gpu2grid/live

Simulates how AI datacenter workloads affect electrical grid stability.

GPU power traces from real LLM inference and training runs are replayed through an IEEE 13-bus distribution feeder solved via OpenDSS, using the openg2g simulation framework. The simulator shows how different models, batch sizes, cluster sizes, and connection points cause voltage violations across the grid.

Try the live simulator at

---

What it does

• Injects real H100 GPU power traces (ML.ENERGY Benchmark v3) into an IEEE 13-bus feeder
• Solves power flow at each 0.1s timestep using OpenDSS via the openg2g framework
• Visualizes per-bus voltage over time, violation rates, and GPU power draw
• Supports multiple LLM models (Llama 3.1 8B/70B/405B, Qwen3), batch sizes, and replica counts

Quick Start

# Backend
cd backend
uv sync
python server.py

# Frontend (in a separate terminal)
cd frontend
npm install
npm run dev

The frontend runs at http://localhost:5173 and expects the backend at http://localhost:8080.

For local development, create frontend/.env.local:

VITE_API_URL=http://localhost:8080

Data

GPU power traces are from the ML.ENERGY Benchmark v3 (H100 measurements). On first run, traces are downloaded and cached under data/offline/.

python examples/offline/run_ofo.py --system ieee13 --mode all

Stack

• Backend — FastAPI + OpenDSS via openg2g
• Frontend — React + Recharts
• Grid model — IEEE 13-bus distribution feeder
• GPU traces — ML.ENERGY Benchmark v3 (H100)