopper et al. (2023), Özcan et al. (2023)
- Next.js 13 App Router
- React-Server Components
- TypeScript, Tailwind CSS, Shad-cn UI
- Vercel Analytics for usage insights
- GPU benchmarking integration
- Batch vs streaming inference toggle
- Export to CSV / JSON
- Your idea? Open an issue!
- Fork &
git clone - Create a feature branch
pnpm run lint && pnpm run test(coming soon)- PR away – we love contributions of all sizes!
If this project helps you understand or reduce AI energy usage, please star this repo and share with a friend. It fuels further development! 🙏
Made with ❤️ by @yourname – MIT license