System-level interconnect architecture exploration for custom memory organizations

For data dominated applications, power consumption and memory bandwidth bottlenecks can be significantly alleviated with a custom memory organization. However, this potentially entails complex memory interconnections and a large routing overhead. This is undesirable for area cost, power consumption, and layout design complexity. By exploiting time-multiplexing opportunities over the long memory buses, this overhead can be significantly reduced. This paper proposes a system-level methodology for automated exploration of the interconnect architecture, which finds the optimal trade-off points for memory bus time-multiplexing. Experiments performed on real-life applications using our prototype tool show that even for very distributed memory organizations, the interconnect complexity can be significantly reduced to a cost-efficient, manageable level.