Designing high performance, power-efficient, reconfigurable compute structures for specialized applications

Abstract

Abstract In this paper, a new approach to design high-performance and power-efficient computing structures are proposed for machine learning tasks. Such structures can be very useful in some specialized applications such as autonomous robots, mobile devices, smart sensors for the Internet of Things (IoT) and so on. This approach is based on the concept of reconfigurable homogeneous computing environments. Major advantages of this approach are discussed. The process of designing a set of elementary operations for such structures is described in detail using an example with a typical Feed-Forward Neural Network (FFNN) and its training module.