Toward a training of CNNs on a multi-FPGA system

Multi-FPGA systems have been receiving a lot of attention as low-cost and energy-efficient systems. Pattern recognition using CNN(Convolutional Neural Network) is a common application of such multi-FPGA systems, and various design examples have been presented, while the training phase of CNN is executed on GPUs in the cloud. However, when a multi-FPGA system is used for the MEC (Multi-access Edge Computing), the execution of the training phase is worth trying even if the absolute performance is smaller than that with GPUs in the cloud. In this paper, we implemented the backpropagation of sparse convolutional neural networks on the FiC system and estimated its performance and resource consumption with HLS simulation.