PaperlayerView exact revision on arXiv

Source-linked AI summary

AI-driven Thermal-aware Data Center Capacity Planning

Yixing Li, Mark Fenton, Matthew Kaufeler, Ka Ming Leung, Xin Ai, Zhiyu Zeng

arXiv:2610.02442v1cs.LG

TL;DR

LLM workloads create localized and transient heat loads, while full CFD is costly and slow for thermal-aware planning. This paper trains and integrates an AI surrogate that predicts cabinet-inlet temperatures from data-center inputs. The authors report millisecond inference, more than 100,000× CFD speedup, and preservation of relative temperature ranking across cabinets.

  • Problem

    LLM GPU workloads create localized hotspots and transient thermal spikes, while high-fidelity CFD can require hours or days per configuration, limiting real-time management and rapid exploration.

    Source: §1 · p. 4

  • Method

    A U-Net surrogate learns cabinet-inlet temperature fields from simulation inputs and CFD-derived inlet outputs, and is integrated into an automated digital-twin workflow.

    Source: §1 · p. 4, §2.2 · p. 7, §2.4 · p. 7

  • Results

    More than 100,000× speedup over full-scale CFD is reported, with millisecond inference, reasonable accuracy, and preserved relative temperature ranking across configurations.

    Source: §5 · p. 13

  • Takeaways & Limitations

    Preserving cabinet temperature ranking helps identify comparatively stressed racks and provides guidance for workload redistribution and thermal-risk mitigation.

    Source: §3.2 · p. 10

    The maximum absolute error for the absolute temperature field remained relatively large.

    Source: §3.2 · p. 9

Abstract

from arXiv · show

The emerging of large language models (LLMs) has posed significant challenges to the thermal management of data center. Intense GPU computation for LLMs results in localized hotspots. Moreover, spiking thermal loads during training and inference bursts make real-time cooling response more difficult to predict and control. Thermal-aware capacity planning of data center requires massive expensive high-fidelity CFD simulations. AI models can perform real-time prediction for unseen designs. However, existing works either have large prediction error, or have over-simplified assumptions for data center operations. This work presents an AI-driven framework that can perform thermal-aware capacity planning for a real-world data center in seconds. The embedded AI model learns from numerous key parameters (rack power, server power, server placement, HVAC settings etc.), and provides temperature prediction within milliseconds. This AI model is tested against high-fidelity CFD simulations, and results show that for unseen data center designs, model can achieve high accuracy with 10000X speedup. Driven by the AI model, the authors design the thermal-aware capacity planning framework. This framework can help data center designers and operators instantaneously optimize both workload distribution and HVAC cooling efficiency.

Author(s) Biography

The biographies describe six Cadence engineers and product leaders with experience spanning CFD, thermal software, data-center design, and AI.

  • Yixing Li combines CFD and FEA expertise with AI work and has led production-grade AI initiatives at Cadence.Source: Author(s) Biography · p. 3
  • Mark Fenton has over 20 years of CFD experience in data-center design and operations and leads product engineering for the Reality Digital Twin Platform.Source: Author(s) Biography · p. 3
  • Matthew is a Principal Product Engineer for Reality DC Digital Twin, with a mechanical engineering degree and experience in cooling-hardware R&D, product development, and technical sales.Source: Author(s) Biography · p. 3
  • Ka Ming Leung leads engineering initiatives and RealityDC and CelsiusEC product lines, and initiated RealityDC Insight.Source: Author(s) Biography · p. 3
  • Xin Ai led the Celsius RD team and initiated development of the Celsius Thermal Solver and Celsius Studio.Source: Author(s) Biography · p. 3
  • Zhiyu Zeng has led teams across Voltus, Celsius, and Reality, with current focus on GPU acceleration and AI.Source: Author(s) Biography · p. 3

1. Introduction

LLM workloads intensify thermal-management demands, while full-scale CFD is too slow for rapid exploration. This work targets accurate cabinet-inlet prediction and integrates an AI surrogate into a digital-twin workflow for capacity planning.

  • LLM GPU clusters create localized hotspots and transient thermal spikes, making fast temperature prediction and real-time workload and cooling adjustments important.Source: §1 · p. 4
  • Hours- or days-long full-scale CFD runs make real-time thermal management and rapid design-space exploration impractical,.Source: §1 · p. 4
  • Existing AI-surrogate approaches often simplify airflow and heat transfer and target specific configurations; one HPE model reported maximum absolute errors of 4 °C to 20 °C across 3D U-Net variants.Source: §1 · p. 4
  • The work prioritizes cabinet-inlet thermal conditions because operators use these locations to assess cooling sufficiency and hotspot risk.Source: §1 · p. 4
  • MAE < 0.25 °C and > 10,000× speedup over full CFD are reported for cabinet-inlet predictions.The model is integrated into Reality Digital Twin for automated data generation, training, and real-time inference.Source: §1 · p. 4
  • A rapid decision engine evaluates workload or equipment placement using an AI-driven strategy rather than traditional heuristic rules.Source: §1 · p. 5

2. Methodology and System Setup

The methodology uses Cadence’s digital-twin and CFD platform to generate simulation data, then trains a U-Net surrogate to predict cabinet-inlet temperatures from multidimensional design and operating inputs.

