ORCID Identifier(s)

0009-0002-6357-2447

Graduation Semester and Year

Summer 2026

Language

English

Document Type

Thesis

Degree Name

Master of Science in Mechanical Engineering

Department

Mechanical and Aerospace Engineering

First Advisor

Dr. Dereje Agonafer

Second Advisor

Dr. Yogesh Fulpagare

Third Advisor

Dr. Miguel Amaya

Abstract

As the power density of artificial intelligence (AI) and high-performance computing (HPC) servers escalates, air cooling is approaching fundamental scaling limits set by air's low convective heat transfer coefficient and rising parasitic fan power. This thesis presents an experimental comparison of direct-to-chip liquid cooling and air cooling for a high-density GPU server, using a Thermal Test Vehicle (TTV) that emulates the device layout and power-density regime of an 8-GPU SXM5-class baseboard (eight emulated GPU blocks and four NVLink Switch modules). The liquid configuration was tested at 8.8 kW (8.0 kW GPU, 800 W switch) across three Coolant Distribution Unit (CDU) inlet setpoints (25, 35, 45 °C) using a 25% propylene-glycol coolant; the air configuration was tested across an airflow sweep (170–370 CFM) at 1.32 kW, with flow measured per ANSI/AMCA 210 / ASHRAE 51-1999. Cooling performance is compared using average per-device thermal resistance, Rth = (mean device temperature − inlet temperature)/(average power per device), with power expressed in watts. Liquid cooling achieved GPU Rth ≈ 0.034 K/W, roughly 4.6 to 8.8 times lower than air cooling (GPU Rth ≈ 0.157–0.303 K/W across the flow sweep). Device temperature tracked coolant inlet temperature with a near-unity slope (≈0.973) across all three liquid setpoints, while GPU thermal resistance remained essentially constant as inlet temperature rose from 25 to 45 °C. This inlet-independence supports the component-level thermal feasibility of elevated-temperature coolant operation over the tested range, a prerequisite for warm-water and economizer-based cooling strategies. Even at the 45 °C inlet, mean GPU temperature remained near 80 °C. These results quantify both the magnitude of liquid cooling's thermal advantage and its robustness to elevated coolant temperatures.

Keywords

direct-to-chip cooling; data center thermal management; GPU thermal management; thermal resistance; warm-water cooling; coolant distribution unit.

Disciplines

Engineering | Heat Transfer, Combustion | Mechanical Engineering

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.