PaperlayerView exact revision on arXiv

Source-linked AI summary

TherMapNet Attention-Guided Runtime Full-Chip Thermal Map Prediction from Performance Metrics

Qin Gu, Chaofang Ma, Mingyu Yang, Yipu Zhang, Jiliang Zhang, Wei Zhang, Lin Jiang

arXiv:2608.21887v1cs.ARcs.LG

TL;DR

Runtime thermal management needs fast, accurate full-chip thermal maps, but conventional approaches add overhead and existing neural simulators have accuracy or dynamic-modeling limitations. TherMapNet predicts maps directly from performance metrics by combining metric-centric Transformer temporal modeling with CNN spatial refinement using DACM and triplet loss. It outperforms compared methods on CPU and GPU datasets while maintaining millisecond-level inference.

  • Problem

    Modern chip thermal management requires fast and accurate temperature profiles, while existing simulators face accuracy, efficiency, or dynamic-data limitations.

  • Method

    TherMapNet directly predicts thermal maps from performance metrics using metric-centric Transformer temporal modeling and CNN spatial refinement with DACM and triplet loss.

  • Results

    TherMapNet achieves the best overall thermal estimation accuracy among compared methods, with mean RMSE of 0.251°C on AMD R7 4800U and 0.165°C on NVIDIA RTX 4060.

  • Takeaways & Limitations

    TherMapNet provides accurate and efficient full-chip thermal-map estimation directly from performance metrics, supporting real-time thermal management.

Abstract

from arXiv · show

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.

I. INTRODUCTION

Modern chips face thermal challenges that require fast, accurate temperature profiles, while existing simulators struggle to provide both accuracy and efficiency. TherMapNet addresses this by predicting thermal maps directly from performance metrics with transformer-based temporal modeling and CNN-based spatial refinement.

  • High chip temperatures degrade performance, impair reliability, and raise power consumption, motivating fast and accurate thermal estimation.
  • Existing thermal simulators generally cannot deliver high accuracy and high efficiency simultaneously, with efficient thermal circuit models sacrificing accuracy.
  • TherMapNet directly predicts chip temperature profiles from performance metrics collected by on-chip monitors using a cooperative transformer–CNN architecture.
  • Representing each metric’s time series as a token enables the transformer to capture temporal correlations and dynamic variations more effectively.
  • DACM extracts and fuses differentiated channel–spatial features through dual branches, while triplet loss improves thermal-map reconstruction accuracy.

II. RELATED WORKS

Prior chip thermal methods trade accuracy, efficiency, or dynamic modeling capability. TherMapNet combines transformer-based temporal processing with dual-branch CNN spatial refinement to estimate thermal maps directly from performance metrics.

  • DNS methods provide high accuracy but are computationally prohibitive for real-time chip thermal estimation because they require extremely fine meshes.
  • Thermal circuit models and Green’s-function methods improve efficiency but trade accuracy for computational speed; some thermal circuit errors exceed 200% versus FEM.
  • POD-based methods reduce thermal problems to low-dimensional spaces and achieve speedups of up to three orders of magnitude while maintaining least-squares errors below 1%.
  • Neural thermal simulators use real-time utilization and monitoring data, but LSTM-based RealMaps struggles with long-range dependencies and cumulative transient-simulation errors.
  • TherMapNet organizes performance metrics into a matrix, uses a Transformer to generate a preliminary map, and applies dual-branch CNN refinement before adaptive fusion.

B. Absolute and Relative Positional Channels Enhanced Transformer

TherMapNet enhances its Transformer by modeling both absolute and relative temporal information, while representing each metric’s time series as a dedicated token. Absolute positional information is concatenated with performance metrics before token embedding.

  • Positional Channels: Absolute and relative positional channels encode temporal order and pairwise temporal relationships in the Transformer encoder.The absolute channel is introduced at the input, while the relative channel is integrated into self-attention.
  • Metric-Centric Tokenization: TherMapNet treats each performance metric’s time series across multiple frames as an individual token, unlike approaches that combine all metrics at one frame.This metric-centric tokenization targets metric-specific temporal patterns and fine-grained thermal dynamics.
  • Absolute Positional Channel: The sinusoidal absolute positional encoding uses a cycle length determined by the performance-metric sampling period and matches normalized input-feature scale.Its values are constrained to [−1, 1], providing temporal order without numerical imbalance among feature dimensions.
  • Absolute Positional Channel: The encoded absolute positional information is concatenated with the collected performance metrics before the resulting matrix is transposed for dedicated metric tokens.The concatenated representation is then used for token embedding.

