Mixed-Size Placement Initialization via Stiefel-Manifold Sequential Subspace Optimization

Pengwen Chen, Chung-Kuan Cheng, Shao-Yun Fang, Chester Holtz, Sheng-Tan Huang, Susmita Sur-Kolay, Yucheng Wang, Ying YuanarXiv 2026 · arXiv

In the high-stakes world of VLSI (Very-Large-Scale Integration) design, "placement" is the task of arranging millions of tiny standard cells and hundreds of large "macros" (like memory blocks) on a silicon die. Modern analytical placers are incredibly powerful, but they have a "cold start" problem: their optimization objectives are non-convex and highly sensitive to where they begin. A poor initial guess can trap the optimizer in a suboptimal state, leading to longer wires, higher power consumption, and manufacturing rule violations.

This paper introduces a mathematically rigorous initialization framework that treats placement geometry as an optimization problem on a Stiefel manifold. By providing a "structure-aware" starting point, the authors demonstrate significant improvements in final design quality across multiple state-of-the-art placement engines.

TL;DR

The authors propose a geometry-driven initialization for VLSI placement that formulates spreading constraints as an optimization problem on a Stiefel manifold. This approach provides a high-quality "drop-in" starting point for analytical placers, reducing final wirelength by up to 13.8% and improving post-route metrics like power and routability.

What Problem Does This Paper Address?

Global placement is a balancing act between minimizing wirelength and ensuring components are spread out enough to be legally manufacturable. Modern "mixed-size" designs—where a single macro might be 10,000 times larger than a standard cell—exacerbate numerical scaling issues.

Existing placers often start with random coordinates or simple heuristics. Because the underlying optimization landscape is "bumpy" (non-convex), these engines are sensitive to their starting points. Previous attempts at "spectral" initialization (using graph theory) often struggled with mixed-size designs, either causing components to clump together or fly off the edges of the die due to poor handling of macro sizes and fixed obstacles.

Methods

The core of the paper is a transformation of the placement initialization problem into a Stiefel-manifold optimization.

  1. QCQP Formulation: The authors start with a Quadratically Constrained Quadratic Program (QCQP). The goal is to minimize a quadratic approximation of wirelength subject to two types of constraints:
  2. First-order: Fixing the area-weighted center (centroid) of the design.
  3. Second-order: Enforcing a specific "spread" of components based on the die's geometry.
  4. Stiefel Transformation: The "aha!" moment comes from realizing that these quadratic spreading constraints, after a process called metric equalization, define a Stiefel manifold (the set of matrices with orthonormal columns). This makes the geometric structure of the problem explicit.
  5. Riemannian SSM Solver: Instead of using generic, fragile solvers, the authors use a Riemannian Sequential Subspace Method (SSM). This solver is designed to stay on the "surface" of the manifold throughout the optimization, ensuring that the spreading constraints are always satisfied.
  6. Utilization-Aware Rescaling: To handle designs where fixed objects (like I/O pins) occupy significant space, the authors introduce a rescaling factor ($\rho^2$) for the constraint matrix. This prevents "pathological spreading" where the solver tries to push cells into occupied regions.
  7. Angle-Sorted Cycle Sparsification: To keep the math efficient for designs with millions of connections, they use a novel way to simplify complex nets (multi-pin connections) into simple cycles that adapt to the current geometry of the placement.

![Fig.

  1. Pipeline of the proposed mixed-size Stiefel-based initialization.](/figures/art_a5640d49861149fd8d73c19b6af67145)

Datasets

The method was evaluated as a drop-in replacement for the default initializers in three major placement engines: RePlAce, DREAMPlace, and AutoDMP. The authors tested on:

  • 16 MMS Benchmarks: A standard set of academic placement challenges.
  • 4 TILOS MacroPlacement Benchmarks: Modern, complex designs including Ariane133 (a RISC-V core) and BlackParrot.
  • Post-Route Evaluation: They took the results all the way through a commercial routing flow (Cadence Innovus) for Ariane133, BlackParrot, and a Mempool group to see if pre-route gains actually mattered in the final silicon.

