Optimization of Energy Consumption in Embedded Systems by Selective Quantization of Convolutional Configurations

[eng] Sustainable power consumption is one of the biggest challenges in computing today. In particular, for embedded systems with machine learning capabilities in the Internet of Things (IoT) era, running convolutional neural network (CNN) models consumes a lot of power and data. Today, heavy workload processing is designed separately into a host processor responsible for general functions and a dedicated accelerator to perform the specific task. The design of accelerators to optimize computational resources for limited energy consumption is the main objective of this thesis. In particular, Int-16 quantization is analyzed not only as a possibility, but also as an element of study to determine a possible characterization of consumption as a function of dynamic quantization, and in parallel, the error generated by quantization and the corresponding processing is monitored.