Transforming Future Data Center Operations and Management via Physical AI

Overview

The study asks what new paradigm is needed to operate and manage AI data centres while cutting cost and keeping reliability. A data centre with 100k B200 GPUs costs 4.5 to 5 billion US dollars in CapEx and its annual electricity OpEx can reach 400 million, so efficient solutions are sought. The proposed framework combines a GPU-accelerated CFD and heat-transfer engine, a physics-informed machine-learning engine and a digital-twin layer.

The case study trains a neural-operator surrogate (GNOT) for 3D temperature and velocity in a roughly 600 m² hot-aisle-contained hall with four rows totalling 60 racks with 3-8 servers each and 6 ACUs. 500 CFD scenarios were generated by Latin hypercube sampling of ACU supply temperature (18-24 °C) and server power (1000-2000 W) on a fixed mesh of about 340k cells, at about 5 min per case on one RTX 4090; the split was 400/50/50 for training, validation and test.

Inputs were server inlet/outlet face coordinates, power and airflow, plus ACU supply face, temperature and airflow. The hall was split into cold-aisle and per-hot-aisle sub-models, with server-level mass and energy conservation added as soft penalty terms. For most of the hall, absolute temperature errors were within 2.5 °C, with a median absolute error of 0.18 °C against CFD and inference latency of about 0.01 s. The authors note that detailed room CFD with fine meshes can take hours per run and that covering the parameter space needs thousands of samples.

Limits: like most surrogates it is benchmarked against CFD output, not measurements, and the text gives the server count as both 317 and 340. It is a preprint and perspective: the future directions (foundation models, physics-informed training, sample-efficient scenario generation, quantum computing, differentiable simulation) are not yet proven.

Key points

1. Case hall: about 600 m², four rows totalling 60 racks with 3-8 servers each, 6 ACUs and hot-aisle containment; the text gives the server count as both 317 and 340.
2. 500 CFD scenarios were generated by Latin hypercube sampling of ACU supply temperature (18-24 °C) and server power (1000-2000 W) on a fixed mesh of about 340k cells, about 5 min per case on one RTX 4090; split 400/50/50.
3. Inputs: server inlet/outlet face coordinates, power and airflow, plus ACU supply face, temperature and airflow; hall split into cold-aisle and per-hot-aisle sub-models with server-level mass and energy conservation as soft penalties.
4. For most of the hall, absolute temperature errors were within 2.5 °C, with a median absolute error of 0.18 °C against CFD and inference latency of about 0.01 s.
5. The authors note that detailed room CFD with fine meshes can take hours per run and that covering the parameter space needs thousands of samples.
6. The CapEx of building a data centre with 100k B200 GPUs is 4.5 to 5 billion US dollars and annual electricity OpEx can reach 400 million.

Using it in design

• For fast what-if analysis consider a CFD-trained surrogate (inference about 0.01 s), but state that the surrogate is only as accurate as the CFD it was trained on and is not validated against measurements.
• Define the surrogate training parameter ranges (for example ACU supply 18-24 °C, server power 1000-2000 W) and sample count (about 500 scenarios) in advance.
• Consider splitting the hall into cold-aisle and hot-aisle sub-models and adding server-level mass and energy conservation as penalty terms for accuracy.

Who should read it

Researchers and operations engineers evaluating CFD surrogates and digital-twin approaches for data-hall operations.

Document contents

1. I. Introduction
2. II. Preliminaries (A. Overview of Data Center Physical Infrastructure, B. Physical AI: Basics and Building Blocks)
3. III. Physical AI for Autonomous DC Operations (A. Framework Design, B. Potential Applications)
4. IV. Case Study: Rapid CFD/HT Simulation (A. Motivation, B. Problem Settings, C. Model Architecture, D. Evaluation Results)
5. V. Future Directions (A. Foundation Models for PhyAI, B. Physics-informed Training and Adaptation, C. Sample-Efficient Scenario Generation, D. Quantum Computing for Accelerated Physical Simulation, E. Differentiable Simulation for Prescriptive Analysis)
6. VI. Conclusions

Read the document

Publication details

Authors

Zhiwei Cao; Minghao Li; Feng Lin; Jimin Jia; Yonggang Wen; Jianxiong Yin; Simon See

Publisher

arXiv preprint (Nanyang Technological University / NVIDIA AI Technology Center)

Year

2025

Document number

arXiv:2504.04982

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.48550/arXiv.2504.04982

Licence

CC BY 4.0 — share with attribution

Licence evidence

arxiv.org

Citation