DC25_PAPER_Track14_DeepReinforcementLearningBasedDesign_Ryu.pdf

1、 Information Classification:General Deep Reinforcement Learning Based Design Optimization of Power/Ground Ball Map in BGA Package in 3D-ICs Considering Multiple Power Domain Environments Seunghun Ryu,KAIST ysh0067kaist.ac.kr Dongryul Park,KAIST Seonghi Lee,KAIST Hyunwoong Kim,Samsung Electronics San

2、guk Lee,KAIST Hyunwoo Kim,KAIST Seongho Woo,KAIST Changmin Lee,KAIST Jaewon Rhee,KAIST Seokbeom Yong,Samsung Electronics Sangsub Song,Samsung Electronics Jiseong Kim,KAIST Seungyoung Ahn,KAIST Information Classification:General Abstract This paper proposes a deep reinforcement learning(DRL)approach

3、to optimize power/ground ball map design in 3D-IC ball grid array(BGA)packages with multiple power domains(MPD).The proposed method addresses challenges such as power noise and diverse power densities by using a U-net-based DRL algorithm that mitigates simultaneous switching noise(SSN)while optimizi

4、ng the power/ground ball map.The U-net preserves both semantic and spatial information,making the method effective for complex MPD designs.It is scalable and reusable,performing well with different ball map sizes and current source positions.Compared to conventional optimization algorithm such as ra

5、ndom search(RS)and genetic algorithms(GA),the proposed method shows superior optimality performance and lower execution time.Author(s)Biography Seunghun Ryu received his B.S.from Koreatech in 2020 and M.S.from Cho Chun Shik Graduate School of Mobility,Korea Advanced Institute of Science and Technolo

6、gy(KAIST)in 2022,where he is currently pursuing a Ph.D.In 2022,he was a visiting scholar at Missouri University of Science and Technology.His research interests include high-speed interconnection design and machine-learning-based power integrity(PI)design for 2.5D/3D ICs.Dongryul Park received his B