Many of the algorithms for performing these kind of analyses have already been explored, but required supercomputers consuming megaWatts to perform. Movidius' boldest claim is that it can equal or exceed the visual awareness of these applications while consuming mere watts of power -- and sometimes even less.
"Our novel micro-architecture starts with cores that are totally optimized for computational imaging in terms of how the delays between stages are structured plus an extremely innovative memory fabric which allows us to maximize the locality of the data and as such drastically reduce power consumption by minimizing the need for external memory accesses," El-Ouazzane told EE Times.
Movidius first generation VPU, Myriad 1 being used by Google's Tango project, was constructed to give its software team an opportunity to create a shared set of tools which are specific to its vector compiler -- essentially a set of advanced vision processing libraries and algorithms that define what Movidius calls a new computing paradigm. El-Ouazzane claims that the next generation VPU to be announced later this year, will be "awesome" regarding all the vision processing capabilities the company has amassed during its work on the first-generation VPU for Google's Tango.
Movidius Myriad system-on-chip (SoC), which it calls a vision processing unit (VPU), has multiple local memory storage bays, each with a different speed and bandwidth.
"There is a reason why computational imaging and visual awareness never took place in robotics, because of the power consumption it needed without a new type of micro-architecture. But the low power consumption of our VPU is opening the door of the kingdom," El-Ouazzane told EETimes. "Because there is a tremendous body of work in computer vision -- the way you do teaching, the way you do walking, the way you sense motion, the different types of classifying technologies all working at light speed -- a combination of computer vision, AI and deep learning."
Only a portion of Movidius's micro-architecture is hard-wired, the rest is programmable because the algorithms are new and changing at a fast pace as the body of knowledge already amassed by computer vision, AI and deep learning gets coded to work with the low-power local-memory-based VPU.
Besides Google, Movidius has two or three lead strategic engagements to bring computational imaging and visual awareness to smartphone devices in 2015. It declined to identify its other customers, but did admit that it is focusing on mobile markets, what it calls a "new category of wearables" and the driverless-car. Yet the closest thing it has to a released product using its VPU technology -- besides Google's Tango -- is one it designed for Centr Camera, originally a Kickstarter project that is now seeking traditional funding sources instead.
Movidius was founded in 2006 and has 65 employees, 60 of which are engineers and 90% of those are software engineers, which tells us how much the company is concentrating on computational imaging algorithms to do the heavy lifting rather than the hardware micro architecture which is unique enough on its own.
— R. Colin Johnson, Advanced Technology Editor, EE Times