Obsessed with technology?
Subscribe to the latest tech news as well as exciting promotions from us and our partners!
By subscribing, you indicate that you have read & understood the SPH's Privacy Policy and PDPA Statement.
News Categories

Computer vision company Movidius just put a deep learning chip inside a USB drive

By Koh Wanzi - on 29 Apr 2016, 3:00pm

Computer vision company Movidius just put a deep learning chip inside a USB drive

Image Source: Movidius

Movidius, the company behind Google’s Project Tango, has announced the Fathom Neural Compute Stick, a thumb drive-sized device that includes its very own deep learning chip. The stick PC resembles products from Intel and ASUS, but it’s actually designed for vision processing and things like advanced object detection and providing the right context for images.

The chip inside the device, which is called the Myriad 2, is actually the same one that gives the new DJI Phantom 4 drone its computer vision and autonomous features. It’s also used in FLIR’s Boson thermal camera module, which can give home security systems, drones, and military equipment smarter and more powerful thermal imaging capabilities.

The Fathom houses the Myriad 2 MA2450 visual processing unit (VPU), along with 512MB of LPDDR3 RAM, and is essentially the first of its kind as a discrete – and truly portable – deep learning accelerator. As Cormac Brick, head of machine learning at Movidius told The Verge, the benefit of this design is that it allows people to implement machine learning in an ad hoc manner.

And because the Fathom plugs into any device with a USB 3.0 port, adoption should be simple as developers don’t need to redesign existing products or build any new prototypes. It is intended to be paired with Linux systems, and can enable developers to quickly create prototype neural networks and take advantage of a 20 to 30x boost in neural compute performance.

Movidius’s device is compatible with neural networks defined in Caffe and Google’s TensorFlow, but you’ll need to use the company’s own tool to execute the network on the Myriad 2 chip, where it will run natively while consuming a mere 1.2 watts of power.

Ultimately, the Fathom makes deep learning more accessible, as you won’t need any ultra-high-end graphics cards and processors to run it. It is common for people to leverage the power of a cloud-based system in order to prototype their own neural nets, but the Fathom’s ability to add a fair amount of computing power to a regular laptop – up to 150 gigaFLOPS actually – could simplify the process and reduce costs.

Furthermore, any robot, drone, or some other device could now easily benefit from the Myriad 2’s ability to process images in real-time. For instance, a Fathom could be hooked up to a Raspberry Pi, which would allow you to add advanced computer vision capabilities to a GoPro.

There’s no news yet on a price for the Fathom, but Movidius is reportedly planning to sell it for under US$100.

Source: Movidius

Our articles may contain affiliate links. If you buy through these links, we may earn a small commission.