​Reducing the energy consumption of communicating devices is a top priority for manufacturing companies. And yet, energy efficiency can be difficult to deliver when integrating certain complex functions—especially those powered by artificial intelligence. CEA Tech institutes Leti and List have built the first-ever demonstrator system that combines a very-low-power platform with a neural network accelerator. Even better, the demonstrator can support end-use scenarios representative of what the market is currently looking for.

The system is built on a L-IoT (low-power IoT) circuit developed by Leti and an architecture that features a low-power wake-up and processing platform, plus an on-demand processing platform that is activated only when complex processing is required. List's PNeuro neural network accelerator is integrated into the on-demand platform to provide the system with parallel processing optimized for artificial intelligence.

The demonstrator, called SamurAI, was used to complete local neural processing to detect, via an infrared camera, human presence and count the number of people at the scale of a room. The goal was to illustrate the system's low energy consumption in sleep mode and its capacity to perform complex processing on demand and to activate this processing very rapidly and with a high degree of flexibility.