Intel has made a highly strategic acquisition, greatly enhancing its position in the race to dominate vision processing chips, a key element in the hot area of mobile VR (virtual reality) driven by AI (artificial intelligence).
It has snapped up Google’s close partner Movidius, which has developed an ultra-low power vision processor which can be incorporated into small devices. This will strengthen some key Intel developments such as its RealSense depth sensing technology, which it has managed to incorporate into mobile devices and drones.
In turn, this will boost Intel’s hand against arch-rival Qualcomm, which is also highly active in this field, and should deepen its ties to Google. All the chip giants, also including Nvidia, are chasing key markets driven by vision processing, including drones and new ultra-context aware mobile user experiences, and potentially autonomous vehicles.
More broadly, this will help Intel to “take the lead in providing computer vision and deep learning solutions, from the device to the cloud,” as Josh Walden, Intel’s SVP for New Technologies said in a blog post about the acquisition. To achieve the power and revenues it needs, devices are not enough. Intel needs its processors to be embedded from end-to-end as networks are deployed based on new architectures – cloud computing, massive connectivity at the network edge for the Internet of Things – to enable new mobile and IoT experiences.
Movidius joins a string of start-ups Intel has acquired in the computer vision and AI fields. They include RealSense itself; deep learning chip provider Nervana Systems; Saffron, which has developed a cognitive computing platform; and Ascending Technologies, a German drone company with sense-and-avoid algorithms.
Movidius, founded in 2006, will now be part of Walden’s New Technologies Group along with its core product, its vision processing unit (VPU), which it describes as “giving sight to smart devices including drones, security cameras, AR/VR headsets and more”.
It will take on rival initiatives such as Qualcomm’s Snapdragon Flight platform for drones and Zeroth mobile ‘brain chip’, which is heavily focused on machine vision. Jim McGregor, principal analyst at TIRIAS Research, told EETimes that the embedded market is becoming “more competitive, especially with the entrance of the mobile chip vendors, who are accustomed to providing complete reference designers for applications”.
The autonomous vehicle market is particularly tough, given the dominance of Mobileye in automotive machine vision for ADAS (advanced driver assistance systems) and future driverless cars. Some analysts believe Intel will seek to adapt Movidius’s technology to take on Mobileye in this potentially important market, though the chip giant already has a partnership with Mobileye and German carmaker BMW – and Movidius itself has been cautious about the car space.
With or without driverless cars, Movidius is a good buy for Intel. It has been emerging as a star of the VR trend, thanks to its close ties to Google. In June, it unveiled the Fathom Neural Compute Stick, a USB stick housing its Myriad 2 chip in a small power package, which powers the computer vision capabilities of the DJI Phantom 4 drone.
At the same time, Movidius announced a deal with Lenovo, which could propel it into the mass consumer space, while it had already signed its agreement with Google in January. This will see the smaller firm’s chips used in Android devices and they have already been part of Google’s Project Tango VR initiative. While smartphones are likely to form the bulk of those, Fathom’s architecture is compatible with Google’s TensorFlow deep learning technology – which Google recently opened up to developers, and so could drive all kinds of form factor innovation.
However, smartphones remain at the heart of Android’s power base and are the best chance to get VR into the true mainstream. Given Lenovo unit Motorola’s reputation as a very ‘pure’ form of Android, and therefore well-liked by Google, it’s not a great leap to think that Motorola might be the first brand to launch Android smartphones using the Myriad 2 vision processing units (VPUs). Lenovo has already, of course, been the first to deliver a mass market commercial Tango smartphone and its close alliance with Google will only get stronger as the pair explore new markets for VR.
Movidius said at the time that the new VPU “features a flexible configuration of 12 high performance programmable vector cores, allowing Lenovo to implement custom algorithms to meet the needs of a wide range of products. With a built-in image signal processor (ISP) and hardware accelerators, the Myriad 2 offloads all vision-related tasks from a device’s CPU and GPU, drastically improving performance while only consuming a single watt of additional power.”
The Fathom stick is an interesting form factor, in that it aims to provide plug-and-play machine learning capabilities to devices in an as-needed capacity. The image recognition functions could be brought to general purpose computing equipment, to quickly augment its abilities with a VR or AR enabling hardware module – bridged to the computer using software.
So while this could include drones, mobile devices like laptops, tablets, and smartphones are also prime candidates for adoption, allowing developers to quickly create hardware adapters that can be used to map a 3D environment, and carry out image processing too.
Providing robots with intuitive real world navigation is a compelling application, as they will not be able to rely on a high speed, low latency connection to a cloud computing instance at all times – and the data plan cost of moving their raw image inputs up to the cloud for processing would begin to get exorbitant rather quickly.
But at a commercial scale, under Intel’s watch, the Fathom chips could also find themselves integrated into all manner of IoT devices. Network edge gateways are a notable leading candidate, but smart home devices that need to identify things within a house are a consumer application that would look to on-board the chips themselves, rather than rely on the USB form factor.
All these partnerships will be attractive to Intel, which has acquired a platform with real market presence rather than a pure R&D play. Movidius will also support its broader goal of driving VR and AI platforms in its own direction, and away from the GPU-focused approach of key graphics chip player Nvidia.
Nvidia has stolen a march on Intel in high end GPUs, threatening the larger firm’s incumbency in supercomputers. If the move away from GPUs and towards specialized processors, for AI, accelerates, it will benefit Intel.
This was the thinking behind the recent acquisition of Nervana – it has already mounted a credible challenge to Nvidia’s CUDA software, which powers its GPUs, with its own Neon cloud service (which is CUDA-compatible). And it has its own deep learning accelerator chip on the drawing board, and set to launch next year.
Both these offerings could gain far greater power with Intel behind them, and could enable the US giant to offer AI accelerate boards to challenge Nvidia’s GPU boards, while enhancing Neon to outperform CUDA too. Intel will incorporate Nervana’s algorithms into its Math Kernel Library for integration with its industry standard frameworks, and it can add Nvidia-compatible deep learning to its portfolio of cloud services.
The upcoming Nervana chip, called Engine, promises an 8Tbps multichip module with terabytes of 3D memory surrounding a 3D fabric of connected neurons, each using low precision floating point units (FPUs). This design enables it to conduct far more deep learning calculations-per-second on a smaller chip than general purpose GPUs, says the firm.
Karl Freund, senior analyst for at Moor Insights & Strategy, told EETimes: “GPUs are the prominent way to train deep learning neural networks and Nvidia is the leader. Intel has its multicore Xeons and Xeon Phi and with the acquisition of Altera, its FPGAs, but it doesn’t have GPUs. Acquiring Nervana is a way of getting into the deep learning market not by copying the general purpose GPU strategy, but by offering a specialized coprocessor specifically designed for neural networks.”
With specialized deep learning chips being deployed from supercomputers to cars to handsets, the competition will only heat up, and Nvidia may respond with its own offering. At the mobile end, Qualcomm has been highly active in this area with its Zeroth machine learning platform, which can be embedded in the Snapdragon SoC.
Initially, Qualcomm’s Zeroth/Snapdragon work focused on bringing AI algorithms and vision processing to cars, but many other applications are in its sights, as highlighted by its recently announced deep learning software developers’ kit (SDK) for the Snapdragon 820 SoC.
Called the Snapdragon Neural Processing Engine, the SDK is an attempt by Qualcomm to expand beyond its core smartphone markets – by enticing automotive, industrial, and IoT developers to its Snapdragon platform. It also sees the chip supplier looking for broader market reach, and accelerated uptake, by enabling a great deal of its brain-like functionality in software.