What is immersion cooling?

Submerging servers or other electronics entirely in an electrically insulating liquid. The liquid takes heat directly from components, removing the need for fans and heat sinks. Single-phase immersion (e.g. dielectric oil) carries heat by warming the liquid; two-phase immersion (fluorocarbons) boils it and carries latent heat.

Why do liquids cool better than air?

Their volumetric heat capacity is orders of magnitude larger: air about 1.2 kJ/m³K, dielectric oil 1.7 MJ/m³K, some 1,500 times more. Carrying 30 kW at a 10 K rise needs 2.6 m³/s of air but 106 L/min of oil. Natural-convection h on a 30 mm surface was 6.7 W/m²K in air versus 121 W/m²K in oil.

What limits two-phase immersion?

The critical heat flux (CHF) of boiling. Fluorocarbons have low latent heat and surface tension, so Zuber's CHF is about 16 W/cm² — roughly an eighth of water's 126 W/cm². Limiting a 2 cm square chip to 70% of CHF gives about 45 W, so high-power CPUs use boiling-enhancement coatings or heat spreaders to enlarge the area.

Immersion Cooling

Category: Thermal – Electronics Cooling | Revised 2026-10-01

Immersion Cooling

Theory: liquid heat capacity and boiling

Overview

🙋

I hear data centres submerge servers in liquid. Does it really cool that well?

🎓

Air and liquids differ by orders of magnitude in heat transport. Dielectric oil holds about 1,500 times air's heat per volume, so modest flow carries a lot of heat and fan power disappears. Two-phase immersion boils a fluorocarbon, using latent heat to hold parts near the boiling point. But boiling has a ceiling, the critical heat flux, much lower for fluorocarbons than for water. Let's compare air, oil and fluorocarbon on heat capacity, natural convection and boiling.

Heat capacity and natural convection

$$ \dot V = \frac{\dot Q}{\rho c_p\,\Delta T},\qquad Nu_L = \left\{0.825 + \frac{0.387\,Ra_L^{1/6}}{\left[1+(0.492/Pr)^{9/16}\right]^{8/27}}\right\}^2 $$

Top: volumetric flow needed to carry $\dot Q$ at a rise $\Delta T$. Bottom: natural convection on a vertical surface (Churchill–Chu).

Boiling critical heat flux

$$ q''_{CHF} = 0.149\,h_{fg}\,\rho_v\left[\frac{\sigma g(\rho_l-\rho_v)}{\rho_v^2}\right]^{1/4} $$

Coffee Break Trivia

It started with supercomputers

Submerging electronics isn't new: the 1985 Cray-2 supercomputer immersed its boards in an inert fluorocarbon liquid. Air cooling later regained dominance, but from the late 2010s AI computing sharply raised per-server heat, and immersion cooling drew attention again in data centres, eliminating fan power and greatly reducing cooling energy.

Worked examples

Example 1: flow to cool a 30 kW rack at 10 K rise

Coolant	Volumetric heat capacity	Required flow
Air	1.17 kJ/m³K	2.57 m³/s
Dielectric oil	1.70 MJ/m³K	106 L/min
Water (reference)	4.16 MJ/m³K	43 L/min

Example 2: natural convection on a 30 mm surface (20 K)

Fluid	Prandtl number	h	Heat flux
Air	0.7	6.7 W/m²K	0.01 W/cm²
Dielectric oil	262	121 W/m²K	0.24 W/cm²
Fluorocarbon (single-phase)	9.2	277 W/m²K	0.55 W/cm²

Example 3: CHF for two-phase immersion

Fluid	CHF (Zuber)
Fluorocarbon (FC-72-like)	16.2 W/cm²
Water (reference, 1 atm)	125.9 W/cm²

For a 2 cm square chip limited to 70% of CHF, the fluorocarbon allows about 45 W.

🙋

Fluorocarbon boiling carries only an eighth of water's heat flux. Why use it anyway?

🎓

Water conducts electricity, so circuits can't be submerged in it. Fluorocarbons insulate well, are chemically stable and can be chosen to boil around 50 °C. They hold component surfaces near the boiling point with little temperature variation. Their low CHF is offset by copper heat spreaders on chips to enlarge the boiling area, or fine surface textures that promote boiling. Single-phase oil doesn't boil, so CHF isn't a concern, but its natural-convection h is only about 120 W/m²K, so high-power parts get heat sinks or pumped flow.

Design tips

Choose single- or two-phase by component heat flux and temperature requirements.
In two-phase, check CHF margin at the highest-flux component.
In single-phase, model tank flow so warmed liquid doesn't stagnate around parts.
Check compatibility of liquid with components, cables and seals (swelling, degradation).
In two-phase, size the vapour condenser and manage evaporative loss.

Coffee Break Trivia

“Only the GPUs ran hot”

In one two-phase immersion trial, CPUs cooled without trouble but high-performance GPUs neared their temperature limits. The GPU dies had high heat per area, so surface flux approached the fluorocarbon's CHF. Fitting copper plates with boiling-enhancing porous coatings onto the dies enlarged the boiling area and dropped temperatures considerably. Even with immersion, local heat flux must be watched.

Common mistakes

Mistakes and fixes

Mistake	Effect	Fix
Using average heat flux	Local CHF exceeded	Check peak flux
Using water's CHF	Capacity overrated 8×	Use the fluid's properties
Ignoring tank flow	Local hot spots	Model the flow
Ignoring material compatibility	Component degradation	Immersion testing
Ignoring oil viscosity vs temperature	Heat transfer wrong	Temperature-dependent properties

🙋

I'd like to learn related topics.

🎓

Rate this article

Thank you for your feedback!

Helpful

More details

Report error

Helpful

More details

Report error

Written by the NovaSolver Editorial Team
Built with engineers and AI (editorial policy) — Sitemap

About the Authors

Immersion Cooling

Theory: liquid heat capacity and boiling

Overview

Heat capacity and natural convection

Boiling critical heat flux

It started with supercomputers

Worked examples

Example 1: flow to cool a 30 kW rack at 10 K rise

Example 2: natural convection on a 30 mm surface (20 K)

Example 3: CHF for two-phase immersion

Design tips

“Only the GPUs ran hot”

Common mistakes

Mistakes and fixes

Related Topics

Related fields