Intel snaps up Movidius to accelerate its AI and RealSense vision efforts

[XAMI]

Custom silicon vendor Movidius has attracted a lot of attention for its high-performance, low-power chips that have powered vision applications like Google Tango, as well as making machine learning possible on mobile devices. Now it has received the ultimate compliment. Chip giant Intel has acquired it to help accelerate its RealSense project and other efforts to provide computer vision and deep learning solutions. Intel is expecting to see Movidius technology deployed in drones, robots, and VR headsets — in addition to more traditional mobile devices such as smartphones and tablets.

The Movidius advantage

Power requirements are the traditional Achilles heel of mobile solutions that require substantial computation, with vision and machine learning being two of the most extreme cases. By creating optimized, custom silicon — its Myriad chip family — Movidius has reduced the power needed to run machine learning and vision libraries by well over an order of magnitude compared to a more-general-purpose GPU.

RealSense

After a lot of initial excitement, Intel’s first-generation RealSense products — designed to provide devices with a 3D view of their surroundings to support mapping, navigation, and gesture recognition — faltered due to technical shortcomings. However, Intel has more than re-doubled its efforts, and is aiming to make RealSense the eyes and ears of the Internet of Things, which Intel believes will comprise over 50 billion devices by 2020. Intel Senior VP Josh Walden likens vision processors such as Movidius’s Myriad to the “visual cortex” of IoT devices.

Intel taking aim at Nvidia’s GPU strategy

This move takes Intel further into Nvidia’s home turf. Nvidia has bet big on high-performance computing for AI, self-driving cars, vision, and VR — the exact markets Intel is trying to move into with its RealSense platform, and now the Movidius acquisition. This pits Nvidia’s strategy of providing the most possible general computing power per watt versus Intel’s custom silicon.

On paper, the advantages of each are fairly straightforward. General purpose GPU (GPGPU) computing provides the most flexibility and adaptability, while custom silicon can be more efficient when running a specific task or library — once it has been developed. In the market, expect to see plenty of design wins for both Intel and Nvidia, and some leapfrogging of each other as subsequent product generations roll out from each.