NVIDIA Unveils Liquid-Cooled GPUs for Sustainable, Efficient Computing

The liquid-cooled NVIDIA A100 PCIe GPU is the first of its kind in a mainstream server GPU to meet customer demands for high-performance green data centers.

To curb climate change, global companies are accelerating the construction of high-performance, energy-efficient data centers, and Zac Smith, head of edge infrastructure at Equinix, is also involved.

He works for Equinix, a global service provider with more than 240 data centers under management, which is committed to leading the industry in achieving climate neutrality.

“Ten thousand customers rely on Equinix to achieve their climate neutrality goals. They need more data and more intelligence, often through AI, and they want to do it in a sustainable way,” said Smith, who got involved with technology as a graduate student at the Juilliard School in the early 2000s, building websites for fellow musicians in New York.

Energy efficiency is steadily improving

As of April, Equinix has issued $4.9 billion in green bonds. Equinix will use these investment vehicles to optimize power usage effectiveness (PUE) in an effort to reduce its environmental impact. PUE is an industry metric that measures how much of the energy used by a data center is directly used for computing tasks.

Data center operators are trying to get PUE down to a level close to the ideal of 1.0. Equinix facilities currently have an average PUE of 1.48, and its new data centers can have PUEs as low as less than 1.2.

Equinix is ​​steadily improving the energy efficiency of its data centers, as measured by PUE (see illustration).

Equinix took this a step further in January when it opened a new facility dedicated to improving energy efficiency. Part of that effort focuses on liquid cooling.

Liquid cooling technology was born in the mainframe era and has matured in the AI ​​era. Today, liquid cooling technology has been widely used in high-speed supercomputers around the world in the form of direct-to-chip cooling.

NVIDIA GPUs are already 20 times more energy efficient than CPUs for AI reasoning and high-performance computing, and it’s only natural that accelerated computing will use liquid cooling technology.

Improving efficiency through acceleration

If all of the world’s CPU servers running AI and HPC were switched to GPU-accelerated systems, up to 11 terawatt-hours of energy could be saved annually. This amount of energy saved could power more than 1.5 million homes for a year.

Today, NVIDIA released the first data center PCIe GPU that uses direct-to-chip cooling technology , contributing to sustainable development.

Equinix is ​​currently qualifying the A100 80GB PCIe liquid-cooled GPU in its data centers as part of a comprehensive approach to sustainable cooling and heat capture. The GPU is currently in the pilot phase and is expected to be officially released this summer.

Save water and electricity

“This is the first liquid-cooled GPU we’ve ever brought to our labs, and we’re excited because customers are eager to leverage AI in a sustainable way,” said Smith.

Data center operators aim to eliminate water chillers used to cool the air inside data centers, which evaporate millions of gallons of water each year. With liquid cooling, the system only circulates a small amount of liquid in a closed system and can focus on the main hot spots.

“We’re turning waste into treasure,” Smith said.

Same performance, less power consumption

In separate tests, Equinix and NVIDIA both found that liquid-cooled data center workloads could match those of air-cooled facilities, while consuming about 30 percent less energy. NVIDIA estimates that liquid-cooled data centers could achieve a PUE of 1.15, well below the 1.6 PUE for air cooling.

Liquid-cooled data centers can achieve twice the amount of computing in the same space because the A100 GPU uses only one PCIe slot, while air-cooled A100 GPUs use two PCIe slots.

NVIDIA saves power and increases density with liquid cooling

At least a dozen system makers plan to use liquid-cooled GPUs in their products later this year, including ASUS, ASRock Rack, Foxconn Industrial Internet, GIGABYTE, H3C, Inspur, Inventec, Nettrix, QCT, Supermicro, Wiwynn, and xFusion.

Global Trends

Regulations setting energy efficiency standards are still undecided in Asia, Europe and the United States, which is driving banks and other large data center operators to evaluate liquid cooling technology.

Liquid cooling isn’t limited to data centers; cars and other systems also use it to cool high-performance systems in enclosed spaces.

The road to sustainable development

“We’re about to start a new journey,” Smith said of the debut of the liquid-cooled mainstream accelerator.

In fact, NVIDIA plans to launch a version of the A100 PCIe card next year that will feature the H100 Tensor Core GPU based on the NVIDIA Hopper architecture. In the near term, NVIDIA plans to apply liquid cooling technology to its own high-performance data center GPUs and the NVIDIA HGX platform.

To drive rapid adoption, the liquid-cooled GPUs announced today reduce power consumption while maintaining performance. In the future, we expect these cards to deliver even better performance for the same power usage, giving users what they need.

“Measuring power alone is meaningless, reducing carbon emissions while improving performance is what we are striving for,” said Smith.

Learn more about NVIDIA's new A100 PCIe liquid-cooled GPU.