Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Novel Frequency-Spatial Domain Aware Network for Fast Thermal Prediction in 2.5D ICs

Dekang Zhang, Dan Niu, Zhou Jin, Yichao Dong, Jingweijia Tan, Changyin Sun

TL;DR

This work tackles fast, accurate thermal prediction for 2.5D chiplet ICs by introducing FSA-Heat, a frequency-spatial dual-domain network that captures global-to-local thermal dissipation features. It combines a high-to-low frequency spatial encoder (FSTE) with a frequency-domain cross-scale interaction module (FCIFormer) and trains with a frequency-spatial hybrid loss to suppress high-frequency gradient noise. The method demonstrates substantial performance gains over GCN-based baselines and conventional solvers, including over 99% RMSE reduction and a 4.23× inference-time speedup, while showing strong generalization to unseen designs. The approach leverages 3D-DCT/IDCT, cross-scale frequency attention, and a meshless end-to-end pipeline to deliver accurate thermal and gradient predictions for complex 2.5D chiplet configurations, enabling faster thermal-aware design optimization.

Abstract

In the post-Moore era, 2.5D chiplet-based ICs present significant challenges in thermal management due to increased power density and thermal hotspots. Neural network-based thermal prediction models can perform real-time predictions for many unseen new designs. However, existing CNN-based and GCN-based methods cannot effectively capture the global thermal features, especially for high-frequency components, hindering prediction accuracy enhancement. In this paper, we propose a novel frequency-spatial dual domain aware prediction network (FSA-Heat) for fast and high-accuracy thermal prediction in 2.5D ICs. It integrates high-to-low frequency and spatial domain encoder (FSTE) module with frequency domain cross-scale interaction module (FCIFormer) to achieve high-to-low frequency and global-to-local thermal dissipation feature extraction. Additionally, a frequency-spatial hybrid loss (FSL) is designed to effectively attenuate high-frequency thermal gradient noise and spatial misalignments. The experimental results show that the performance enhancements offered by our proposed method are substantial, outperforming the newly-proposed 2.5D method, GCN+PNA, by considerable margins (over 99% RMSE reduction, 4.23X inference time speedup). Moreover, extensive experiments demonstrate that FSA-Heat also exhibits robust generalization capabilities.

A Novel Frequency-Spatial Domain Aware Network for Fast Thermal Prediction in 2.5D ICs

TL;DR

This work tackles fast, accurate thermal prediction for 2.5D chiplet ICs by introducing FSA-Heat, a frequency-spatial dual-domain network that captures global-to-local thermal dissipation features. It combines a high-to-low frequency spatial encoder (FSTE) with a frequency-domain cross-scale interaction module (FCIFormer) and trains with a frequency-spatial hybrid loss to suppress high-frequency gradient noise. The method demonstrates substantial performance gains over GCN-based baselines and conventional solvers, including over 99% RMSE reduction and a 4.23× inference-time speedup, while showing strong generalization to unseen designs. The approach leverages 3D-DCT/IDCT, cross-scale frequency attention, and a meshless end-to-end pipeline to deliver accurate thermal and gradient predictions for complex 2.5D chiplet configurations, enabling faster thermal-aware design optimization.

Abstract

In the post-Moore era, 2.5D chiplet-based ICs present significant challenges in thermal management due to increased power density and thermal hotspots. Neural network-based thermal prediction models can perform real-time predictions for many unseen new designs. However, existing CNN-based and GCN-based methods cannot effectively capture the global thermal features, especially for high-frequency components, hindering prediction accuracy enhancement. In this paper, we propose a novel frequency-spatial dual domain aware prediction network (FSA-Heat) for fast and high-accuracy thermal prediction in 2.5D ICs. It integrates high-to-low frequency and spatial domain encoder (FSTE) module with frequency domain cross-scale interaction module (FCIFormer) to achieve high-to-low frequency and global-to-local thermal dissipation feature extraction. Additionally, a frequency-spatial hybrid loss (FSL) is designed to effectively attenuate high-frequency thermal gradient noise and spatial misalignments. The experimental results show that the performance enhancements offered by our proposed method are substantial, outperforming the newly-proposed 2.5D method, GCN+PNA, by considerable margins (over 99% RMSE reduction, 4.23X inference time speedup). Moreover, extensive experiments demonstrate that FSA-Heat also exhibits robust generalization capabilities.

Paper Structure

This paper contains 14 sections, 7 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Methodology
  4. Model Inputs and PPNet
  5. Frequency-spatial aware prediction network: FSA-Heat
  6. Frequency-spatial domain thermal encoder: FSTE
  7. Frequency domain cross-scale interaction module: FCIFormer
  8. Decoder
  9. Frequency-spatial hybrid loss (FSL)
  10. Experiments and Results
  11. Experiments setting
  12. Accuracy, speed and generalization comparisons
  13. Ablation Experiments
  14. Conclusions

Figures (5)

  • Figure 1: Proposed end-to-end chiplet thermal and its gradient prediction network (FSA-Heat) and detailed structures of each module.
  • Figure 2: A visualization example of thermal and its gradient prediction comparisons. (a)–(d) represents the heat source, TIM, heat spreader, and heatsink layer, respectively.
  • Figure 3: Thermal and its gradient prediction comparisons between ours and GCN+PNA under unseen thermal conductivity $k_i$.
  • Figure 4: Thermal prediction comparisons between ours and GCN+PNA under unseen varying heat source numbers.
  • Figure 5: Ablation method error visualizations.