01 / MEET THE COLD PLATE
Bring heat transfer close to the chip
A cold plate is a small heat exchanger mounted against a high-power component such as a CPU or GPU. Heat moves from the chip into the plate, then into liquid flowing through channels inside it. In a typical direct-to-chip design, the coolant stays inside sealed passages; it does not touch the electronic component.
This is called direct-to-chip cooling because the heat-transfer device is attached directly to the chip package. The liquid carries heat away from the source through hoses and a liquid loop.
- CPU / GPUHeat is produced
- Cold plateReceives chip heat
- CoolantCarries the heat away
- Liquid loopMoves heat onward
- Heat removalFacility systems reject heat
WHY LIQUID?
A useful path as compute gets denser
Liquid can transport significant heat through relatively compact flow paths and collect it close to the component producing it. That can help when chip power and rack density rise, AI and GPU workloads grow more demanding, or moving enough air through a rack becomes harder to manage.
This does not mean every data center needs liquid cooling. The right approach depends on equipment, workload, facility design, operating needs, and how heat can be rejected outside the IT room.
02 / TWO WAYS TO MOVE HEAT
Air cooling and liquid cooling have different paths
Air cooling remains mature and widely deployed, with a simple physical interface to servers. At high density it may require substantial airflow, fan energy, and careful distribution. Liquid cooling captures some heat close to the source and can reduce how much chip heat must first travel through room air. Neither approach is a universal winner; facilities may use one or both.
Air cooling
Chip → Heatsink → Air → Room cooling system
- Mature and widely deployed
- Simple physical interface to servers
- High density can increase airflow, fan-energy, and distribution challenges
Liquid cooling
Chip → Cold plate → Liquid → Liquid cooling system
- Captures heat close to high-power components
- Can support high-density compute
- Reduces dependence on moving all chip heat through room air
MANY DESIGNS USE BOTH
Liquid cooling does not always replace air
In a hybrid cooling design, liquid may remove heat from CPUs, GPUs, or accelerators while air still cools memory, storage, NICs, power supplies, motherboards, and other rack electronics. The exact split depends on the server design. Both paths ultimately connect to facility systems that remove heat from the data center.
One server · two heat paths
A liquid loop moves heat; it does not make the heat disappear. Facility cooling still needs to transfer that heat onward and reject it to the environment.
03 / FOLLOW THE RACK LOOP
AI racks concentrate the heat-removal challenge
AI and accelerated-compute racks can concentrate large electrical and thermal loads in a small footprint. The question expands from “How do we cool the room?” to “How do we remove heat efficiently from this dense equipment?” Not every AI rack uses the same cooling architecture; the workload, hardware, and facility all matter.
What is a rack manifold?
A manifold distributes supply and return coolant connections between the rack liquid loop and multiple liquid-cooled servers. It helps organize where coolant enters and leaves the server branches.
- Liquid supply
- Rack manifold
- Server cold plates
- Return via manifold
- High-power chipCPU / GPU
- Cold plateTransfers chip heat
- CoolantMoves heat in the loop
- CDUConnects liquid loops
- Facility coolingHeat rejection
The cooling distribution unit (CDU) helps connect rack-side liquid cooling to facility heat removal. Continue with CDU & the Liquid Cooling Loop · Lesson 07 to follow that boundary.
THREE THINGS TO REMEMBER
Keep the whole heat path in view
- Liquid cooling captures heat closer to high-power components.
- Cold plates transfer chip heat into a circulating liquid loop.
- Many deployments are hybrid: liquid removes some heat while air cools other parts.
PUT THE HEAT PATH IN CONTEXT
Now design a high-density AI rack
In the AI Rack & Liquid Cooling lab, increase rack-1 AI equipment demand, add a liquid loop, and inspect pump feeds, branch flow, heat-exchanger capacity, and the remaining air/liquid split. The mission checks that the rack has an active liquid branch and retains its thermal service; it does not claim every watt is removed by liquid.
To explore failure diagnosis after the loop is familiar, open the separate CDU Failure Diagnosis lab from the workspace lab selector.
PASS IT ON