The drive toward heterogeneous 2.5D chiplet packages promises unmatched performance and customization, yet the resulting density amplifies thermal challenges. To facilitate rapid design space exploration, we present ThermLeT, a multimodal transformer-based, scale invariant surrogate that integrates power density maps, spatial coordinate embeddings, material properties and heat transfer rate to predict steady state temperature in 13 ms, about four orders of magnitude faster than conventional finite element solvers such as ANSYS. To train the transformer model, we also create a dataset of 6,221 chiplet layouts perturbed from three representative architectures. On a 150 sample test set, ThermLeT achieves an RMSE of 3.16°C and an MAE of 2.24°C, improving accuracy by 35% over ThermEDGe and by 71% over HotSpot, while running approximately 6x and 11x lower inference latencies respectively. Ablation studies confirm robust generalization to unseen floorplans. Code and dataset available at https://github.com/MicroSystemsDesignLab/ThermLeT
Figures & tables
Fig. 1: Design parameter perturbations in 2.5D chiplet thermal analysis. Four categories of perturbations: spatial, power, material, and boundary conditions require complete thermal recalculation. This iterative evaluation process, while essential for design space exploration, becomes computationally prohibitive when evaluating thousands of configurations using traditional finite element methods.
Related Work
Input
Approach
Steady-State or Transient
ANSYS
Geometry Modeling
FEM/CFD
Both
HotSpot [ 13 ]
Layout, BC
FEM
Both
DeepOHeat [ 6 ]
Layout, BC, Power Map
Operator Learning ML
Transient
CoMeT [ 14 ]
Layout, BC, Power Map
FEM
Both
U-Net [ 15 ]
Layout, Power Map
Image-to-image and PINN
Steady-State
ThermEDGe [ 12 ]
Layout, Power Map
Encoder-Decoder Network
Both
TABLE I: Summary of related works with their inputs, approach, and simulation type.
Components
Materials
Thermal Conductivity (w/m.k)
CTE (ppm/°C)
Young’s Modulus (gpa)
Substrate
FR-4
0.3
13.0
20
C4 Bump
Tin-Lead (60-40)
50
20.5
20
μ Bump
SAC (SnAgCu)
50
20
50
Interposer
Silicon Anisotropic
148
2.6
150
Chiplets
Silicon Anisotropic
150
3.1
130
TIM
Indium
86
29
10
TABLE II: Material Properties for Key Components
Fig. 2: Cross-sectional schematic of the 2.5D chiplet integration stack. The structure includes key components such as C4 bumps, thermal interface materials (TIM), underfill, interposer, microbumps, chiplets, and a heatsink. This model serves as the foundation for finite element analysis (FEA) simulations
Multi-GPU
Micro150
Ascend 910
Chiplet
CPU
GPU
HBM
CPU
DRAM
Virtuian
Nimbus
HBM
Widths [mm]
12
18.2
7.75
8.25
8.75
31.4
10.5
7.75
Height [mm]
12
18.2
11.87
9
8.75
14.5
16
11.87
Power [W]
105
295
20
150
20
256
14
20
TABLE III: Chiplet dimensions and powers in 2.5D examples.
Fig. 3: Overview of the proposed ThermLeT architecture. The model takes three input modalities: power map, spatial coordinates, and material properties and embeds them using dedicated modules. These embeddings are fused and passed through a Transformer encoder and convolutional decoding blocks to predict the final temperature map. The design enables the model to capture spatial and physical dependencies for accurate thermal prediction across diverse chiplet architectures.
Fig. 4: Visual comaprision of temperature predictions: (a) Input Power map with different color representing different power values shown III (b) Ground truth generated using Ansys simulation (c,d) Temperature predictions by baseline ThermEDGe [ 12 ] and Hotspot [ 13 ] (e) Proposed transformer based temperature predictor.
Method
RMSE ↓
MAE ↓
ThermEDGe [ 12 ]
4.893
3.691
HotSpot [ 13 ]
10.734
9.886
ThermLeT (Power Only)
4.1186
3.307
ThermLeT (Ours)
3.160
2.238
TABLE IV: Evaluation metrics on the test set of 150 samples. Lower values indicate better performance.
Model
Latency
Throughput
FLOPs
(s/sample)
(samples/s)
(GFLOPs)
Ansys steady-state thermal
95.000*
-
ThermEDGe [ 12 ]
0.0762
13.11
11.42
HotSpot [ 13 ]
0.1442*
-
-
ThermLeT (Ours)
0.0130
76.77
155.79
TABLE V: Latency, throughput, and FLOPs for ThermLeT and baselines on the full dataset.
