Four-node desktop cluster aims to bring HPC work to individual labs
NVIDIA announced the DGX Spark HPC Developer Kit at the SC26 supercomputing conference in Chicago on November 16, 2026. It is aimed at researchers, academics and developers, and is on display and for sale at the show. DGX Spark products and cables are also sold through OEM partners including Acer, ASUS, Dell Technologies, GIGABYTE, H3C, HP, Lenovo, MSI and xFusion, and through NVIDIA Marketplace.
The kit is four DGX Spark systems, each built on the GB10 Grace Blackwell Superchip. Each unit connects directly to its two neighbors in a ring over ConnectX-7 networking, so no network switch is required. It scales from one node to four, and at four nodes it offers 512GB of combined memory and 20TB of NVMe storage. With NVFP4, models of up to 700B parameters can run locally across the cluster, with DeepSeek V4.1 given as an example.
Early uses span several fields. Demonstrated four-system workloads include LAMMPS with Kokkos for materials simulation and NAMD for molecular dynamics. Kvantify reports simulating quantum-chemistry circuits of up to 70 qubits and 975 gates on a single unit. A UC Berkeley lab says a genome-based bacterial-trait prediction came back 15x sooner than waiting through an 11-hour cluster queue. Another Berkeley group fine-tuned a 120B-parameter model on one unit.
The abundance angle is access to knowledge-making tools. Much research compute is rationed by queues and institutional allocations. A desk-sized system that handles large models and simulation work lets smaller groups iterate without waiting, which widens who can do the work.
Caveats: this comes from NVIDIA's own announcement, and the performance claims are vendor or user-reported, not independently verified. The 15x figure depends on that one cluster's queue times. The HPC SDK installation playbook is only listed as coming soon, and no price was given, so how cheap this really is remains unclear.
Source: NVIDIA
Many Minded