Skip to main content
Back

Huawei's EIM Learns Chip Design From Experts, Cuts Routing Congestion by 92%

A new framework trains RL agents for chip placement by learning reward signals directly from expert layouts, closing the quality gap with human designers

Pro

Paper, Full Access
  • Core idea: EIM learns reward signals for RL-based chip placement directly from finished expert layouts, bypassing manual reward engineering.
  • Two variants: EIM-D (inverse RL from demonstrations) generalizes better; EIM-P (RLHF-style preferences) fits seen designs more precisely but overfits.
  • Key results: EIM-D cuts routed congestion overflow by up to 75-92% and wirelength by ~15% vs. analytical baselines on unseen ICCAD 2015 chips.
  • Data efficiency: The reward model trains on as few as 50 layouts from a single chip design and still generalizes to unseen designs without retraining.
  • Benchmarks: Tested on ICCAD 2015 (8 designs) and OpenROAD-flow-scripts (6 RTL designs), outperforming MaskPlace, EfficientPlace, and DREAMPlace 4.1.0.
  • No code release mentioned; paper accepted at DAC 2026.

Chip placement -- deciding where to physically position memory blocks and logic gates on a silicon die -- is one of the most consequential steps in chip design. Get it wrong and you pay in power, performance, and area (PPA) for the entire lifetime of the chip. Chip placement determines the physical positions of macros (e.g., memory blocks) and standard cells (e.g., logic gates) on the chip canvas, and macro placement has a great impact on the final power, performance, and area (PPA) metrics of the circuit. Reinforcement learning has been chasing human expert quality for years -- and a new paper from Nanjing University and Huawei Noah's Ark Lab argues the field has been solving the wrong problem.

While RL-based methods have recently emerged, their training primarily focuses on wirelength optimization, and therefore often fail to achieve expert-quality layouts. The new paper, accepted at DAC 2026, introduces a framework called the Expert Imitation Model (EIM) that sidesteps reward engineering entirely -- instead learning what "good" looks like directly from finished expert layouts.

Why Wirelength Isn't Enough

To understand why this matters, you need to understand what existing RL-based chip placers actually optimize. The dominant approach, pioneered by Google's AlphaChip, treats macro placement as a Markov Decision Process (MDP): an agent places one macro at a time on a grid, collecting a reward signal at each step. AlphaChip receives no reward feedback until all macros have been placed, making the reward signal sparse for policy training; MaskPlace solves this problem by introducing a dense macro half-perimeter wirelength (HPWL) reward.

HPWL (half-perimeter wirelength) is a proxy metric -- it estimates how long the wires connecting components will be by measuring the half-perimeter of the bounding box around each net. It's fast to compute, but it's a simplification. The quality of macro placement is influenced by various factors, such as wirelength, routability, macro halo, power delivery, and electrical characteristics. How these factors affect the final PPA metrics remains an open problem. Therefore, an inadequately crafted reward function may prevent existing RL-based methods from achieving expert-level results.

Human experts, by contrast, apply tacit knowledge that resists formalization. Much of the expertise is inherently difficult to formalize and integrate into learning-based frameworks. For example, macro regularity in terms of dead space management and notch avoidance is crucial but difficult to quantify. Simplified formulations often fail to accurately describe the complex nature of such regularity. The EIM paper's key insight: rather than trying to write down all that expertise as a reward function, just learn it from the outputs experts already produce.