Network-on-chip (NoC) is used for routing packets among multiple functional modules such as memory blocks, processors, communication blocks, and so on in 3D integrated circuits (ICs). A 3D IC installs vertical links on routers using through-silicon vias to transfer packets among different layers. Traditionally, each vertical link includes two virtual channels for ascending and descending packets to avoid deadlock. Since a virtual channel takes space, energy, and delay cost, there is a need to reduce the number of required virtual channels.

The major contribution of this paper is the algorithm to eliminate the dependency on the virtual channels (VCs) for deadlock-free routing. The authors proposed routing rule sets under the assumption of mesh network topology that restricts the selection of a vertical link so that no two packets are transferred through the same vertical link. Based on the rule sets, deadlock-free routing is achieved without VCs. The rule sets are extended for irregular partial vertical links and heterogeneous mesh layers. To evaluate the performance and energy consumption, the proposed solution is compared with a baseline elevator-first algorithm and the proposed algorithm with VCs. Simulation results show that the proposed algorithm with VCs outperforms the proposed one without VCs and the baseline because the proposed algorithm uses the virtual channels efficiently. Overall, the proposed algorithm shows benefits in performance and energy saving.

Since the algorithm restricts the selection of the vertical links, it is possible to select a non-shortest path. It is interesting to see the non-shortest path selection policy results in reduced delay and energy consumption. The major reason for this is that having no competition in a vertical link reduces the overall running time. As noted in the paper, the relative benefit decreases as the number of VCs increases because the possibility of competition in a vertical link decreases.

Another interesting point in this paper is that the VCs are advantageous to the proposed algorithm even though the algorithm was designed to eliminate the dependency on the VCs. The original algorithm without VCs even shows worse performance than the baseline. For better understanding about the impact of the VCs, it would be interesting to know if there are simulation results showing the performance and energy consumption when VCs are installed in only part of vertical links.