A collaborative team led by Ph.D. student Fangzhou Liu and supervised by Prof. Yu Bei from CUHK, in partnership with National Tsing Hua University (NTHU) and Giga Design Automation, has secured the first place in Problem C at the 2025 ICCAD Contest. Sponsored by NVIDIA, this year’s challenge focused on incremental placement optimization—a critical step in modern chip design that involves fine-tuning cell locations, gate sizing, and buffering within highly constrained spaces to balance performance, power, and area (PPA). To overcome the limitations of traditional discrete flows and the instability of existing differentiable methods, the team introduced a novel hybrid solution. This approach seamlessly integrates discrete sign-off capabilities with differentiable gradient-based optimizations, delivering superior PPA results compared to other winners and state-of-the-art techniques, all while maintaining competitive runtime efficiency. This work demonstrates the method’s potential for optimizing chip design flows while highlighting how high-level competitions broaden students’ industry perspectives and enhance their practical skills.
