Preprint / Version 1

Numerical Study of Workload-Dependent Air-Side Convective Heat Transfer over a High-Density AI Server Board under Large Language Model Inference

##article.authors##

  • Nafisa Anjum Tanisha Bangladesh University of Engineering & Technology
  • Anindya Guha Bangladesh University of Engineering & Technology

DOI:

https://doi.org/10.31224/8852

Keywords:

Convective heat transfer, AI server, Large language model inference, Thermal management

Abstract

Large language model (LLM) inference has pushed data center servers toward higher and more uneven heat loads. System-level thermal models usually give each component a fixed convective heat transfer coefficient, although the heat load changes several-fold with the model being served. This work tests that assumption on a high-density AI server board based on the NVIDIA DGX A100 layout, modelled with separate regions for two compute boards (CB1 and CB2), four DIMM banks and two PCIe cards, and solved in ANSYS Fluent as a steady three-dimensional conjugate case with the k-ε turbulence model at inlet velocities of 1, 3 and 5 m/s. Block heat loads were derived from token-level power estimates for LLaMA 7B, LLaMA 13B, GPT-3 Small and GPT-3 175B at 1 to 70 generated tokens. At 70 tokens the load on one compute board ranged from 64 to 359 W, and at 1 m/s the coefficient changed with the workload in every block: from 63.6 to 64.4 W/m2K on CB1, 60.2 to 61.8 on CB2, 222 to 288 on the finned CB1, 11.0 to 12.7 on the DIMM banks and 11.0 to 17.9 on the PCIe cards, while the CB1 temperature rise went from 72 to 400 K. The coefficient rose with velocity as Nu = C Rem Pr1/3, with m = 0.37 to 0.39 for the compute boards and 0.6 to 0.7 for the DIMM and PCIe blocks and it depended strongly on position: the DIMM bank lying in the wake of a PCIe card and a compute board had less than half the coefficient of the others. The downstream compute board was 3 to 6 percent below the upstream one without fins and 30 to 39 percent below it with fins. Rectangular fins with a thermal paste interface raised the effective coefficient 3.5 to 4.5 times on CB1 and 2.6 to 2.8 times on CB2, cutting the temperature rise by 71 to 78 and 61 to 65 percent. The study characterizes the air-side coefficient of each block and the reported ranges show how far a coefficient found for one workload can be carried to another.

Downloads

Download data is not yet available.

Downloads

Posted

2026-11-20