Nvidia’s Always-On Chip Detects Faces in Less Than a Millisecond

Always-on vision systems might be used in autonomous vehicles, robotics, or to help consumer electronics save power by turning screens off when no one’s around. But to be used in such a way, these systems need to minimize their own power consumption.

An always-on computer vision system developed by Nvidia researchers can detect human faces in less than a millisecond. The face detector, which is part of a chip that could be integrated into robots, autonomous vehicles, or laptops, saves power by storing all data locally and “racing to sleep” after detections. Nvidia electrical engineer Ben Keller presented the system on 18 February at the IEEE International Solid State Circuits Conference in San Francisco.

Efficient Vision-Processing Technology

According to the researchers, this kind of vision processing typically requires about 10 watts. But that’s too much power to leave a face-detection system on continuously. The Nvidia system on chip (SoC) uses less than 5 milliwatts with a frame rate of 60 frames per second.

The system refreshes to process a new image every 16.7 milliseconds, and is fully powered on only for 5 percent of that time, says Keller. Within 787 microseconds, the SoC calls on a deep-learning accelerator to determine whether or not a human face is present, with about 99 percent accuracy.

The Nvidia team carefully designed the system to perform the detection rapidly and save power. Most parts of the SoC are powered off by default. A subsystem that uses less than 10 mW remains on. This subsystem is dubbed Always-on Low-Power Accelerator, or Alpha-Vision. It consists of a deep learning accelerator, a small CPU, and a subsystem to do certain computations physically near where data is stored.

Alpha-Vision uses a deep neural network to recognize faces, which requires a lot of data—in other words, a potential power drain. To save power and speed up detections, all the necessary data is stored locally in a relatively large area of SRAM adding up to 2 megabytes. To prevent SRAM leakage from dominating power usage, the face-recognition system rushes through its work and then quickly puts the SRAM into a low-power sleep mode. The researchers call this approach “race to sleep.”

The Nvidia team proposed several possible uses of such a system. A laptop computer integrating the face sensor could save energy by turning its display off when the user walks away, and then turning it back on when they return. The goal would be to provide a seamless experience—no need to type in a password. Keller says systems based on these designs might also be used to provide always-on vision for autonomous vehicles, drones, and robotics.

This post was corrected on 18 March to more accurately describe the state of the SRAM.