Deep Dive: Liquid-Cooled AI Switches for the Next Generation

AI-driven Liquid Cooling: the Right Fit for Today

At the end of 2022, OpenAI sparked an AI boom, and topics related to liquid cooling were also widely discussed.

In 2024, NVIDIA launched the NVL72/NVL36 solutions, further increasing the demand for constructing fully liquid-cooled data centers.

In the data center field, the two liquid cooling methods in large-scale use are cold plate cooling and immersion cooling. Other liquid cooling methods such as spray cooling are not deployed on a large scale.

Why Do Switches Require Liquid Cooling?

For networks, liquid cooling of switches is crucial. For servers, the emphasis is on GPU liquid cooling.

In the past 12 years, the power consumption of typical switch systems has increased by 22 times, optics by 26 times, and ASIC SerDes by 25 times. The rate of power consumption growth has outpaced the improvement in air cooling capabilities.

Switch liquid cooling demand 1: The device 128 x 3.2T DR16-DSP is used as an example. The power consumption of optical transceivers is higher than that of GPUs, while their temperature resistance is lower than that of GPUs. Radiator space scalability is much more constrained for optical transceivers compared with GPUs.

Switch liquid cooling demand 2: The power consumption and heat flux density of switching chips are higher than those of GPUs.

→Click here to submit your detailed product solution requirements.

Difficulties in Designing Liquid-Cooled Switches

Difficulty 1: The cold plate system of switches is assembled in three dimensions, while that of GPUs is assembled in two dimensions. The assembly difficulty in three dimensions is far greater than that in two dimensions.

Difficulty 2: Hardware inevitably experiences wear and tear, and liquid cooling systems are no exception. The impact of a single switch failure is much greater than that of a single server. How can we ensure that a switch remains operational even in the case of liquid leakage?

Difficulty 3: The cold plate design of belly-belly optics is easier, but the system cost is higher. The cost of the stack optics solution is lower. However, as of 2024, no cold plate solution of bottom optics in stack cage is available on the market.

Liquid Tigatron Switches Jointly Developed by Alibaba and Ruijie

For Difficulty 1: The customized automated liquid-cooling assembly platform based on the S³IP UNP concept has been verified through mass delivery. This platform can realize three-dimensional rapid assembly and airtight testing of the liquid-cooling kit for TH5 Tigatron switches.

For Difficulty 2: The TH5 Liquid Tigatron switch adopts a design that separates liquid circuits and electric circuits. Liquid leakage in high-risk areas will not leak onto the circuit board, thereby improving availability.

For Difficulty 3: A low-cost solution with water pipes embedded in aluminum plates is used. This solution supports cascaded liquid cooling of multiple 2 x 4 stack cage upper and lower optics, addressing industry difficulties. (One cold plate covers 22 optical transceivers to ensure good contact and heat dissipation of all transceivers. The thermal performance does not degrade after 50 to 100 plug-in and unplugging cycles.)

What Else Does Liquid Cooling Bring Besides Cooling?

Liquid cooling is not only a simple heat dissipation method, but more like a lubricant for the "innovation engine" in the network device field.

For example, the recently popular CPO "corrosion inhibitor" in vertical switch design requires liquid cooling to continue evolving.

For the CPO switch that has been controversial in the industry for many years, cold plate liquid cooling is also a mandatory requirement.

The Era of Liquid Cooling Is Upon Us

Liquid cooling is about to embrace a larger-scale market explosion. Let us witness the rise of more liquid cooling design cases.