Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Escaping Local Optima in Global Placement

Ke Xue, Xi Lin, Yunqi Shi, Shixiong Kai, Siyuan Xu, Chao Qian

TL;DR

Global placement is a challenging non-convex optimization problem where gradient-based methods like DREAMPlace can stall in local optima. The authors propose Hybro, a hybrid framework that alternates gradient-based placement with perturbation steps (Shuffle, Shuffle (all), WireMask) over $N$ iterations, selecting the best HPWL outcome. Across ISPD 2005 and ICCAD 2015 benchmarks, Hybro demonstrates significant HPWL improvements and improved PPA metrics, with Hybro-WireMask often delivering the strongest performance. This approach offers a robust, GPU-efficient means to enhance analytical placement quality and stability across seeds.

Abstract

Placement is crucial in the physical design, as it greatly affects power, performance, and area metrics. Recent advancements in analytical methods, such as DREAMPlace, have demonstrated impressive performance in global placement. However, DREAMPlace has some limitations, e.g., may not guarantee legalizable placements under the same settings, leading to fragile and unpredictable results. This paper highlights the main issue as being stuck in local optima, and proposes a hybrid optimization framework to efficiently escape the local optima, by perturbing the placement result iteratively. The proposed framework achieves significant improvements compared to state-of-the-art methods on two popular benchmarks.

Escaping Local Optima in Global Placement

TL;DR

Global placement is a challenging non-convex optimization problem where gradient-based methods like DREAMPlace can stall in local optima. The authors propose Hybro, a hybrid framework that alternates gradient-based placement with perturbation steps (Shuffle, Shuffle (all), WireMask) over iterations, selecting the best HPWL outcome. Across ISPD 2005 and ICCAD 2015 benchmarks, Hybro demonstrates significant HPWL improvements and improved PPA metrics, with Hybro-WireMask often delivering the strongest performance. This approach offers a robust, GPU-efficient means to enhance analytical placement quality and stability across seeds.

Abstract

Placement is crucial in the physical design, as it greatly affects power, performance, and area metrics. Recent advancements in analytical methods, such as DREAMPlace, have demonstrated impressive performance in global placement. However, DREAMPlace has some limitations, e.g., may not guarantee legalizable placements under the same settings, leading to fragile and unpredictable results. This paper highlights the main issue as being stuck in local optima, and proposes a hybrid optimization framework to efficiently escape the local optima, by perturbing the placement result iteratively. The proposed framework achieves significant improvements compared to state-of-the-art methods on two popular benchmarks.
Paper Structure (13 sections, 3 equations, 5 figures, 3 tables, 1 algorithm)

This paper contains 13 sections, 3 equations, 5 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. Preliminaries
  4. DREAMPlace
  5. Non-Convex Optimization
  6. Method
  7. Hybro Framework
  8. Perturbation strategy
  9. Experiment
  10. Experimental Settings
  11. ISPD 2005 Results
  12. ICCAD 2015 Results
  13. Conclusion

Figures (5)

  • Figure 1: Hybro-powered design flow, which can help in escaping local optima in global placement. If the gradient-based global placer (e.g., DREAMPlace) converges, Hybro adds perturbation to the placement result. This procedure will be repeated for $N$ iterations.
  • Figure 2: HPWL vs. iterations of different methods on the ISPD 2005 benchmarks, where the shaded region represents the standard error derived from 3 independent runs.
  • Figure 3: Runtime breakdown of Hybro-Shuffle (all) and Hybro-WireMask on the ISPD 2005 benchmark adaptec3.
  • Figure 4: Comparison of full placement HPWL and macro HPWL curves of Hybro-WireMask on the ISPD 2005 benchmarks adaptec4 and bigblue3.
  • Figure 5: Placement layouts and congestions of Multiple-DMP (top row) and Hybro-WireMask (bottom row) on the ICCAD 2015 benchmarks, superblue1, superblue3, superblue4, and superblue10. The congestion results are obtained by Cadence Innovus, where red points indicate the congestion critical regions.