Fast Thermal-Aware Chiplet Placement Assisted by Surrogate

Qinqin Zhang, Xiaoyu Liang, Ning Xu, Yu Chen · 2025

Read Fast Thermal-Aware Chiplet Placement Assisted by Surrogate by Qinqin Zhang, Xiaoyu Liang, Ning Xu, Yu Chen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-03808.txt.

Introduction

Qinqin Zhang and three co-authors speed heat-aware placement of chiplets in stacked packages, in a 2025 paper. Chiplets are small dies packed side by side, a route past the limits of single chips. Placing them to keep wires short and heat low is hard, since heat checks are slow. Their method pairs a classic annealing search with a learned stand-in for the heat field.

The stand-in, a radial basis network, guesses heat maps in a flash. The search then settles on better layouts in less time. Tests show results that rival top heat-aware placers. The authors pitch the tool for fast design cycles in advanced packages.

Catalog record

CREATORQinqin Zhang, Xiaoyu Liang, Ning Xu, Yu Chen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-03808

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Fast Thermal-Aware Chiplet Placement Assisted by Surrogate Qinqin Zhang, Xiaoyu Liang, Ning Xu, Yu Chen Abstract With the advent of the post-Moore era, the 2.5-D advanced package is a promising solution to sustain the development of very large-scale integrated circuits. However, the thermal placement of chiplet, due to the high complexity of thermal simulation, is very challenging. In this paper, a surrogate-assisted simulated annealing algorithm is proposed to simultaneously minimize both the wirelength and the maximum temperature of integrated chips. To alleviate the computational cost of thermal simulation, a radial basis function network is introduced to approximate the thermal field, assisted by which the simulated annealing algorithm converges to the better placement in less time. Numerical results demonstrate that the surrogate-assisted simulated annealing algorithm is competitive to the state-of-the-art thermal placement algorithms of chiplet, suggesting its potential application in the agile design of 2.5D package chip. I Introduction The integrated circuit (IC) has been following Moore’s Law in the first decades of its development. However, with the progress of science and technology and the development of the semiconductor industry, the chip process size has approached the physical limit, which results in a decline in chip yield and a rise in chip cost [ 1 ] . Therefore, IC has begun to enter the ”post-Moore era” guided by non-Moore’s law, and Chiplet technology has emerged. Chiplet technology is a technique that divides a complex monolithic integrated circuit into several smaller chips that can realize sub-functions as required and then integrates these small chips through advanced packaging technology. Nowadays, regarding the packaging forms of chiplets, the most common ones are 2D multi-chip module, 2.5D fan-out wafer-level packaging, 2.5D silicon interposer and 2.5D embedded bridge. The silicon interposer is usually divided into passive interposer and active interposer. There are transistors in the active interposer, and the process technology applied to the active interposer is more complex than that applied to the passive interposer [ 2 ] . Therefore, the production cost of the active interposer is higher, and the yield is lower. Considering cost and other issues, the studies focus on chips based on 2.5D passive interposer integration technology. The implementation of chiplet technology can usually improve the process yield and reduce the cost [ 3 ] . Chiplets can be reused in different products to improve the scalability of chips. In general, chiplet has lower combined costs and higher benefits. However, chiplet technology also faces many challenges, such as thermal effects, electromagnetic challenges, capacitive coupling, inductive coupling, mechanical stress and so on. Among them, the influence of thermal effects is the most prominent [ 1 ] . With increasing demand for processing large amounts of data, the power consumption of the 2.5D system is increasing, so the possibility of thermal problems in the 2.5D system is also increasing. However, the aim of traditional placement methods is to reduce the total wirelength and minimize the area, so the final results of traditional placement algorithms are usually compact [ 2 ] . This also means that it is difficult for traditional placement methods to avoid thermal problems for 2.5D systems with high power density. Therefore, we need to consider how to reduce the maximum operating temperature of 2.5D systems. Due to the high computational complexity of thermal simulation, simultaneous minimization of both wirelength and temperature is time-consuming, making it challenging to design an efficient algorithm for the thermal-ware placement of chiplet. In this paper, we propose a thermal-aware chiplet placement algorithm assisted by surrogate (TACPAs) to optimize the wirelength and reduce the maximum operating temperature of the 2.5D system. The algorithm is mainly based on simulated annealing algorithm (SA). For the thermal simulation of each placement result in the optimization process, we do not choose commercial thermal simulation software to get the maximum temperature of placement, but make use of the idea of surrogate model to predict the maximum temperature by training a radial basis function network (RBFN). Consequently, the proposed TACPAs can converge to optimized placement results in less time. This paper is structured as follows. Related work is reviewed in Section II . Section III presents the details about the proposed placement algorithm, and Section IV presents the experimental results with analysis. Funially, Section V summarizes this paper. II Related Work Nowadays, an increasing number of people have started to pay attention to Chiplet. Li et al. [ 4 ] proposed a reusable general interposer architecture to amortize non-recurring engineering costs and it is capable of significantly speeding up integration flows of interposers across different chiplet-based systems. Zhi Li et al. [ 5 ] combined the differential evolution (DE) and the particle swarm optimization (PSO) algorithms with the Metropolis rule that can jump out of the local optimum to optimize the impedance of chiplet-based 2.5D integrated circuits. Raikar and Stroobandt [ 6 ] developed a modular placement algorithm that has achieved a very good balance between runtime efficiency and placement quality. These studies have carried out optimizations on the placement of the chiplet-based systems from various perspectives, but the thermal effect of chiplet is not considered. Given the substantial impact of thermal effects on the chiplet-based systems, numerous placement algorithms have emerged recently with the aim of lowering the maximum operating temperature of the chiplet-based systems. Ma et al. [ 2 ] presented the TAP-2.5D algorithm to jointly minimize the temperature and total wirelength, and it is the first open-source network routing and thermally-aware chiplet placement methodology for heterogeneous 2.5D systems. The SP-CP algorithm proposed in [ 7 ] combines the sequence-pair based tree, branch-and-bound method, and advanced placement/pruning techniques. It can find the result fast with the optimized total wirelength. It also proposes a post placement procedure to lower the maximum operating temperature of 2.5D integrated circuits. Duan et al. [ 8 ] presented an efficient early-stage floorplanning tool named RLPlanner for chiplet-based systems. Making use of the advanced reinforcement learning, the RLPlanner can jointly minimize the wirelength and the maximum operating temperature. Based on the RLPlanner, Deng et al. [ 9 ] proposed a learning to rank approach with graph representation to choose the best chiplet placement order for each chiplet-based system. Actually, it achieves a 10.05\% reduction in wirelength and a 1.01\% improvement in peak system temperature. In [ 10 ] , a reinforcement learning (RL) framework is employed to reduce thermal hotspots, improve heat dissipation and enhance system performance. Some pioneering research has been done on fast thermal analysis of chiplet-based systems. The HotSpot compact thermal modeling approach in [ 11 ] is able to provide detailed static and transient temperature information and it is computationally efficient. Chen et al. [ 12 ] deployed a novel graph convolutional networks (GCN) architecture to generate the thermal map of the chiplet-based systems and the time it takes to run is less than that of HotSpot. Wang et al. [ 13 ] introduced a multiscale anisotropic thermal model, which takes the feature-scale thermal conductivities of different materials into account to predict the package-scale steady-state temperature fields. III Thermal-Aware Chiplet Placement Assisted by Surrogate III-A Surrogate Approximation of Thermal Map Via the Radial Basis Function Network Commercial thermal simulation software usually outputs the thermal map of the entire 2.5D system by solving

Read the complete text