Integrating Downstream Routing Constraints into Macro Placement

17 concepts7 papers1 proposal

In the physical design of integrated circuits, layout automation is traditionally split into two sequential steps: placement, where components are assigned physical coordinates on a canvas, and routing, where physical wires are drawn to connect them. Because routing is computationally expensive, placement algorithms traditionally rely on geometric proxies like Half-Perimeter Wirelength (HPWL) to evaluate layout quality.

However, this decoupled approach creates a critical blind spot. Optimizing purely for geometric distance frequently produces layouts that are physically impossible to route. Modern research focuses on bridging this gap by making early-stage placement algorithms explicitly aware of downstream routing constraints.

The Blind Spot of Early-Stage Layout Design

To make placement computationally tractable, traditional algorithms assume that if two components are placed close together, connecting them will be straightforward. But in reality, wires cannot simply take the shortest path; they must navigate a dense maze of other components and pre-allocated routing channels.

When placement engines operate without foresight regarding the routing stage, they generate layouts that look optimal on paper but fail during routing. As noted by Davide Basso et al. in Effective Analog ICs Floorplanning with Relational Graph Neural Networks and Reinforcement Learning, ignoring these constraints early on leads to layouts that are difficult or impossible for automated tools to route.

This fundamental tension spans decades of electronic design automation. Even in early design stages like fanout optimization, Hirendu Vaishnav demonstrated in Routability-Driven Fanout Optimization that neglecting routing considerations leads to severe congestion and bloated chip areas, because routing structures consume a massive portion of the available chip canvas.

Why Simple Proxies Fail: The Congestion Problem

Why exactly does a layout with a short wirelength proxy fail during routing? The answer lies in the spatial distribution of large functional blocks, known as macros.

When an algorithm optimizes solely for wirelength, it tends to cluster macros tightly together in the center of the chip to minimize the distance between them. However, this creates a massive physical bottleneck. As Feng Wu et al. explain in LaMPlace: Learning to Optimize Cross-Stage Metrics in Macro Placement, the positioning of these large functional blocks dictates the success of all subsequent routing and clock tree synthesis stages. If macros occupy the center of the chip haphazardly, they block the routing channels required by millions of smaller standard logic gates. This creates localized routing congestion and timing violations that no downstream tool can resolve, ultimately leading to manufacturing failures.

Traditional Wirelength Optimization       Routing-Aware Periphery Placement
┌───────────────────────────────────┐     ┌───────────────────────────────────┐
│      ┌───────┐ ┌───────┐          │     │ ┌───────┐               ┌───────┐ │
│      │ Macro │ │ Macro │          │     │ │ Macro │               │ Macro │ │
│      └───────┘ └───────┘          │     │ └───────┘               └───────┘ │
│          CONGESTION ZONE          │     │        AVAILABLE CENTER           │
│      ┌───────┐ ┌───────┐          │     │        FOR ROUTING AND            │
│      │ Macro │ │ Macro │          │     │        STANDARD CELLS             │
│      └───────┘ └───────┘          │     │ ┌───────┐               ┌───────┐ │
│                                   │     │ │ Macro │               │ Macro │ │
└───────────────────────────────────┘     └─└───────┘───────────────└───────┘─┘

Compounding this issue, Yunqi Shi et al. point out in Macro Placement by Wire-Mask-Guided Black-Box Optimization that standard wirelength proxies cannot model these complex, non-linear routing blockages. Without integrating explicit routability metrics into the initial optimization loop, the placement engine remains blind to the physical feasibility of the layout.

Preserving the Center via Periphery-Biased Regularity

To prevent central congestion without running slow, full routing simulations at every placement step, researchers have turned to structural inductive biases. In industrial chip design, human experts traditionally push large macros toward the edges (periphery) of the chip canvas, leaving the center open for complex standard-cell routing.

The authors of Reinforcement Learning Policy as Macro Regulator Rather than Macro Placer formalized this practice by introducing a geometric "regularity metric" into their optimization objective. By explicitly minimizing the distance between large modules and the chip periphery, Ke Xue et al. demonstrated that enforcing a periphery bias preserves the central region for routing channels. This alignment with industrial regularity principles dramatically reduces routing blockages and improves final timing metrics compared to models that optimize for wirelength alone.

Cross-Stage Learning and Routing-Informed Frameworks

Beyond geometric heuristics, recent frameworks leverage machine learning to predict downstream routing success directly during the placement phase.

In LaMPlace, Feng Wu et al. trained a predictor on offline data to generate a "Learnable Mask." This mask allows the placement engine to make fast, greedy placement decisions that actively align with final post-routing performance metrics, such as Worst Negative Slack, rather than relying on local geometric proxies.

This routing-first philosophy is even more pronounced in specialized domains like silicon photonics. In Apollo: Automated Routing-Informed Placement for Large-Scale Photonic Integrated Circuits, Hongjian Zhou et al. demonstrated that for photonic circuits, routing feasibility must be treated as the primary driver of placement rather than an afterthought. By integrating physical routing constraints directly into the placement loop, the authors raised the automated routing success rate from roughly 50% in traditional tools to 94.79%.

Generalizing to Complex Topologies

The shift toward routing-aware placement also addresses the strict performance demands of analog circuits. Analog layouts are highly sensitive to layout-dependent effects (LDEs) and parasitic wiring constraints.

As explored by Supriyo Maji et al. in Late Breaking Results: Breaking Symmetry - Unconventional Placement of Analog Circuits using Multi-Level Multi-Agent Reinforcement Learning, traditional placement strategies that ignore active routing optimization often result in suboptimal circuit performance during final simulations. By moving away from rigid geometric assumptions and utilizing reinforcement learning to navigate routing constraints dynamically, contemporary placement engines ensure that early-stage floorplans remain viable all the way through to final manufacturing.

Key Foundational Work

Routability-Driven Fanout Optimization (1993) — contributed 6 concept elements to this direction.
Macro Placement by Wire-Mask-Guided Black-Box Optimization (2023) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_periphery_biased_placement_efadac16,
  title = {periphery-biased placement},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/periphery-biased-placement/rdr_efadac167f9444b0b3019ddfb0460b87},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_periphery_biased_placement_efadac16_v1,
  title = {periphery-biased placement},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/periphery-biased-placement/rdr_efadac167f9444b0b3019ddfb0460b87/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake