As AI workloads continue to drive unprecedented rack power densities, the limits of traditional air cooling are becoming increasingly visible. At IEEE ITherm Conference (Orlando, FL – May 26-29), Cadence and Microsoft will jointly present new research that takes a physics‑based look at how air‑cooled data centers behave under real‑world operating conditions—and what that means for the future of thermal design.
The joint presentation, "Towards Physics‑Based CFD Digital Twins for Hyperscale Data Centers – Part I," will be delivered as part of the Methods for Data Centers session on Thursday, May 28, from 11:15am to 11:30pm.
Establishing Modeling Best Practices for Physics‑Based Digital Twins
As hyperscale data centers evolve into AI factories, operators must manage non‑uniform rack power, airflow variability, leakage, and transient operating events. This Cadence – Microsoft research focuses on a foundational requirement for truly predictive digital twins: accurate, physics‑based CFD models that capture how air‑cooled data halls behave under both steady‑state and transient conditions. This was based on a generic hyperscale conceptual model.
Using detailed 3D CFD simulations in the Cadence Reality Digital Twin Platform, the study investigates air‑cooled data hall performance across multiple scenarios, including uniform and mixed-rack power densities, incomplete rack sealing, and airflow ramp‑down events. The work highlights how seemingly small effects—such as cabinet leakage or local pressure imbalances—can have an outsized impact on thermal reliability.
Key Findings from the Study
The analysis reveals several important behaviors that data center designers and operators must account for:
- Venturi‑driven recirculation near AHUs: Racks located close to air‑handling units (AHUs) consistently experience reduced inlet pressure and elevated inlet temperatures as accelerated airflow induces hot‑air recirculation through cabinet leakage.
- Network vs IT racks: some racks can particularly be sensitive to airflow imbalance and transient airflow reduction, showing the fastest temperature rise during low‑airflow events. Other racks can be less sensitive to these changes.
- Impact of rack leakage: Even modest leakages (e.g., eight RU openings) enable bypass airflow that recirculates hot exhaust air back into the server inlets, degrading local cooling effectiveness without changing total airflow‑to‑power ratios.
- Transient analysis: During airflow ramp‑down scenarios, temperature spikes occur rapidly, underscoring the importance of modeling dynamic—not just steady‑state—conditions.
Learnings for Next-Generation AI Data Centers
Looking ahead, the authors will outline future work extending these digital twins to support:
- Hybrid and liquid‑assisted cooling integration
- Experimental validation of transient CFD behavior
- Reduced‑order models for fast prediction
- Real‑time predictive thermal control
Together, these advances enable physics‑based digital twins for predictive monitoring, failure forecasting, and cooling optimization—providing a scalable foundation for next‑generation, high‑density AI data centers.
Join Cadence and Microsoft at ITherm to explore how these thought leaders are advancing CFD and physics‑based digital twin methodologies to shape the future of data center thermal design and to establish the rulebook for data center design and simulation.
Learn more about the Cadence Reality Digital Twin Platform.