Data Center Liquid Cooling: The AI Heat Solution

Stroll right into a typical knowledge middle and one of many first issues that jumps out at you is the noise—the low, buzzing sound of hundreds of followers: followers subsequent to particular person computer chips, followers on the again panels of server racks, followers on the community switches. All of these followers are pushing scorching air away from the temperature-sensitive pc chips and towards air-conditioning models.

However these followers, whirr as they could, are now not reducing it. Over the previous decade, the power density of essentially the most superior pc chips has exploded. In 2017, Nvidia got here out with the V100 GPU, which attracts 300 watts of energy. Most of that energy dissipates again out as warmth. Three years later, in 2020, Nvidia’s A100 got here out, drawing as much as 400 W. The now-popular H100 arrived in 2022 and consumes as much as 700 W. The latest Blackwell GPUs, revealed in 2024, devour as much as 1,200 W.

“Highway maps are taking a look at over 2,000 watts [per chip] over the following 12 months or two,” says Drew Matter, president and CEO of the liquid-cooling firm Mikros Technologies. “In truth, the business is getting ready for 5-kilowatt chips and above within the foreseeable future.”

This energy explosion is pushed by the apparent wrongdoer—AI. And all the additional computations consuming all that added energy from superior chips are producing unmanageable quantities of warmth.

“The typical energy density in a rack was round 8 kW,” says Josh Claman, CEO of the startup Accelsius. “For AI, that’s rising to 100 kW per rack. That’s an order of magnitude. It’s actually AI adoption that’s creating this actual urgency” to determine a greater solution to cool data centers.

Particularly, the urgency is to maneuver away from followers and towards some type of liquid cooling. For instance, water has roughly 4 occasions the particular warmth of air and is about 800 occasions as dense, that means it may soak up round 3,200 occasions as a lot warmth as a comparable quantity of air can. What’s extra, the thermal conductivity of water is 23.5 occasions as excessive as that of air, that means that warmth transfers to water way more readily.

“You possibly can stick your hand right into a scorching oven and also you received’t get burned. You stick your hand right into a pot of boiling water and you’ll immediately get third-degree burns,” says Seamus Egan, normal supervisor of immersion cooling at Airedale by Modine. “That’s as a result of the liquid transfers warmth a lot, a lot, a lot, way more rapidly.”

The information-center business by and huge agrees that cooling chips with liquid is the longer term, at the very least for AI-focused knowledge facilities. “As AI has made racks denser and warmer, liquid cooling has turn out to be the de facto answer,” Karin Overstreet, president of Nortek Data Center Cooling, stated through e mail.

However there are a selection of how to do liquid cooling, from the easy and easy to the advanced and barely bizarre.

On the easy finish, there’s circulating chilled water by means of chilly plates connected to the most well liked chips. Then there’s circulating not water however a particular dielectric fluid that boils contained in the chilly plate to remove the warmth. A 3rd strategy is dunking your complete server right into a fluid that retains it cool. And, final and most splashy, is dunking the server right into a boiling vat of liquid.

Which technique will find yourself being the business customary for the high-end AI factories of the longer term? At this level, it’s anybody’s guess. Right here’s how the 4 strategies work, and the place they could discover essentially the most use.

#1: Single-Section Direct-to-Chip Cooling

Probably the most technologically mature strategy is to make use of water. Already, many AI knowledge facilities are using such direct-to-chip liquid cooling for his or her hottest chips.

On this scheme, metallic blocks, referred to as chilly plates, with channels in them for coolant to flow into, are positioned immediately on prime of the chips. The chilly plates match the scale of the chips and go contained in the server. The liquid is normally water, with some glycol added to forestall bacterial development, stabilize the temperature, defend in opposition to freezing and corrosion, and improve the viscosity of the liquid. The glycol-water combination is compelled by means of the chilly plate, whisking away warmth proper from the supply.

