The Race to Cool the World’s Data

“Managing heat generated by data centers will become progressively more challenging – and, at the same time, necessary.” Chris Frye, Director of Chiller Sales, LG Electronics U.S.A. Inc.

According to McKinley, global demand for data center capacity could rise between 19% and 22% annually from 2023 to 2030, reaching an annual demand of 171 to 219 gigawatts (GW). This is quite a difference from the current use of 60 GW annually.

The primary reason for this growth? Demand for AI-ready datacenter capacity. McKinley predicts that around 70% of total demand for data center capacity will be for data centers equipped to host advanced-AI workloads by 2030, specifically generative AI.

To meet increasing demand, data centers’ power consumption has ballooned. Ten years ago, a 30-megawatt (MW) datacenter was considered large. Today, a 200-MW facility is considered normal.

One of the most significant issues arising from more, bigger, and more energy-consuming data centers is heat. According to Julia Borgini of Spacebarpress Media, “The more equipment packed into a facility, the greater the heat generated.” As chips generate more heat, efficient cooling solutions will become a critical focus.

Air Cooling

Air remains the most common cooling method, specifically using hot and cold aisles. This setup alternates rows of server racks to separate hot exhaust air from cool intake air. It’s simple, scalable, and works best when rack space is fully utilized. Systems like Computer Room Air Handlers (CRAHs) function much like commercial HVAC systems. But traditional air cooling alone is quickly hitting its limits.

“Air cooling is not enough,” said Christian Belady, distinguished engineer and vice president of Microsoft’s datacenter advanced development group in Redmond.

Related to air cooling, but not in the traditional sense, KyotoCooling is a strong contender. KyotoCooling uses a rotating thermal wheel to exchange hot internal air with cool outside air. Obviously, it only works in locations with consistently cool temperatures. It eliminates water usage, reduces carbon emissions, and cuts power needs by up to 92% compared to conventional CRAH systems. HP and United Airlines already use KyotoCooling in several data centers.

Liquid Cooling

Better than air is liquid, as liquid is more efficient than air at transferring heat. Think about a blow dryer versus a cold shower. The hot hair dryer will warm you up, but more slowly. In contrast, stepping into a cold shower creates an instant freezing sensation. Water has way more heat transfer power than air.

Options in this category include:

1. Cold Plates

These are metal plates (usually copper or aluminum) mounted directly on chips. Coolant flows through internal channels to draw heat away. Cold plate systems can be customized to fit specific chip designs and sizes, offering precision cooling for high-performance setups.

2. Direct-to-Chip Cooling

Flexible tubing delivers non-conductive coolant directly to the hottest components, such as CPUs or GPUs. The coolant absorbs heat and carries it away in a closed-loop system.

3. Immersion Cooling

This cutting-edge method submerges entire servers in a dielectric fluid specially made by 3M. The fluid absorbs heat, boils into vapor, then condenses back into liquid. It creates a highly efficient, two-phase cycle. Microsoft’s Quincy data center already uses this approach. Results show energy savings of 5% to 15% per server.

“If done right, two-phase immersion cooling will meet cost, reliability, and performance goals with a fraction of the energy used in air cooling.”

—Ioannis Manousakis, Microsoft Azure

4. Direct-to-die Cooling

Direct-to-die cooling is an advanced form of direct-to-chip liquid cooling. Instead of transferring heat through a heat spreader or cold plate, the coolant flows as close as physically possible to the actual silicon die—the core of the processor. The idea is that the closer you get to the die, the less thermal resistance there is. However, there is little room for error in this approach. Today, it’s mostly used in experimental, high-performance computing, or hyperscale AI deployments.

Cooling MethodDescriptionCooling Efficiency (1-5)Deployment Complexity (1-5)Maintenance Difficulty (1-5)Current Adoption (1-5)
Two-Phase Immersion CoolingSubmerges servers in a dielectric fluid that boils to remove heat; vapor condenses and recirculates.5552
Direct-to-Die CoolingLiquid coolant flows directly over the silicon die, bypassing thermal interfaces.5551
Cold Plate CoolingCoolant flows through a metal plate attached to the chip, removing heat via conduction.4334
Rear-Door Heat ExchangerChilled water flows through a heat exchanger on the back of the rack to remove hot air.3224
Hot and Cold AislesSeparates hot and cold airflow paths using containment to improve air cooling efficiency.2224
Geothermal CoolingUses stable underground temperatures to aid in cooling the data center with minimal energy.4432
Kyoto CoolingUses a large thermal wheel to transfer heat from hot exhaust air to cooler intake air efficiently.3322
Evaporative CoolingUses water evaporation to lower air temperature; commonly used in dry climates to reduce cooling load.3333
Solar CoolingUtilizes solar thermal energy to power absorption chillers or desiccant cooling systems for sustainable cooling.3441

1 = Low and 5 = High

Renewable Cooling

1. Geothermal

By tapping into the earth’s steady underground temperature, data centers can maintain cooling year-round with minimal energy input. Businesses using this technique include Iron Mountain (Pennsylvania), Verne Global (Iceland), and Green Mountain (Norway).

2. Evaporative Cooling

Also known as swamp cooling, this technique uses water evaporation to lower temperatures. As you may guess, it’s effective in dry climates but less so in humid environments. Hello Ina, Peru, goodbye Houston, Texas.  

3. Solar Cooling

This approach uses solar thermal energy to power absorption chillers. It’s still considered a niche cooling option but represents an eco-friendly path forward for sustainable cooling, in very sunny locations.

Hybrid Cooling

Between the worlds of liquid and air cooling is a rear-door heat exchanger (RDHx), a system designed for high-density data centers. It involves a liquid-cooled door attached to the back of a server rack that pulls heat directly from hot exhaust air.

Instead of flooding the data center, hot air passes through a radiator-style panel on the rear door. Chilled coolant circulates through this panel and absorbs the heat. Cooler air is then discharged into the room. It can retrofit existing racks, but still requires high-quality plumbing and airflow.

So Hot It’s Cold

The significance of cooling extends beyond infrastructure, affecting energy supply, climate, land use, the societal value of AI. As chips grow more powerful and workloads more demanding, innovators must develop smarter and more energy-efficient solutions.

Who knows? AI cooled by one of these methods may generate an entirely new cooling process.

Penney Berryman

Penney Berryman, MPH (she/her), is the Content Editor for Planet Mainframe. She also writes health technology and online education marketing materials. Penney is based in Austin, Texas.

Connect with her on LinkedIn.

Leave a Reply

Your email address will not be published. Required fields are marked *