Core network interface architecture and latency constrained on-chip communication

This paper proposes a Core Network Interface (CM) architecture to interface IP cores with on-chip networks. Besides the basic functionality of packetizing communication requests and responses, we expect the CNI to provide additional services critical to communication in complex systems-on-a-chip (SoC). A CNI for interfacing with OCP-compliant core interfaces was developed for architecture validation. With the support of a modified on-chip router, the CNI was setup to bound on-chip communication latency jitter. We observed that jitter due to inefficient virtual channel allocation in a particular configuration of an on-chip interconnection network lead to latency variations of up to 400%. Using a class-based virtual channel allocation scheme in a 2D torus network, we provide for predictable end-to-end latency. While the proposed scheme does not guarantee least possible end-to-end latency, it provides for constrained bounds.