AFCNNM: Accelerating Fully Connected Neural Network on Mesh-based Optical Network-on-Chip
Wen Chuan Yang, Yunting Liao · 2024
Fully Connected Neural Network (FCNN) are widely used in image recognition and natural language processing. However, the time cost of training large datasets is high. Optical Network-on Chip (ONoC) has been proposed to accelerate the parallel computing of FCNN because of its advantages. Therefore, this paper proposes an accelerated FCNN model based on ONoC. We first design an FCNN-aware mapping strategy, and then propose a group-based inter-core communication scheme with low wavelength requirements according to the distribution of mapping cores. The optimal number of cores in each period is obtained by achieving the trade-off between the communication and computation time. The simulation results show that the proposed scheme has the advantages of low wavelength requirement, short training time and good scalability.