Fang Juan, Cai Huayi, Lv Xin
Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China.
Micromachines (Basel). 2023 Mar 9;14(3):628. doi: 10.3390/mi14030628.
Increasing the number of cores on a chip is one way to solve the bottleneck of exponential growth but an excessive number of cores can lead to problems such as communication blockage and overheating of the chip. Currently, networks-on-chip (NoC) can offer an effective solution to the problem of the communication bottleneck within the chip. With current advancements in IC manufacturing technology, chips can now be 3D-stacked in order to increase chip throughput as well as reduce power consumption while reducing the area of the chip. Automating the mapping of applications into 3D NoC topologies is an important new direction for research in the field of 3D NoC. In this paper, a 3D NoC partitioning algorithm is proposed, which can delineate the 3D NoC region to be mapped. Additionally, a double particle swarm optimization (DPSO) based heuristic algorithm is proposed, which can integrate the characteristics of neighborhood search and genetic algorithms, and thus solve the problem of a particle swarm algorithm falling into local optimal solutions. It is experimentally demonstrated that this DPSO-based hybrid optimization algorithm has a higher throughput rate and lower energy loss than the traditional heuristic algorithm.
增加芯片上的内核数量是解决指数增长瓶颈的一种方法，但过多的内核会导致诸如通信阻塞和芯片过热等问题。目前，片上网络（NoC）可以为解决芯片内通信瓶颈问题提供有效的解决方案。随着当前集成电路制造技术的进步，现在可以对芯片进行3D堆叠，以提高芯片吞吐量、降低功耗并减小芯片面积。将应用程序自动映射到3D NoC拓扑结构是3D NoC领域研究的一个重要新方向。本文提出了一种3D NoC分区算法，该算法可以划定要映射的3D NoC区域。此外，还提出了一种基于双粒子群优化（DPSO）的启发式算法，该算法可以整合邻域搜索和遗传算法的特点，从而解决粒子群算法陷入局部最优解的问题。实验表明，这种基于DPSO的混合优化算法比传统启发式算法具有更高的吞吐率和更低的能量损耗。
