Qualcomm has a vested interest in smarter mobile devices: it wants to power even more of them with its own Snapdragon chips. To that end, today sees the launch of the Qualcomm Snapdragon Neural Processing Engine an SDK for the Snapdragon 820 chipset that promises a simpler path to deep-learning software and artificial intelligence on phones, tablets, wearables, and even in cars.
It’s all powered by Qualcomm Zeroth Machine Intelligence Platform, and while Zeroth may sound like an enemy for the Power Rangers to fight, it’s actually a way to leverage formally cloud-trapped processing entirely on a local device.
Traditional machine learning – the ability to assess the world and categorize elements within it – generally requires masses of processing grunt and a big database against which to check the potential results. What Zeroth does is condense that down into a single chipset, taking advantage of the various components of Snapdragon like the Hexagon DSP, Adreno graphics, and more.
Qualcomm had been offering Zeroth technologies before – its recent chips could do things like automatically recognize scenes to switch between certain camera settings, for instance – but according to Gary Brotman, director of product management at Qualcomm, the demands from OEMs turned out to be much broader.
In fact, Brotman explained to me, phone makers and other manufacturers saw much more expansive possibilities for the system, and wanted to use it not just with Qualcomm’s potted collection of skills, but with their own proprietary data sets. In effect, each will be training its own, specialist neural network using Qualcomm hardware and the Zeroth SDK.
There are advantages in privacy, too, since the data can stay on your device rather than being transmitted to the cloud.
The potential applications are broad-reaching. Imaging is an obvious one, but not just for smartphone and point-and-shoot connected cameras: there’s interest from the security sector, for instance, which could use Zeroth technology to automatically identify threats and flag potential crimes.
Automotive is likely to be another significant area, as cars gain the ability to spot pedestrians and other traffic, identify road signs, track lane markings, and eventually transition to semi-autonomous and fully autonomous driving. There, speed of processing is of the essence: the lag introduced between the vehicle’s onboard cameras sending data to the cloud, it being processed there, and the results fed back could mean the difference between a crash and a safe lane-change.
To that end, Qualcomm announced the Snapdragon 820A, the automotive-specific variant of the 820, earlier this year, and that too supports Zeroth. NAUTO, a connected camera company which uses the technology for monitoring safe driving to cut insurance costs, as well as developing autonomous vehicles, has already signed up to use Snapdragon and Zeroth to power its new cameras.
The Zeroth SDK launches with support for scene detection, text recognition, object tracking and avoidance, gesturing, face recognition, and natural language processing. Further skills are likely to follow on in due course.
Initially, meanwhile, Qualcomm expects to work primarily with device makers: after all, they’re the ones buying Snapdragon chips for their devices. Along the way, however, it plans to pick up a variety of developers on the Zeroth journey, as the potential for locally-processed machine learning becomes more clear.
The Snapdragon Neural Processing Engine SDK will be released in the second half of 2016.