A preprint measures edge LLM speed, token timing, hardware use, power and energy across Jetson generations, backends and formats.