Firms like Mikros Applied sciences are pursuing single-phase direct-to-chip liquid cooling. On this method, a chilly plate is positioned on prime of the most well liked chips. Liquid is circulated by means of the chilly plate, whisking away warmth. Marvell Technology

The glycol water is often saved in a closed loop, circulating from the chilly plates to a heat-exchange unit, which cools the liquid again down, after which again to the chilly plate. Contained in the heat exchanger, a separate loop of “facility water” is used to chill down the glycol water. The power water is in flip cooled by both a chiller—an electrically powered refrigeration unit—or a dry cooler, an outside unit that makes use of followers to blow ambient air over the water because it strikes by means of pipes. A dry cooler is far less complicated and extra vitality environment friendly than a chiller, however it works solely in cooler climates—it may’t cool the water beneath the ambient temperature.

One issue with this strategy is that placing a chilly plate on each single heat-producing part in a server is unfeasible. It solely is sensible to place chilly plates on essentially the most energy-dense elements—specifically GPUs and a few CPUs—leaving smaller elements, like power supplies and reminiscence models, to be cooled the old school means, with followers.

“The development is transferring towards a hybrid-cooling answer,” Overstreet says. “So liquid cooling does about 80 % of the cooling for the server room or the information corridor, and about 20 % is the present air-cooling answer.”

#2: Two-Section Direct-to-Chip Cooling

With GPU energy densities displaying no indicators of leveling off, direct-to-chip water cooling is hitting a restrict. You possibly can, in fact, improve the circulation of water, however that may use extra vitality. Or you possibly can function the chips at the next temperature, which can minimize into their efficiency and in the long term degrade the chips. Thankfully, there’s a 3rd choice: to squeeze a bit extra out of the physics of warmth alternate.

The additional cooling energy provided by physics comes from latent warmth—that’s, the vitality it takes to alter part, on this case from liquid to gasoline. Because the liquid boils off the GPU, it absorbs that further latent warmth because it turns into gasoline, with out rising temperature.

Firms like Accelsius are proposing two-phase direct-to-chip liquid cooling. Right here, a chilly plate can also be positioned on prime of the most well liked chips, and the liquid circulating by means of the chilly plate boils immediately atop the chip. Huge Concept Productions

That’s mainly how two-phase direct-to-chip cooling works. On this scheme, a specifically formulated dielectric liquid circulates by means of chilly plates sitting atop high-energy chips and boils into vapor. The vapor is then fed again to a warmth exchanger, which cools the fluid utilizing facility water.

“It’s actually boiling to chill,” says My Truong, chief expertise officer of the startup ZutaCore, which makes two-phase direct-to-chip cooling techniques.

Water boils at 100 °C (at atmospheric stress), which is just too excessive for correct chip operation. So that you want a specifically formulated fluid with a decrease boiling level. ZutaCore’s chief evangelist, Shahar Belkin, explains that the fluid they use is sourced from chemical suppliers like Honeywell and Chemours, and boils at a temperature as little as 18 °C, which might be adjusted up or down by tweaking the stress within the loop. As well as, the fluid is dielectric—it’s not electrically charged until polarized by an exterior electric field. So, not like water, if a few of the fluid spills onto the electronics, it received’t harm the pricey gear.

With water, the temperature will increase drastically because it flows over the hot chips. Meaning the incoming water must be saved chilly, and so the ability water requires cooling with chillers in most climates.

With boiling dielectric fluid, nonetheless, the fluid stays roughly the identical temperature and easily modifications part right into a vapor. Meaning each the liquid and the ability water might be saved at the next temperature, leading to important vitality financial savings.

When liquid boils on prime of a scorching chip, the chip is cooled not solely by means of contact with the cooler liquid, but in addition by means of the latent warmth it takes to induce a phase change. Accelsius

“Due to the actually environment friendly boiling course of that occurs on the chilly plate, we will settle for facility water that’s 6 to eight levels hotter than [with] single part,” says Lucas Beran, director of product advertising and marketing at Accelsius, one other startup engaged on two-phase direct-to-chip liquid cooling.

