2D thermal CFD solver for data center rack cooling

Overview

Data center cooling is one of the largest operational costs in modern computing infrastructure. Poor airflow can create hot spots, reduce hardware life, and waste energy. To reason about these flows quantitatively, we built a 2D thermal CFD solver on top of NVIDIA Warp and packaged it as an interactive Streamlit application.

The solver starts from the classic lid-driven cavity benchmark and extends it to:

Incompressible Navier-Stokes with fractional-step projection
Thermal advection-diffusion with buoyancy
Porous zones representing server racks with heat generation and flow resistance
Perforated floor tiles supplying cold air into cold aisles
Directional rack fan airflow pushing air front-to-back through each rack
ASHRAE A1 compliance overlay at 27°C

This post covers two parts: first, validating the solver against established benchmarks; second, applying it to a realistic 4-rack data center layout.

Part 1: Solver Validation

Lid-Driven Cavity

The lid-driven cavity is the “hello world” of CFD: a square domain with a moving top wall and no-slip walls on the other three sides. We compare our centerline velocity profiles against the benchmark data from Ghia, Ghia, and Shin (1982).

Centerline velocity profiles compared against Ghia et al. (1982) at Re = 100.

On a 64×64 grid the solver achieves:

u-velocity centerline: ~2.4% max error
v-velocity centerline: ~5.6% max error
Overall: PASS within 6% tolerance

Error drops to roughly 3% on finer grids, confirming the expected spatial convergence.

Velocity Field

A snapshot of the velocity magnitude shows the characteristic primary vortex and the corner eddies that define the benchmark flow.

Velocity magnitude and streamlines for the lid-driven cavity at Re = 100.

Thermal Benchmarks

Beyond the isothermal cavity, we run two canonical thermal cases:

Rayleigh-Bénard convection: heated bottom, cooled top, buoyancy-driven roll cells
Heated cavity: differential side-wall temperatures driving a natural-convection loop

These cases exercise the coupled temperature and momentum equations before we add the data center geometry.

Part 2: Data Center Rack Cooling

Layout

The data center model is a 12 m × 3 m domain containing four server racks:

Each rack is 0.6 m wide and 2.0 m tall
Racks alternate front-facing direction (left / right) to form realistic hot and cold aisles
Perforated floor tiles inject cold air upward from the floor
Hot air exits through the ceiling

Directional Fan Airflow

In a real rack, server fans push air front-to-back. We model this as a horizontal momentum source inside each rack, with direction alternating to match the rack orientation. The result is visible cold-air intake on the front face and hot exhaust on the back face.

Temperature Field and ASHRAE A1 Compliance

The temperature field below shows the hot exhaust plumes and the 27°C ASHRAE A1 compliance line. Cold air from the perforated tiles keeps the inlet regions below the limit, while hot spots appear behind the racks where exhaust air recirculates.

Temperature field for a 4-rack data center layout. The dashed line marks the ASHRAE A1 27°C limit.

Interactive Demo

The Streamlit GUI lets you adjust rack spacing, power, cooling CFM, and supply temperature, then run the GPU solver and inspect thermal and airflow results in real time.

Key Takeaways

Layout matters: racks far from perforated tiles or with poor aisle spacing see higher peak temperatures
Fan direction matters: without front-to-back momentum, hot air recirculates and inlet temperatures rise
ASHRAE compliance is local: a domain can be 95% compliant while individual racks still exceed 27°C

Implementation

The solver is open source under the MIT license:

Repository: github.com/tech-microcosm/2d-thermal-cfd
Baseline tag: v0.1.0
Key files:
- solver.py — core Navier-Stokes solver
- thermal_solver.py — thermal, porous zones, and rack fan model
- datacenter_config.py — layout and boundary-condition setup
- gui_interface.py — Streamlit interactive interface
- run_benchmark.py — CLI benchmark runner

Conclusion

Starting from a validated 2D Navier-Stokes solver, we added thermal physics, porous racks, perforated tiles, and directional fan airflow to build a practical data center cooling simulator. The Streamlit GUI makes it easy to explore trade-offs between rack spacing, cooling CFM, and supply temperature. Future work includes 3D extension, turbulence modeling, and calibrated fan curves for real server hardware.

Experiments conducted June 2026. Solver built with NVIDIA Warp and visualized with Matplotlib inside Streamlit.