DC25_PAPER_Track14_FlowModelCanGiveYou_HyunjunAn_V3.pdf

1、 Information Classification:General Flow Model Can Give You Optimal and Diverse Design Guidance for TSV Array placement in HBM Hyunjun An,KAIST anhyunjunkaist.ac.kr Keunwoo Kim,KAIST keunwookimkaist.ac.kr Haeyeon Rachel Kim,KAIST haeyeonkimkaist.ac.kr Junghyun Lee,KAIST junghyunleekaist.ac.kr Inform

2、ation Classification:General Taein Shin,KAIST taeinshinkaist.ac.kr Keeyoung Son,KAIST keeyoungkaist.ac.kr Joungho Kim,KAIST jounghokaist.ac.kr Information Classification:General Abstract This paper proposes a novel framework for addressing the combinatorial optimization problem in hardware domain,sp

3、ecifically targeting the through silicon via(TSV)array placement in high bandwidth memory.The TSV array placement problem involves optimizing the positions of differential signal,single-ended signal,and ground TSVs considering signal integrity.Given that the TSV array placement problem is a highly c

4、omplex combinatorial optimization problem,it is challenging to solve manually.Thus,artificial intelligence(AI)based methodology such as reinforcement learning(RL)has been predominantly used.However,RL often gets trapped in local maxima and is limited to generate diverse solutions.Therefore,we propos

5、e a new model structure,which is hardware problem specialized-generative flow network(HPS-GFN),to address these critical issues.Unlike RL,HPS-GFN reinforces all paths leading to high reward states,ensuring a variety of optimal solutions.By providing design preferences for all possible actions at eac

6、h step,HPS-GFN generates optimal and diverse solutions,making it a valuable tool for the industry.Our experimental results demonstrate that HPS-GFN achieves an average score of 61.9,significantly outperforming other methodologies including deep reinforcement learning(60.3),genetic algorithm(59.4)and