Can air-cooling handle the heat from a 1-GW data centre?

UPSC Syllabus: Gs Paper 3- Science and Technology

Introduction

Large data centres consume huge amounts of electricity, and most of this energy ultimately becomes heat that must be removed continuously. A 1-GW data centre can, in principle, generate 1 GW of heat, creating a major cooling challenge. Air-cooling uses less water and costs less initially, but the rising heat density of modern AI chips makes air-cooling alone increasingly difficult for such large facilities.

Why Do 1-GW Data Centres Generate Enormous Heat?

1. Processors inside servers: Data centres contain multiple servers, and each server contains processors made up of billions of transistors that process data.
2. Heat from transistors: Each transistor generates some heat because electrical resistance and charging or discharging release energy during data processing.
3. Heat accumulation: A data centre may contain several million processors and trillions of transistors, making the combined heat generation considerable.
4. 1-GW heat load: Assuming the entire 1 GW powers the data centre, the facility must remove roughly 1 GW of heat continuously.
5. First stage of heat movement: Heat first moves from the silicon die to a heat sink, such as a cooling plate, before further removal.

What Are the Major Cooling Technologies for Data Centres?

1. Air-Cooling

• Working: Fans move cool air through or around server racks and remove the resulting hot air.
• Main configurations: Computer Room Air Conditioners (CRACs), Computer Room Air Handlers (CRAHs), hot-aisle/cold-aisle containment and free-cooling are common air-cooling arrangements.

2. Direct Liquid Cooling

• Working: A liquid is brought directly to heat-generating components and carries the heat to a heat exchanger.
• Cold-plate cooling: A metal plate with liquid-carrying channels is placed against the processor to transfer heat efficiently.
• Advantage: Water-based liquids generally have higher heat capacity than air, allowing them to absorb more heat.

3. Immersion Cooling

• Working: Electronic components are immersed in a non-conductive liquid that absorbs their heat.
• Types: Single-phase systems pump the liquid through a heat exchanger, while two-phase systems allow the liquid to boil and later condense.
• Advantage and limitation: It can handle very high power densities, but needs specialised hardware and complex coolant maintenance.

4. Evaporative Cooling

• Working: Heat is transferred from the servers to water, which is then evaporated through cooling towers or evaporative condensers.
• Limitation: It can be highly energy-efficient in low-humidity conditions but requires large quantities of water.

5. Dry-Cooling

• Working: A finned heat exchanger releases heat directly into ambient air without using evaporating water.
• Limitation: It avoids water demand but needs larger heat-exchange surfaces, especially in warm weather.

6. Heat-Rejection Options

• Chilled-water systems: A central chilled-water plant provides cooling for the facility.
• Geothermal heat rejection: Heat is transferred to underground heat exchangers, but the technology is experimental and difficult to scale.
• Natural-water cooling: A nearby lake or another large natural water body acts as the heat sink.
• Heat reuse: Waste heat can support industrial processes or desalination, although heat pumps may be needed because the heat is usually at a low temperature.

Can Air-Cooling Handle the Heat from a 1-GW Data Centre?

1. Lower initial cost: Air-cooling generally requires lower upfront infrastructure costs, while liquid-cooled facilities can cost 7–10% more because of specialised systems.
2. Mature technology: Air-cooling has established technicians and maintenance practices, making it a proven and widely understood cooling method.
3. Suitable for lower-density racks: Engineers often prefer air-cooling for racks producing 20 kW or less, regardless of the cooling system used elsewhere.
4. Advantage of cool environments: In naturally cool or arid conditions, air-side economisers can reduce energy used for mechanical cooling by up to 70%.
5. Thermal wall: Air-cooling can remove only around 40 kW per rack, while modern AI racks can generate 120–150 kW.
6. AI chip heat density: Nvidia Blackwell GPUs can generate 700–1,000 W each, making high-density AI clusters difficult to cool using air alone.
7. Rising operational burden: Cooling high-density racks through air may require wind-tunnel-like arrangements, sharply increasing power consumption and total operating costs.
8. Noise problem: Large numbers of fans, chillers and air-handling units can produce up to 100 dB, requiring acoustic control.

Why Is Hybrid Cooling Emerging as a Practical Solution?

1. Matching cooling to heat density: Different parts of a data centre produce different amounts of heat, so one cooling method may not suit every rack.
2. Air-cooling for low-density areas: Conventional air-cooling can continue serving servers and components whose heat output does not justify liquid-cooling.
3. Rear-door heat exchange for medium density: Rear-door systems can extend air-cooling capacity by using chilled water to absorb heat from outgoing server air.
4. Direct-to-chip cooling for AI: DTC liquid-cooling can directly remove heat from high-density AI processors, including Nvidia Blackwell systems.
5. Better heat transfer: Water-based liquids generally absorb more heat than air for the same volume because they have higher heat capacity.
6. Immersion for high power density: Immersion cooling can handle very high power densities because the cooling liquid directly contacts the electronics.
7. Specialised requirements: Liquid and immersion systems need specialised hardware, piping, leak detection, coolant distribution and, in some cases, complex coolant maintenance.

Google and TCS: Contrasting Cooling Approaches

1. Google’s air-cooling plan: Google has reportedly said its planned 1-GW Visakhapatnam data centre will use air-cooling, addressing concerns about water demand.
2. Google’s water commitment: Google has said it will replenish 120% of water consumed for non-cooling needs by 2030, measured through the Volumetric Water Benefit Accounting standard.
3. Watershed management: Google also plans rainwater catchment, groundwater recharge infrastructure, water pumps and water ATMs around the data-centre project.
4. TCS Hyper Vault approach: TCS’s 1-GW Hyper Vault is being designed for higher-density infrastructure and greater use of liquid cooling.
5. Direct-to-chip cooling: The Hyper Vault will feature DTC cooling, which is suited to high-density AI workloads.
6. Renewable energy: TCS’s Hyper Vault will also use renewable energy, linking its cooling and power infrastructure with energy-efficiency concerns.

Way Forward

1. Choose cooling according to workload: Cooling systems should match the heat density of different servers instead of relying on one technology across the entire facility.
2. Expand hybrid cooling: Combining air-cooling, rear-door heat exchangers and DTC liquid-cooling can balance performance, water use and infrastructure costs.
3. Consider local conditions: Naturally cool or arid conditions can make air-side economisers more effective and reduce energy used for mechanical cooling.
4. Improve water efficiency: Where water-based systems are used, watershed management, rainwater collection and groundwater recharge can reduce pressure on local water resources.
5. Explore heat reuse: Waste heat can serve industrial processes or desalination, although higher temperatures may require heat pumps to make the heat useful.
6. Assess alternative systems carefully: Geothermal cooling, natural-water cooling and dry-cooling can help in suitable locations, but their scalability, water needs or infrastructure requirements differ.

Conclusion

Air-cooling has clear advantages in cost, maturity and water use, but it faces a major thermal limit as AI rack densities rise. A 1-GW facility relying only on air-cooling is therefore increasingly impractical. Hybrid cooling, using air for lower-density workloads and liquid systems for high-density AI clusters, offers a more practical balance between heat removal, energy use, water demand and cost.

Question for practice:

Examine the challenges of air-cooling in data centres and discuss the practical cooling solutions for managing heat efficiently.

Source: The Hindu