Movidius unveils $100 USB neural network co-processor

April 29, 2016 // 10:43 a.m.

Tags: #accelerator #co-processor #movidius #movidius-fathom #neural-network #visual-processing-unit #vpu

Movidius, the company responsible for the co-processor inside Google's Project Tango, is looking to bring its technology to a wider market with the announcement of a $100 USB-connected accelerator for neural computation.

Based on the company's Myriad 2 co-processor and backed by a dedicated 512MB of LPDDR3 memory, the Movidius Fathom is designed with a singular task in mind: accelerating the performance of neural networks without vastly increasing the heat output and power draw of the host system. With the company's previous chip having found a home in a battery-powered mobile device, it's no surprise to hear that Movidius is targeting a very small power envelope: driven entirely from a system's USB port, the Fathom accelerator is claimed to offer 150 gigaflops - 150 billion floating-point operations per second - while drawing a mere 1.2W.

Designed primarily for computer vision tasks - Movidius positions the co-processor as a Visual Processing Unit or VPU - the device is claimed to boost neural computing computation by around 20 to 30 times, without the increase in bulk, power requirements, and heat output associated with using GPU-based accelerators. Better still, the price is certainly right: although Movidius is currently shipping the Fathom only to selected partners, the company has confirmed that when the device becomes generally available it will cost a mere $100 (around £69 plus taxes).

Movidius has yet to set a formal date for a public launch of the Fathom stick.
Discuss this in the forums

QUICK COMMENT

SUBSCRIBE TO OUR NEWSLETTER

WEEK IN REVIEW

TOP STORIES

SUGGESTED FOR YOU