Scalding hot AI accelerators have put datacenters in hot water. Two-phase cooling could chill them out

Liquid cooling makes it possible to operate multi-kilowatt accelerators and ultra -dense datacenter racks, but new research commissioned by startup Accelsius suggests that we've been using the wrong liquids for the job. The company found different cooling fluids can shave millions off datacenter operating costs by allowing bit barns to run their water systems up to 14 C higher than existing systems. By swapping single-phase coolants used by existing GPU servers for two-phase fluids, like those used in refrigerators and air conditioning units, Accelsius claims it can support facility water temperatures up to 54 C, or up to 59 C if the operator is willing to double a facility's flow rate. While running at 59 C rather than 45 might not sound like a big difference, Accelsius claims that every 1 C increase in facility water temp corresponds to roughly 4 percent in annual energy savings. Given the datacenter building boom and electricity supply constraints pushing energy costs higher, the savings could be significant. There's not that much you can do with a lot of 30 C waste heat, but as you get to 50-60 C, now you're talking about shower temperature levels of heat," Accelsius CTO Rich Bonner told El Reg in an exclusive interview. Two-phase coolants are nothing new in datacenter cooling. As far back as 2021, Microsoft was dipping entire server chassis into vats of boiling fluoroketones and hydrofluoroethers to cool them. ZutaCore, meanwhile, is using these two-phase coolants for its own direct-to-chip cooling tech. Compared to these designs, Accelsius' approach could easily be confused for a conventional datacenter liquid cooling system. The company's approach is to pump fluids through rack manifoldstowards datacenter hardware. The fluids flow through finned cold plates attached to servers and other kit, then carry heat back to the coolant distribution unit. There, heat captured from the system is exchanged with facility water and eventually pumped out to the atmosphere. The main difference is that Accelsius is using two-phase coolants that are much more efficient at pulling heat off the accelerators. In a whitepaper shared with El Reg, Accelsius reports that by modifying a Dell PowerEdge XE9680L, a liquid-cooled server with eight Nvidia B200 accelerators, to use its NeuCool two-phase coolant and cold plates, the company was able to drop temps as much as 19 C at the cold plate and 9-14 at the system level. What you choose to do with that is really up to you and your end goals," Bonner said. For instance, you can run at higher facility water temperatures, and maybe not run your chillers at all and take that power that you're saving and now put more compute in your datacenter," he explained. Or, maybe, you're trying to overclock or get as much performance out of that chip." The higher water temperature inside a facility, the less heat you need to push out to the atmosphere. Instead of needing evaporative coolers, which consume water, dry coolers - essentially industrial radiators - can be used. Dry coolers typically only work during colder months out of the year, as the temperature gradient between the facility water and air temperature has to be sufficiently high for them to be effective. During warmer months, evaporative coolers and other refrigerant-based coolers may be necessary to maintain operating temperatures. Barriers to adoption Despite the purported benefits of its two-phase coolant tech, Accelsius still faces some hurdles to widespread adoption. While universal cold plates are the subject of ongoing research, Accelsius' B200 test system used custom cold plates and a coolant distribution unit (CDU) specifically designed for two-phase coolants. Cold plate design is a bigger issue than you might think. While swapping the plates on a B200 system is relatively straightforward, Nvidia's NVL72 is a very different beast, packing 72 GPUs, 36 CPUs, 18 switch ASICs, and dozens of NICs, all of which are liquid-cooled. Bonner tells us universal cold plates could help simplify matters by enabling OEMs to build one platform compatible with either single- or two-phase coolants. In fact, because Accelsius' technology does not rely on boiling refrigerants in wells or tanks, its cold plates don't look that different from the cold plates used in existing systems. Our technology uses skived fins and we rely on flow boiling to remove heat. We're pumping it through small channels... it's exactly what single-phase is doing," he said. Two-phase coolants, however, still rely on compatible CDUs large enough to support modern AI systems. CDUs are the heart of any liquid-cooled system and are essentially a pump and heat exchanger that sits between the system loop and facility water system. While it may be possible to build servers that can work with either single or two-phase coolants, the latter requires the use of specialized CDUs. Nvidia rack systems feature up to 72 GPUs with system TDPs approaching 250 kW this year and pushing 600 kW beginning in 2027. At the time of writing, Accelsius' largest in-row CDU can only support one 250 kW rack. Bonner tells us that the company has a 350-plus kW unit in the works, and is working on a 2.4 megawatt unit in the lab that he expects to be available sometime early next year. Then there are the coolants themselves. Early two-phase systems often relied on coolants originally developed for applications outside the datacenter. One of the most commonly used is 3M Novec, a two-phase coolant based on per- and polyfluoroalkyl substances (PFAS), a group of synthetics linked to major health complications. In late 2022, 3M announced it would halt production of the chemicals, putting immersion and liquid cooling suppliers dependent on the stuff in a bind. Since then, researchers have developed less harmful fluids. Bonner tells us the blend of refrigerants used in its NeuCool formula meets ASHRAE A1 safety requirements, which means it's non-flammable and non-toxic. Setting aside the infrastructure requirements, one final challenge that may not be so easily overcome is OEM buy-in, which has long plagued liquid-cooling vendors. Retrofitting servers for liquid cooling has never been the issue; it's getting them serviced that's the problem. If something fails in the system, who is to blame? Unless the OEM, whether it be Dell, HPE, Lenovo, Supermicro or someone else, has given their blessing, there's a good chance they'll point the finger at the liquid-cooling supplier. Because of this, many liquid-cooling providers have had to offer pass-through support contracts to convince prospective enterprise buyers they won't end up with a $250,000 paperweight if something goes wrong. Our end goal is to have a two-phase system you can get from a Dell or a Supermicro," Bonner said. (R)