White Paper : In-Row vs. In-Rack vs. Room Cooling — Choosing the Right Precision Cooling for AI Density

1. Why cooling selection is now a critical decision

AI racks are pushing power density beyond 30 kW per cabinet. At this density, traditional room-level cooling can no longer guarantee stable inlet air temperatures. As a result, operators see thermal throttling, reduced GPU performance, and shortened hardware life. This is why data center cooling types must now be matched to actual rack density, not room average.

Three cooling architectures dominate the discussion:

• Room cooling — CRAC/CRAH units serving the entire room. Best for legacy low-density deployments.
• In-row cooling — units placed between server racks, close to the heat source. Best for medium-density AI racks.
• In-rack cooling — units mounted directly in or beside the rack. Best for high-density AI and edge deployments.

The confusion is understandable: all three are called "precision cooling," yet they perform very differently depending on rack density, room layout, and redundancy requirements. Understanding precision cooling selection is now a core skill for data center planners.

2. Standards and comparison framework

Relevant standards

These standards define thermal guidelines and energy requirements for data centers. They do not prescribe a specific cooling architecture. Instead, that decision must be based on rack density, layout, and redundancy goals. Key references include ASHRAE TC 9.9 for thermal guidelines, ASHRAE Standard 90.4 for energy efficiency, ISO 14644-1 for air cleanliness, NFPA 75 for fire protection, and U.S. DOE FEMP for data center energy efficiency guidance.

Cooling architecture comparison: in-row vs in-rack vs room cooling

In short, as rack density increases, the cooling source must move closer to the load. Room cooling is adequate below 10 kW per rack. In-row cooling becomes the practical choice from 10–30 kW. Above 30 kW, in-rack cooling or liquid cooling is typically required. This is the core principle behind AI data center cooling design.

3. A real-world example: high-density GPU deployment

To make this concrete, consider a recent ACwatt project. A research institute in Southeast Asia was deploying a small GPU cluster for AI training. The initial rack density was around 15 kW per rack — higher than the facility's original design of 8 kW per rack.

The existing room cooling system could not maintain stable inlet temperatures. During peak training loads, the inlet air temperature rose above the ASHRAE recommended range. This caused thermal throttling, and GPU performance dropped by more than 10% during extended runs.

The project team evaluated two options: upgrading the room cooling system, or deploying in-row cooling units between the racks. They chose in-row cooling because it provided localized cooling close to the heat source, and it could be installed without a full room redesign. The chosen product was the HyperAir in-row cooling system.

The result: inlet air temperature stabilized within the ASHRAE recommended range, and GPU throttling events were eliminated. This is a real example of how cooling architecture selection should follow rack density, not room average.

4. Decision matrix and deployment guidance

Deployment checklist

• Confirm actual rack power density (kW per rack), not average room density.
• Verify sensible cooling capacity matches the load.
• Confirm hot aisle / cold aisle containment requirements.
• Verify redundancy level (N+1 or 2N) matches business criticality.
• Confirm SNMP/Modbus support for DCIM integration.
• Plan condensate removal and pipe routing for in-row/in-rack systems.

5. Expert commentary

"The biggest mistake we see in AI data center projects is designing cooling based on average room density. AI racks are not average. A single 30 kW rack surrounded by 8 kW racks will still overheat if the cooling source is too far away. In-row and in-rack cooling solve this by moving the cooling unit right next to the load. That is why we recommend selecting in-row vs in-rack cooling based on peak rack density, not room average."

— ACwatt Technical Team

6. Video walkthrough (coming soon)

We are preparing a short video walkthrough that compares room cooling, in-row cooling, and in-rack cooling in a live data center environment. The video will cover:

• Airflow path comparison between room cooling and in-row cooling
• Inlet temperature measurement at different rack densities
• Installation footprint and maintenance access
• DCIM integration and monitoring

Check back soon, or contact our pre-sales team to request an early preview.

7. Frequently asked questions

Can I mix room cooling and in-row cooling in the same data center?

Yes. In fact, a hybrid approach is common in AI data centers. Room cooling handles the base load and low-density areas, while in-row cooling handles high-density racks. This allows phased upgrades without a full cooling system replacement.

How do I know if my racks need in-row cooling?

Check the actual power draw per rack, not the average. If any rack exceeds 10 kW and the inlet air temperature rises above the ASHRAE recommended range during peak load, in-row cooling should be evaluated.

Is in-rack cooling suitable for all AI workloads?

In-rack cooling is well suited for high-density AI racks above 30 kW. However, it requires sufficient rack space and careful condensate management. For densities above 50 kW, liquid cooling may be more appropriate.

What is the typical lead time for in-row cooling deployment?

In-row cooling units are typically factory-assembled and tested. Lead time depends on capacity and configuration, but standard units can usually be delivered within 6–10 weeks. Site installation is faster than room cooling upgrades because it does not require a full room redesign.

How does in-row cooling affect PUE?

In-row cooling can improve PUE by reducing the distance air must travel and allowing higher return air temperatures. However, the actual PUE improvement depends on the overall design, including containment, chilled water temperature, and ambient conditions.

What is the difference between in-row cooling and in-rack cooling?

In-row cooling units are placed between server racks, typically at the end of a row. In-rack cooling units are mounted directly inside or beside a single rack. In-rack cooling provides the shortest cooling path, but it also takes up rack space and has lower cooling capacity per unit. In-row cooling handles higher total heat loads per unit and is easier to maintain.