AI, GPU Create New Data Center Cooling Challenges
One of the hottest topics in the data center and IT industries right now is Artificial Intelligence (AI) and the use of GPUs as opposed to the traditional CPUs. As technology evolves, more players are utilizing GPUs to process the performance required for deep learning and AI. As AI becomes more widespread and mainstream, the question of cooling takes center stage.
As with most IT advances, infrastructure is often late to the party. Software and hardware is developed for hyperscale customers at the onset and then infrastructure – such as power and cooling – are left with the challenge of devising solutions that will not shortchange the performance of, in this case, AI. A recent article posted on Data Center Frontier by Rich Miller (“AI Boom Boosts GPU Adoption, High-Density Cooling”) noted the increasing need for power and cooling. He says, “The rise of artificial intelligence and the GPU computing hardware that often supports it, is reshaping the data center industry’s relationship with power density.” He adds, “The trend is challenging traditional practices in data center cooling, and prompting data center operators to adapt new strategies and designs.”
The good news for many of the early AI and GPU adopters is the availability of a data center cooling strategy that not only provides the high-density cooling required for this new technology, but allows for scalability and higher densities in the future. The ChilledDoor® Rack Cooling System by Motivair® provides 100% heat removal for its customers for densities up to 75kW per rack.
ADOPTION OF AI/GPUs IN DATA CENTERS
The adoption of AI hardware is still in the infancy stages and primarily focused on hyperscale users at this point; however, experts are already reporting on the growing investments by other companies and industries wanting to step into the new technology. In the same article from Data Center Frontier, Diane Bryant of Intel reiterates “’AI is the fastest-growing segment of the data center…that 7 percent of servers sold in 2016 were dedicated for AI workloads. While that is still a small percentage of its business, Intel has invested more than $32 billion in acquisitions of Altera, Nevana and MobilEye to prepare for a world in which specialized computing for AI workloads will become more important.”
Even more notable is the increased activity at NVIDIA, the leader in GPU computing. According to reports, its revenue from data center customers had grown almost 200 percent over the past year. GPUs allow for accelerated computer applications, which is used most notably for analytics and deep learning. Early adopters of GPU technology include government labs, universities and hyperscale data center clients.
HOW TO COOL AI, GPU TECHNOLOGY
So just how dense are these racks that are currently utilizing GPUs and AI? Some users are seeing up to 30kW to 40kW per rack, which obviously requires specialized power and cooling. Motivair has worked with hyperscale data center clients, such as LinkedIn, to utilize ChilledDoors as a solution which will not only cool their workload for today but build in additional capacity for future growth for when the time comes they add GPUs or AI into their workload.
Additionally, ChilledDoors are rack agnostic meaning that can fit on any standard server rack. As companies look to implement AI and GPU hardware into their server racks and cabinets, the doors require very little future infrastructure work. For example, the door can be upgraded and swapped to reach even higher density or simply an interface frame change may be all that is required down the road, which gives a greater ROI on your initial infrastructure investment.
To learn more about Motivair’s ChilledDoor Rack Cooling System for your high-density cooling need, visit www.ChilledDoor.com