2) Relative Positional Channel:

TherMapNet’s relative positional channel augments standard multi-head self-attention with learnable encodings of pairwise temporal relationships. These encodings modify key and value representations so attention can model relative timing between tokens.

  • Motivation: Absolute positional information alone records time-frame locations but does not directly represent relative temporal distances between tokens.Adjacent tokens may have stronger correlations than tokens separated by larger intervals, regardless of absolute position.
  • Self-Attention: TherMapNet uses standard multi-head self-attention, projecting each input vector into query, key, and value vectors.The input sequence is mapped to an output sequence through these attention components.
  • Relative Position Encoding: The relative positional channel introduces learnable relative position encodings into both key and value representations.This design explicitly incorporates pairwise temporal relationships into self-attention.
  • Relative Position Encoding: Relative-distance clipping bounds the encoded distance to [−k, k], limiting parameter growth while preserving local temporal relationships.The clipped relative distance is used to define the relative key and value embeddings.
  • Self-Attention: The self-attention output aggregates value vectors together with their corresponding relative position encodings.This combines content information with relative temporal information in the output.

C. Dual Branch Channel Spatial Attention Convolution Module (DACM)

The DACM refines the Transformer-generated preliminary thermal map by extracting complementary spatial representations through two attention-equipped CNN branches and adaptively fusing them.

  • C. Dual Branch Channel Spatial Attention Convolution Module (DACM): The Transformer encoder models temporal dependencies among performance metrics and generates a preliminary thermal map, Hpre.
  • C. Dual Branch Channel Spatial Attention Convolution Module (DACM): The preliminary map is processed by convolutional layers and two parallel CNN branches that learn complementary high-level spatial representations.
  • C. Dual Branch Channel Spatial Attention Convolution Module (DACM): CBAM applies sequential channel and spatial attention in both branches to emphasize thermally relevant regions and suppress redundant information.
  • C. Dual Branch Channel Spatial Attention Convolution Module (DACM): The two branch outputs are fused through spatially adaptive weighting to generate the final thermal map Hmap.
  • C. Dual Branch Channel Spatial Attention Convolution Module (DACM): Adaptive fusion dynamically exploits each branch’s strengths at different spatial locations, leading to more accurate thermal estimation.

D. Triplet Loss

TherMapNet uses triplet-loss constraints on branch residual maps to encourage complementary spatial representations while jointly training for thermal-map reconstruction accuracy.

  • D. Triplet Loss: The dual branches may learn similar feature spaces, so triplet-loss constraints explicitly encourage feature diversity between their spatial representations.
  • D. Triplet Loss: Triplets are constructed from each branch’s residual map relative to the measured thermal map rather than directly from feature embeddings.
  • D. Triplet Loss: An ideal zero-error residual serves as the anchor, while the two branch residuals serve as positive and negative samples for Branch 1’s triplet loss.
  • D. Triplet Loss: The margin α enforces minimum separation between positive and negative representations, encouraging each branch toward the ideal state while remaining distinct from the other branch.
  • D. Triplet Loss: RMSE loss is employed alongside triplet loss during training, with α and β adjusting their respective contributions in the overall objective.

IV. DEMONSTRATION AND VALIDATION

TherMapNet is demonstrated for real-time thermal estimation on an eight-core AMD CPU and a many-core NVIDIA GPU, with accuracy evaluated against infrared measurements and neural-network simulators.

  • IV. DEMONSTRATION AND VALIDATION: TherMapNet is applied to the AMD Ryzen 7 4800U multi-core CPU and NVIDIA GeForce RTX 4060 many-core GPU.
  • IV. DEMONSTRATION AND VALIDATION: The evaluation compares predicted thermal maps with thermal infrared measurements and conducts comprehensive comparisons with state-of-the-art neural-network thermal simulators.

A. Experimental setup

The experiments use public CPU and GPU datasets pairing time-series performance metrics with thermal maps, including infrared-captured maps and preprocessing for GPU thermal artifacts.

