Nvidia introduced the DGX Spark HPC Developer Kit, a four-node cluster of its DGX Spark desktop systems, at the SC26 supercomputing conference in Chicago on November 16, 2026.

The move targets a persistent bottleneck in computational science as described by Nvidia: researchers waiting hours or days in shared cluster queues, slowing the iteration required to develop and debug multi-node scientific workloads. The company’s pitch is that placing a four-system cluster on a researcher’s desk shifts development work off shared infrastructure and onto local hardware.

Automatica reported on Monday that Nvidia broadened its Vera CPU line into an HPC rack system3 and announced the Rubin HPC GPU, a simulation-focused chip expected in 20284, at the same event.

The kit, built on Nvidia’s GB10 Grace Blackwell Superchip, links four DGX Spark systems directly to one another over ConnectX-7 networking in a ring topology, according to Nvidia’s announcement1. No external network switch is required. The configuration scales from one node to four; at full scale it provides 512 GB of combined memory and 20 TB of NVMe storage. Nvidia said models with up to 700 billion parameters, such as DeepSeek V4.1, can be deployed locally using NVFP4 quantization across the four nodes.

Researcher Claims Hinge on Local Iteration

Nvidia said two groups at UC Berkeley are already using single DGX Spark units to bypass cluster queues. The Rubin Lab at UC Berkeley’s Innovative Genomics Institute used one DGX Spark alongside the Nvidia Evo 2 model to predict bacterial traits from genome sequences. With no 11-hour cluster queue times, results arrived 15 times sooner on the DGX Spark, according to Nvidia.

The Hartwig Group at UC Berkeley used a single Spark to perform low-rank adaptation fine-tuning on the Nemotron 3 Super 120B-A12B model, fitting the 120-billion-parameter-class model on one device via NVFP4 quantization. Nvidia said the lab bypassed multi-day cluster queue times and ran continuous training fits of eight to 24 hours back-to-back. None of these performance claims have been independently verified, and Nvidia’s announcement included no independent throughput, power-draw or scaling-efficiency data for the four-node kit.

Software Stack Tied to a Forthcoming Playbook

The developer kit is eventually designed to run the same software stack used on Nvidia-powered supercomputers, but Nvidia said an installation playbook for the Nvidia HPC SDK—which includes GPU-accelerated C, C++ and Fortran compilers, libraries, and message passing interface tools—is “coming soon.” The playbook is intended to let researchers build and test multi-node scientific applications locally, then scale to larger systems.

Nvidia listed demonstrated four-system workloads including LAMMPS with Kokkos for materials simulation and NAMD for molecular dynamics. The company also cited the broader DGX Spark community for validated workloads in GROMACS, OpenFOAM, WRF, and Nvidia NIM microservices such as OpenFold3 and Boltz-2.

Quantum-chemistry software firm Kvantify published its own account of using a single DGX Spark with Nvidia’s cuQuantum libraries on its company blog2. The researchers reported running tensor-network simulations of quantum-chemistry circuits with up to 70 qubits and 975 gates, using circuits generated for penicillin ground-state energy calculations. Nvidia’s post also cites the 70-qubit, 975-gate figures. The blog describes results from a Dell-branded DGX Spark unit and does not evaluate the four-node configuration announced at SC26.

Availability Without Pricing

The developer kit is on display at SC26 and available for sale, Nvidia said. The company listed partners including Acer, ASUS, Dell Technologies, GIGABYTE, H3C, HP, Lenovo, MSI, and xFusion, and said DGX Spark products and cables can be purchased through those OEMs or via the Nvidia Marketplace. Nvidia did not disclose pricing for the four-node kit, nor did it state when the HPC SDK playbook would ship.

The Nvidia Nemotron for Science5 release, which includes the Nemotron 3.5 Super VL open model, the LitmusBench v0.2 chemistry dataset with nearly 50,000 questions, and post-training recipes, said KTH Royal Institute of Technology built an AI agent on Nemotron that repaired broken GROMACS molecular dynamics setups autonomously, fixing 81% of test cases overnight on a single DGX Spark system.