A Multi-Objective Optimization Algorithm Based on Deep Learning for Circuit Partition

Circuit partition is a critical step during Three-dimensional integrated circuit(3D IC) physical design. The quality of circuit partition directly affects the subsequent processes(placement, clock-tree synchronization, route, etc.). Circuit partition is the process that divides a circuit composed of logic gates into subsets with the objective of minimizing the inter connection of subsets and balancing the sum of modules weights between subsets. It can be modeled as a graph partition problem, which is known to be NP-hard. Compared with the traditional graph partition problem, in order to make full use of the space resources of 3D chips, circuit partition pays more attention to the balance between subsets. In this paper, we propose a deep learning-based multi-objective optimization algorithm. Our algorithm uses multiple branch feature extractors(based on GraphSAGE) to learn the features and topological relationships of nodes, and then generates node embedding. Finally, our algorithm obtains the circuit partitioning strategy by using node allocator(based on FC). Compared to the state-of-the-art graph partition(metis), our algorithm improves the balance of partition(2×), and slightly reduces cut-edge quality.