Design Tradeoffs of Long Links in Hierarchical Tiled Networks-on-Chip

Hierarchical topologies are frequently proposed for large Networks-on-Chip (NoCs). Hierarchical architectures utilize, at the upper levels, long links of the order of the die size. RC delays of long links might reach dozens of clock cycles in advanced technology nodes, if delay reduction techniques (e.g. wire sizing and repeater insertion) are not applied. Some proposals assume that long links can be adjusted to satisfy timing requirements but lack a deep evaluation of the tradeoffs and costs. Other proposals assume that long links must be pipelined, but do not provide a comprehensive justification. In this paper we evaluate the efficiency and the system costs of wire sizing and repeater insertion as methods to reduce link delays in hierarchical NoCs. We present a unified interconnect cost function that accounts for power and wiring overheads of these methods. Then, we quantify the costs of modifying long links in typical hierarchical NoCs for different target clock frequencies and technology nodes. Although long links might undergo aggressive adjustments, we find these overall costs to be low at the system level for many typical cases, taking into account that there are only a few long links in most proposed hierarchical NoC architectures.