  • A. Experimental setup: The public datasets pair time-series performance metrics with corresponding thermal maps collected from representative benchmark applications.
  • A. Experimental setup: Performance metrics are collected with AMD uProf 4.0 and NVIDIA System Management Interface, while thermal maps are captured using a thermal infrared camera.
  • A. Experimental setup: 14,718 AMD R7 4800U samples include 158 performance metrics and 223 × 280-pixel thermal maps, with 11,774 samples used for training.
  • A. Experimental setup: For three AMD R7 4800U time steps, estimated-map RMSE values are 0.071, 0.104, and 0.062 °C.
  • A. Experimental setup: GPU thermal maps are cropped to 183 × 205 pixels, and printed text and logos are treated as missing thermal data because they introduce thermal artifacts.

B. Validation Against IR Measurements

TherMapNet’s predicted thermal maps closely agree with infrared measurements on both evaluated chips. Across 200 time steps, it maintains accurate transient predictions without noticeable error accumulation or fluctuation.

  • TherMapNet’s predicted thermal maps show strong agreement with infrared measurements and small absolute errors on the AMD Ryzen 7 4800U and NVIDIA RTX 4060.Figures 6 and 7 compare estimated maps, measured maps, and absolute error maps across time steps.
  • TherMapNet maintains consistently high estimation accuracy over 200 consecutive time steps during dynamic heating.The temporal comparisons show no noticeable error accumulation or fluctuation over time.
  • 0.027, 0.140, and 0.006 °C are the RMSE values for three NVIDIA RTX 4060 time steps.
  • Figure 8 compares temporal temperature evolutions at representative pixel locations on the AMD Ryzen 7 4800U and NVIDIA GeForce RTX 4060.

C. Comparisons with State-of-the-Art NN-Based Simulators

Against state-of-the-art neural thermal simulators, TherMapNet achieves the best overall accuracy while retaining millisecond-level inference. Ablation studies further examine positional encoding and other architectural components.

  • TherMapNet achieves the best overall thermal estimation accuracy among the compared neural simulators on both datasets.RMSE is calculated across corresponding predicted and measured thermal-map pixels.
  • 0.251 °C is TherMapNet’s mean RMSE on AMD R7 4800U, with 30.3%, 57.9%, and 88.5% reductions versus ThermTransformer, ThermGAN, and RealMaps.Its maximum RMSE is 1.429 °C and RMSE standard deviation is 0.185 °C.
  • 13.2% and 39.4% are TherMapNet’s reductions in mean RMSE and RMSE standard deviation versus GPUThermalMap on NVIDIA RTX 4060.TherMapNet’s mean RMSE on this dataset is 0.165 °C.
  • 2.32 ms and 1.95 ms are TherMapNet’s inference latencies on AMD R7 4800U and NVIDIA RTX 4060, respectively.The paper reports a trade-off between substantially improved accuracy and near-real-time inference capability.
  • Positional Encoding: Removing positional encoding produces the largest RMSE on both datasets, while combining absolute and relative encoding achieves the best performance.The ablation evaluates different positional encoding strategies.

2) Token Representation:

TherMapNet’s metric-centric token representation and Transformer–CNN architecture are evaluated through ablations. The results support modeling temporal dependencies with metric time series and spatial characteristics with dual-branch CNN processing.

  • 2) Token Representation: Treating each metric’s complete time series as a token improves temporal-dependency modeling over treating all metrics at one instant as a token.The conventional representation causes degraded accuracy in the ablation study.
  • 3) Model Architecture: The Transformer-only model has the largest RMSE, while adding CNN branches improves accuracy by enhancing local spatial feature extraction.
  • 3) Model Architecture: 0.251 °C and 0.165 °C are the best RMSE values achieved by the dual-branch CNN architecture on AMD R7 4800U and NVIDIA RTX 4060.The improvement supports the DACM design’s ability to capture complementary spatial features.
  • Conclusion: TherMapNet combines metric-centric tokenization, positional encoding, a Transformer, and a DACM-enhanced CNN with triplet loss for full-chip thermal-map estimation.
  • Conclusion: Ablation studies verify contributions from positional encoding, metric-centric token representation, and the cooperative Transformer–CNN architecture.
Loading 2608.21887v1…