Nvidia has introduced a four-node DGX Spark kit for researchers who need to develop multi-system AI and simulation workloads before moving them to a shared supercomputer. The nodes connect in a ring through ConnectX-7 networking, avoiding a separate network switch.

Together, the systems provide 512GB of memory and 20TB of NVMe storage. Nvidia says the cluster can run models with as many as 700 billion parameters when they use NVFP4, a low-precision format that reduces memory use. That capacity describes deployment rather than full-precision training, and actual performance will depend on the model and workload.

The company demonstrated scientific software for molecular dynamics, materials research, fluid dynamics and weather modeling. An upcoming HPC SDK playbook will add GPU-accelerated C, C++ and Fortran compilers, libraries and message-passing tools. The intended workflow is to build and debug distributed jobs locally, then reserve larger clusters for production runs.

Nvidia cited early users reporting shorter queues and local control of proprietary data. A UC Berkeley lab said one DGX Spark delivered results 15 times sooner by avoiding an 11-hour shared-cluster wait, while another fit a quantized 120-billion-parameter chemistry model on one device. Those are selected examples rather than general benchmarks. The kit is on sale through Nvidia and hardware partners; the HPC software playbook is still pending.