Realizing the Potential of AI on the Edge.

• Author: Chew, James S.B.
• Position: Viewpoint

* In the late 1970s, the outlook for the internal combustion engine was bleak. Faced with complying to both aggressive fuel economy standards and stringent emission regulations, the American driver seemed destined for a future of small, underpowered cars.

But something funny happened. Engineers and researchers began to study the technological state of those engines. They discovered that while those engines performed adequately, the technology had not progressed much since the development of the Otto cycle. These engines were designed to perform optimally at one set of operating conditions--open road, nonstop driving. Internal combustion research-and-development efforts were executed to develop technologies and components, which would allow the internal combustion engine to operate optimally at all common operating conditions.

These efforts followed a systems approach to re-examine the basic engineering principles to understand and optimize subsystems, then optimize subsystem to subsystem to system interactions. Because of this system-of-systems approach, today we have four-cylinder engines that produce more horsepower and torque than the 1960s-era muscle car engines while achieving average fuel efficiencies of over 30 miles to the gallon. We have V-8 engines that produce horsepower and torque once reserved only for special-purpose racing vehicles--all while achieving fuel efficiencies that are far better than the 1980s-era sub-compact cars.

The same system-of-systems optimization approach must be followed to fully realize the potential of artificial intelligence. With the advent of AI at the edge and the associated need to make intelligent decisions in an untethered mobile environment, upwards of three Tera-MACs per watt are needed. To achieve that kind of performance, industry must examine new AI hardware architectures and software solutions that are tailored to the broad range of missions across the military and aerospace landscape. To create these new solutions, new design methodologies, architectures and innovation in the development and verification of AI-related hardware and software is required that go well beyond the frameworks of today.

Today, continued advancements in microprocessor technology and the availability of big data provide a natural foundation that fuels the interest in AI-enabled devices, as well as machine learning-enhanced design and verification processes. Most of the Defense Department's artificial intelligence...