AluminatiAI – per-job GPU cost tracking (Nvidia-smi shows watts, not dollars)

1 point

3 months ago

With H100 clusters running $2–3/hr per GPU and training runs lasting hours or days, it's surprisingly hard to answer "what did that job actually cost?" nvidia-smi gives you instantaneous wattage. Your cloud bill gives you a monthly total. Neither tells you which experiment was the expensive one.

I built AluminatiAI to fill that gap. A lightweight Python agent (pynvml, ~0% CPU overhead, ~50MB RAM) samples power draw every 5 seconds and streams to a Next.js + Supabase dashboard. The dashboard converts watts → dollars per job, per GPU, per day. Works on A100, H100, RTX 3090/4090, and Google Colab.

Install is `pip install` + one env var. Under 2 minutes.

As H100 time gets more expensive and teams are running larger sweeps against frontier models, not knowing which run cost $800 is a real problem — not just a nice-to-have dashboard.

Open source: https://github.com/AgentMulder404/aluminatai-landing Website: https://www.aluminatiai.com

Happy to answer questions about the sampling approach or the watt-to-dollar conversion logic.