Workshops Archive

Towards Direct-to-Chip Liquid-Cooled Megawatt Racks: Characterization of a Multi-Kilowatt Chip-Scale Cooler & The Implications on Facility Energy Use


Workshop: EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices

Authors: Daniel Lorenzini and Yogendra Joshi (EMCOOL Inc)

Abstract: The power density of artificial-intelligence (AI) and high-performance-computing (HPC) accelerators continues to increase. Although the HPC community adopted liquid cooling years ago, the broader computing industry is now undergoing a similar transition. With 1, 3, and 5 kW accelerators appearing on technology roadmaps, removing multi-kilowatt chip-level heat loads without excessive facility-cooling energy or hydraulic demand remains a challenge. This study characterizes a liquid-cooled thermal stack and identifies the design parameters needed to support next-generation AI devices. Repeatable steady-state measurements yielded an inlet-referenced thermal resistance of 0.0383 K/W. A reduced-order model separated coolant warming, cold-plate convection, and flow-independent stack resistance. At a 100 °C average heater limit and 5 L/min, the baseline stack dissipated 3.04 kW, while a controlled thin liquid-metal interface increased the projected capacity to 6.41 kW. At 3 kW, this interface reduced the required flow from 4.61 to 0.92 L/min and increased allowable inlet-temperature margin by 40.5 °C.


Back to EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices Archive Listing

Back to Full Workshop Archive Listing