DC25_PAPER_Track10_NovelMultiPowerDomainPCB_Manoharan_v2.pdf

1、 Information Classification:General DesignCon 2025 Novel Multi Power Domain PCB Decap Layout Synthesis using Multi-Agent Reinforcement Learning Haran Manoharan,Missouri S&T hm6h6mst.edu Hanfeng Wang,Google Jingnan Pan,Google Xu Gao,Google Kelvin Qiu,Google Chulsoon Hwang,Missouri S&T Information Cla

2、ssification:General Abstract Modern Systems on Chips(SoC)have more than one power domain for efficient power management across multiple functional blocks.In the design of Printed Circuit Board(PCB)pre-layout stage,engineers use physical insight and trial and error methods to place decaps to satisfy

3、the target specifications needed for the SoC for efficient operation in the PCB.This is a time-consuming and ineffective method since there could be numerous possibilities.This could sometimes lead to overdesign or under-design for some power domains;repetitive iterations are required to overcome th

4、is challenge.This paper addresses a challenge not tackled at this scale beforespecifically,managing multiple power domains in a practical setting.For the first time,we proposed a hybrid algorithm that combines multi-agent reinforcement learning with a genetic algorithm to effectively handle the comp

5、lexity of decap pre-layout synthesis for multiple power domains.This work optimizes the placement,orientation,value,and number of decaps,offering a solution that accounts for the intricate dependencies and constraints inherent in PCB pre-layout designs,where no initial layout is available to begin t

6、he optimization process.This algorithm computes the PDN impedance using an in-house tool making it fast and completely free of other commercial simulation tools.The algorithm has been designed in a way to take inputs directly from a board file or with just ball map and PCB stackup information making