AI优化芯片布局，设计阶段即考虑最终性能，中科大华为诺亚新方法入选ICLR 2025 Oral

Core Insights - The article discusses the introduction of LaMPlace, a new chip macro placement optimization method developed by a team from the University of Science and Technology of China, Huawei Noah's Ark Lab, and Tianjin University, which utilizes AI to enhance chip design efficiency [1][2]. Group 1: LaMPlace Overview - LaMPlace allows for performance considerations during the placement phase, potentially speeding up the design process and improving overall efficiency [1]. - The method aims to facilitate the intelligentization of domestic EDA tools and accelerate design processes, promoting a trend of "early optimization" in the chip design industry [2]. Group 2: Methodology - The methodology involves a shift from "optimizable" to "should optimize" in EDA goals, focusing on macro placement as a critical step in physical design that influences timing performance, power consumption, and area (PPA) [4]. - LaMPlace introduces a structured predictor to directly connect layout decisions with final performance metrics, enabling a new paradigm of "left-shift optimization" in chip design [5]. Group 3: Key Steps in LaMPlace - The optimization process consists of three main steps: 1. A structured predictor is trained offline to estimate multiple cross-stage metrics based on the current macro unit layout [8]. 2. The metrics are modeled as a Laurent polynomial, allowing for efficient computation and structural interpretability [12]. 3. A learnable mask is generated to guide the greedy placement strategy based on the predicted metrics [14]. Group 4: Technical Features - LaMPlace's modular design allows it to be integrated into existing systems as a plug-in module, enhancing various layout optimization paradigms [15]. - Experimental results demonstrate LaMPlace's superior performance on standard chip layout benchmarks, showing improved timing convergence and congestion metrics while maintaining stability on unseen designs [16][17]. Group 5: Implications for AI in Chip Design - LaMPlace represents an effective paradigm combining structural modeling and learning guidance, providing new insights for the deeper application of AI in chip design [18].