Cooling Fundamentals

Data center cooling removes heat generated by IT equipment to maintain safe operating temperatures. Servers generate heat as a byproduct of electrical power consumption — a server drawing 500W generates 500W of heat that must be removed from the facility.

The fundamental challenge: as rack densities increase, the heat generated per unit of floor space increases proportionally. Traditional air cooling becomes less effective as density increases because air has limited heat capacity — moving enough air to cool a 40 kW rack requires significant airflow that creates noise, pressure differentials, and bypass airflow problems.

ASHRAE (American Society of Heating, Refrigerating and Air-Conditioning Engineers) publishes thermal guidelines for data center equipment. ASHRAE A1 class equipment (the most common) operates at inlet temperatures of 15–32°C. Modern equipment increasingly supports ASHRAE A3/A4 classes, allowing higher inlet temperatures and reducing cooling energy consumption.

CRAC & CRAH Units

CRAC (Computer Room Air Conditioning)

Self-contained units that cool, dehumidify, and filter air using a direct expansion (DX) refrigeration cycle. The compressor is integral to the unit — no external chiller required. Simpler to install and operate than chilled water systems. Typical capacity: 10–50 kW per unit. Most common in smaller data centers and server rooms.

CRAH (Computer Room Air Handler)

Fan-coil units that use chilled water from a central chiller plant to cool air. No integral compressor — requires a chilled water infrastructure. More efficient than CRAC at scale; chiller plants achieve higher COP (Coefficient of Performance) than distributed DX systems. Typical capacity: 20–200 kW per unit. Standard for large data centers.

Limitations

Both CRAC and CRAH units cool the room air, not the equipment directly. As rack density increases, hot spots develop where equipment heat output exceeds the local cooling capacity. Effective up to approximately 5–8 kW per rack without containment; 15–20 kW per rack with hot/cold aisle containment.

Hot/Cold Aisle Containment

Hot/cold aisle containment is the most cost-effective cooling improvement for most data centers. By separating hot exhaust air from cold supply air, containment eliminates mixing and allows cooling units to operate more efficiently.

Cold Aisle Containment (CAC)

Encloses the cold aisle with doors and a ceiling, preventing cold supply air from mixing with hot exhaust air. Cold air is delivered to the enclosed aisle and drawn through servers. Simpler to implement than hot aisle containment; does not require a plenum ceiling.

Hot Aisle Containment (HAC)

Encloses the hot aisle, capturing hot exhaust air and directing it back to cooling units. More efficient than cold aisle containment because it captures heat at the source. Requires a return air plenum (raised floor or overhead) to route hot air back to cooling units.

Effectiveness

Containment typically improves cooling efficiency by 20–40% and allows rack densities of 15–20 kW with standard CRAC/CRAH cooling. A relatively low-cost improvement that should be implemented before considering more expensive cooling technologies.

In-Row Cooling

In-row cooling units are installed within the server rack rows, directly adjacent to high-density racks. They cool air at the point of heat generation rather than at the room perimeter, reducing the distance hot air must travel and improving cooling effectiveness for high-density deployments.

In-row units use chilled water or DX refrigeration. Capacity: 20–60 kW per unit. Effective for rack densities of 20–30 kW. Requires careful airflow management to prevent hot air recirculation.

Advantages: targeted cooling for high-density racks, can be added incrementally as density increases, works with existing room cooling infrastructure. Disadvantages: requires chilled water or refrigerant piping to the rack row, adds complexity to the data center floor.

Rear-Door Heat Exchangers

Rear-door heat exchangers (RDHx) replace the standard rear door of a server rack with a heat exchanger that captures heat as it exits the rack. Chilled water flows through the heat exchanger, cooling the exhaust air before it enters the room.

Passive RDHx (no fans) rely on server fans to push air through the heat exchanger. Active RDHx include fans to supplement server airflow. Capacity: 20–40 kW per rack for passive; 40–60 kW for active.

Advantages: can be retrofitted to existing racks without server modifications, contains heat at the rack level, does not require changes to room cooling infrastructure. Disadvantages: requires chilled water piping to each rack, condensation risk if chilled water temperature is too low.

Direct Liquid Cooling (DLC)

Direct liquid cooling brings coolant directly to heat-generating components — CPUs, GPUs, and memory — via cold plates attached to the components. Coolant (water or dielectric fluid) absorbs heat at the source and carries it to a facility cooling system.

Cold Plate Systems

Cold plates are attached directly to CPUs and GPUs. Coolant flows through microchannels in the cold plate, absorbing heat with much higher efficiency than air. Remaining components (memory, storage, power supplies) are cooled by residual airflow.

DLC systems can handle 40–80% of server heat load via liquid, with the remainder handled by air. This hybrid approach allows DLC to be used with standard data center air cooling infrastructure for the residual heat.

Requirements for AI Infrastructure

NVIDIA GB200 NVL72 systems require direct liquid cooling — air cooling is not supported. DLC is also strongly recommended for H100/H200 DGX systems in high-density deployments. Facilities deploying next-generation AI hardware must have DLC infrastructure in place before hardware delivery.

Facility Requirements

DLC requires: chilled water or facility water loop at appropriate temperature (typically 18–45°C supply), leak detection systems, manifolds and quick-disconnect fittings for server connections, and trained staff for maintenance.

Immersion Cooling

Immersion cooling submerges servers directly in dielectric fluid (non-conductive liquid) that absorbs heat from all components simultaneously. The heated fluid is cooled by a heat exchanger and recirculated.

Single-Phase Immersion

Servers are submerged in a tank of dielectric fluid that remains liquid throughout the cooling process. Fluid is pumped through a heat exchanger to remove heat. Capacity: 100+ kW per tank. PUE as low as 1.03. Requires purpose-built tanks and modified server configurations (fans removed).

Two-Phase Immersion

Servers are submerged in a fluid that boils at low temperature (30–60°C), absorbing heat as it vaporizes. Vapor condenses on a condenser coil and returns to liquid. Extremely efficient heat transfer. Higher fluid cost than single-phase; more complex system design.

Advantages and Disadvantages

Advantages: highest density (100+ kW per tank), lowest PUE (1.03–1.05), silent operation, extended hardware life (no fans, lower operating temperatures). Disadvantages: highest infrastructure cost, requires modified servers (fans removed), complex fluid management, limited vendor support for warranty with immersion-modified hardware.

Immersion cooling is increasingly adopted for AI training clusters and cryptocurrency mining where density and efficiency are paramount. Not yet mainstream for general enterprise workloads.

Selecting the Right Cooling Approach