🤖 AI

Order Matters: Unveiling the Hidden Impact of Macro Placement Sequences via Proxy-Guided LLM Evolution

This paper introduces OrderPlace, a proxy-guided LLM evolution framework that automatically discovers optimal macro placement sequences to overcome the limitations of static heuristics, achieving significant wirelength reductions on ISPD 2005 benchmarks compared to state-of-the-art methods.

Original authors: Shibing Mo, Jing Liu, Jianchu Xu, Ruilin Wu

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shibing Mo, Jing Liu, Jianchu Xu, Ruilin Wu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to pack a moving truck. You have a mix of fragile vases, heavy sofas, and awkwardly shaped boxes. The goal is to fit everything in without breaking anything and to make sure the wires connecting your electronics (like a TV and a sound system) aren't stretched too tight.

In the world of computer chip design, this "moving truck" is the chip itself, the "boxes" are the macro components (large blocks of circuitry), and the "wires" are the electrical connections. The process of deciding where to put these blocks is called Macro Placement.

For a long time, engineers and computer programs have focused on where to put the blocks. But they largely ignored the order in which they put them in. They assumed the order didn't matter much, or they just picked a random order.

This paper, titled "Order Matters," argues that the order is actually the most critical part of the puzzle. Here is the breakdown of their discovery and solution:

1. The "Domino Effect" of Bad Ordering

The authors explain that placing these blocks is a sequential process. You place one, then the next, then the next.

The Analogy: Imagine trying to build a house of cards. If you place the first few cards in the wrong spots, you might block yourself from placing the rest of the cards later. You might end up with a pile that looks okay at first but is impossible to finish, or you have to stretch the "wires" (the connections) across the whole room just to connect two pieces that got separated early on.
The Claim: The paper proves mathematically that if you pick a bad order at the very beginning, it creates a "domino effect." This forces the rest of the placement into a suboptimal state, making it impossible to find the best solution, no matter how smart the computer is later on.

2. The Problem: Testing is Too Slow

If you want to find the perfect order to pack your truck, you might think, "Let's just try every possible order!"

The Reality: There are too many combinations. Trying every single order to see which one works best would take a supercomputer years to finish. It's too expensive and slow to test every idea.

3. The Solution: OrderPlace (The "AI Chef")

The authors built a new system called OrderPlace. Instead of a human guessing the order, or a simple rule (like "put big things first"), they used a Large Language Model (LLM)—the same kind of AI that writes code and answers questions—to act as an "AI Chef" inventing new recipes for the order.

Here is how it works:

The Evolutionary Kitchen: The AI starts with a few basic "recipes" (rules for ordering, like "put the biggest blocks first" or "put the most connected blocks first").
The Taste Test (Proxy Evaluation): Instead of cooking the whole meal (running the full, slow chip design process) to see if a recipe is good, the AI uses a lightweight proxy. Think of this as a "quick taste test." It simulates a simplified version of the packing process to see if the order looks promising. This is fast and cheap.
The Evolution: The AI looks at the results of the taste tests. It takes the best recipes, mixes them up, and asks itself, "How can I make this even better?" It then generates new, more complex rules (code) to try. It repeats this process, evolving better and better ordering strategies over time.

4. What They Found

The AI didn't just find a slightly better rule; it discovered entirely new ways of thinking about the order that humans hadn't thought of.

The Results: When they tested this on standard chip design benchmarks (the "ISPD 2005" datasets), their AI-discovered strategies beat the current best methods (like EGPlace and WireMask-EA).
The Improvement: They reduced the total length of the wires (wirelength) by 14% to 34% compared to the previous best methods. In chip design, even a 1% improvement is huge; 34% is a massive leap.

5. The "Secret Sauce"

The AI found that the best strategies weren't static (fixed rules). Instead, they were dynamic.

Static Rule: "Always put the biggest block first."
Dynamic Rule (Discovered by AI): "At the start, put the biggest blocks. But as we get halfway through, switch to putting blocks that are connected to the ones we just placed. If a group of blocks is almost fully connected, finish that group immediately."

The AI learned to adapt its strategy based on how the "packing" was progressing, much like a skilled packer who changes their approach as the truck fills up.

Summary

The paper claims that the order in which you place components on a chip is a hidden lever for optimization that has been ignored. By using an AI to automatically invent and evolve new ordering strategies—using a fast "quick test" to filter out bad ideas—they found a way to pack chips much more efficiently than ever before, significantly reducing the length of wires needed to connect them.

They did not claim this fixes all chip problems or applies to other fields; they specifically showed that optimizing the sequence of placement is a powerful, previously untapped dimension for making better computer chips.

Technical Summary: OrderPlace

Problem Statement

Macro placement is a critical step in modern chip physical design, fundamentally dictating the quality, performance, and manufacturability of the final integrated circuit. Formally, it is a large-scale combinatorial optimization problem requiring the determination of physical positions for macros within a fixed canvas without overlap. While recent machine learning advancements have accelerated coordinate determination, the temporal dimension of placement sequencing (i.e., the order in which macros are placed) remains largely governed by static heuristics or random permutations.

