# Cold plates, manifolds and separated coolant loops

Canonical: https://semiaifoundry.com/learn/atlas/coolant-loop-separation/
Revision: 2026-10-07
Rights: https://semiaifoundry.com/ai-atlas/licenses.html

Data-center physical systems · Atlas v1.3.0

# Cold plates, manifolds and separated coolant loops

Transports heat through controlled fluid circuits without treating circulation as consumption.

[Explore in Atlas](https://semiaifoundry.com/ai-atlas/#coolant-loop-separation)Chip voltage rails to facility heat rejection

Learning objectiveSeparate thermal transport, loop ownership and water consumption.

## The mechanism

A cold plate conducts chip heat into a fluid. A rack manifold distributes that fluid among components, and a coolant distribution unit can exchange heat with a separate facility loop. Separating loops allows different pressure, chemistry and service requirements on each side. Heat captured by liquid may be only part of rack heat; fans and room air still handle other components in many systems. The NVIDIA rack guide explicitly describes mixed cooling. Fluid flow, temperature rise and pressure drop determine transport capability, while the facility’s final heat-rejection method determines how much energy or water is consumed beyond the rack.

1.  Identify which components transfer heat to liquid and which remain on an air path.
2.  Follow thermal interfaces into the cold plate and measure inlet and outlet temperatures.
3.  Distribute fluid through manifolds while respecting pressure and chemistry requirements.
4.  Exchange heat with the facility circuit through the specified CDU arrangement.
5.  Balance captured heat, residual air heat and pump demand before declaring cooling readiness.

## Science and assumptions

### Single-phase heat transport

$$
\dot Q=\dot m c_p\Delta T
$$

Single-phase transported heat is mass flow times specific heat times stream temperature rise.

Symbols and units

Q̇

Heat transported by the selected single-phase stream

W

ṁ

Stream mass-flow rate

kg/s

c\_p

Specific heat capacity

J/(kg·K)

ΔT

Stream outlet-to-inlet temperature rise

K

Q̇ is watts, ṁ kg/s, cp J/(kg·K) and ΔT kelvin.

Assumptions Single-phase fluid with approximately constant properties; pressure limits and local thermal resistances are separate constraints.

## Engineering example

Synthetic water loop carrying 100 kW with a 10 K rise and cp ≈ 4,180 J/(kg·K) needs about 2.39 kg/s. This is circulating mass flow, not a consumption estimate.

## Distinctions and tradeoffs

Liters per minute circulating through a closed loop are not liters of water consumed by the facility.

Greater flow can reduce temperature rise but increases pumping requirements; hotter coolant can aid heat rejection while narrowing component thermal margin.

Dense accelerator racks require sufficient heat transport before their installed arithmetic can operate continuously.

## Specifications and interfaces

Transport rate

Mass flow × specific heat × temperature rise

Heat rate in watts; flow is not consumed water.

Loop boundary

IT-side fluid versus facility-side fluid

A heat exchanger transfers heat without requiring fluid mixing.

### Interfaces

Chip and component heat Input

Conducted heat entering cold plates or air paths

Facility heat exchanger Output

Heat carried in a controlled secondary loop

## Check your understanding

### What happens when heat load rises at fixed flow and fluid properties?

1.  The heat exchanger stops needing a sink
2.  The required fluid temperature rise increases
3.  The circulating water disappears

Answer and reasoning

The required fluid temperature rise increases

The energy balance requires greater temperature gain unless another variable or heat path changes.

## Evidence and scope

This is a heat-balance example, not a qualified coolant recipe. Product-specific pressure, chemistry and temperature limits remain required.

No current calculator directly models this lesson’s main mechanism. Use the library for related bounds while retaining their stated omissions.

## Sources

-   [NVIDIA DGX GB Rack Scale Systems: Hardware](https://docs.nvidia.com/dgx/dgxgb200-user-guide/hardware.html)
-   [Cooling Water Efficiency Opportunities for Federal Data Centers](https://www.energy.gov/cmei/femp/cooling-water-efficiency-opportunities-federal-data-centers)

[Published lesson data](https://semiaifoundry.com/ai-atlas/lesson-data/coolant-loop-separation-cc65e3a89d10352f.json) · [Atlas releases and licensing](https://semiaifoundry.com/ai-atlas/licenses.html)

## Continue exploring

### Foundations

-   [Liquid-cooling coldplate](https://semiaifoundry.com/learn/atlas/coldplate/)
-   [Coolant distribution unit and separated loops](https://semiaifoundry.com/learn/atlas/cdu/)

### Related lessons

-   [Facility heat rejection and cooling tradeoffs](https://semiaifoundry.com/learn/atlas/facility-heat-rejection/)