MUGNoC

Abstract

Current communication infrastructures for convolutional neural networks (CNNs) only focus on specic transmission patterns, not applicable to benet the whole system if the dataow changes or dierent dataows run in one system. To reduce data movement, various CNN dataows are presented. For these dataows, parameters and results are delivered using dierent trac patterns, i.e., multicast, unicast, and gather, preventing dataow-specic communication backbones from beneting the entire system if the dataow changes or dierent dataows run in the same system. Thus, in this paper, we propose MUG-NoC to support typical trac patterns and accelerate them, therefore boosting multiple dataows. Specically, (i) we for the rst time support multicast in 2D-mesh software congurable NoC by revising router conguration and proposing the ecient multicast routing; (ii) we decrease unicast latency by transmitting data through the dierent routes in parallel; (iii) we reduce output gather overheads by pipelining basic dataow units. Experiments show that at least our proposed design can reduce 39.2% total data transmission time compared with the state-of-the-art CNN communication backbone.