Nvidia electrical engineer Ben Keller unveiled an always-on computer vision system has demonstrated a face detection system on a chip that can identify you in less than a millisecond.
The face detector sits in a system on chip pitched for robots, autonomous vehicles and laptops, and it saves juice by keeping data local and “racing to sleep” after detections.
Typical vision processing for this sort of job runs about 10 watts, which makes it largely useless for something that needs to run all day.
Nvidia says its SoC uses less than 5 milliwatts while pushing 60 frames per second, with a fresh image processed every 16.7 milliseconds.
Keller says the chip is fully powered for only five per cent of that cycle, then it drops back into its own nap routine.
Within 787 microseconds, the SoC calls a deep-learning accelerator to decide if a human face is present, with about 99 per cent accuracy.
The trick is brutal power gating, where most of the SoC is off by default, and only a small subsystem stays awake.
That always-on bit uses less than 10 mW and is called “Always-on Low-Power Accelerator,” or Alpha-Vision, which combines a deep learning accelerator, a small CPU, and compute parked close to storage.
Face recognition wants chunky model data, which usually means a steady drain on power and patience.
Nvidia dodges that by stuffing what it needs into local SRAM, with 2 megabytes set aside so it is not dragging data in from elsewhere.
SRAM leakage can become the real enemy in always-on designs, so the system quickly detects it and then shoves the memory into a low-power sleep state.
The researchers call that “race to sleep,” and one suggested use is a laptop that turns off its display when the user walks away, then lights it back up when they return, aiming for a smooth experience without the usual password faff.
Keller says the same design style could bring always-on vision to autonomous vehicles, drones and robots, which is great news if you fancy a future where everything notices your face before you notice it noticing.
