Design of multicycle path accelerator for neural network

Abstract

In this paper, MAC computation, which is the most used operation in neural network computing, is accelerated by multicycle_path. When using a pipeline, we can not divide the operation into fixed delays. Thus, exiting pipelines determine the operating frequency with the lowest frequency. We proposed the method confirmed that confirmed that the operation frequency of the whole process is improved by using multicycle_path to divide the operation into a certain delay and to perform the operation. In this paper, the experiment was divided into SingleCycle, Conventional Multicycle, MultiCycle_Ex and MultiCycle_Slice. Conventional Multicycle achieved 2.23 times higher operating frequency than SingleCycle, but resource usage tripled. MultiCycle_Ex achieved a 2.67 times higher operating frequency while maintaining the resource usage of SingleCycle, and MultiCycle_Slice achieved operating frequency about 8.2 times higher than SingleCycle.