Edge AI for Critical Real Time System Control
Faster Smarter
Control systems are often limited by their inability to measure directly at the point of interest. Alternatively, some control situations have too many unknown dynamics and no fundamental structure on which to develop a control model. We then must resort to numerical analysis methods, fluid dynamics computed on a numerical grid would be an instance of such an application. The matrices derived from this numerical application have to be resolved and interpreted using a learned intelligence to derive actionable data.
Real Time Analytics
Engaging this level of numerical processing has been done in the past, but it is restricted by long computation times or large-scale computing platforms. Real time control on the other hand requires millisecond response latency and numerical analysis use cases typically produce massive amounts of data. When most of the data processing is done locally, on the edge, centralized service or data transfer will not become a bottleneck. If you have to make a decision on the result of your numerical analysis representation (offen a large-scale array) thousands of times per second, transferring the data to a cloud service is not a viable solution. This implies that suitable hardware and electronics devices must be designed to make independent decisions faster and more safely without having to connect to the Internet or a high bandwidth server.
Efficiency Becomes the Solution
Operating at the edge complex numerical algorithm brings about the following struggles, you have to compute with short latency (faster), using less power, in less space. It is easy to miss the importance of power efficiency in this effort. Indeed, if power is reduced for a finite amount of effective compute, power supply needs are reduced, heat dissipation and accordingly size is reduced. This enables compute power to be further increased. This self-reinforcing cycle is achieved by focusing on processors tailored to the very specific type of use task. It is additionally achieved by focusing on the efficiency of data manipulation. Every data movement cost energy and both the hardware and software work to simplify, minimize and shorten data transfers.