Quickstart¶
Verify Setup¶
After installation, confirm telemetry is working:
If the energy monitor reports readings, your platform's collector is working. If it fails, see Platform Support.
Single-Turn Profiling¶
Start your inference server, then profile:
ollama serve && ollama pull llama3.2:1b
ipw profile \
--client ollama \
--model llama3.2:1b \
--client-base-url http://localhost:11434 \
--max-queries 50
Analyze and plot:
ipw analyze ./runs/profile_*
ipw analyze ./runs/profile_* --analysis regression
ipw plot ./runs/profile_*
Agentic Profiling¶
Install an agent extra and run a multi-turn workload:
uv pip install -e 'intelligence-per-watt[react]'
export OPENAI_API_KEY=sk-...
ipw run \
--agent react \
--model gpt-4o \
--dataset gaia \
--max-queries 10
Analyze and plot:
Listing Components¶
ipw list all # Everything
ipw list clients # ollama, vllm, openai
ipw list datasets # ipw, mmlu-pro, supergpqa, gaia, frames, ...
ipw list analyses # accuracy, regression
ipw list visualizations # regression, output-kde
Choosing a Dataset¶
ipw profile --dataset ipw ... # General knowledge (built-in)
ipw profile --dataset mmlu-pro ... # Multiple-choice academic
ipw run --dataset frames ... # Multi-hop reasoning (agentic)
ipw run --agent terminus --dataset terminalbench ... # Terminal tasks
ipw run --agent openhands --dataset swebench ... # Software engineering
Next Steps¶
- Profiling Guide -- all
ipw profileoptions - Datasets Overview -- benchmark details
- Cost Tracking -- cloud API cost estimation
- Platform Support -- hardware setup