Why Two-Phase, Direct-to-Chip Liquid Cooling is Not the Same Thing as Immersion Liquid Cooling!

         With the massive build-out of AI factories and HPC data centers, liquid cooling has suddenly become one of the most critical requirements of the decade as hyperscalers scramble to control their heat and footprint while still contributing to global sustainability goals. There is no question that this technology is being adopted at an unprecedented rate, but there are still a lot of questions in the market over what the differences are between two-phase direct-to-chip liquid cooling and immersion liquid cooling and they are not the same!

At a high level, both types of cooling use dielectric liquid. However, with Immersion liquid cooling, all the servers and other components are completely submerged in dielectric liquid in big heavy tanks. In contrast, the two-phase direct-to-chip process brings only a small amount of dielectric liquid to a cold plate that is placed directly on top of the high heat flux source such as CPUs and GPUs.

As an example, the below picture shows ZutaCore cold plates sitting on NVIDIA H100 GPUs. As you can see, the dielectric liquid never touches the equipment, since it flows in a closed loop system. If this were showing immersion cooling, the entire rack would be completely submerged in the liquid. To put in perspective how much fluid each solution uses….a 100kW rack using two phase direct-to-chip technology uses less than 4 gallons of dielectric fluid, compared to immersion cooling that needs over 100 gallons per rack.

In contrast to two-phase direct-to-chip cooling, immersion cooling requires significant data center infrastructure investment because large and heavy tanks filled with liquid are now needed to hold the equipment. There are several disadvantages of this approach:

• All the equipment needs to be compatible to the dielectric liquid so that it is not damaged by the fluid itself. This often requires the selection of specialized equipment or a modification to the servers.
• Requires significant data center infrastructure and layout changes.
• As some of the components on the server such as fiber optics connectors cannot function when immersed, the servers require to be mechanically reconfigured.  
• Maintenance of the servers is difficult where any required server maintenance requires ‘pulling’ the single server out of the tank using a crane and letting it drip for 30 minutes before starting to service.
• Each time the tank is opened for service, vapors that contain PFAS are released into the environment, resulting in a 10 percent loss of liquid (100s of liters) per year and the release of large amount of PFAS vapors into the atmosphere.
• The type of fluid used in single-phase immersion cooling features hydrocarbons which is flammable and combustible. This could cause catastrophic damage if a fire were to occur inside the data center.
• With single-phase immersion cooling, the thermal design power (TDP) is limited. If a GPU has a TDP over 700 watts, the single-phase immersion method is unable to effectively cool the hardware.

The Two-Phase, Direct to Chip Liquid Cooling Advantage

In contrast to immersion cooling, the two-phase direct-to-chip approach such as ZutaCoreÒ HyperCoolÒ uses a highly efficient, two-phase boiling and condensation process moving large amounts of heat off the processors and away from servers. This technology is scalable and can be deployed in new or retrofitted data centers, with an ability to cool 2,800 watts and beyond. Other advantages include:

• The dielectric fluid never needs to be replaced and does not get released into the atmosphere during server and rack maintenance.
• The liquid used has been engineered with an ozone depletion potential (ODP) of 0 and very low global warming potential (GWP).
• Because the liquid maintains a constant temperature, the heat from the servers can be harvested for heat re-use applications such as heating adjacent offices, other parts of the data center, or even schools and office buildings in proximity.
• Maintains 1U server form factor even as the heat increases for next-gen GPUs, allowing increased compute density of up to 150kW per rack.