White Paper No. 133
Navigating Liquid Cooling Architectures for Data Centers with AI Workloads
Navigating Liquid Cooling Architectures for Data Centers with AI Workloads
White Paper No. 110
This white paper explores the necessity of liquid cooling systems in data centres, particularly for AI workloads that generate significant heat. Traditional air cooling is often inadequate for AI servers with high thermal design power (TDP), making liquid cooling essential for maintaining performance and reliability.
Liquid cooling offers several benefits such as enhanced energy efficiency and lower noise levels. The paper categorizes liquid cooling into two main types: direct-to-chip and immersion cooling. It then breaks down the cooling ecosystem into three critical elements: heat capture within the server, the coolant distribution unit (CDU) type, and the method of rejecting heat outdoors. The focus is primarily on the CDU types and heat rejection methods.
The CDU is a crucial component, isolating the IT fluid loop from the broader cooling system and managing functions like temperature, flow, and pressure control, fluid treatment, and heat exchange. CDUs vary based on two main attributes: the type of heat exchange (e.g., liquid-to-air, liquid-to-liquid) and their capacity and form factor (e.g., rack-mounted or floor-mounted).
The paper identifies six common liquid cooling architectures, each combining a heat rejection method with a CDU type. The selection process involves two steps: choosing the heat rejection method and then selecting the CDU based on capacity and form factor.
Three primary heat rejection methods are discussed:
- Rejecting Heat to Air in the IT Space: This method uses a liquid-to-air CDU, compatible with existing air-cooled infrastructure, making it ideal for small-scale deployments and scenarios where rapid deployment is essential. However, it is less efficient and can be space-consuming in larger deployments.
- Rejecting Heat to Facility Water Systems: A liquid-to-liquid CDU transfers heat to an existing facility water system, offering higher energy efficiency and suitability for mid to large-scale deployments. This method integrates well with existing infrastructure but requires more installation work.
- Dedicated Heat Rejection System: This approach uses an independent system optimized for liquid cooling, providing the highest energy efficiency. It is ideal for large-scale deployments but involves significant investment and longer deployment times.
Rack-mounted CDUs are suitable for individual racks, offering easier integration, but become costly with larger deployments. Floor-mounted CDUs are better for extensive deployments, providing higher density and efficiency but requiring more space.
In conclusion, the paper provides a framework for selecting the appropriate liquid cooling architecture, helping data centre operators make informed decisions based on their specific needs, infrastructure compatibility, and priorities like deployment speed and energy efficiency.
