Skip to content
Neural Daily

ToolsNews

NVIDIA DGX Spark HPC kit: a four-node cluster for the desk

🇭🇺 Olvasd el magyarul →

At SC26 NVIDIA introduced the DGX Spark HPC Developer Kit: four DGX Sparks in a ring with 512GB of memory and 20TB of storage, enough to run models up to 700B parameters locally.

· 3 min read

Four stacked DGX Spark systems linked by cables on a desk next to a coffee mug
Image: NVIDIA · Press kit · Source

What was announced?

At SC26, the supercomputing conference taking place this week in Chicago, NVIDIA introduced the DGX Spark HPC Developer Kit (NVIDIA). It is a four-node cluster of DGX Spark systems for researchers, academics and developers. The nodes form a ring, each connected directly to its two neighbors. The goal is simple: run multi-node scientific workloads, AI models and agents at the desk instead of waiting in the queue for a shared GPU cluster.

In October we covered GTC Berlin, where NVIDIA said up to eight DGX Sparks can be clustered. What is new is a ready-made four-node configuration aimed at research work, plus HPC software.

What do four nodes give you?

  • Four GB10 Grace Blackwell Superchips linked over ConnectX-7 networking, with no switch to buy or set up.
  • 512GB of combined memory and 20TB of NVMe storage across the four nodes.
  • With NVFP4, models up to 700B parameters run locally; NVIDIA's example is DeepSeek V4.1.
  • Scales from a single node up to four.

What software comes with it?

An installation playbook for the NVIDIA HPC SDK is coming soon. It brings GPU-accelerated C, C++ and Fortran compilers, libraries and MPI tools, the same toolchain used on NVIDIA-powered supercomputers. Teams can test how computation and data are distributed across four nodes and catch communication bugs before scaling to larger systems.

On four systems NVIDIA demonstrated LAMMPS with Kokkos for materials simulation and NAMD for molecular dynamics. The wider DGX Spark community has also validated GROMACS, OpenFOAM and WRF, plus the OpenFold3 and Boltz-2 NIM microservices. Nemotron 3.5 Super VL gets a new open dataset and post-training recipes for life sciences and chemistry.

Who is already using it?

  • UC Berkeley, Rubin Lab: uses Evo 2 to predict bacterial traits from genome sequences. Without an 11-hour cluster queue, results arrived 15x sooner.
  • UC Berkeley, Hartwig Group: LoRA fine-tuning of Nemotron 3 Super 120B-A12B for chemistry on a single device with NVFP4, running 8 to 24 hour training fits back to back.
  • Kvantify: cuQuantum tensor-network simulations with up to 70 qubits and 975 gates on a single DGX Spark, for penicillin ground-state energy calculations.
  • OpenFold: fine-tuning OpenFold3 on DGX Spark to improve protein-ligand structure predictions.

Where can you get it?

The kit is on display at SC26 and available for sale. DGX Spark systems and cables are sold by OEM partners (Acer, ASUS, Dell, GIGABYTE, H3C, HP, Lenovo, MSI, xFusion) and through NVIDIA Marketplace. The announcement does not list a price.

Why it matters

In universities and research labs, GPU time is the bottleneck. A four-node desktop cluster lets teams develop and debug locally and save their server allocation for production runs. Unpublished results and proprietary datasets also stay on premises. The HPC SDK playbook is not out yet, so the full HPC workflow is worth testing once it ships.

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles