PORTLAND, Ore. — You've heard about the Movidius Myriad 1 vision processing unit (VPU) inside Google's Tango prototype -- a handset that, for one thing, can map 3D spaces just by walking through them. But what about this VPU at its heart, that Movidius Ltd. in San Mateo, Calif., claims is 10-times faster and bears little resemblance to the graphics processing units (GPU) with which we are all familiar. Well for one thing, it goes between the camera and the application processor, instead of between the application processor and the display like a GPU. But that is only the beginning of the story, since if you take Movidius at their word, the VPU is destined to become an essential new component to bring visual awareness to anything with a camera.
"We believe that cameras in general and mobile cameras in particular are going through a revolution which we call computational imaging which brings in new functionality," Movidius CEO Remi El-Ouazzane told EE Times. "What we are working on is additional processing units we call VPUs which essentially are trying to reproduce the functions that the visual cortex in the brain is doing -- which will bring to our devices the same kind of awareness and realism that the eye and brain bring to a human being."
Most graphics processing units (GPU) are merely bit-bangers -- vector processors that perform identical operations on every pixel on the screen at rip torn speeds. The VPU, on the other hand, interprets the data coming from the camera -- the way the eye and visual cortex do -- sending to the application processor high-level metadata, rather than raw pixels, identifying where objects begin and end, which ones are in front of each other, what kind of object each is, where its shadow is, what trajectory it is following, and dozens of other smarty pants information that the application processor can use to make its job not only markedly easier, but which enables nuevo applications that were never possible before.
Movidius Myriad silicon platform makes massive use of parallelism to do the kinds of things that the human visual cortex does.
"Our methodology is to the convert the photons the sensor captures into metadata which expresses and understanding of the scene," El-Ouazzane told EE Times. "The metadata can be used in many different ways depending on the application, but in the long-run what we want to provide is total visual awareness of the relevant details in a scene."
Next Page: Algorithms and Movidius' bold claim