Design Space Exploration of Reinforcement Learning Configurations for Post-Legalization Placement
Post-legalization placement refinement seeks to reduce wirelength after global placement without breaking site constraints. This paper audits whether a GNN-PPO agent can handle that task by systematically exploring 18 training configurations on three ISPD2015 benchmarks, varying rewards, action spaces, GNN architectures, and graph-partitioning strategies. The study finds that RL lags behind classical tools, and it dissects why—pinning blame on credit-assignment bottlenecks, GNN over-smoothing, and netlist regularity—while proposing a lightweight netlist classifier to guide training decisions.
The paper is a candid exploratory study that transparently reports negative results. The proposed GNN-PPO configurations consistently underperform OpenROAD, which achieves $4.3\%$--$6.4\%$ HPWL reduction versus the best RL inference result of $0.43\%$ on the irregular design and even less on the regular ones. The authors correctly attribute much of the gap to structural limitations of single-cell local moves, but the proposed netlist classifier is suggestive at best because its thresholds are calibrated on the same three designs used for evaluation.
Its chief virtue is honesty: instead of claiming state-of-the-art wirelength, the authors systematically expose why RL falters on standard-cell refinement. The analysis of three compounding pitfalls—credit assignment bottlenecks, GNN over-smoothing on regular arrays, and tightly placed cells—offers actionable intuition for future researchers. The link between netlist micro-architecture and partitioning utility is particularly insightful, showing that subregion training can clean up noisy rewards for irregular bus-style circuits while destroying the policy on regular butterfly topologies where long-range connections are severed.
The empirical scope is too narrow to support general conclusions: only three ISPD2015 benchmarks, all below $155$K cells, are studied, and the classification thresholds in Algorithm 1 are fitted to those exact three designs. Absolute gains are minuscule for most configurations, with evaluation improvements below $0.5\%$, and the best regular-design result reaches only $0.03\%$ HPWL reduction. The paper also omits key hyperparameters—PPO learning rate, GNN hidden dimensions, batch size, and optimizer settings—so it is unclear whether the poor results reflect a fundamental RL limitation or merely insufficient tuning.
The internal evidence supports relative claims convincingly: partitioning clearly helps the irregular bus design but fails on the regular FFT, with explanations grounded in concrete netlist properties. The OpenROAD comparison is fair, though the paper would benefit from additional baselines such as analytical local-search refiners. The framing of prospects and opportunities is slightly overstated because the sole configuration that comes closest to classical quality is the partitioned SAGE run on the irregular design ($0.43\%$ versus OpenROAD's $4.3\%$), which is still an order of magnitude worse. The reward-engineering ablations are a strength: they clearly demonstrate that simple reward shaping cannot overcome structural RL limitations.
Reproducibility is weak. While standard ISPD2015 benchmarks and OpenROAD improve transparency, no code, checkpoints, or training scripts are released. The paper omits PPO hyperparameters, GNN hidden dimensions, learning rates, batch sizes, and hardware details. Training duration is given only as $400$ or $800$ episodes without defining steps per episode or reporting wall-clock time, and no variance metrics across random seeds are provided. An independent group would have to guess substantial implementation details to reproduce these results.
Graph Neural Networks (GNNs) and Reinforcement Learning (RL) have recently been explored for VLSI placement optimization, but their effectiveness for post-legalization placement refinement remains unclear. This paper presents a systematic design space exploration of a GNN-Proximal Policy Optimization (GNN-PPO) framework across three ISPD2015 benchmarks. We evaluate 18 configurations covering reward functions, GNN architectures, action spaces, candidate-set sizes, and training strategies. We analyse the results we obtain to identify pitfalls and to assess the prospects and opportunities of the RL methods for post-legalization placement. On top of that we propose a design classification approach to guide the design decisions of RL models.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.