- AWS built custom NVIDIA cooling after rejecting existing liquid solutions for the scale
- IRHX conforms to AWS racks without changes in existing infrastructure
- Amazon could extend this cooling approach to graviton chips in the future
Amazon Web Services (AWS) has introduced a patented cooling system created to handle the newest NVIDIA GPU demands.
The heat exchanger in the row, or IRHX, was developed in response to the growing power and heat requirements of the hardware such as the NVIDIA GB200 NVL72.
AWS evaluated existing liquid cooling solutions, but discovered that they did not meet the needs of the company.
AWS Graviton Next?
“They would take too much space in the data center, they would still require important modifications for the data centers, or substantially increase the use of water,” said Dave Brown, vice president of computation and ML in AWS, in a presentation published on YouTube, which can be seen below.
“And although some of these solutions could work for lower volumes in other suppliers, they would simply be sufficient liquid cooling capacity to support our scale.”
The IRHX system consists of a pumping unit, a water distribution and fan distribution cabinet.
The fluid cools the chips through a coldly designed cold plate and Nvidia, then tour the IRHX, where it cools and releases.
“With the IRHX we do not need to design the data center around the shelf,” said Brown.
The system admits the most powerful EC2 instance of AWS, the P6E Ultraver, which includes the GB200 NVL72. This Blackwell GPU Scale configuration will work together as a single unit.
Brown said the GB200 NVL72 “allows 72 GPU Nvidia Blackwell acting as a single mass GPU.”
Amazon has previously built personalized hardware, including chips and network systems. The IRHX extends that strategy to cooling, which allows AWS to implement new GPU racks without redesign its facilities.
The company said the system conforms to the dimensions and infrastructure of the existing frame, which makes it scalable in global data centers.
While IRHX is currently combined with Nvidia Blackwell headquarters, it is likely to be used with Amazon Graviton chips if your cooling needs increase.
For now, the system is feeding AI workloads that require both the scale and speed.
Attend