2.1 Cadence Reality DC Design and Reality Digital Twin Platform

  • Cadence Reality DC Digital Twin combines detailed digital-twin models with a high-fidelity CFD solver, and Figure 1 shows a simulated cabinet thermal-temperature plane.The platform supports data-center design and operational workflows, including capacity visualization, deployment evaluation, and simulation scheduling.Source: §2.1 · pp. 5–6, Figure 1 · p. 6

2.2 Data acquisition

  • Commander automates CFD simulation runs across customizable rack, cooling-unit, and IT-equipment parameters; initial datasets typically include about 500 cases and around ten randomized parameters.Source: §2.2 · pp. 6–7
  • Training inputs aggregate simulation parameters, while outputs are CFD-derived temperature, velocity, and pressure at cabinet inlets, selected for their operational relevance to cooling sufficiency and hotspot formation.Source: §2.2 · p. 7

2.4 AI model architecture

  • The U-Net maps multi-channel inputs for cabinet, equipment-placement, and HVAC parameters to a single-channel cabinet-inlet temperature field, using skip connections to retain fine spatial detail.Source: §2.4 · p. 7

3. Case Setup and Results

The raised-floor case varies IT power and ACU fan speeds to train a surrogate on 1,000 CFD cases. The model predicts cabinet-inlet temperatures rapidly and preserves thermal rankings useful for workload placement.

3.1 Case Setup

  • The raised-floor design has 32 cabinets in four rows and four row-end ACUs supplying a fully contained cold aisle; IT-unit powers are sampled uniformly from 0 to 10 kW, with ACU fan speeds varied.Zero-power assignments represent idle or partially populated racks.Source: §3.1 · pp. 7–8

3.2 AI Prediction Results

  • Maximum absolute error for the absolute temperature field remains relatively large, although the model preserves relative temperature ordering across cabinets.That ranking can help identify hotter cabinets for workload redistribution even when some absolute errors are modest.Source: §3.2 · pp. 9–10, Figure 4 · p. 10
  • The model averages 9 milliseconds per inference, about 150,000× faster than full-scale CFD at roughly 25 minutes per configuration.Figure 3 compares CFD simulation and AI-predicted cabinet-inlet temperature fields; the accompanying text says the visualization indicates the model captures the distribution.Source: §3.2 · p. 9, Figure 3 · p. 9

3.3 Capacity Planning Framework

  • AI and CFD both rank the placement candidates M7, M8, M10, then M11 by maximum cabinet inlet temperature, despite AI errors of approximately 1–3 C.The model evaluates each candidate in milliseconds and updates its predicted cabinet maximum as power is assigned.Source: §3.3 · p. 11, Figure 5 · p. 11, Table 2 · p. 11

4. AI-Driven Capacity Planning Framework

The framework automates data acquisition, model training, and real-time inference within the Reality Digital Twin platform, enabling immediate cabinet-inlet thermal feedback during design exploration.

4. AI-Driven Capacity Planning Framework

  • The framework integrates data acquisition, model training, and real-time inference into a fully automated Reality Digital Twin workflow.Users can move from simulation to surrogate modeling without manual intervention or external scripting.Source: §4 · p. 12
  • Engineers configure varying parameters, ranges, and simulation counts; jobs run locally or via LSF while maximum data-center temperature is monitored.The interface monitors job-level metrics while simulations run.Source: §4 · p. 12
  • Training tasks run through LSF, with mean and maximum absolute error tracked each epoch; trained models are then available in the Digital Twin.Users can evaluate the trained model in the Digital Twin environment.Source: §4 · p. 12
  • Changes at IT-equipment, rack, ACU, or room-layout level trigger immediate cabinet-inlet predictions rendered in 3D for drag-and-drop design exploration.Figure 6 contrasts baseline inference with inference after assigning a new cabinet.Source: §4 · p. 12, Figure 6 · p. 12
  • Automating dataset generation, training, and deployment removes manual scripting and pipeline management, reducing the effort and expertise required.Engineers can focus on design exploration and decision making rather than managing computational pipelines.Source: §4 · pp. 12–13

5. Conclusion and Future Work

Across multiple data center configurations, the surrogate delivers millisecond-level prediction and more than 100,000 times speedup over full-scale CFD while preserving relative temperature ranking. The automated digital-twin workflow supports design exploration and operational decisions, while future work targets broader generalization and additional prediction and optimization capabilities.

5. Conclusion and Future Work

  • More than 100,000 times speedup over full-scale CFD, with millisecond-level inference, reasonable accuracy, and preserved relative temperature ranking across multiple data center configurations.The surrogate predicts cabinet inlet temperatures.Source: §5 · p. 13
  • Integration into the Cadence Reality Digital Twin enables drag-and-drop exploration of design variants, automated data acquisition and training, and rapid evaluation of optimization scenarios.The framework combines data generation and AI-based surrogate modeling in an automated workflow.Source: §5 · p. 13
  • The reported capabilities support design-stage analysis and live operational decision making, including workload placement and cooling-strategy exploration.The authors state this can reduce reliance on repetitive CFD simulations and support more agile management of thermally dynamic environments such as LLM hosting facilities.Source: §5 · p. 13
  • Future work includes a foundational thermal model adaptable to broader data-center architectures, transient prediction for time-varying workloads, and multi-objective thermal and energy optimization.The proposed adaptation direction is transfer learning for new designs.Source: §5 · p. 13

Related papers

Loading 2610.02442v1…