- AWS built personalized Nvidia cooling after rejected existing liquid solutions for the scale
- IRHX is part of AWS racks without modifying the existing infrastructure
- Amazon could prolong this approach to cool graviton fleas in the future
Amazon Web Services (AWS) has introduced a owner cooling system designed to manage the requirements of the most recent Nvidia GPUs.
The heat exchanger in the row, or IRHX, was developed in response to the growing power and the heat requirements of the material such as the NVIDIA GB200 NVL72.
AWS evaluated existing liquid cooling solutions, but found that they did not meet the needs of the company.
AWS Graviton then?
“They would occupy too much floor space in the data center, would always require major changes in data centers or would considerably increase water consumption,” said Dave Brown, VP Compute and ML of services in AWS, in a presentation published on Youtube, which you can see below.
“And although some of these solutions can work for lower volumes at other suppliers, they simply would not be enough liquid cooling capacity to support our scale.”
The IRHX system consists of a pumping unit, a water distribution cabinet and fan coils.
The liquid cools the chips through a cold plate co-designed by AWS and NVIDIA, then retreats through the IRHX, where it is cooled and released.
“With the IRHX, we don’t need to design the data center around the rack,” said Brown.
The system supports the most powerful EC2 instance of AWS, the Ultraserver P6E, which includes the GB200 NVL72. This Rack scale configuration allows 72 GPU Blackwell to operate together in a single unit.
Brown said that the GB200 NVL72 “allows 72 GPU Nvidia Blackwell to act as a single massive GPU.”
Amazon has already built personalized equipment, including chips and networking systems. The IRHX extends this cooling strategy, allowing AWS to deploy new GPU racks without rethinking its installations.
The company said that the system corresponded to the existing dimensions and infrastructure, which makes them scalable in global data centers.
Although the IRHX is currently associated with NVIDIA Blackwell -based systems, it is probably used with Amazon’s own graviton chips if their cooling needs increase.
For the moment, the system feeds the workloads of the AI which require both the scale and the speed.
To watch
