Auto-PPA: An Adaptive Deep RL Agent for VLSI Physical Design Optimization

Hussain Ali Mutar; Ibtihal Razaq Niama ALRubeei; Omar Hashim Yahya; Naseer Ali Hussien; Haider TH. Salim AlRikabi; Abdul Hadi M. Alaidi

doi:10.31838/JCVS/08.01.02

Authors

Hussain Ali Mutar Computer Department, College of Computer Science and Information Technology, Wasit University, Wasit, Iraq https://orcid.org/0009-0004-4934-416X
Ibtihal Razaq Niama ALRubeei Electrical Engineering Department, College of Engineering, Wasit University, Wasit, Iraq https://orcid.org/0009-0003-8958-8487
Omar Hashim Yahya Department of Computer Engineering Technology, Technical Engineering College for Computer and AI, Northern Technical University, Mosul, Iraq. https://orcid.org/0000-0003-0277-562X
Naseer Ali Hussien Department of Computer Engineering Technology, College of Technical Engineering, Al-Ayen Iraqi University, AUIQ, Nasiriyah, Iraq. https://orcid.org/0000-0001-9499-6694
Haider TH. Salim AlRikabi Electrical Engineering Department, College of Engineering, Wasit University, Wasit, Iraq https://orcid.org/0000-0002-5201-2084
Abdul Hadi M. Alaidi Computer Department, College of Computer Science and Information Technology, Wasit University, Wasit, Iraq. https://orcid.org/0000-0001-9120-708X

DOI:

https://doi.org/10.31838/JCVS/08.01.02

Keywords:

VLSI, Optimization, AI model, Reinforcement Learning, Real-Time PPA Co-Optimization

Abstract

The physical design phase of Very-Large-Scale-Integration (VLSI) is notoriously difficult since it must strike a balance between PPA, power, and performance. Computationally costly design cycles and less-than-ideal Pareto fronts are common challenges of using traditional optimization methods to tackle these metrics in order. As part of physical design, this study suggests a new reinforcement learning (RL) framework that can optimize all three PPA measures in real time. In the proposed method, commercial electronic design automation (EDA) tools were used in conjunction with a deep deterministic policy gradient (DDPG) agent to make routing and placement decisions incrementally. Guided by a customized reward function that dynamically balances PPA trade-offs based on design stage priorities, the agent operates on a continuous action space that represents geometric coordinates and constraint modifications. While conventional sequential optimization methodologies reduce optimization runtime by about 35%, the proposed RL agent improves the power-performance product by 18.7% and the area reduction by 12.3%, according to simulation results on the ISPD 2015 benchmark suite. An innovative approach to optimize intelligent, adaptable physical designs that successfully traverse the high-dimensional PPA trade-off space is presented by the suggested framework.