acceptodds
Under review as a conference paper at ICLR 2027

HoliPlace: Holistic Macro Placement via Global Planning and Entropy-Guided Ordering

Abstract

Macro placement is a critical step in chip design that significantly affects power, performance, and area (PPA). Recent reinforcement learning (RL) methods have shown promising results in this task. However, existing RL-based methods place macros in a fixed order, making each decision myopic and prone to error propagation, as early misplacements irreversibly constrain subsequent placements. To address these limitations, we propose HoliPlace (Holistic Macro Placement via Global Planning and Entropy-Guided Ordering), a Transformer-based RL framework for holistic macro placement. HoliPlace first leverages a Graph AutoEncoder (GAE) to extract netlist and macro features, and then employs an encoder-only Transformer to simultaneously predict position distributions for all macros, enabling global dependencies modeling. To bridge global planning and sequential execution, we introduce an entropy guided ranker that dynamically determines the placement order based on prediction uncertainty. Finally, we adopt Group Relative Policy Optimization (GRPO) to jointly optimize macro positions and their ordering. Experimental results on the ICCAD2015 benchmark demonstrate that HoliPlace improves placement quality across PPA metrics, outperforming existing methods.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Physical Design, Macro Placement, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tyawOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes HoliPlace, a reinforcement-learning framework for macro placement in chip physical design. The method first pretrains a GVAE/GNN encoder on circuit netlists, then uses an encoder-only Transformer to predict placement distributions for all macros in parallel. The method also introduces an entropy-guided ordering strategy: macros whose predicted position distributions have lower entropy are placed earlier. The policy is trained with a task-adapted group-relative policy optimization (GRPO) objective. Experiments on eight ICCAD2015 benchmark circuits report improvements over recent RL/meta-heuristic baselines on global HPWL and downstream PPA metrics from Cadence Innovus, with ablations for ordering, position predictor design and group size.

Critical strengths
  1. The paper's critique of fixed-order sequential RL macro placement is well motivated: early macro decisions can constrain later ones, and ordering bias is a real concern in constructive placement methods.

  2. Combining global netlist encoding, parallel position-distribution prediction, entropy-guided legal placement, and group-relative reward optimization is conceptually sound, as it directly incorporates global context before finalizing the placement order.

  3. On ICCAD2015, the method obtains the best average rank across the reported proxy and PPA metrics in Table 1 when compared with existing methods.

  4. The paper compares entropy-guided ordering against descending-entropy and area/connectivity ordering, evaluates position predictor variants, and studies group size, thus attributing the gains to the proposed components.

Critical weaknesses
  1. "Parallel prediction" is overstated. HoliPlace predicts all macro position distributions in parallel, but the final legal placement is still sequential because macros are placed one by one according to entropy order.

  2. The self-attention mechanism of Transformers incurs O(N^2) computational complexity when handling a large number of macros or standard cells.

Questions for authors
  1. Why not directly take the argmax position from each macro’s predicted distribution and place all macros simultaneously, which would make the method truly parallel?

  2. As the number of macros or standard cells increases, this cost may become prohibitive and potentially unacceptable for large-scale placement. How would HoliPlace scale to substantially larger designs, especially if extended beyond macro placement to include standard-cell placement as well?

Reviewer_zMr1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents HoliPlace, a holistic framework for RL-based macro placement that addresses the myopic decisions and error propagation caused by fixed-order sequential prediction. It first uses a Graph AutoEncoder to extract macro netlist features, and then employs an encoder-only Transformer to jointly predict position distributions for all macros, instead of sequential prediction. However, since the predictions from Transformer's cannot guarantee non-overlap constraints, HoliPlace employs a sequential legalization process that ranks macros by prediction entropy and places low-entropy, high-confidence macros first while masking illegal locations. A simplified GRPO algorithm evaluates groups of complete placement solutions using macro wirelength, regularity, and legality rewards. Experiments on ICCAD2015 show improved global HPWL and downstream PPA metrics, strong generalization, and faster inference than conventional sequential RL placers.

Critical strengths
  1. The paper addresses the limitations of sequential macro placement by predicting the macro potition distributions in parallel, enabling full cross-macro information exchange.

  2. Although a sequential legalization process is also needed to obtain the final layout, the order is decided by the prediction entropy instead of a fixed heuristic, which can evolve with the position distributions during the GRPO training process.

Critical weaknesses
  1. The comparison with other generative methods [1,2] that jointly produce macro positions is incomplete. Generative placers such as ChipDiffusion [1] and FlowPlace [2] share the same idea of global positioning without a fixed order, which is relevant to the motivation of moving beyond sequential prediction. However, experimental comparisons with these methods are missing in this paper.

  2. The need for separate GNN and Transformer encoders is not established. The ablation studies in this paper compare global and sequential prediction, and evaluate GAE pretraining. However, neither comparison shows whether the two encoder stages are necessary. In my view, a single graph-aware encoder can also capture both netlist structure and interactions across macros. The paper should justify the added architectural complexity.

  3. The issue of fixed order is not that critical for macro placement. Many other problems, such as unawareness of cell connectivity and mismatch with final PPA metrics, are much more significant for current learning-based placers.

References:

[1] Fang et al., "GraphDiffusion: A Graph-conditioned Diffusion Model for Chip Placement," International Symposium on Circuits and Systems (ISCAS), 2025.

[2] Xie et al., "FlowPlace: Flow Matching for Chip Placement," Design Automation Conference (DAC), 2026.

Questions for authors

Could macro locations from an analytical placer serve as a global prior or prototype for the legalization policy? Analytical placement considers macros and standard cells together when optimizing global HPWL, so its output may provide useful information about standard-cell connectivity. A comparison with this alternative would help clarify the value of the learned global planning stage.

Reviewer_zM7iOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HoliPlace, Transformer-based RL framework for holistic macro placement. Specifically, it first predicts position distributions for all macros in parallel, and then places macros sequencially in ascending order of prediction entropy.

Critical strengths

This paper proposes a better structure, i.e., prediting position contributions in parallel, for RL-based macro placement.

Critical weaknesses
  • This paper mainly considers wirelength and regularity as reward, but does not explicitly considers post-routing PPA as rewards. It also does not consider standard cells. So it is not convincing that it benefits final PPA. The authors also did not compare with methods like DREAMPlace on PPA metrics.
  • The main contribution is that it considers placement order in parallel for all macros, which seems not noval. Theoretically this should not affect the performance because ideally the model can learn this in RL.
  • How to determine alpha? The wirelength and regularity may not be in a same scale, and it is unclear why alpha=0.7.
Questions for authors

See weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.