Developing of models of dynamically reconfigurable neural network accelerators based on homogeneous computing environments

Nowadays, machine learning algorithms are widely used in many intelligent systems. As a result, the development of high-performance, power-efficient, flexible and reliable computing devices is becoming a major challenge. This is especially important for autonomous and mobile systems that use several different machine learning algorithms at the same time. Reconfigurable hardware accelerators are one possible solution to the problem. A key feature of these accelerators is the ability to be dynamically configured by external configuration signal to implement a required at the moment neural network model. In this paper reconfigurable accelerators based on the concept of reconfigurable homogeneous computing environments are proposed. Benefits of the reconfigurability and the usage of reconfigurable environments are discussed. Possible models of a computing environment and its elements, simulation results and their comparison with a classical non-reconfigurable solution are presented.