CoolIT, a company specializing in liquid cooling solutions for data centers, has debuted a new liquid cooling cold plate, advertised to dissipate 4kW of heat, which is four times what Nvidia’s B200 accelerator is rated at. The company claims a doubling of heat dissipation over current industry standards regarding single-phase Direct Liquid Cooling (DLC) solutions. This announcement falls right before GTC, where Nvidia is expected to introduce its GB300 Blackwell Ultra offerings and discuss plans for the next-generation Rubin architecture. The timing may be more than a coincidence.
As the name suggests, Direct Liquid Cooling brings the liquid coolant in direct contact with the heat-generating components, such as a CPU or GPU. The single-phase part refers to the state of the coolant; in single-phase, the coolant remains a liquid while in two-phase it changes state by boiling / condensing. In general, this is quite similar to the AIOs (All-In-Ones) and custom liquid cooling loops your PC might have. Both rely on the same principles and fall under the umbrella of the direct-to-chip (D2C) cooling approach.
A cold plate absorbs heat from the processor and transfers it to the coolant in the tubes, which CoolIT mentions is either water or a mixture of water-glycol. In standard testing, at a flow rate of six Liters Per Minute (LPM), CoolIT’s new cold plates reportedly achieved 97% heat removal from a 4,000W Thermal Test Vehicle (TTV) which indicates excellent heat transfer efficiency. The company also claimed a thermal resistance of under 0.009 Celsius/Watt while incurring a full flow loop pressure drop to 8 PSI (Pounds per Square Inch). For a quick understanding of these numbers, a lower thermal resistance value for a given material means it conducts heat more effectively. Similarly, in a closed-loop, a lower pressure drop means the pump doesn’t have to work as hard to circulate the coolant.
Latest Videos FromTom's Hardware
Server-grade CPUs such as AMD’s EPYC and Intel’s Xeon can chug up-to 500W of power. While liquid-cooling is deployed with certain high-performance setups, this is a number well within reach of air-cooling solutions. AI accelerators on the contrary, like AMD’s Instinct MI325X and Nvidia’s Blackwell family, are designed with TDPs over 1,000W. Training and running AI is both computationally taxing and power-intensive.
With the manifestation of a practical arms-race between tech-giants across the globe, chipmakers have been pushing the boundaries of technology for higher performance. Sadly, one of those aspects is increased power consumption, and that gets progressively worse as more of these racks are deployed. Consequently, we are witnessing a growing trend towards liquid cooling. Immersion cooling seems like a viable solution, but it limits vertical rack stacking and not every data center is designed with such infrastructure in mind, so for the time being, we’re likely to see D2C solutions increase in adoption.
