Intro You rented a GPU instance, kicked off a training or inference job, and now you're staring at nvidia-smi. The utilization number bounces between