Desktop Supercomputers: Past, Present, and Future

Summary

Today’s supercomputers are following the same path that past supercomputers followed: toward massive, centralized, shared resources. Users must create job scripts, submit them to a resource manager, and wait for the applications to run. Although some applications certainly can use an entire TOP500 system (or more) for a single application, a number of applications and a great deal of research do not need that level of computer power and can get by with just a few nodes or interactive applications. Interactivity is becoming extremely important as AI application usage grows. Rather than have these applications sit in a job queue, why not run them on a desktop supercomputer that is controlled by the user?

Desktop supercomputers achieve two things. First, they give individual users more compute power to run applications locally, under their control, whenever they want. They can be used for application development, testing, pre- and postprocessing of data, applications and problems that do not require large core counts or memory capacity, interactive applications that are extremely popular because of Jupyter notebooks, and a myriad of other cases. More power to the user! Second, users can remove small node count jobs to desktop supercomputers from the large centralized systems, which collectively can result in more time for the very large scale applications that need the entire system or a significant part of it.

Although past attempts have been made at desktop/deskside supercomputers, a combination of factors and events did not allow them to succeed. A key reason was the prohibitive cost of these systems, limiting the market. Current desktop supercomputers such as Limulus are less than a quarter the cost of the least expensive past desktop supercomputer. Moreover, Limulus allows you to turn nodes on and off as needed, making it very energy friendly. If a user does not need absolute performance, then an SBC cluster is very affordable and requires very low power. A simple Cluster HAT system can be carried in your bag and uses less than 10W.

Given that the desire for more compute power in the hands of the user is growing exponentially, driven by Jupyter-style notebooks and AI, it is easy to see why desktop supercomputers are becoming more popular.