CoolIT Systems’ 4000W Breakthrough: Redefining single-phase DLC for ultra-high wattage AI processors

The world of artificial intelligence seems to be getting more exciting every day. As applications abound, GPU companies are working hard to produce more powerful processors and fulfill demand.

While chips like the NVIDIA Blackwell line and AMD’s MI355X enable users to unlock incredible computing power, they come at the cost of exponentially increasing power consumption.  Data center operators must address major thermal challenges with next-generation GPUs exceeding 1400W and rack densities growing to 600kW per rack.  To address this, CoolIT Systems has developed the groundbreaking 4000W coldplate – a single-phase direct liquid cooling (DLC) solution designed to cool these ultra-high-wattage AI processors efficiently.

CoolIT’s 4000W coldplate attaches directly to high-performance processors like GPUs or CPUs and uses a liquid coolant to absorb heat efficiently. This innovative approach is crucial as traditional air cooling becomes increasingly inadequate for AI deployments, making DLC solutions the primary option for large-scale AI tasks.

CoolIT’s 4000W coldplate and TTV

Solving AI Thermal Challenges

Single-phase DLC solutions like CoolIT’s 4000W coldplates work by removing heat from semiconductors with either water or a water-glycol mixture running through coldplates attached to the chips.  Until now, the industry assumed that the upper limits for single-phase coldplates were between 1,500W to 2,000W and managing thermals above 2000W would require more complex two-phase DLC using refrigerants or immersion cooling.

As the name suggests, this system can cool thermal loads in excess of 4000W. In fact, in its tests, CoolIT’s 4000W coldplate captured over 97% of heat from a 4000W thermal test vehicle (TTV) with a flow rate of 6 LPM. This translates to 1.5 liters per minute per kW, which is the semiconductor industry’s recommended flow rate for high-wattage chips.

CoolIT’s 4000W coldplate also incorporates two key technical breakthroughs that can assist in AI deployments: CoolIT’s patented Split-Flow™ technology and the OMNI™ All-Metal Coldplate design.

Split-Flow technology can be found in all CoolIT Systems’ coldplates. This patented coldplate architecture allows the coolant to enter the coldplate, where it is then distributed throughout arrays of microchannels to target hotspots while simultaneously providing uniform cooling.  By reducing the coolant’s distance across the microchannels by half, this design enhances thermal performance and lowers flow impedance.

The Split-Flow system is proven to provide 30% better thermal and flow performance compared to standard coldplates. With a lower pressure drop, the technology also reduces the energy needed to run the coolant distribution unit. What’s more, this system can handle thermal loads exceeding 4000W thermal design power (TDP) and thermal fluxes over 300 W/cm².

The OMNI All-Metal Coldplate design features a single-metal, unibody construction using aerospace-grade materials. This design eliminates the need for brazing materials, eliminating differential coefficients of thermal expansion, making it highly reliable, especially in fluctuating temperatures.

Looking to the Future

When designing a computing environment, future-proofing is an important aspect to consider. While data centers and other large computing environments may not swap out technology as often as consumers change their phones, having the latest and greatest chips will always be a priority for those who can afford to upgrade.

Right now, the following chips are what AI operators are focused on:

• NVIDIA GB300 has a TDP of 1400W
• AMD MI355X has a TDP of 1100W
• NVIDIA’s Blackwell B200 tensor core chip consumes 1200W
• NVIDIA’s GB200 (combining two B200 chips and a Grace CPU) has a total power draw up to 2,700W

However, these are just the current requirements. The 4000W coldplate is designed for processors exceeding 2000W, anticipating future AI chip developments. Thus, installing these coldplates will allow computing decision-makers to ensure that investments in DLC systems today will mean they are prepared whatever thermal challenges the future may have in store for them.

Outside of the actual cooling requirements, something else to consider here is how much energy AI is currently consuming and will consume in the near future. According the U.S. Department of Energy, the energy consumed by AI data centers could triple by 2028. Considering that up to 40% of data center power use comes from cooling high-power chips, ensuring for efficient liquid cooling can significantly reduce energy consumption and operating expenses.

“With the 4000W coldplate, we’re addressing the thermal challenges faced by data centers as they deploy more powerful AI systems,” said Reza Najjari, Thermal Mechanical Engineering Manager at CoolIT Systems. “Our goal is to provide reliable and efficient cooling that supports the scalability of AI infrastructure for many years to come.”

The CoolIT 4000W coldplate is setting a new standard for single-phase DLC in the age of AI. To explore the technology’s capabilities in greater detail, including specific use cases and performance benchmarks, download the official CoolIT technical brief. Discover how this breakthrough can transform your data center cooling strategy by reading the brief.

Article originally posted on HPCwire.