Key Points of Cooling and Power Consumption Management for Data Centre GPUs

1. Root Cause: the Leap in Power Density

Conventional rack power density is generally 5-8 kW. A server fitted with eight high-end GPUs can reach 6-10 kW or more on its own - a single AI rack easily exceeds 30 kW, and in some cases 80 kW.

This has two direct consequences:

• Cooling: conventional air cooling struggles to remove this much heat within a limited space
• Power supply: distribution, cable sizing and PDU specifications all need to be reassessed

2. Comparison of Cooling Options

Cold-plate liquid cooling is currently the mainstream transitional option for AI server rooms: cooling capacity is sufficient, the retrofit is relatively manageable, and the impact on the existing server room structure is comparatively small.

3. Key Points of Cooling Planning

1. Cold aisle containment is the foundation

Whichever option is used, mixing of cold and hot airflow greatly reduces efficiency. The airflow direction must be uniform (the standard being front intake and rear exhaust).

2. Inlet temperature is the core metric

As recommended by ASHRAE, the inlet temperature should be kept within the 18-27 °C range. Too high causes throttling; too low wastes energy.

3. Focus on hot spots rather than the average

A healthy average does not mean there is no problem. Local hot spots in a single rack or a single device are the direct cause of throttling and failure.

4. Plan high-density zones separately

Do not mix AI racks with general-purpose racks; otherwise the local heat load will drag down the cooling efficiency of the whole area.

5. Confirm compatibility in advance for liquid cooling

Cold plate, quick-connect coupling and coolant specifications must be confirmed with the server manufacturer, and leak detection and redundancy should also be considered.

4. Key Points of Power Consumption Management

1. Recalculate power distribution capacity

Base the calculation on full-load power consumption and leave headroom. The inrush current of an 8-GPU server at start-up must be taken into account.

2. Three-phase balance

For high-power racks, ensure that the three-phase load is balanced to avoid single-phase overload.

3. Power redundancy

Dual power feeds should come from different distribution circuits, and switchover tests should be carried out regularly - untested redundancy is equivalent to no redundancy.

4. Monitoring and power limiting

• Monitor whole-rack power, per-server power and GPU power consumption trends
• Where necessary, set a power cap to avoid overload
• Set alarm thresholds so that action is taken before protection is triggered

5. PUE and cost

PUE = total data centre energy consumption / IT equipment energy consumption. The cooling method directly affects PUE. Liquid cooling requires higher up-front investment, but its long-term energy cost advantage is clear in high-density scenarios.

5. Implementation Checklist

• Whether rack power density has been recalculated against the full-load configuration
• Whether power distribution capacity, cable sizing and PDU specifications match
• Whether the cold aisle is contained and the airflow direction is uniform
• Whether the inlet temperature is within the recommended range
• Whether power and temperature monitoring and alarms have been deployed
• Whether redundant power supply has actually been switchover tested
• Whether high-density zones are separated from general-purpose racks
• If liquid cooling is adopted, whether compatibility, leak detection and maintenance procedures have been confirmed

6. Summary

The planning sequence for compute infrastructure should be: work out power consumption and cooling first, then discuss compute scale. Cooling and power supply are hard constraints; once the server room has been built, changes cost several times more. For enterprises that are expanding AI compute capacity, the value of early-stage planning here is far greater than that of the equipment itself.

Note: This article is provided for technical exchange and experience sharing. Always refer to the original manufacturer's technical documentation and the conditions of your own environment. Maintenance and change operations should be carried out by qualified engineers.