The Rack is the New Server


Published May 2024


The supercomputer industry and the Web infrastructure industry – what are sometimes called hyperscalers because they need millions of servers support billions of users on their applications – have much in common. Rackscale infrastructure is one of them.

And soon, this rackscale approach to building and deploying datacenter equipment will be common not only among HPC centers, hyperscalers and their cloud computing groups, and other large infrastructure service providers, but among enterprises who are mainstreaming generative AI to extend their applications.

As is often the case with technology, those who are doing high performance computing – pushing the limits of compute and storage scale – blaze the trails for fast-following enterprises. Supercomputing centers and hyperscalers both use distributed computing systems for a handful of applications that need to be run at a scale that is not typical in a normal enterprise. Both also have strong traditions of designing and building their own systems, but it has been a long time since either actually cobbled together servers from parts and then networked the resulting servers into a distributed system. What was possible with dozens to hundreds of servers is no longer possible with thousands to tens of thousands of servers.