Reinforcement Learning-based 3D Floorplanner with Die Swapping Mechanism

Recommended citation: Qin Luo, Hailiang Li, Evangeline F.Y. Young. Reinforcement Learning-based 3D Floorplanner with Die Swapping Mechanism. International Symposium of EDA (ISEDA), 2026.

3DICs offer significant advantages in performance, power and area over their 2D designs but introduce floorplanning challenges requiring the co-optimization of block positions and die assignments. While reinforcement learning (RL) has shown promising performance in 2D floorplanning, existing RL-based 3D methods rely on static and pre-assigned die placements, preventing true simultaneous optimization. To overcome this, we propose a novel RL-based 3D floorplanner that features a die-swapping mechanism, enabling the dynamic co-optimization of block placement and die assignment in RL-based approach for the first time. We develop an efficient spatial checking algorithm for available space checks in large layouts to support the die-swapping mechanism. Evaluation on the GSRC benchmark and a new benchmark generated by the OpenROAD suite demonstrates that our approach outperforms existing methods by significantly reducing cross-die wirelength, showcasing the critical importance of joint optimization and strong potential of RL for 3D physical design.