Students at Southern Methodist University have built a “baby” supercomputer capable of running and training artificial intelligence (AI) and machine learning (ML) models. Using 16 NVIDIA Jetson Nano modules, four power supplies, more than 60 handmade wires, a network switch, and some cooling fans, the miniature system provides the students the opportunity to administer their own supercomputer and test their machine learning models.
The project aims to help educate those who may never get hands-on with a normal-sized supercomputer, which can sometimes fill a warehouse, or be locked in a data center or in the cloud. The mini supercomputer fits comfortably on a desk, allowing students to tinker with it and learn about what makes up a cluster. A touch screen displays a dashboard with the status of all of its nodes.
“We started this project to demonstrate the nuts and bolts of what goes into a computer cluster,” says Eric Godat, team lead for research and data science in the internal IT organization at SMU.
The DIY challenge was taken on as a STAR research project by senior computer science major Conner Ozenne ’23, who built the computer from scratch.
“It was a great opportunity to apply what I’ve learned in computer science classes,” says Ozenne. “It also helped further my understanding of some foundational principles like networking and parallel computing.”
The first iteration was a mess of wires on a table connecting the NVIDIA Jetson Nano developer kits, with cardboard boxes as heatsinks. One of the initial challenges was how to fit all the components together as a cluster supercomputer inside a 14 x 14 x 16-inch transparent acrylic housing.
“Initially, a lot of it was just calculations,” says Ozenne. “How can we make this fit in the box?”
A precision laser cutter on campus was used to fabricate the airtight container. In just four months, says Ozenne, the project went from nothing to something that resembled a supercomputer.
The students are now developing the mini cluster’s software stack, with the help of the NVIDIA JetPack software development kit, and prepping it to accomplish some small-scale machine learning tasks. Plus, the baby supercomputer could level up with the recently announced NVIDIA Jetson Orin Nano modules.
“Our NVIDIA DGX SuperPOD just opened up on campus,” says Godat. “So we don’t really need this baby supercomputer to be an actual compute environment. But the mini cluster is an effective teaching tool for how all this stuff really works — it lets students experiment with stripping the wires, managing a parallel file system, reimaging cards and deploying cluster software.”
The University’s NVIDIA DGX SuperPOD, which includes 160 NVIDIA A100 Tensor Core GPUs, is in an alpha-rollout phase for faculty, who are using it to train AI models for molecular dynamics, computational chemistry, astrophysics, quantum mechanics and a slew of other research topics. Godat collaborates with the NVIDIA DGX team to flexibly configure the DGX SuperPOD to support tens of different AI, machine learning, data processing and HPC projects.
As far as the baby supercomputer, says Godat, the plan is to take it on the road for workshops around campus and beyond to teach how to build models and do AI research.