How to cope with the increasing demand for cooling technology in AI servers
At present, the heat dissipation module is mainly composed of active and passive hybrid heat dissipation technology containing thermal pipes. The heat pipe heat dissipation module is designed and combined with components such as air diffusers, heat sinks, and thermal pipes, which can provide a uniform temperature heat dissipation operating environment for internal electronic components, making electronic equipment operation more stable. With the trend of multifunctional and lightweight terminal electronic products, the thermal module factory has turned to design thermal solutions mainly based on vapor chamber and heatpipe .
The heatsink module is divided into two types: "air-cooled heatsink" and "liquid cooled heatsink". Among them, air-cooled solution is the use of air as a medium, through intermediate materials such as heat interface materials, vapor chamber (VC) or heat pipes, and is dissipated by convection between the heat sink or fan and the air. "Liquid cooled is mainly achieved through convection with the liquid, thereby cooling the chip, However, as the heat generation and volume of the chip increase, the thermal design power consumption (TDP) of the chip increases, and the use of air-cooled heat dissipation gradually becomes insufficient.
With the development of the Internet of Things, edge computing, and 5G applications, data AI has driven global computing power into a period of rapid growth. According to research firm TrendForce, the shipment volume of AI servers equipped with GPGPUs (General Purpose GPUs) accounted for about 1% in 2022. However, in 2023, driven by ChatGPT applications, it is expected that the shipment volume of AI servers will grow by 38.4%, and the overall compound annual growth rate of AI server shipments from 2022 to 2026 will reach 29%.
There are two main directions for the design of the next generation of heatsink modules. One is to upgrade the existing heat dissipation modules with 3D vapor chamber (3DVC), and the other is to introduce a liquid cooling system, using liquid as the convective medium to improve thermal efficiency. Therefore, the number of liquid cooling test cases will significantly increase in 2023, but 3DVC is only a transitional solution. It is estimated that from 2024 to 2025, we will enter the era of parallel gas cooling and liquid cooling.
With the rise of ChatGPT, generative AI has driven up server shipments, along with requirements for upgrading the specifications of heatsink modules driving them towards liquid cooling solutions to meet the strict requirements of servers for heat dissipation and stability. At present, the industry mostly uses single-phase immersion cooling technology in liquid cooling to solve the heat dissipation problem of high-density heating servers or parts, but there is still an upper limit of 600W, because ChatGPT or higher-order servers need a heat dissipation capacity of more than 700W to cope.
Based on the fact that the cooling system accounts for approximately 33% of the total energy consumption in the data center, reducing total electricity consumption and reducing power usage efficiency includes improving the cooling system, information equipment, and using renewable energy. Water has a heat capacity four times that of air. Therefore, when introducing a liquid cooling cooling system, only 1U of space is needed for the liquid cooling plate. According to NVIDIA testing, to achieve the same computing power, the number of cabinets required for liquid cooling can be reduced by 66% Energy consumption can be reduced by 28%, PUE can be reduced from 1.6 to 1.15, and computational efficiency can be improved.
Fast computing leads to continuous improvement in TDP, and AI servers have higher requirements for heat dissipation. Traditional heat pipe cooling is approaching its limit, and it is inevitable to introduce liquid-cooled thermal solutions.
