Dual Diagonal Mesh: An Optimal Memory Cube Network Under Geometric Constraints
Masashi Oda, Kai Keida, Ryota Yasudo · 2023
Recent important applications such as deep learning and big data analysis require fast and large memory systems. For such applications, the memory-cube networks has been proposed. These networks, composed of 3D stacked memories connected via packet-switched interconnection networks, differ from traditional interconnection networks due to geometric constraints and the limited number of ports. Consequently, the hop count increases as the number of nodes increases. To solve this issue, this paper proposes the dual diagonal mesh (DDM) memory cube network. The basic concept of DDM is to employ two disconnected diagonal mesh networks to eliminate unnecessary links between memory cubes. This approach optimizes the average shortest path length and the diameter under geometric constrains. We further propose the deadlock-free deterministic shortest path routing methodology with a routing bypassing technique. The experimental results demonstrate that DDM outperforms conventional memory cube networks.