First Direct-to-Chip liquid cooling cold plates for Nvidia Grace Blackwell Superchip
Zutacore has developed a single monolithic dielectric cold plate capable of cooling 2.8kW for the next generation Grace Blackwell superchip.
The direct to chip liquid cooling system developed by ZutaCore will support the GB200 superchip, which combines the Nvidia Grace ARM processors and next generation Blackwell GPU.
The dielectric cold plates can cool up to 120kW of rack power with little or no modifications to current real estate, power, or cooling systems.
With over 400 billion transistors, the Grace Blackwell superchip features two Blackwell GPUs, a Grace CPU, and two I/O chips. A single ZutaCore HyperCool dielectric cold plate will sit directly on top of this superchip, taking up very little space while providing significant cooling capabilities.
Nvidia pushes Blackwell GPU as start of a whole new industry
In a typical GB200 2RU with two GB200 modules, a single, monolithic ZutaCore cold plate can deliver 2.8kW of cooling with a low pressure, closed loop system. The monolithic cold plate ensures a simple design with only one input and one output for the dielectric coolant, which reduces the potential points of failure to deliver a higher level of reliability. The module is also equipped with OCP-compliant blind mate connectors that support 5.600kW for the two modules.
Direct-to-chip system cooling save energy in AI data centers
“ZutaCore’s groundbreaking technology enables us to cool the NVIDIA GB200 Superchip with a single cold plate, delivering an industry-first waterless direct-to-chip liquid cooling solution,” said Erez Freibach, Co-founder and CEO at ZutaCore. “This breakthrough frees AI data center owners and operators from the risks associated with water-based cooling, empowering them to deploy the processing power needed for compute-intensive workloads like generative AI without the fear of leakages.”
“The next generation of high-performance GPUs, such as NVIDIA’s GB200, that are required to meet the demanding needs of AI model training will generate a significant amount of heat, making efficient cooling solutions crucial for optimal system performance and sustainable operation,” said Mark Nossokoff, Research Director for Hyperion Research. “Innovative cooling solutions such as ZutaCore’s HyperCool waterless, direct-to-chip liquid cooling approach that eliminates the need for water in the cooling process, aim to optimize performance, maximize energy efficiency, and reduce operation costs associated with running the most advanced AI computing systems.”
In addition to HyperCooling the GB200 Grace Blackwell Superchip, ZutaCore previously showed support for the Nvidia H100 and H200 Hopper GPUs with the col plate technology. It has signed a strategic OEM deal with UNICOM Engineering to deliver warrantied HyperCooled AI servers for the H100 and H200 GPUs.
The first platform will be the Dell Technologies XE9680, supporting 120kW per rack of computing power, while also increasing rack processing density by 300 percent.
“This partnership is a significant milestone for driving ZutaCore’s waterless, direct-to-chip liquid cooling into the volume market globally,” said Erez Freibach, Co-founder and CEO at ZutaCore. “The future of AI hinges on the ability to deliver the performance of the next-generation GPUs, while also adhering to strong sustainable practices within the data centre. This is where ZutaCore sets itself apart from any other cooling solution with its ability to cool the hottest processors with no water used and little to no modifications to current real estate, power, or cooling systems.”
www.zutacore.com; www.unicomengineering.com
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