Key Insights and Ideas

  • Geometry as a Manifold: The most significant insight is that the "spreading" required for a good placement is mathematically equivalent to staying on a Stiefel manifold. This allows designers to use powerful tools from Riemannian geometry to solve a traditionally messy engineering problem.
  • Stability is Key: By deriving explicit spreading targets that account for the individual sizes of every macro and cell, the initializer remains stable even when component sizes vary by orders of magnitude.
  • Drop-in Compatibility: The framework doesn't require changing the downstream placer. It simply provides better "starting coordinates," making it easy to integrate into existing industrial flows.

Key Results

The improvements in weighted Half-Perimeter Wirelength (wHPWL) were consistent across different placers:

  • RePlAce: 13.8% average improvement (with a peak of 36.2% on the NVDLA benchmark).
  • DREAMPlace: 3.5% average improvement across 20 benchmarks.
  • AutoDMP: 4.8% average improvement, and it enabled convergence on the "adaptec3" benchmark where the default flow failed.
  • Post-Route: The benefits persisted after routing. For Ariane133, routed wirelength decreased by 2.3%, and Design Rule Check (DRC) violations dropped by 40%.

Limitations and Future Work

  • Runtime: While efficient, the solver can take significant time for massive designs. For the million-cell Mempool group, initialization took about 87 minutes. The authors suggest GPU acceleration as a future direction.
  • Timing and Congestion: The current objective focuses on wirelength. Future work could incorporate timing paths and local congestion directly into the Stiefel-manifold framework.
  • 3D Placement: The authors note that the framework could potentially be extended to 3D IC placement, which is becoming increasingly relevant in modern chip design.

For a summary of the mathematical notation used in the paper, see the table below:

Table I: Summary of Key Notations

Summary updated

Abstract

Mixed-size global placement simultaneously optimizes the locations of large macros and millions of standard cells in modern very-large-scale integration (VLSI) physical design. State-of-the-art analytical placers achieve strong scalability, but their nonlinear objectives and density handling make them sensitive to the initial placement. This paper presents a geometry-driven framework that formulates placement initialization as a Quadratically Constrained Quadratic Program (QCQP) with first-order centroid and second-order spreading constraints. Explicit mixed-size spreading targets are derived from the die geometry and object sizes, to obtain stable constraint values even under extreme size variation. A utilization-aware rescaling of the constraint matrix prevents pathological spreading when fixed objects occupy significant die area. The QCQP is then transformed into a Stiefel-manifold optimization through metric equalization and centering projection, making the underlying geometric structure explicit. The resulting Stiefel-constrained quadratic objective admits efficient large-scale optimization via a Riemannian sequential subspace method (SSM). An angle-sorted cycle sparsification for multi-pin nets periodically keeps the Laplacian approximation sparse yet geometry-consistent. As a drop-in initializer for existing analytical placers, our method yields improvements in weighted half-perimeter wirelength (wHPWL): 13.8% average under RePlAce on the four TILOS benchmarks supported by its flow, 3.5% under DREAMPlace on all 20 benchmarks, and 4.8% under AutoDMP on 19 benchmarks with converged baselines. Post-route evaluation shows that these pre-route improvements carry through to routed designs: routed wirelength decreases by 1.1 to 2.3% on all three benchmarks. Standard-cell area and power also decrease on all three. Design-rule-check (DRC) violations decrease on two of the three.
BibTeX
@misc{chen2026mixed,
  author = {Pengwen Chen and Chung-Kuan Cheng and Shao-Yun Fang and Chester Holtz and Sheng-Tan Huang and Susmita Sur-Kolay and Yucheng Wang and Ying Yuan},
  title = {{Mixed-Size Placement Initialization via Stiefel-Manifold Sequential Subspace Optimization}},
  year = {2026},
  eprint = {2611.22402},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22402},
}
Figure