URMP: using reconfigurable multicast path for NoC-based deep neural network accelerators

Network-on-chip (NoC) exists with the advantages of high communication efficiency, scalability and reliability. In recent years, NoC-based deep neural network (DNN) accelerators have been proposed. Although existing NoC research solutions can solve the problem of the existence of one-to-one traffic in the network and transmit unicast traffic efficiently. However, due to the traffic characteristics of neural networks, there exists a large amount of one-to-many traffic, and if unicast is used to transmit multicast traffic, it may rapidly exhaust the network bandwidth and greatly degrade the performance of the platform. To solve the problem of a large amount of one-to-many multicast traffic existing in the network, we propose a path-based multicast mechanism that greatly exploits the traffic characteristics of neural networks and has excellent scalability. Also a router architecture that can efficiently replicate multicast packets and provide single-cycle per-hop transmission for multicast packets was designed. Detailed simulation results indicate that our proposed scheme can effectively reduce the classification delay, the average packet delay and the number of packets transmitted by the network.