To alleviate these problems, the goal is to split the tasks of processing and analytics between the cloud and the edge, which would reduce the end-to-end latency to levels suitable for real-time applications at the edge and reduce the amount of data sent to the cloud. Most of the attention to this approach has focused on large-scale IoT applications such as industrial production facilities and “smart” cities, but it will be a major component of 5G as well, and for mostly the same reasons.
As it applies to 5G, most of the talk about AI (and its subsets machine learning and deep learning) focuses on network management and other high-level applications to reduce operating costs through precision network planning, capacity expansion forecasting, autonomous network optimization, dynamic cloud network resource scheduling, among others. However, it will eventually further expand its reach even to smartphones that today rely on the massive resources in the cloud. For this to occur the semiconductor industry will need to develop “on-device AI” realized by dedicated coprocessors or accelerators, a market that has just emerged and is growing rapidly with more than 40 start-up companies working on the problem, along with the usual cohort of deep-pocketed silicon vendors.
The need for AI at the edge is perhaps most obvious for the autonomous transportation environment, as when it arrives this application inherently requires decisions to be from data produced by sensors in a few milliseconds or even less. Latency this low can only be achieved over a very short distance, which effectively mandates placing intelligence locally, in the vehicles and the roadside infrastructure that supports them. As the technology used for intelligent transportation system communication is most likely to be the cellular industry through its “Cellular-Vehicle to Everything” (C-V2X) architecture, AI will become a fundamental element of AI at the edge in this application.