The paper argues that placement sequence is not merely a preprocessing step but a decisive factor in optimization. In sequential placement processes, early decisions define the feasible boundary conditions for all subsequent macros. A single suboptimal placement early in the sequence can trigger an irreversible "domino effect," constraining the solution space and leading to poor local optima. Despite this, existing methods predominantly focus on the spatial question of "Where to place?" while treating the temporal question of "Who goes first?" as a fixed prior. Furthermore, exploring the impact of sequencing is notoriously difficult because the ordering signal is often masked by powerful refinement methods, and evaluating a single sequence typically requires running a full, computationally prohibitive placement cycle.

Methodology: OrderPlace

To address these challenges, the authors propose OrderPlace, a proxy-guided Large Language Model (LLM) evolution framework designed to automatically discover macro placement order strategies. The framework operates through three main components:

1. Initial Placement Order Strategies

The framework initializes a population of strategies to provide a diverse search space. These are categorized into:

Static Strategies: Compute priorities based solely on macro-level geometric and topological features (e.g., area, degree, net connectivity).
Dynamic Strategies: Adapt priorities based on the evolving placement state, often inspired by physical dynamics (e.g., spring potentials, field models, entropy reduction).

2. LLM-Driven Strategy Evolution

OrderPlace utilizes LLMs (e.g., GPT-4, DeepSeek-V3, Claude) to evolve code-level ordering strategies. The process involves:

Prompt Engineering: The LLM is provided with system prompts defining the role of a VLSI expert and user prompts containing evaluation results of top-performing strategies (Top-K) from the previous generation.
Code Generation: The LLM generates new Python classes representing placement strategies, either as static sorters or dynamic priority functions.
Evolutionary Loop: The framework maintains a population of strategies across generations. It selects the top-K strategies based on fitness scores and uses them as context to generate new candidate strategies, fostering the discovery of novel combinations and dynamic mechanisms.

3. Proxy-Guided Evaluation

To mitigate the prohibitive cost of full-scale optimization for every candidate, OrderPlace introduces a lightweight proxy evaluation mechanism:

Wire-Mask-Guided Greedy Procedure: Instead of full optimization, strategies are evaluated using a deterministic greedy placement algorithm guided by precomputed wire-length masks.
Parallel Monte Carlo: Each strategy is executed on multiple random initial seeds in parallel to generate a distribution of Half-Perimeter Wire Length (HPWL) results.
Fitness Scoring: The fitness of a strategy is the mean HPWL over successful runs. A timeout mechanism and syntax validation ensure efficiency and robustness.

Once the evolutionary search concludes, the top-K strategies undergo fine-tuning via the wire-mask-guided greedy procedure to produce the final placement results.

Key Contributions

Perspective Shift: The paper provides theoretical and empirical evidence that placement sequence is a critical, yet historically overlooked, dimension of optimization. It demonstrates that suboptimal ordering leads to asymptotic degradation in solution quality (e.g., $\Omega(l)$ degradation for chain topologies and $\Theta(g)$ for star topologies).
Methodological Innovation: OrderPlace is the first LLM-driven evolutionary framework specifically designed for macro placement sequencing. It successfully integrates proxy evaluation to solve the challenge of expensive feedback loops in floorplanning optimization.
Discovery of Novel Strategies: The framework discovers generalizable sorting algorithms that range from static heuristics to complex, dynamic, physics-inspired rules (e.g., gravitational models, entropy reduction, and net closure prioritization) that outperform manually crafted heuristics.

Experimental Results

The authors evaluated OrderPlace on the standard ISPD 2005 benchmark suite (eight chip designs).

Performance: OrderPlace achieved the best (lowest) HPWL values on 7 out of 8 circuits.
Comparisons:
- Compared to WireMask-EA, OrderPlace reduced wirelength by 34.04%.
- Compared to the state-of-the-art method EGPlace, OrderPlace reduced wirelength by 14.08%.
- In a direct comparison using the Wilcoxon rank-sum test, OrderPlace performed significantly better on 7 circuits compared to EGPlace.
Convergence: OrderPlace demonstrated remarkable computational efficiency, converging significantly faster than EGPlace. The best discovered strategy consistently reached lower convergence points earlier in the optimization trajectory.
Congestion and Mixed-Size: OrderPlace also showed superior routability (congestion) and achieved leading performance in mixed-size placement tasks when combined with a global placer (DreamPlace).
LLM Impact: Ablation studies confirmed that LLM-generated strategies consistently outperformed all hand-crafted initial strategies, indicating that the performance gain stems from the discovery of non-trivial ordering policies beyond existing heuristics.

Significance and Claims

The paper claims that OrderPlace represents the first systematic evidence that placement sequencing, traditionally viewed as a static preprocessing step, serves as a critical determinant in placement optimization. By automating the discovery of ordering strategies via LLMs and proxy evaluation, the method effectively mitigates suboptimality in sequence-decision-based placement.

The authors emphasize that their approach is distinct from previous works that treat sequencing as a fixed prior or rely on RL agents to implicitly learn sequences. Instead, OrderPlace explicitly optimizes the input sequence as a hyperparameter. The work suggests that the optimization potential of placement sequencing has been a "dormant" dimension that, when unlocked, can fundamentally accelerate greedy solvers and unlock better global solutions.

Limitations: The study is currently limited to evaluating the impact of placement order on mask-guided greedy methods. The interaction with stochastic, learning-based methods remains unexplored, and future work is proposed to investigate how placement ordering can enhance deep learning-driven placement methods.