NVIDIA Unveils Four-Node DGX Spark Kit

NVIDIA Unveils Four-Node DGX Spark Kit
Tuesday, November 17, 2026
  • •NVIDIA introduced a four-system DGX Spark HPC Developer Kit at SC26 in Chicago.
  • •Four nodes provide 512GB memory and 20TB storage, supporting local models up to 700B parameters.
  • •Berkeley researchers report results 15x sooner without 11-hour cluster queues using DGX Spark.

At SC26, the supercomputing conference taking place this week in Chicago, NVIDIA introduced the DGX Spark HPC Developer Kit for researchers, academics and developers. The kit links four DGX Spark systems in a ring, each connected directly to two neighbors. NVIDIA says an installation playbook for its HPC SDK is coming soon. The system brings the GB10 Grace Blackwell Superchip to researchers’ desks, where they can run multi-node scientific workloads, AI models and agents without waiting for shared GPU cluster time.

The cluster scales from one to four systems using NVIDIA ConnectX-7 networking, with no network switch required. Four nodes provide 512GB of combined memory and 20TB of NVMe storage for distributing models and scientific workloads. Using NVFP4, researchers can deploy models with up to 700B parameters locally across the four systems, including DeepSeek V4.1. NVIDIA says this can help teams with fixed compute allocations develop and debug locally, reserve server hours for production runs, and keep unpublished results and proprietary datasets on premises.

An upcoming NVIDIA HPC SDK playbook will bring GPU-accelerated C, C++ and Fortran compilers, libraries and message passing interface tools to the cluster. Teams can use them to build and test applications across four nodes, inspect how computation and data are distributed, and debug communication before scaling to larger systems. Demonstrated workloads include LAMMPS with Kokkos for materials simulation and NAMD for molecular dynamics. DGX Spark users have also validated GROMACS, OpenFOAM, WRF and NVIDIA NIM services including OpenFold3 and Boltz-2. NVIDIA says new open datasets and post-training recipes for Nemotron 3.5 Super VL let researchers adapt it for life sciences and chemistry on infrastructure they control.

NVIDIA describes uses including agents that prepare simulation inputs, launch runs and summarize results for scientists to check, and AI models trained to approximate slow simulations for refining wing designs toward minimum drag. Teams can narrow a 10-million-cell simulation to 50-500 relevant dimensions before using supercomputer time. Kvantify reports running tensor-network simulations of quantum-chemistry circuits with up to 70 qubits and 975 gates on one DGX Spark. CYRAN AI uses its PixelFlux library to accelerate JPEG 2000 decoding for large satellite images and local geospatial analysis. OpenFold demonstrates fine-tuning OpenFold3 for protein-ligand predictions in drug discovery.

At UC Berkeley, the Rubin Lab at IGI uses DGX Spark and NVIDIA Evo 2 to predict bacterial traits from genome sequences, comparing genomic and protein language model embeddings with classical baselines. It reports results arrived 15x sooner without 11-hour cluster queues. The Hartwig Group uses one system to LoRA fine-tune Nemotron 3 Super 120B-A12B on experimental chemistry values; with NVFP4 quantization, it fit the 120B-class model on one device and ran continuous eight- to 24-hour training fits without multi-day queue waits. The kit is on display at SC26 and available for sale through OEM partners and NVIDIA Marketplace.

Subscribe to the AIB Newsletter

The fastest, most reliable AI news, straight to your inbox

Companies related to this article