The 2-phase setup additionally requires decrease liquid circulation charges than the normal single-phase water strategy, so it makes use of much less vitality and runs much less threat of damaging the gear. The circulation price of two-phase cooling is about one-fifth that of single-phase cooling, Belkin says.

With single-phase water cooling, he says, “you’ll should circulation a gallon per minute into the chilly plate” for essentially the most superior chips working at 2,000 W. “This implies very, very excessive stress, very, very excessive circulation. It signifies that pumping might be costly, and [the cooling system] will really hurt itself with the excessive circulation.”

#3: Single-Section Immersion Cooling

Direct-to-chip liquid cooling gives way more cooling capability than simply blowing air, however it nonetheless depends on chilly plates as intermediaries to do the cooling.

What if you happen to may bypass the chilly plate altogether and simply dunk your complete pc server in coolant? Some corporations are doing simply that.

On this strategy, the information middle is organized round immersion tanks somewhat than racks, every tank roughly the scale of a fridge. The immersion tanks are full of a dielectric fluid, normally an oil, which should be nonconductive and have sturdy thermal switch properties, says Rachel Bielstein, international gross sales supervisor of immersion cooling at Baltimore Aircoil Co. The fluid additionally requires long-term stability and low environmental and fireplace threat.

Sustainable Steel Cloud is advocating for single-phase immersion cooling, during which a complete server is submerged in a vat of liquid to maintain it cool.Firmus Technologies

With immersion cooling, every little thing will get cooled by the identical fluid. After the oil has whisked away the warmth, there are numerous approaches to cooling the immersion fluid. Baltimore Aircoil, for one, has designed a warmth exchanger that circulates facility water by means of coils and plates contained in the tank, Bielstein explains. “The heated water is then pumped to an out of doors cooler that releases the warmth into the air, cools the water, and sends it again to the warmth exchanger to soak up extra warmth from the tank. This course of makes use of as much as 51 % much less vitality versus conventional designs.”

The crew at Singapore-based Sustainable Metal Cloud (SMC), which builds immersion-cooling techniques for knowledge facilities, has discovered the modifications that have to be made to servers to make them appropriate with this cooling technique. Past eradicating the built-in followers, the corporate swaps out the thermal-interface supplies that join chips to their heat sinks, as a few of these supplies degrade within the oil. Oliver Curtis, co-CEO of SMC and its sister firm Firmus, instructed IEEE Spectrum the modifications they make are small however vital to the functioning of SMC’s setup.

“We’ve created the right working atmosphere for a pc,” Curtis says. “There’s no mud, no motion, no vibration, as a result of there’s no followers. And it’s an ideal working temperature.”

There are some chips whose energy density remains to be too excessive to be fully cooled by the slow-moving oil. In these instances, it’s obligatory so as to add chilly plates to extend the oil circulation over them. “Single-phase immersion has already hit the bounds” for cooling these superior chips, says Egan of Airedale by Modine. Including chilly plates to immersion cooling, he says, “will certainly present assist for extra superior chip architectures and cut back the warmth load on the single-phase dielectric fluid. The brand new problem is that I now want two separate cooling-loop techniques.”

#4: Two-Section Immersion Cooling

If nobody cooling technique is sufficient by itself, how about placing all of them collectively, and dunking your knowledge middle right into a vat of boiling oil?

Some corporations already are.

“Two-phase immersion might be essentially the most moon-shot expertise in terms of data-center liquid cooling,” says Beran, of Accelsius.

However Brandon Marshall, international advertising and marketing supervisor of data-center liquid cooling at Chemours, says that is the place the business is headed. “We consider from the analysis that we’ve finished that two-phase immersion goes to return up in a fairly cheap means.”

