Multi-FPGA implementation of feedforward network and its performance analysis

Multi-FPGA system for the design of the spiking neural network is a great challenge for hardware acceleration. Multilayer feedforward neural networks (FNNs) are vitally important for the study of the coding problems in sensory organs. In this paper a multilayer FNN is implemented on a multi-FPGA-based system, which can guarantee both the high computational efficiency and the large network scale. The time-division multiplexing technology is employed in the proposed platform to be cost-efficient. In addition, since a parallel structure is used in the hardware design, the proposed hardware implementation can speedup approximately 4.8×10⁴ times faster than the real-world biological behaviors in a high computational precision. Clock synchronization is considered in the design of the FNN to guarantee the accuracy of the signal transmission between layers. The proposed system can be applied to a broad kinds of fields such as the control of motors and the artificial intelligence, and the presented neurons can be replaced by more complicated neurons for the study of other dynamical characteristics of the neural networks.