Abstract
The design of package-level power delivery networks (PDNs) has become increasingly challenging as modern high-performance systems demand higher currents. Existing PDN design flows treat ball map assignment, stackup selection, and power plane routing as separate, largely manual steps, leading to long iteration times and limited scalability. This work proposes a unified and automated package PDN design framework based on multi-agent reinforcement learning (MARL). Each power domain is modeled as an agent, with specialized agents responsible for ball map assignment, routing layer selection, and power plane synthesis. A central controller coordinates agent decisions using a global reward that captures electrical and physical objectives, including target DC resistance and layout constraints. An in-house resistance engine is integrated to accurately evaluate DC resistance using planes and via modeling combined with a node-voltage method. Experimental results on packages with up to 8 power domains demonstrate that the framework efficiently generates manufacturable designs that satisfy all domain-specific resistance targets, highlighting the practicality and scalability of the proposed approach.
Recommended Citation
H. Manoharan and C. Hwang, "Multi-Agent Reinforcement Learning Driven Package PDN Design Automation," Proceedings Electronic Components and Technology Conference, pp. 1865 - 1869, Jan 2026.
The definitive version is available at https://doi.org/10.1109/ECTC51846.2026.00306
Department(s)
Electrical and Computer Engineering
Keywords and Phrases
MARL; Package PDN design; power plane; stackup synthesis
International Standard Serial Number (ISSN)
0569-5503
Document Type
Article - Conference proceedings
Document Version
Citation
File Type
text
Language(s)
English
Rights
© 2026, All rights reserved.
Publication Date
01 Jan 2026
Comments
National Science Foundation, Grant IIP-1916535