2 July 2025 in ACM Transactions on Design Automation of Electronic Systems
Traditional supervised single-task learning models are used in timing-driven placement exploration to improve both effectiveness and efficiency by predicting wire length, wire delay, and cell delay separately. However, these metrics are interdependent, with the two delays being timing-based and wire length non-timing, which makes it difficult for single-task models to capture their complex relationships. Moreover, the limited existing multi-task learning methods can only predict either multiple timing or non-timing metrics. To address these limitations, this paper introduces DLGNN, a novel multi-task graph learning model that simultaneously predicts these three metrics through an embedder-predictor architecture featuring two residual connections, a combination of both soft and hard parameter sharing, and a geometric loss strategy. Cross-design experimental results on the Nangate 45nm library demonstrate that DLGNN outperforms baseline models in terms of both predictive performance and time efficiency. Additionally, ablation studies emphasize the critical roles of the residual connections, the combination of soft and hard parameter sharing, and the geometric loss strategy in improving DLGNN’s predictive performance. The generalization experiment on the ASAP 7nm library further confirms DLGNN’s advantages for more advanced technology nodes.