Immersion cooling: a promising cooling technology for servers, provided it is mastered

Immersion cooling consists of completely submerging the electronic components of servers in a dielectric liquid (a mineral oil to prevent short circuits and oxidation) to cool them. Server cooling is ensured by the circulation of the oil through convection (natural, or forced by a pump), which limits the use of energy-intensive air conditioning systems. This system is all the more effective as oil is an excellent heat conductor (much more so than air).

In practical terms, the heat produced by the electronic components of the servers (CPUs, GPUs, etc.) is transferred to the oil, which itself evacuates this heat to a water loop via a heat exchanger. The temperatures of the water loops are much higher than in traditional data centers. As a result, the need for air conditioning (chillers) is considerably reduced.

Dry coolers, often adiabatic (using evaporation), allow this heat to be rejected directly outside the data center. Or, better yet, the heat from the water loops can be used to heat housing, offices, or even a swimming pool. Finally, the absence of a fan on each server also makes them less energy-intensive.

In data centers, the main use case for immersion cooling is a situation of high heat density, i.e., maximum computing power in a minimum of space. This is the case, for example, for solutions dedicated to artificial intelligence (AI) and online gaming, but also for all data centers with HPC computers (research laboratories, aeronautics, finance for certain specific operations, healthcare, etc.).

Some racks can indeed accommodate 50 kW, or even 100 kW or more today. In this context, immersion cooling should make it possible to drastically limit electrical energy consumption: the synthetic oils used would be capable of absorbing 1,500 times more thermal energy than air, at isometric volumes and temperatures.

Furthermore, the first environmental studies (simplified life cycle assessments) carried out at Lab-by-APL using the NegaOctet database on a theoretical case study show that immersion cooling with a silicone-based dielectric liquid would generate between 30% and 50% fewer CO2 emissions than traditional cooling (direct expansion) over the entire life cycle (manufacturing, transport, use, and end-of-life) of the equipment and the liquid.

Existing rack servers are not all compatible with an immersive solution: on the one hand, their format does not necessarily adapt to immersion cooling units (rectangular oil baths) and, on the other hand, the design of traditional servers does not allow for oil circulation.

In this second case, modifications are necessary, such as removing fans or replacing certain materials, such as thermal paste, which must be replaced by another component because it dissolves in any dielectric liquid.

As for hard drives, while SSDs are natively compatible with immersion cooling, traditional rotating hard drives are not, unless they are designed to be airtight or certified by the manufacturer.

Standard fiber optics are generally not compatible with immersion. To be able to use fiber connections, specific covers must be used.

However, several manufacturers are already starting to offer immersion-ready servers in their catalogs.

This mainly concerns data centers that require very high agility, as immersion cooling systems can be less mobile than traditional cooling infrastructures. This may be the case, for example, for infrastructure or cloud providers who need to change 'small' servers quickly and regularly. This is not necessarily possible with immersed servers, due to the necessary handling precautions involved.

In addition, the maintenance of immersion cooling systems appears to be more delicate due to the need to handle the dielectric liquid and maintain the integrity of the sealed systems. It requires the intervention of trained personnel to maintain this type of equipment, and could require more time and materials (drip trays, cloths and gloves, lab coats, service carts, etc.).

Furthermore, while some dielectric liquids are designed to minimize corrosion, a thorough evaluation of materials is always recommended before implementing an immersion cooling solution.

Finally, there are no norms or standards governing immersion cooling yet.

The performance of immersion cooling is very interesting from a theoretical point of view, particularly in high-density data centers. However, this technology involves constraints, particularly in terms of operation and maintenance, which make it more suitable for certain specific use cases. Thus, air cooling remains, to date, an alternative solution that fulfills its role for traditional uses.

In addition, there are few in-depth (techno-economic) studies or feedback on successful and mastered immersion cooling use cases on a significant scale. It is therefore advisable to carry out a thorough assessment of the risks and the means implemented to remedy them before implementing this solution.

Research and innovation continue while standardization initiatives such as that of the OCP (Open Compute Project) have been launched. In the long term, these initiatives will undoubtedly allow the data center sector to master the design, operation, and maintenance of immersion cooling infrastructures from end to end.