Input Modalities
RMSE ↓
MAE ↓
Power + Coord + MatProp (Full Model)
3.160
2.238
Power + MatProp
5.424
4.271
Power + Coord
6.162
4.858
TABLE VI: RMSE and MAE for different input combinations. The full model uses power, coordinates, and material properties.
Method
RMSE ↓
MAE ↓
ThermEDGe [ 12 ]
9.6493
7.6973
ThermLeT (Ours)
8.908
7.3745
TABLE VII: RMSE and MAE on unseen chiplet architectures (Ascend910 + Micro150) after training only on MultiGPU.
Runtime thermal management of high-performance chips depends on fast and accurate full-chip thermal maps. Conventional simulators typically estimate power traces from performance metrics first, which adds overhead. This work proposes TherMapNet, an attention-guided thermal simulator that predicts full-chip thermal maps directly from performance metrics. A Transformer encoder captures temporal evolution by treating the time series of each metric as a token, improving modeling of dynamic workloads. A CNN then extracts fine-grained spatial features. For the CNN, a dual-branch channel-spatial attention convolution module (DACM) and a triplet loss are used to improve spatial learning and reconstruction accuracy. TherMapNet is applied to a multi-core CPU (AMD Ryzen 7 4800U) and a many-core GPU (NVIDIA GeForce RTX 4060). Experiments show that it outperforms prior thermal simulators, with RMSE below 0.26 C and inference under 2.4 ms on an NVIDIA GeForce RTX 3090 GPU. These results indicate that TherMapNet can support high-quality runtime thermal management of modern multi-core chips.
Qin Gu, Chaofang Ma, Mingyu Yang +4
College of Information Science and Engineering, Northeastern University, Shenyang, China · Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology, Hong Kong
Transformer-based surrogate models are increasingly used to replace expensive first-principles simulation in engineering design. But conventional transformer architectures are often over parameterized for the small, low-dimensional datasets typical of engineering design spaces, where large simulation data is expensive to generate. Under these conditions, excess parameter capacity leads to overfitting rather than improved accuracy, while also incurring unnecessary memory and compute overhead. This motivates a shift towards architectures that focus on additional compute rather than additional learnable parameters. This paper presents a hardware-aware evaluation of three recursive transformer paradigms for surrogate thermo-mechanical analysis of advanced packages: a)Tiny Recursive Model, b) our proposed Depth Recursive transformer, c) and a simple recursive transformer. We systematically compare their predictive performance (Recall, Mean Reciprocal Rank), parameter count, computational complexity (FLOPs), providing practical design guidelines for selecting recursive transformer architectures under resource-constrained scenarios. We validate this principle on two low-dimensional engineering prediction tasks: 1) thermo-mechanical reliability analysis of advanced semiconductor packages, where stress and warpage from thermal cycling must be evaluated repeatedly across a design-of-experiments sweep under costly finite element analysis (FEA). 2) Laplace PDE iterative numerical solver for capacitance field. Overall, recursive weight-sharing transformers provide an effective and generalizable trade-off between prediction accuracy, parameter efficiency, and computational cost for small data engineering surrogate modeling.
Kart-leong Lim
Institute of Microelectronics (IME), Agency for Science, Technology and Research (A*STAR), Singapore
Accurate thermal analysis of heterogeneous 2.5D/3D-IC packages is essential yet computationally prohibitive. A single full-package FEM simulation can take hours, while AI-based surrogates treat the entire stack as a monolithic prediction target and must be retrained whenever the die count or topology changes. To address this limitation, this work proposes Domain-Decomposed AI-Accelerated Iterative Solver for Thermal Analysis (DAIST), a composable thermal solver that decomposes the global package simulation into block-level subdomain problems, replaces subdomain solvers with neural operators, and couples them through iterative exchanges of interfacial temperature and heat flux. This local-to-global architecture eliminates the topology lock-in of monolithic models: block-level neural operators can be directly reused in unseen package assemblies without retraining. The iterative coupling strategy further provides a controllable accuracy-runtime tradeoff, where the iteration budget can be adjusted to trade accuracy for runtime. Evaluated on a multi-chiplet system and an advanced packaging system, DAIST achieves up to 178× speedup over traditional FEM solvers with mean temperature errors of 0.068% and 0.323%, respectively, while demonstrating cross-topology reuse of block-level models across structurally distinct package assemblies.