Research on the Liquid Cooling Industry for Servers

 

IDC is a typical high-energy consumption industry and faces energy consumption policy constraints. Developed regions have introduced energy consumption restriction policies, and the energy consumption of IDC is mainly concentrated on cooling. The thermal power consumption of AI servers is more than five times that of ordinary servers, requiring more efficient liquid cooling to achieve cooling and reduce PUE levels.

 

Driven by Nvidia and domestic operators, the penetration rate of liquid cooling (cold plate) in servers will exceed 10% in 2024, entering a rapid growth stage. It is expected that the domestic liquid cooling penetration rate will reach 30% by 2028, with a market size of over 20 billion RMB.

 

The most valuable components in a liquid cooling system are the liquid cooling plates and pipes. It is estimated that the domestic market will be around 10 billion RMB by 2028. Since the product's technical barriers are not high, companies that enter the market early and establish partnerships with core chip and server manufacturers will have significant advantages.

 

Companies in the supply chain of Nvidia and Huawei will have better investment opportunities, with special attention to Inovance Technology, Gaolan Co., Ltd., Jingyan Technology, and Siquan New Materials.

 

 

I AI Servers/IDCs Require Liquid Cooling Systems

 

1) High Power Consumption of AI Servers Requires Liquid Cooling

A standard CPU server with dimensions of 4U (737mm × 437mm × 178mm) typically consumes around 800W of power, with a heat output of approximately 200W, corresponding to 50W/U of heat output per unit space. In contrast, an AI server equipped with 8 GPU chips in the same size format has a total power consumption of about 5KW and heat output of around 1000W, corresponding to 250W/U of heat output per unit space, which is approximately four times that of an ordinary server. The maximum power dissipation of a typical 4U air-cooled server is 600W, far below the 1000W heat dissipation requirements of an AI server, making liquid cooling necessary.

 

▲ AI Server Heat Dissipation Needs are about 5 Times Higher than Ordinary Servers

 

2) AI Server IDCs Must Use Liquid Cooling Systems

To maintain the server room temperature between 20-25°C, a general server IDC using air cooling can achieve a minimum PUE of 1.3, barely meeting policy requirements. However, an AI server IDC using air cooling can only achieve a PUE of 1.6. Therefore, only by using a liquid cooling system can the PUE of the server room be kept below the required 1.3.

 

▲ Comparison of Various Cooling Solutions

 

According to the latest industry chain information, Nvidia will fully adopt liquid cooling starting with the B100 and will launch a fully customized liquid cooling solution in 2024, significantly accelerating the development of liquid-cooled servers.

 

 

II Liquid Cooling Plates/Liquid Cooling Pipes

 

1)Optimizing Thermal Management with Liquid Cooling Plates

Liquid cooling plates transfer the heat from heat-generating components indirectly to the cooling liquid circulating within a closed-loop system via the cold plate, which is typically made of thermally conductive metals like copper or aluminum. The cooling liquid then carries the heat away. There are four key technical requirements for the quality of liquid cooling plates: first, high cooling capacity; second, high reliability to ensure the cold plate's seal; third, precise heat dissipation design to avoid large temperature differences within the system; and fourth, strict control of the weight of the cold plate to prevent the cooling system from significantly reducing the energy density of the overall system.

 

Liquid cooling plates are manufactured using machining processes, allowing the internal channel dimensions and paths to be freely designed. This makes them suitable for thermal management products with high power density, irregular heat source layouts, and space constraints. They are mainly applied in the design of cooling products for wind power converters, photovoltaic inverters, IGBTs, motor controllers, lasers, energy storage power supplies, supercomputing servers, and other fields, but are less commonly used in power battery systems.

2) Cold Plate Liquid Cooling is the Mainstream Solution

 

 

▲ Cold Plate and Immersion Liquid Cooling Solutions

 

There are two main liquid cooling solutions: cold plate and immersion. The cold plate solution uses a hollow liquid cooling plate attached to the CPU and GPU surfaces. The heat is carried away by the coolant flowing inside, reducing the IDC's PUE to 1.2. The immersion solution directly submerges the entire server in a special insulated cooling liquid, removing heat through the circulation of the liquid. This method can achieve a PUE of less than 1.1 and is mainly used for supercomputers with extremely high power density. It is expected that cold plate liquid cooling will become the mainstream liquid cooling solution in the future, accounting for more than 90% of liquid cooling system usage.

 

3) Liquid Cooling Plates and Liquid Cooling Pipes

 

 

▲ Schematic of Cold Plate Liquid Cooling System

 

A cold plate liquid cooling system consists of three main parts: the internal heat conduction system of the server chassis (liquid cooling plates and pipes), the external circulation system of the chassis (pumps, valves, pipes), and the external heat exchange system of the server room (cooling towers, etc.). Among these, the heat exchange system and cooling towers are considered common electromechanical equipment and have low value. Therefore, the value of the liquid cooling system is mainly concentrated on the liquid cooling plates and pipes inside the chassis.