NEURAL NETWORKS FOR FAST ARBITRATION AND SWITCHING NOISE-REDUCTION IN LARGE CROSSBARS

A neural network-based controller is presented for the real-time arbitration of routing paths in large crossbar switches constructed from one-sided crosspoint chips. This controller is suitable for a synchronous environment where a number or connection requests are simultaneously presented to the switch. The controller aims to maximize the effective bandwidth of the switch and to minimize the simultaneous-switching noise in the individual chips. The controller uses multiple winner-take-all networks coupled with some competitive-cooperative mechanisms to achieve the joint optimization. The effects of various network parameters are studied through simulation, and cases leading to nonoptimal solutions analyzed. The results show that the arbitration complexity and time scale well with the size of the switches, and the throughput achieved is close to the theoretically maximum attainable. We also introduce a hierarchical neural network controller for a packet-switched environment where connections are established and broken asynchronously. This controller provides almost the same level of performance as the first one, but with significantly reduced computation for each connection request.