By Willem Weber, mechanical engineering consultant, Digital Parks Africa
The rapid adoption of artificial intelligence (AI) and high-performance computing (HPC) is changing the engineering requirements of data centres. As computing power increases, rack densities are rising sharply, creating heat loads that conventional air-cooling systems are increasingly difficult to manage.
Liquid cooling changes that balance by allowing considerably more heat to be transported through a comparatively compact distribution system. Kues1 | Magnific.com
This is part one of a two-part series.
For data-centre operators, cooling is no longer simply a supporting utility. It is becoming a fundamental consideration in determining how much computing capacity a facility can accommodate, how efficiently it can operate and how readily it can scale.
Traditionally, data-centre cooling has been based on a relatively straightforward principle: move sufficiently cool air through the data hall and remove the heat generated by the servers.
Conventional racks typically operated in the region of 5kW to 15kW. As densities increased, chilled-water systems could be introduced at facility level, but the fundamental approach remained the same – cool the air and move it to where the heat is being generated.
AI is changing that equation.
When air becomes impractical
AI workloads can produce rack densities of 60kW to 100kW, with some deployments reaching approximately 150kW in a standard cabinet. At these densities, simply increasing airflow is no longer a practical solution.
Air has a relatively low heat-carrying capacity compared with liquid. Attempting to remove the heat from a 100kW rack using air requires enormous airflow volumes and pressure, increasing both the physical complexity and energy consumption of the cooling system.
This is driving the industry towards liquid cooling. Liquids such as water and dielectric fluids can transport substantially more heat than air, allowing heat to be removed much closer to the source. In direct-to-chip cooling, for example, cold plates are mounted directly to high-power processors, transferring heat into a liquid loop.
The result is a fundamental change in the cooling architecture. Fans and large volumes of moving air become less important, while pumps, heat exchangers and coolant distribution systems take on a greater role. Cooling Distribution Units (CDUs) can manage secondary liquid loops supplying coolant to high-density racks.
Liquid cooling therefore isn’t simply another type of air-conditioning system. It represents a change in the way the data centre’s thermal load is managed.
Cooling closer to the chip
The attraction of direct-to-chip cooling is that it removes heat at its source rather than relying on the surrounding air to transport it. This can significantly reduce the amount of air that has to be moved through a high-density data hall. Conventional air cooling can still serve lower-density equipment and other areas of the facility, but the highest-density AI workloads increasingly require a more targeted approach.
This does not mean that air cooling disappears. Many data centres will continue to operate hybrid environments in which conventional air-cooled racks coexist with liquid-cooled high-density equipment. The appropriate solution depends on the computing load, rack density, equipment configuration and the requirements of the facility.
For African data-centre operators, this distinction is particularly important. The continent’s data-centre market is expanding, but facilities must also contend with electricity availability and cost, high ambient temperatures in many locations and, in some markets, constrained water resources.
The cooling system therefore needs to be designed around the location rather than simply adopting a technology because it is becoming fashionable globally.
A new design threshold
The increasing density of AI racks effectively creates a new threshold for data-centre cooling design. At relatively low rack densities, air remains an effective and well-understood heat-transfer medium. As the heat generated within each cabinet increases, however, the physical infrastructure required to move sufficient air becomes progressively more demanding.
Liquid cooling changes that balance by allowing considerably more heat to be transported through a comparatively compact distribution system.
But this introduces another engineering challenge. Once liquid becomes part of the primary cooling architecture, the implications extend beyond the rack itself and into the structure, pipework, heat-rejection plant and overall building design.
That is where the move to liquid cooling becomes more than a change in cooling technology. It becomes a change in the physical design of the data centre – with consequences for structure, heat rejection, energy efficiency and future expansion.
Source:
