海外消息网
As the safeguard for artificial intelligence data centers (AIDCs), Huawei Thermal Management Unit (TMU) incorporates four layers of safety protection for liquid cooling systems, eliminating failure risks.
Arenowned AI research institute was training multiple leading foundation models in its AIDC. During a critical training phase, the liquid cooling system failed, causing coolant to leak across multiple critical servers. The servers immediately overheated, triggering alarms and emergency shutdown on some devices. Despite the staff’s emergency measures, some chips sustained irreversible damage from overheating. This incident halted a critical model training project, wasting significant prior investment in computation and time.
This is neither an isolated incident nor an exaggeration. With the boom of AI and big data, demand for data centers and computing power has grown exponentially, driving a substantial increase in device power density. In this context, traditional air cooling is no longer sufficient, and liquid cooling technology has emerged as an industry solution. At the core of the liquid cooling system, the Cooling Distribution Unit (CDU) is at a pivotal point in its transition from a “passive distributor” to an “active controller.” The evolution of the CDU from a basic “cooling system component” into an “intelligent core” with management and control capabilities is critical for the stability, efficiency, and safety of the entire system.
Challenges: The Urgent Need for Liquid Cooling System Evolution in the Era of Intelligent Computing
The computing power race is intensifying. User demands for greater computing power are driving chip thermal design power (TDP) consumption to soar. Traditional air cooling technology has been insufficient to meet the cooling requirements of scenarios with high power density. Therefore, the liquid cooling system is seeing rapid large-scale adoption.
As liquid cooling technology accelerates its penetration, the market has grown rapidly in recent years. The global data center liquid cooling market reached US$870 million in 2024 and is expected to grow at a compound annual growth rate (CAGR) of 51.93% from 2024 to 2030, reaching US$10.7 billion by 2030, according to Arizton, a market research firm. Among liquid cooling technologies, the cold plate liquid cooling system is a widely used and mature solution. However, as intelligent computing demands greater safety, efficiency, and cost-effectiveness, cold plate liquid cooling systems struggle to meet the requirements of large-scale operations due to their weaknesses.
• Safety: Emergencies such as coolant leakage and coolant supply interruption may directly cause device breakdown, jeopardizing the continuity of intelligent computing services.
• O&M: The traditional mode of manual inspection and shutdown maintenance is inefficient. It also increases O&M costs and the risk of service interruptions.
• Energy efficiency: The heat-exchange efficiency of plate heat exchangers is low, and the device operation is inconsistent. Therefore, cooling efficiency is limited, and energy consumption remains high.
• Deployment: CDU cleaning and other preparations alone take 7 to 15 days. This significantly slows down the construction and capacity expansion of AIDCs.
These core pain points hinder the large-scale deployment and efficient operation of AIDCs.
Key to Breakthroughs: CDU Transformation
A liquid cooling system consists of three core components.
• The primary-side system: It includes the cooling tower, hydronic module, and chilled water pipes. It is technologically mature and highly standardized, but with limited optimization potential.
• The secondary-side system: It includes the coolant pipe and cold plate. The cold plate is closely coupled to the server chip layout. It is highly customized, making standardization difficult.
