Hewlett Packard Enterprise (HPE) and Supermicro have separately announced new liquid cooling products to keep up with the heat generated by large-scale AI deployments.
In the case of HPE, it showed off a design that offers direct liquid cooling without an expensive component: the fans.
Liquid cooling is done through one of three ways: chillers, or coolants that cool the pipes as the hot liquid travels through it; fans that blow on the pipe but don’t produce cold air, which is lower power but also doesn’t cool the water as much as chillers; and simple heat dissipation that allows liquid to cool naturally as it travels along the length of the pipe.
The second example is fairly common. Hot liquid traveling through the pipes away from the CPU can be adequately cooled through a combination of natural dissipation of heat and keeping the pipes cool through fans. That’s how a car radiator works, after all. But that still means an awful lot of fans that need to be powered and take up space.
HPE architecture’s doesn’t rely on using fans. Instead, it uses an 8-element cooling design that covers various server components such as the server blade, network fabric, and coolant distribution unit (CDU).
It’s not a new design – HPE has been using this technology for some time in its Cray EX supercomputers. Now it’s expanding the technology to offer it to more non-supercomputing systems. HPE claims that this approach effectively reduces the required data center floor space by 50% and reduces the cooling power necessary per server blade by 37%.
“As organizations embrace the possibilities created by generative AI, they also must advance sustainability goals, combat escalating power requirements, and lower operational costs. The architecture we unveiled today uses only liquid cooling, yielding greater energy and cost efficiency advantages than the alternative hybrid cooling solutions on the market,” said Antonio Neri, president and CEO of HPE, in a statement.
Supermicro announces liquid cooling offering
Server maker Supermicro called its latest liquid cooling product “a complete liquid cooling solution.” The offering includes CDUs, cold plates, coolant distribution manifolds (CDMs), cooling towers, and end-to-end management software.
The liquid cooling solution includes cold plates that dissipate up to 1600W, enough to handle even the hottest Nvidia chips. Supermicro’s liquid-cooled SuperClusters are specially designed for Nvidia‘s GB200 Grace Blackwell Superchip NVL72 platform, which consists of one CPU and two Blackwell GPUs on a single die. The clusters feature the new advanced in-rack or in-row coolant distribution units and custom cold plates designed for the compute tray housing two GB200 chips in a 1U form factor.
The rack solution incorporates up to 72 Blackwell GPUs and 36 Grace CPUs, interconnected by Nvidia’s NVLink network, which can handle up to 130 terabytes per second (TB/s) of total GPU communication with extremely low latency and enhanced performance.
The new cooling solutions have started sampling to select customers for full-scale production late in the fourth quarter.
Read more about liquid cooling
- Vertiv and Nvidia define liquid cooling reference architecture: Jointly designed architecture is intended for GPU-loaded AI factories, which will generate tremendous amounts of heat.
- Data center liquid cooling market heats up: Liquid cooling technologies for data centers are transitioning from niche options deployed in specific market segments to mainstream applicability, according to research firm Dell’Oro Group.
- Pros and cons of air, liquid and geothermal systems: Whether it’s to save money, reduce carbon emissions, comply with regulations or accommodate high-powered AI workloads, enterprises are looking to operate more energy-efficient data centers.
- Schneider Electric shares liquid cooling guidelines: As high-density data centers continue to add AI workloads, there’s growing interest in liquid cooling, thanks to its ability to transfer heat more efficiently than air cooling.
- Data centers warm up to liquid cooling: AI, machine learning, and high-performance computing are creating cooling challenges for data center owners and operators. As rack densities increase and temperatures rise, more data centers are finding ways to add liquid cooling to their facilities.
- Accelsius offers liquid cooling without a data center retrofit: Accelsius, a relative newcomer in the liquid cooling market, has launched its NeuCool dual-phase, direct-to-chip liquid cooling technology, which is designed to be deployed without having to do a massive retrofit of your data center.
- ZutaCore launches liquid cooling for advanced Nvidia chips: The HyperCool direct-to-chip system from ZutaCore is designed to cool up to 120kW of rack power without requiring a facilities modification.
- Is immersion cooling ready for mainstream? Liquid cooling started as a fringe technology but is becoming more common. Proponents hope the same holds true for immersion cooling.
- Supermicro has a new liquid-cooled server for AI: Supermicro’s new server uses a self-contained liquid cooling system, and vendors outside of the IT infrastructure market are showing interest in the technology.
Source:: Network World