Students at the Southern Methodist University in Dallas, Texas, have put together a "baby supercomputer" — by clustering 16 NVIDIA Jetson Nano boards into a single cohesive hole designed as a gentle introduction to the university's somewhat bigger supercomputing systems.
"We started this project to demonstrate the nuts and bolts of what goes into a computer cluster," says team lead Eric Godat of the miniature high-performance computing cluster, which will be on display at the International Conference for High Performance Computing, Networking, Storage, and Analysis (SC22) next week. "We’re bringing the baby supercomputer to the conference to get people to stop by and ask, 'Oh, what's that?'"
The answer: 16 NVIDIA Jetson Nano systems-on-modules (SOMs), in their bundled carrier boards, in a custom laser-cut acrylic housing with integrated cooling, four power supplies, and 60 hand-made interconnecting wires.
The latter, Godat jokes, provided one of the only hiccups in the build process: "They had to learn how to strip wires and not shock themselves — they put together everything from the power supplies to the networking all by themselves," he explains of the two students who build the system. "We only started one small fire."
"We chose to use NVIDIA Jetson modules because no other small compute devices have onboard GPUs," says Conner Ozenne, one of the two students behind the project, "which would let us tackle more AI and machine learning problems."
Clustering single-board computers into a more powerful whole isn't a new concept, of course: BitScope's Cluster Blade can cram in up to 96 Raspberry Pi 4 Model B nodes into a single whole, while Wei Lin's Broccoli project creates a functional cluster from low-cost Espressif ESP32 microcontroller boards.