다중 신경망 가속기의 설계 최적점 탐색

Abstract

The huge design space of NPU dataflows of neural networks has led to the development of several exploration tools for the dataflows. However, these tools is limited by its ability to find the optimal design point for the multi neural network applications due to its lack of expressiveness in the syntax of dataflows. Therefore, we suggest a Design Space Exploration tool for multi neural network applications with a novel syntax to express a multi neural network dataflow. The mappability of various networks of this work comes from extracting the hardware model solely and abstracting algorithm dimensions, which ensures the consistent hardware model without any conflicts between diverse neural networks. It is expected to help design the accelerators for target applications of complex and heterogeneous structures of neural networks.