NVIDIA launches DGX Spark kit for desktop scientific computing
NVIDIA introduced the DGX Spark HPC Developer Kit at SC26 in Chicago, offering researchers, academics and developers a desktop platform for scientific workloads, AI models and agents. The four-node cluster connects systems in a ring using NVIDIA ConnectX-7 networking, without requiring a network switch. Built around the GB10 Grace Blackwell Superchip, the kit is available for sale through hardware partners and NVIDIA Marketplace.
The cluster provides 512GB of combined memory and 20TB of NVMe storage. Using NVFP4, researchers can deploy models with up to 700B parameters locally. The setup is intended to let teams develop and debug applications without waiting for shared GPU resources, while keeping unpublished research and proprietary datasets on premises. Demonstrated workloads across the cluster include LAMMPS with Kokkos for materials simulation and NAMD for molecular dynamics.
An NVIDIA HPC SDK installation playbook is coming soon, bringing GPU-accelerated C, C++ and Fortran compilers, libraries and message passing interface tools. Researchers will be able to test workload distribution and debug communication before moving applications to larger systems. Potential uses include agent-assisted simulation workflows and AI models that approximate slower simulations. Existing DGX Spark projects span quantum chemistry, satellite imagery, protein structure prediction and genomic analysis.