This tiny USB stick can bring image-based AI processing power to your computer

By Gary Elinoff, contributing writer

Intel just made it a lot easier for developers and researchers to quickly incorporate artificial intelligence (AI) into their designs. Intel’s Movidius Neural Compute Stick (NCS), available in the form of a USB 3.0 thumb drive, incorporates the Myriad 2 Visual processing Unit (VPU) to serve as an AI accelerator delivering deep neural network processing without the need for cloud computing. Intel is making the NCS available through select distributors for $79 to accelerate what it calls the “democratization” of deep-learning application development.

The drive’s core VPU, available from Intel, has already found industry-wide acceptance as a component of many commercially available machine-vision applications. It is able to perform a variety of computer vision tasks not limited to facial recognition and the detection of specific objects. But machine vision is only one of many applications for deep learning that the Myriad 2 can handle, despite the VPU designation. It’s the device’s processing power that gives it this flexibility.

“The Myriad 2 VPU housed inside the Movidius Neural Compute Stick provides powerful yet efficient performance — more than 100 gigaflops of performance within a 1-W power envelope — to run real-time deep neural networks directly from the device,” said VP of Movidius, Remi El-Ouazzane, as reported in an article on Engadget. It’s a surprisingly low power consumption, given that the Myriad 2 VPU itself incorporates 12 parallel cores. The units can also be paralleled to increase the available computing power.

Software for the device is centered around Caffe, an open-source deep-learning framework developed at UC Berkeley. Widely accepted in the AI community, Caffe has, at last count, been worked on by over 1,000 developers. There’s also a caffe-users group and a presence on Github as well. With the Movidius NCS, a Caffe-based neural network can be made to run on the Myriad 2 VPU, and it can be done completely offline.

In addition to Caffe, there has been some speculation about the possibility of the Compute Stick VPU also being able to work with Google’s TensorFlow framework.

The Movidius Neural Compute Stick will be a great help to AI developers as they endeavor to develop neural nets. With it, designers can quickly prototype AI and machine vision into their designs without the need for any troublesome cloud connection other than the time spent configuring the NCS at the start of the process. The stick itself can even be incorporated into deliverable, small-run products. But in most cases, it will be used to allow developers an easy way to train a neural network, while manufacturers will ultimately design just the Myriad 2 VPU chip, rather than the NCS, into their finished devices.

The Movidius Neural Compute Stick. Source: Movidius/Intel.

Movidius provides a quick guide to using the device — NCS Getting Started — and Intel has published a fact sheet describing its Myriad 2 VPU itself. As you might expect, you’ll need to work in Linux, Ubuntu 16.04 specifically, rather than in Windows.