Towards Cognitive Networking Using FPGA-Based Acceleration

The cognitive network controller (CNC) is a brain-inspired computing architecture that has been applied to the decentralized outbound link selection for data bundles in challenged networks. To this end, cognitive entities perceive the environment, reason using spiking neural networks (SNN), and decide routing actions that attempt to optimize pre-assigned performance goals. In this paper, a software architecture and SoC/FPGA implementation are proposed that enable decoupling the SNN computation tasks from the network gateways allowing the provision of accelerated CNCs as a network service to multiple network gateways. The design was experimentally evaluated using a 10-node network testbed and a Zynq Z7020 (PYNQ-Z2) SoC/FPGA board, which show the feasibility of computing the SNN state faster-than-real time despite the shared use of resources and the network communication overhead involved in remotely accessing the CNC service.