Schneider Electric publishes new White Paper on Liquid Cooling for AI Data Centres

Schneider Electric has released white paper 133 titled "Navigating Liquid Cooling Architectures for Data Centres with AI Workloads." The paper provides a thorough examination of liquid cooling technologies and their applications in modern data centres, particularly those handling high-density AI workloads.

The demand for AI is growing at an exponential rate. As a result, the data centres required to enable AI technology are generating substantial heat, particularly those containing AI servers with accelerators used for training large language models and inference workloads. This heat output is increasing the necessity for the use of liquid cooling to maintain optimal performance, sustainability, and reliability.

Schneider Electric’s latest white paper guides data centre operators and IT managers through the complexities of liquid cooling, offering clear answers to critical questions about system design, implementation, and operation.

Understanding Liquid Cooling Architectures

Over the 12-pages authors Paul Lin, Robert Bunger, and Victor Avelar identify two main categories of liquid cooling for AI servers: direct-to-chip and immersion cooling. They describe the components and functions of a coolant distribution unit (CDU), which are essential for managing temperature, flow, pressure, and heat exchange within the cooling system.

“AI workloads present unique cooling challenges that air cooling alone cannot address,” said Robert Bunger, Innovation Product Owner, CTO Office, Data Centre Segment, Schneider Electric. “Our white paper aims to demystify liquid cooling architectures, providing data centre operators with the knowledge to make informed decisions when planning liquid cooling deployments. Our goal is to equip data centre professionals with practical insights to optimise their cooling systems. By understanding the trade-offs and benefits of each architecture, operators can enhance their data centres’ performance and efficiency.”

The white paper outlines three key elements of liquid cooling architectures:

1. Heat Capture Within the Server: Utilising a liquid medium (e.g. dielectric oil, water) to absorb heat from IT components.

2. CDU Type: Selecting the appropriate CDU based on heat exchange methods (liquid-to-air, liquid-to-liquid) and form factors (rack-mounted, floor-mounted).

3. Heat Rejection Method: Determining how to effectively transfer heat to the outdoors, either through existing facility systems or dedicated setups.

Choosing the Right Architecture

The paper details six common liquid cooling architectures, combining different CDU types and heat rejection methods, and provides guidance on selecting the best option based on factors such as existing infrastructure, deployment size, speed, and energy efficiency.

With the increasing demand for AI processing power and the corresponding rise in thermal loads, liquid cooling is becoming a critical component of data centre design. The white paper also addresses industry trends such as the need for greater energy efficiency, compliance with environmental regulations, and the shift towards sustainable operations.

“As AI continues to drive the need for advanced cooling solutions, our white paper provides a valuable resource for navigating these changes,” added Bunger. “We are committed to helping our customers achieve their high-performance goals while improving sustainability and reliability.”

Providing the Industry with AI Data Centre Reference Designs

This white paper is particularly timely and relevant in light of Schneider Electric's recent collaboration with NVIDIA to optimise data centre infrastructure for AI applications

This partnership introduced the first publicly available AI data centre reference designs, leveraging NVIDIA's advanced AI technologies and Schneider Electric's expertise in data centre infrastructure.

The reference designs set new standards for AI deployment and operation, providing data centre operators with innovative solutions to manage high-density AI workloads efficiently.