Domain-Specific Design Abstraction with Emphasis on Neural Networks

The increasing complexity in digital system design has introduced significant challenges, particularly in maintaining productivity relative to the rapid advancement of technology. This paper addresses the issue by presenting a new level of abstraction, termed Data Transaction Level (DTL), which operates at a higher level than the traditional Register Transfer Level (RTL). The proposed DTL design methodology draws inspiration from RTL but is tailored for the growing complexity of modern systems, particularly in neural network implementations. As neural networks are becoming essential in various applications, this paper applies the DTL methodology to hardware implementations of these networks. A SystemC-based hardware description language is introduced to facilitate the design process. The results section evaluates the effectiveness of the proposed methodology, highlighting increased productivity using the new abstraction and description language.