Liquid Cooling Heat Exchanger Units
Transforming Air-Cooled Datacenters for AI Workloads
Bringing liquid cooling to air-cooled datacenters.
Microsoft is continuously architecting and optimizing every layer of the cloud and AI infrastructure stack to meet the demands of our AI advancements. Modern AI systems powering AI workloads demand higher power at higher densities, leading to a need to develop new methods of cooling to manage heat and power consumption.
To address this challenge, liquid cooling has emerged as the preferred solution, using circulating fluids to dissipate heat and prevent overheating to ensure efficient operation of AI hardware. At Microsoft, as we continue to expand our datacenter fleet to help enable the world’s AI transformation, we’re also developing methods for using air-cooled datacenters to provide liquid cooling capabilities for AI.
Microsoft is enabling advanced cooling innovations for modern AI systems, revolutionizing how datacenters housing GPUs and AI accelerators are cooled. We are integrating liquid cooling in legacy air-cooled datacenters, retrofitting existing facilities and quickly deploying to meet the cooling needs of AI systems.
The introduction of the dedicated liquid cooling “sidekick” for Microsoft’s Maia 100 marked the first use of direct-to-chip (DTC) liquid cooling units in existing air-cooled datacenters. Building on this innovation, we continue to evolve our liquid cooling technology to enhance cooling capacity for modern AI processors. At the 2024 Open Compute Project (OPC) Global Summit, we showcased our advanced liquid cooling heat exchanger units (HXUs), designed to support next-generation GPUs and AI accelerators from Microsoft and other industry leaders.
Our approach focuses on innovating the entire cloud system—from silicon to software—to deliver improved performance, greater choice, enhanced resiliency, and broader availability for our customers. By co-designing and co-optimizing hardware, software, networks, and datacenters, we tackle the critical challenges of performance, energy efficiency, and capital investment.
Through the OPC, we are opensourcing our HXU designs, sharing our development work to help enable closed-loop liquid cooling across AI datacenters industry-wide. This collaborative effort ensures that innovations in thermal management not only benefit Microsoft but also drive progress for the entire computing ecosystem.
Liquid Cooling Heat Exchanger Unit
To support higher densities of AI accelerators per rack, we are utilizing standalone liquid-to-air heat exchanger units. These units enable legacy datacenters—traditionally unequipped for using direct-to-chip liquid cooling—to efficiently handle the demands of new and advanced AI workloads.
Carbon Negative Goals
By 2030, we’ll be carbon negative, and by 2050, we’ll remove our historical emissions since our founding in 1975.
By 2030, we’ll replenish more water than we consume across our direct operations.
Heat generated by high-performance chips is transferred directly to a liquid coolant via cold plates attached to the processors. The coolant effectively transfers the heat from the chip, bypassing reliance on traditional airflow for component-level cooling.
These hoses connect the cold plates in the server rack to the HXU and help maintain the flow of hot and cold liquid between the HXU and the server racks.
Components
- Server rack
- Heat exchanger core
The process begins with single-wide racks designed to enhance heat dissipation capacity. These racks are securely packaged in leak-proof materials to ensure safe transport and installation, preserving the system’s integrity before it’s brought online.
Cool air, usually supplied from underfloor vents or overhead ducts, is drawn into the front of the servers in the cold aisle. The servers’ internal fans guide this cool air over their components, absorbing the heat generated by processors, memory, and other high-performance hardware.
Server Rack Airflow Design
High-efficiency pumps are activated to control fluid flow precisely, adapting the cooling system to the specific thermal demands of each rack. The modular fluid connections between HXUs and server racks allow for flexible configurations, optimized to match the setup requirements.
By 2030, we’ll be zero waste across our direct waste footprint.
Fans
Fans are placed at the rear of each HXU to maintain uniform airflow. They prevent preheating within the HXU and help maintain consistent cooling performance.
Microsoft designs and operates our datacenters to support our climate goals: to become carbon negative, water positive, and zero waste all by 2030, while enabling others to do the same.
Protect and Preserve Ecosystems
Finally, strategically placed sensors—including quick disconnects, leak detection ropes, and drip pans at the HXU base—are engaged to detect, contain, and manage any potential leaks, ensuring that temperatures remain precisely regulated for consistent performance.