- AWS constructed customized Nvidia cooling after rejecting current liquid options for scale
- IRHX matches into AWS racks with out adjustments to current infrastructure
- Amazon may prolong this cooling method to Graviton chips sooner or later
Amazon Net Providers (AWS) has launched a proprietary cooling system constructed to deal with the calls for of Nvidia’s latest GPUs.
The In-Row Warmth Exchanger, or IRHX, was developed in response to the rising energy and warmth necessities of {hardware} just like the Nvidia GB200 NVL72.
AWS evaluated current liquid cooling options however discovered they didn’t match the corporate’s wants.
AWS Graviton subsequent?
“They might take up an excessive amount of knowledge middle flooring area, would nonetheless require main modifications to knowledge facilities, or improve water utilization considerably,” Dave Brown, VP Compute and ML Providers at AWS, stated in a presentation posted on YouTube, which you’ll be able to see beneath.
“And whereas a few of these options may work for decrease volumes at different suppliers, they merely would not be sufficient liquid cooling capability to help our scale.”
The IRHX system consists of a pumping unit, a water distribution cupboard, and fan coils.
Liquid cools the chips via a chilly plate co-designed by AWS and Nvidia, then cycles again via the IRHX, the place it’s cooled and launched.
“With the IRHX we don’t have to design the information middle across the rack,” Brown stated.
The system helps AWS’s strongest EC2 occasion, the P6e UltraServer, which incorporates the GB200 NVL72. This rack-scale setup permits 72 Blackwell GPUs to work collectively as one unit.
Brown stated the GB200 NVL72 “allows 72 Nvidia Blackwell GPUs to behave as a single large GPU.”
Amazon has beforehand constructed customized {hardware}, together with chips and networking techniques. The IRHX extends that technique into cooling, permitting AWS to deploy new GPU racks with out redesigning its services.
The corporate stated the system matches current rack dimensions and infrastructure, making it scalable throughout world knowledge facilities.
Whereas IRHX is at present paired with Nvidia’s Blackwell-based techniques, it’s seemingly for use with Amazon’s personal Graviton chips if their cooling wants rise.
For now, the system is powering AI workloads that demand each scale and pace.
