News

NVIDIA Introduces Four-Node DGX Spark HPC Developer Kit

NVIDIA introduced a four-node DGX Spark cluster for desk-side development of distributed scientific and AI workloads and plans an HPC SDK installation guide.

D
Nov 16, 2026 · 1 min read

NVIDIA introduced a four-node DGX Spark HPC Developer Kit at SC26 for desk-side development of distributed scientific-computing and AI workloads before moving them to shared or larger systems. NVIDIA says the cluster combines 512GB of unified memory and 20TB of NVMe storage.

The four DGX Spark systems connect directly in a ring over ConnectX-7 networking, with each node linked to two neighbors and no network switch required. NVIDIA’s current clustering instructions specify four QSFP cables for that topology and require users to configure the four-device ring manually.

The 512GB total matches the 128GB of unified memory listed for each system. By contrast, the same hardware guide lists 1TB or 4TB of storage per DGX Spark. Those published specifications do not explain how four standard nodes reach the announced 20TB total.

An installation playbook for building and running the NVIDIA HPC SDK on the cluster is coming soon, the company said, without giving a release date. The HPC SDK includes C, C++ and Fortran compilers, GPU-accelerated libraries, multi-node communication tools, and profiling and debugging software. NVIDIA positions the setup as a local environment for testing how computation and data move across nodes and troubleshooting communication before applications are scaled to data-center systems.

NVIDIA said it demonstrated four-system workloads using LAMMPS with Kokkos for materials simulation and NAMD for molecular dynamics. It did not provide independent benchmark results in the materials reviewed. The company said the kit was on display at SC26 and available through DGX Spark systems and cables sold by its marketplace and OEM partners. It did not announce a kit-specific SKU or package price.

More news