Mouser Electronics White Paper
become evident. Air has a low specific heat capacity, so removing
higher thermal loads requires higher airflow velocity; however,
higher fan speeds increase energy consumption and acoustic
noise. On top of this, the tightly packed arrangements of AI
servers further reduce efficiency by restricting flow paths.
The term liquid cooling may evoke images of equipment
completely submerged in fluid, but this is only one type of
liquid cooling, known as immersion cooling, and it supports
extremely high densities. With the right preparation, servers
can operate fully submerged in a dielectric liquid that
transfers heat directly to an external heat exchanger.
But immersion cooling requires infrastructure planning. Facilities
must be built to support fluid containment, handling, and
long-term maintenance. Retrofitting existing installations can
be complex. For greenfield sites designed around immersion
from the outset, this liquid cooling method can support
long-term thermal objectives. However, for many operators,
incremental alternatives offer a more practical starting point.
Liquid cooling does not require the submersion of hardware.
Rear door heat exchangers represent a more easily adopted
transition. Installed on the back of a rack, they capture hot
exhaust air before it disperses into the server room. A liquid
loop within the door extracts heat via an integrated exchanger,
reducing the need for room-level cooling without redesigning
internal server layouts. An advantage of this approach is that
operators can deploy it selectively on high-density racks.
Direct-to-chip cooling is an even more focused liquid cooling
approach. Cold plates mount directly onto processors or accelerators,
and coolant flows through internal channels within the plate to remove
heat at its source (Figure 1). This technique reduces dependence on
large heatsinks and high-speed fans. Moreover, the space created
within the server becomes available for additional compute modules
or improved airflow management for the remaining components.
Figure 1: Direct-to-chip liquid cooling architecture showing rack-level manifolds
supplying cold plates within dense server configurations. (Source: Amphenol LTW)
