Estimating LLM Training FLOPs on the Nvidia Jetson Orin Nano

According to new research on LessWrong, verifying the compute used in frontier AI model training could become a key part of proposed international agreements. Right now, AI policy relies on companies' self-reported training numbers, with no independent check. A researcher working with the University of Chicago Existential Risks Laboratory is prototyping a way to estimate training compute through side-channel GPU readings—measuring power draw and memory usage patterns. Tested on Nvidia's Jetson Orin Nano, the approach could allow regulators to verify claims without exposing proprietary model code or data.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton