The AI genAIRation | Part 1: Taking Air-Cooled AI Servers to the Next Level

"Air has reached its limits"

This emblematic phrase is now echoing across the entire IT cooling industry.

For years, air cooling was the industry default. It’s cheap, modular and easy to implement in traditional air cooled data centers. This gives IT operators a high level of implementation flexibility and thermal autonomy. Today, hyperscalers are rapidly shifting to liquid cooling to handle the extreme rack densities of AI and HPC workloads.

But where does that leave everyone else?

Enterprises, air-cooled data centers and edge operators  face a dilemma: many of them desire AI computing, but liquid cooling introduces investment, complex new infrastructure and new critical cooling components not present before.

Operators deserve ultimate flexibility in deployment types. Choice of liquid or air.

This article presents where, exactly, does air fall short and how Calyos can help to address this problem without rebuilding everything.

"Air has reached its limits"…but where?

The cooling process inside a server can be broken into three distinct steps:

1. Heat absorption (extracting heat from the processor)
2. Heat transport (moving heat from the processor to the server perimeter)
3. Heat rejection (releasing heat out of the server)

In traditional servers, air manages all three. See picture below:

When air-cooled servers (that can be assisted by vapor chambers and standard heat pipes) implement high TDP processors  , maintaining an acceptable chip temperature becomes challenging :  

1. at heat absorption : the capacity of air to receive the heat load remains too low due to its poor convective thermal coefficient 
2. at heat transport : air volumetric heat capacity remains too poor to transport the heat

In order to accommodate air cooled servers for AI GPUs, engineers have to drastically increase both: 

1. the size 
2. and the airflow rate of their servers.

This is illustrated by the air-cooled HGX B300 server platforms available today from Dell, Cisco, Supermicro, Lenovo, and other vendors.

‍

‍

The HGXB300 air-cooled solution presents up to 15 fans and a height up to 8U for the GPU tray.

But this solution comes with two design trade-offs:

1. Large chassis volume : Server form factor is an important criterion in the IT industry. The larger the server, the more rack space it occupies, and increase the IT room space requirements.

2. High fan count : Fans are moving components, and their power consumption increases roughly with the cube of the required airflow rate. This impacts operational costs as well as the acoustic noise levels.

One way to address these challenges is to identify a more effective medium than air for heat absorption and heat transport.

From Heat Pipes and Vapor chambers to the Loop Heat Pipe

Calyos' technology allows us to build on the foundation from heat pipes and vapor chambers and take this to the next level. By leveraging the HTC of 2-phase in a simple, passive, closed loop system, there are opportunities to enable either: 

1. higher power GPUs for air cooled systems 

2. smaller systems for current GPUs. 

This enables a new generation of air-cooled AI infrastructure, combining higher compute density with improved thermal efficiency.

Partner with Calyos for AI server Cooling

Calyos is actively developing and deploying advanced cooling solutions for server and racks applications. We are working with industry leaders to ensure the new generation of AI servers operates with greater efficiency, reliability, and sustainability.

If you are designing AI servers , or tackling thermal management challenges in data center applications, we want to hear from you. Let’s work together to create robust, maintenance-free cooling solutions that push the boundaries of performance and reliability.

Any question? Contact our technical team : info@calyos-tm.com

‍