At their lab in Newark, Del., the Chemours crew is creating a specifically formulated liquid for two-phase immersion cooling. On this strategy, the server is dunked right into a vat of liquid, and the liquid boils atop the new elements, cooling the system. Chemours

Marshall argues {that a} two-phase—also referred to as boiling—liquid has 10 to 100 occasions as a lot cooling capability as a single-phase liquid, as a consequence of its latent warmth. And whereas two-phase direct-to-chip cooling may fit for the chips of at this time, it nonetheless leaves many elements, equivalent to reminiscence modules and energy provides, to be air cooled. As CPUs and GPUs develop extra highly effective, these reminiscence modules and energy provides will even require liquid cooling.

“That record of issues isn’t going anyplace,” Marshall says. “I believe the immersion-cooling piece goes to proceed to develop in curiosity as we transfer ahead. Individuals are going to get extra snug with having a two-phase fluid within a rack identical to they’ve [with] placing water in a rack by means of single-phase direct-to-chip expertise.”

Of their lab in Newark, Del., the Chemours crew has positioned a number of high-power servers in tanks full of a proprietary, specifically formulated fluid. The fluid is dielectric, in order to not trigger shorts, and it’s additionally noncorrosive and designed to boil on the exact temperature at which the chips are to be held. The fluid boils immediately on prime of the new chips. Then the vapor condenses on a cooled floor, both on the prime or the again panel of the tank.

Of their lab in Newark, Dela., the Chemours crew is testing their two-phase immersion cooling fluid. On this strategy, the entire server is dunked right into a tank with dielectric liquid. The warmth from the server boils the liquid, leading to cooling. Chemours

That condenser is cooled with circulating facility water. “All we want is water despatched on to the tank that’s about 6 levels decrease than our boiling level, so about 43 °C,” Marshall says. “The fluid condenses [back to a liquid] proper within the tank. The temperature required to condense our fluid can remove the necessity for chillers and different advanced mechanical infrastructure normally.”

In line with a latest case study by Chemours researchers, two-phase immersion cooling is more economical than single-phase immersion or single-phase direct-to-chip in most climates. For instance, in Ashburn, Va., the 10-year whole price of possession was estimated at US $436 million for a single-phase direct-to-chip setup, $491 million for a single-phase immersion setup, and $433 million for a two-phase immersion-cooling setup, principally as a consequence of decrease energy necessities and a simplified mechanical system.

Critics argue that two-phase immersion makes it laborious to keep up the gear, particularly for the reason that oils are so specialised, costly, and susceptible to evaporating. “Once you’re in an immersion tank, and there’s greenback indicators evaporating from it, that may make it a little bit of a problem to service,” Beran says.

Nonetheless, Egan of Airedale by Modine says his firm has developed a solution to principally keep away from this situation with its immersion tanks, that are meant for edge purposes. “Our EdgeBox is particularly designed to keep up the vapor layer decrease down within the tank with a layer of air above it and nearer to the tank lid. When the tank is opened (for a brief upkeep interval), the vapor layer doesn’t ‘circulation out’ of the tank,” Egan wrote through e mail. “The vapor is far heavier than air and subsequently stays decrease within the tank. The minimal vapor loss is offset by a buffer tank of fluid inside the system.”

For the foreseeable future, individuals within the business agree that the facility calls for of AI will maintain going up, and the necessity for cooling together with them.

“Until the ground falls out from underneath AI and all people stops constructing these AI clusters, and stops constructing the {hardware} to carry out coaching for large language models, we’re going to want to maintain advancing cooling, and we’re going to want to unravel the warmth drawback,” Marshall says.

Which cooling expertise will dominate within the coming AI factories? It’s too quickly to say. However the quickly altering nature of knowledge facilities is opening up the sphere to loads of inventiveness and innovation.

“There’s not solely an ideal marketplace for liquid cooling,” says Drew Matter, of Mikros Applied sciences, “however it’s additionally a enjoyable engineering drawback.”