Ultra-NoC: Unified Low-Transmission Routing Assisted NoC for High-flexible DNN Accelerator

With the advancement of Deep Neural Network (DNN) accelerators in recent years, the efficiency of neural network computations has significantly improved. However, the varying layer's shapes and sizes in DNN models have posed challenges to the DNN accelerator design. This challenge increases the complexity of designing flexible and scalable accelerators. DNN computation requires substantial memory access and computational resources. Various dataflows have been proposed to address this demand to enhance data reuse and enable parallel computing. Nevertheless, existing DNN accelerators often face limitations in efficiently supporting diverse data transmission requirements across various dataflow simultaneously. These limitations stem from the flexibility of interconnection, which can restrict efficiency and data reuse opportunities, thereby increasing memory access and transmission latency. To overcome these challenges, we propose an unified low-transmission routing method to design a highly flexible NoC-based DNN accelerator. This unified low-transmission routing assisted NoC (Ultra-NoC) supports hybrid data transmission (i.e., unicast, multicast, and broadcast) requirements to leverage different data reuse methods in various DNN operations. Compared with the related work, we can reduce memory access times by 15% to 56% and total energy consumption by 24% to 55% because of the efficient data transmission mechanism.