I spoke with two engineers trying to prevent AI from overheating the future
- Advertisement -
Related Posts
- Advertisement -
- AI -data centers overwhelm air cooling with rising power and heat
- Liquid cooling becomes essential as serverness rises with AI growth
- New hybrid cooling cuts electricity and water, but is confronted with hesitation
Since AI transforms everything, from search engines to logistics, the hidden costs are becoming increasingly difficult to ignore, especially in the data center. The power needed to perform generative AI is pushing the infrastructure that goes beyond traditional air cooling.
To explore the scale of the challenge, I spoke with Daren Shumate, founder of Shumate Engineeringand Stephen Spinazzola, director of the Mission Critical Services company.
With decades of experience in building large data centers, they are now aimed at solving AI’s energy and cooling requirements. From the failure of air systems to the promise of new hybrid cooling, they explained why AI forces data centers in a new era.
What are the biggest challenges in cooling a data center?
Stephen Spinazzola: The biggest challenges in cooling data centers are strength, water and space. With computer use with high density, such as the data centers that perform artificial intelligence, there is enormous heat that cannot be cooled with a conventional air cooling system.
The typical cabinet loads have doubled and tripled with the use of AI. An air cooling system simply cannot catch the heat generated by the high kW/ cabinet loads generated by AI cabinet clusters.
We have performed Computational Fluid Dynamic (CFD) on numerous data center rooms and an air cooling system has high temperatures above acceptable levels. The air flows that we map with CFD Show Temperature levels above 115 degrees F. This can lead to servers being closed.
Water cooling can be done in a smaller room with less power, but it requires an enormous amount of water. A recent study found that a single hyper-scaling facility would need 1.5 million liters of water per day to offer cooling and humidification.
These limitations are major challenges for engineers, while planning the new generation of data centers that can support the unprecedented demand that we see for AI.
How does AI change the standard when it comes to heat disiperation from data center?
Stephen Spinazzola: With CFS modeling with potential servers that are switched off with conventional air cooling within AI cabinet clusters, the need for direct liquid cooling (DLC) is required. AI is usually used in 20-30 cabinet clusters at or above 40 kW per cupboard. This represents a four -time increase in KW/ Kabinet with the deployment of AI. The difference is amazing.
A typical chat-gpt-query uses about 10 times more energy than a Google search assignment and that is only for a basic generative AI function. More advanced questions require considerably more power that an AI-cluster farm must process large-scale computer between multiple machines.
It changes the way we think about power. Consequently, the energy needs shift the industry to use more techniques for cooling liquids than traditional air cooling.
We talk a lot about cooling, what about delivering real strength?
Daren Shumate: There are two umbrella new challenges to supply electricity to AI Computing: how to move electricity from UPS output signs to racks with a high density and how you can deliver creatively high densities of UPS stream from Utility.
Power to-racks is still reached with both branching circuits from distribution PDUs to rack PDUs (plug strips) or with plug-in busway about the racks with the in-rack PDUs that connects to the busway at each rack. The nuance is now what the ampacity of Busway makes sense with the striping and what is commercially available.
Even with plug-in busway available on an amplifier of 1,200 A, the density of electricity forces the use of a larger amount of individual busway circuits to meet the density and the strip requirements. Further complicating power distribution is a specific and varying requirement of individual data center -end users of branch circuit monitoring or preferences for distribution.
Datacenter -cooling designs can have average voltage -ups depending on the site limitations. Driven by voltage falling problems, the MV -ups solves the concern for the need to have very large feeding channel banks, but also introduces new medium -sized voltage/tension stations in the program. And when considering medium -sized tension -ups, another consideration is the applicability of MV Rotary UPS systems versus static MV solutions.
What are the advantages/disadvantages of the different refrigerators?
Stephen Spinazzola: There are two types of DLC on the market today. Emersiem cooling and cold plate. Emersion cooling uses large tanks of a non-coonging liquid with the servers placed vertically and are fully stolen in the liquid.
The heat generated by the servers is transferred to the liquid and then transferred to the cooled water system of the buildings with a heat exchanger with closed loop. Emersion tanks take up less space, but require servers that are configured for this type of cooling.
Voudige cooling uses a cooling body that is attached to the bottom of the chip pile that transfers the energy from the chip pile to a liquid led through the cupboard. The liquid is then led to an end of the rich cooling unit of the Government cooling (CDU) that transfers the energy to the cooled water system of the building.
The CDU contains a heat exchanger to transfer energy and 2N pumps on the secondary side of the heat exchanger to guarantee a continuous liquid flow to the servers. Cold plate cooling is effective when cooling the server, but it requires a huge amount of liquid pipe connectors that must have broken the leakage technology.
Air cooling is a proven technique for cooling data centers, which has been around for decades; However, it is inefficient for the racks with high density needed to cool AI data centers. As the taxes increase, it becomes more difficult to fail with the help of CFD modeling.
You present another cooler, how does it work and what are the current challenges for adoption?
Stephen Spinazzola: Our patent pending hybrid-dry/adiabatic cooling (HDAC) design solution offers unique for two temperatures of coolant from a single closed loop, making a higher temperature fluid possible to cool DLC servers and a lower temperature fluid for conventional air cooling.
Because HDAC uses 90 percent less water at the same time than a cooling tower system and 50 percent less energy than an air -cooled chiller system, we have succeeded in getting the most important power use effectiveness (PUE) to around 1.1 annually for the type of Hyperscale Datacenter needed to process AI. Typical AI data centers produce a PUE ranging from 1.2 to 1.4.
With the lower PUE, HDAC offers around 12% more usable IT power from the same size of the utility company. Both economic and environmental benefits are considerable. With a system that offers both an economic and environment benefit, HDAC only requires “a sip of water”.
The challenge for adoption is simple: nobody wants to go first.
Maybe you like it too
- Advertisement -
