Researchers who wait in a queue for shared GPU time now have another option: NVIDIA’s DGX Spark HPC Developer Kit links four DGX Spark systems into a desktop cluster with 512GB of combined memory and 20TB of NVMe storage. NVIDIA announced it at SC26, the supercomputing conference taking place this week in Chicago.

Each system is built on the GB10 Grace Blackwell Superchip. The four nodes connect to their two neighbors in a ring over NVIDIA ConnectX-7 networking, so there is no network switch to buy or set up. The kit is on display at SC26 and, according to NVIDIA, available for sale.

What is in the DGX Spark HPC Developer Kit?

The kit scales from one system, called a node, up to four. The main points from NVIDIA:

  • Four DGX Spark nodes, each linked directly to its two neighbors
  • 512GB of combined memory at four nodes
  • 20TB of NVMe storage at four nodes
  • NVFP4 support for running models of up to 700B parameters locally across the four systems, with DeepSeek V4.1 given as an example
  • An NVIDIA HPC SDK playbook, listed as coming soon

In practice, the ring layout means a lab can add a node without shopping for switch hardware. NVIDIA says the aim is to let teams develop and debug on their own machines, save server hours for production runs and keep unpublished results and proprietary datasets on premises.

DGX Spark systems from NVIDIA and its OEM partners arranged in a pyramid
Photo: NVIDIA

What will the HPC SDK playbook add?

The playbook is not out yet. NVIDIA says it will bring GPU-accelerated C, C++ and Fortran compilers, libraries and message passing interface tools to the cluster, the same development tools used on NVIDIA-powered supercomputers.

That would let a team build and test a scientific application across four nodes, check how computation and data are distributed and debug communication problems before scaling to a larger system.

Which scientific workloads run on it?

NVIDIA demonstrated two four-system workloads: LAMMPS with Kokkos for materials simulation and NAMD for molecular dynamics. The wider DGX Spark community has validated GROMACS for molecular dynamics, OpenFOAM for fluid dynamics and WRF for weather modeling, along with NVIDIA NIM microservices such as OpenFold3 and Boltz-2 for predicting biomolecular structures.

NVIDIA also points to early users:

  • The Rubin Lab at IGI, University of California, Berkeley, uses one DGX Spark and the Evo 2 model to predict bacterial traits from genome sequences. With no 11-hour cluster queue, results arrived 15x sooner, according to NVIDIA.
  • The Hartwig Group at UC Berkeley fine-tuned the Nemotron 3 Super 120B-A12B model with LoRA on a single device using NVFP4 quantization, running eight- to 24-hour training fits back to back.
  • Kvantify reports tensor-network simulations of quantum-chemistry circuits with up to 70 qubits and 975 gates on a single DGX Spark.
  • CYRAN AI, an IIT Delhi spin-out, uses DGX Spark to speed up JPEG 2000 decoding of large satellite images.

New open dataset and post-training recipes for Nemotron 3.5 Super VL also help researchers adapt that model for life sciences and chemistry on infrastructure their organization controls.

How much does it cost?

NVIDIA’s announcement lists no price for the kit and gives no pricing by country. DGX Spark products and cables are sold through NVIDIA Marketplace and through OEM partners Acer, ASUS, Dell Technologies, GIGABYTE, H3C, HP, Lenovo, MSI and xFusion.

NVIDIA’s DGX Spark product page describes each system as offering up to 1 petaFLOP of FP4 AI performance and up to 128 GB of system memory. For the current price and the countries a partner ships to, check the Marketplace or partner listing directly.

Who it suits

University and lab researchers. A fixed server allocation means every test job waits its turn. Four local nodes give a place to run and debug multi-node work first.

Teams with sensitive data. Unpublished results and proprietary datasets stay on premises, which matters for groups that cannot send them to a shared cluster.

Developers of HPC code. Once the playbook arrives, C, C++ and Fortran tooling on the desk should make it easier to test communication between nodes before a production run.

Who should wait. Anyone who needs the HPC SDK playbook today, or needs a firm price first. NVIDIA frames the kit as a place to prepare work for larger systems, not a replacement for them.

What is still unknown?

NVIDIA has not given a price, a ship date by country or a date for the playbook, and the announcement does not spell out what is in the box beyond the four systems and their cables.

For the specifics, read NVIDIA’s announcement post and check the DGX Spark page on NVIDIA Marketplace for current listings.