Artificial intelligence is changing the data center game. If data centers are to host AI engines that demand massively dense racks and hundreds of GPUs, things will have to change. Existing data centers and traditional designs won’t be able to cope with AI requirements. The solution emerging is to customize the data center to address AI needs. In this TechRepublic Premium feature, Drew Robb discusses how these data centers will need to be architected to host AI successfully.
Featured text from the download:
BEST PRACTICES FOR AI DATA CENTERS
There are many factors that those planning or building AI data centers must take into account if they are to succeed:
Consider outsourcing AI
Outsourcing is a fact of life in the data center space. The large hyperscalers and colocation providers lease space from other data centers in local markets where they lack presence, for example. Enterprise data centers, too, have offloaded many functions over the years, keeping certain gear and applications in house and using other data centers to handle everything else. Why should AI be any different? Where the in-house capabilities are thin, or the power, cooling, and overall infrastructure are found wanting, AI services can be delivered by using the services of nearby data centers. For some, outsourcing of AI may be the only way forward if they don’t have the means to buy a completely new data center.
Phased approach
Those with deep pockets may be able to slam dunk an entire AI data center in one fell swoop. For existing data centers, though, the ability to adapt them or retrofit them to accommodate higher density may be difficult or even impossible due to many variables such as power availability and economics. For most enterprises that don’t go the AI outsourcing route, a phased approach is needed. If the demand is there, they can add a rack or two of AI servers and leave the rest of the data center at a relatively low rate of compute density. Even if power is constrained, many data centers may be able to manage to add at least a few AI servers. Further, only having to implement liquid cooling for a few pieces of equipment lowers the learning curve on liquid technology and keeps cost down. It must be understood that moving a data center from 200 watts per square foot to 400 or more is a major undertaking. It could also be disruptive to existing data center traffic that currently pays the bills. It may not be wise to sacrifice the existing business model in the hope of capitalizing on AI. A phased approach is a better strategy in many cases.
Enhance your AI and data center knowledge with our in-depth 10-page PDF. This is available for download at just $9. Alternatively, enjoy complimentary access with a Premium annual subscription.
TIME SAVED: Crafting this content required 20 hours of dedicated writing, editing, research, and design.
Source link
lol
