Abstract
Edge computing is a recent paradigm of distributed computing that addresses the problem of massive data, particularly in the context of connected objects. These objects are becoming increasingly important in our lives. Examples include connected watches, smart homes, and connected cars. For reasons of reactivity due to network overload and energy efficiency, the processing of data generated by these objects has progressively moved from centralized infrastructures in the cloud to distributed systems integrating powerful servers and embedded systems that can be used as close as possible to the data sources. Today, the data processing is gradually integrating more and more artificial intelligence algorithms (typically for data analysis and decision making) in edge computing. To make this viable on embedded systems, it is important to study new architectures with sufficient performance and low power consumption.This thesis addresses the issue of embedded computing dedicated to edge computing. In particular, it focuses on the design of low-power architectures for processing machine learning algorithms. First, it explores a solution based on a heterogeneous multicore architecture in order to see to what extent it can meet a large algorithmic demand. This innovative architecture is based on the processor technology proposed by the French company Cortus S.A. Then, the thesis focuses on the acceleration of deep networks by proposing a new MAC (multiply-accumulate) unit that is both flexible and energy efficient. The gains provided by this MAC unit are evaluated through high-level modeling in convolutional neural network accelerators architectures. More generally, the work presented in this thesis offers interesting insights into the choice between general-purpose multicore architectures and dedicated artificial intelligence accelerators architectures for energy-efficient edge computing nodes.