Friday, April 3, 2015

Robotics & AI - Silicon, Part 1


Your task is not to foresee the future, but to enable it. - Antoine de Saint Exupery

In previous installments on this topic, we've focused on what some of the leading systems and software houses were doing in AI. There are some notable differences in approach between them - for instance, Apple's emphasis on the individuality and autonomy of the smartphone user in contrast to others treating the mobile phone as almost a dumb terminal that sources functionality from a server farm. One can also see how competition between these industry hopefuls is already providing evidence of who are the early winners or losers. For example, it is obvious that Microsoft is outperforming Google in almost every aspect of AI research, despite their nearly identical approaches (which parallels Microsoft's growing leadership over Google in other areas as well.)

Yet the companies that source the basic computing power to these systems houses - the chip firms - are, for a change, not simply following in the footsteps of their system level customers. They are, in many respects, equal to them in knowledge and seem to even be pulling ahead in certain areas. In this new editorial series, we'll look at a selection of them to understand in what direction the chip companies are going.

Qualcomm Zeroth

While still clearly the dominant player in both baseband and applications processors for mobile telephony, Qualcomm has demonstrated its strategic acumen over the last year by pouring resources into positioning itself for an eventual breakout of the IoT market. The company has a very clear conception of how pervasive the IoT will be, with smart homes, smart cities, smart cars, smart medical facilities, wearables, robotics, etc.. penetrating nearly every facet of our personal and professional lives. 

Qualcomm, in fact, seems to be counting on the IoT becoming an endemic feature of modern life. What is especially impressive is that Qualcomm understands that something akin to a converged tablet and smartphone will become the ubiquitous portable processor that will be the individual user's centerpiece for interacting with the IoT.

The Qualcomm AI team has a well defined vision of what that personal processor should be able to support and exactly how it should perform when employing AI functionality. How it breaks down this list of functions is itself quite revealing. Per the company's Zeroth group, a mobile-based AI needs to deliver the following capabilities:
1. Visual perception
2. Intelligent connectivity
3. Intuitive security
4. Always-on awareness
5. Immersive multimedia
6. Speech and audio recognition
7. Natural interaction

Putting it more simply, Qualcomm wants machines that perceive, reason and act intuitively. They would not, however, be autonomous. A Qualcomm AI would be smartphone/personal processor-based, operating as an extension of human senses to increase user awareness of their surroundings, anticipate needs thru understanding individual preferences and provide timely additional information that is contextually relevant (which involves an intuitive understanding of the immediate user environment as absorbed and interpreted by device sensory input.)

The AI work at Qualcomm encompasses all the usual tasks that are being researched elsewhere, including Machine Vision, Image Recognition and Voice Activation. But there are some other duties Qualcomm expects an AI to perform on behalf of a user, such as:
1. Intelligent Connectivity - the ability to sense networks with different bandwidths and protocols and choose among them with considerations for battery life and performance
2. Always-On Sensing - staying awake and aware of a user's surroundings while the user is potentially distracted by their own thoughts and concerns.
3. Intuitive Security - sensing when a threat is present and acting on it without user intervention being required.


Back in my junior high school days in northern New Jersey, some of the most memorable moments I can recall occurred in Shop class. Instruction included the basics of metalworking and carpentry. The teachers were grouchy but well-meaning fellows who left an impression on me from their sparse yet effective manner of speech. An oft-repeated phrase which stuck with me was their definition of a tool - "an extension of the hand." In their view, a tool was only so good as it made its existence unnoticeable to the user as they employed it.

This is evidently how Qualcomm sees AI as well. They are pursuing the creation of a machine equivalent to the human mind, but intend it to be completely subservient to man's needs - to serve, in effect, as an extension and enhancement of innate human capabilities.

The possibilities are indeed intriguing. Think of your Qualcomm AI-equipped mobile device serving as a universal translator while you travel, translating the spoken and written word for you - whether you are reading a road sign in Hong Kong, chatting with a fruit & vegetable vendor in Bologna or watching TV in Sao Paolo. A blind person could get audio alerts to potholes, curbs and other hazards, while a deaf person could receive visual warnings of approaching traffic.

Security is a primary consideration for Qualcomm's AI effort. They clearly believe that an individual user has a basic right to privacy. Besides the hardware-based biometric security features the company is developing, they want an AI to intuitively grasp when the user may be under either physical or electronic threat and act accordingly.

The richness of Qualcomm's IP portfolio is paying off for the company in its new IoT and AI pursuits. Snapdragon has evolved into a heterogeneous processing architecture, sporting multicore CPUs either directly sourced from or based on ARM, as well as home-grown GPUs, DSPs and hardware accelerators/offload engines for multimedia. It is a premier SoC architecture that can handle dynamic workloads and process data with industry leading performance and power efficiency. The Hexagon DSP, originally targeted at audio, video and imaging applications is now being tasked with text and image recognition (including facial recognition) and image enhancement. The Adreno GPU is now also being employed for machine vision.

In a time of drastic change it is the learners who inherit the future. The learned usually find themselves equipped to live in a world that no longer exists. - Erif Hoffer

Qualcomm is, as is to be expected, acutely aware of the severe and contradictory physical constraints of mobile computing - the need for a small form factor, minimal power to extend battery life and reliably operate with a very limited airflow, and yet requiring cutting edge computing capability for communications and multimedia. They are probably at least as good and arguably better than Apple in anticipating the enormous increase in computational demands from artificial neural networks for AI and the implications to mobile computing hardware. In preparation for those future computing burdens, Qualcomm seems to be ahead of everyone regarding methods to sustain Moore's Law beyond 28nm, as they are at the absolute bleeding edge of monolithic 3D-IC, as can be seen here:


Above all, Qualcomm seems to be much more realistic than the systems houses who, in their attempts to merely emulate the human brain, are struggling to develop linear approximations of what we know to be a very non-linear system. That very non-linearity is a processing problem that the company appears ready to tackle head-on, as they evidently have no fear of trying to learn new things. 

It's safe to assume that Qualcomm will maintain its pre-eminent position in the semiconductor industry by its obvious preparedness to lead in the emerging markets of AI, Robotics and the IoT. This does not mean, however, that there won't be challengers. We'll look at a few of them in next week's editorial.
Special Note to My Readers:
This editorial was at first intended to be the closing chapter on the "I, Robot" series. However, my research into semiconductor companies on this topic is revealing a level of sophistication on their part which surpasses that of the systems houses, as seen above. It was enough to provoke me into starting a new series on AI and Robotics focused solely on the chip sector.
This series will continue for at least one and possibly two more installments. After that, we will revisit the "State of the Union" financial series and update it with Q1 2015 numbers, as the "silly season" of financial reporting is nearly upon us again.

Dear readers,
Don't forget the Amazon and FlipKart banners! Many thanks to those who already have clicked thru them to do their shopping, and I hope more will join in to help keep this blog going. :-)
p.s. If any of you would like those banners to hilite particular products or categories, please let me know.