Accelerating Computer Vision with VPUs

April 22, 2020

Sight is a sense humans often take for granted. We can identify faces and objects in a photograph with little thought. Teaching a machine to replicate human sight with the same speed and accuracy is a much more difficult task.

What is Computer Vision?

Computer vision is how a machine uses algorithms to analyze visual information such as video for patterns, identification and other information. Consumers might recognize computer vision in how their digital camera automatically focuses on human faces, or how the display of their smartphone lights up when they reach for it. Computer vision has industrial applications like creating heat maps for environmental monitoring, identifying product issues for quality assurance, object detection and tracking for surveillance. It also enables gesture tracking for virtual and augmented reality and obstacle detection for robot guidance.

What is a Vision Processing Unit (VPU)?

A VPU is a microprocessor designed to accelerate machine vision algorithms pertaining to learning and analyzing. Vision processing units are different from video processing units, which handle video encoding and decoding. A VPU can operate independently or as a coprocessor to a central processing unit (CPU). They consume a fraction of the power required by GPUs and have low thermal generation, which makes them ideal for edge computing.

Edge computing is meant to locally store and process data, saving bandwidth and improving latency. However, compared to the cloud or data center, edge computing has slower inference power. Accelerators, such as a VPU, help speed up learning models and analysis. Their compact structure makes VPUs ideal for use in small devices, such as cameras, phones and USB sticks.

Current and Future VPU Technology

In 2017, Intel announced the world’s first VPU, the Myriad 2. The product had applications for drones, smart cameras, security systems, VR & AR programs and other products. The upgraded version, released in 2018, the Myriad X, is a 16nm device that includes vision accelerators, Neural Compute Engine, imaging accelerators, 16 SHAVE vector processors and CPU. It supports up to 16 video streams per device.

Keem Bay is Intel’s next VPU, expected to launch sometime this year, at reportedly 10 times the AI performance of the Myriad X. But since 2017, many other companies have joined the VPU market, including Samsung and Google. The vision processing unit market is expected to grow to 3.2 billion by 2024.