> data limits

The data required is small. Each embedding might be 1/2 kB per face.

> power budget

To process a video for biometric feature extraction, it might take 0.5% to 2% of the total power used to record a video. Video uses a lot of power (compression, screen, etc)

Assuming you've got a modern device (e.g. with Apple Neutral Engine). Disclosure: Googled info (Gemini).