Computing

NVIDIA Launches DGX Spark HPC Developer Kit to Bring Supercomputing Power to Researchers' Desks

Tendela Briefing ·
4 views · 0 Comments
NVIDIA Launches DGX Spark HPC Developer Kit to Bring Supercomputing Power to Researchers' Desks

Image: NVIDIA · Source

NVIDIA unveiled the DGX Spark HPC Developer Kit at SC26, enabling researchers to run multi-node scientific and AI workloads locally with a four-node cluster of DGX Spark systems linked for high performance without external switches.

At the SC26 supercomputing conference in Chicago, NVIDIA introduced the DGX Spark HPC Developer Kit, targeting researchers, academics, and developers seeking dedicated high-performance computing resources. The kit consists of a four-node cluster of NVIDIA DGX Spark systems, interconnected in a ring topology over NVIDIA ConnectX-7 networking. This design eliminates the need for external network switches, simplifying setup and reducing overhead.

Each DGX Spark node integrates the GB10 Grace Blackwell Superchip, and combined, the four-node cluster offers 512GB of memory and 20TB of NVMe storage. This enables researchers to distribute large AI models and scientific workloads across the cluster. For example, NVIDIA demonstrated deployment of models with up to 700 billion parameters locally across the cluster using NVFP4 quantization formats. This capacity allows local execution of demanding workloads such as scientific simulations and large AI agents, removing dependence on shared GPU clusters and reducing queue times.

NVIDIA is soon releasing a playbook for the HPC Software Development Kit (SDK) that supports GPU-accelerated C, C++, and Fortran compilers, libraries, and message-passing interfaces. These tools facilitate development, debugging, and optimization of multi-node scientific applications on the DGX Spark cluster before scaling to larger data center systems.

The kit supports workloads demonstrated with popular scientific simulation software including LAMMPS with Kokkos for materials science, NAMD and GROMACS for molecular dynamics, OpenFOAM for fluid dynamics, and weather modeling software WRF. NVIDIA’s microservices like OpenFold3 and Boltz-2 enable biomolecular structure prediction, illustrating the broad applicability of the system.

Multiple research groups are leveraging DGX Spark for accelerating diverse scientific projects. For instance, the Rubin Lab at UC Berkeley reported 15 times faster results predicting bacterial traits from genomic data by avoiding cluster queue delays. Similarly, the Hartwig Group at Berkeley achieved efficient fine-tuning of a 120-billion parameter model for chemistry predictions locally, bypassing multi-day waits on shared clusters.

The DGX Spark Developer Kit is commercially available through OEM partners like Acer, ASUS, Dell Technologies, GIGABYTE, HP, Lenovo, MSI, and others, as well as via the NVIDIA Marketplace. By placing supercomputing power directly on researchers’ desks, NVIDIA aims to transform workflows, enabling iterative development, debugging, and secure management of proprietary data and experimental results locally.

In parallel, NVIDIA also announced Nemotron for Science, a platform incorporating its advanced 3.5 Super VL language models, scientific datasets, and post-training recipes to accelerate discovery in fields such as chemistry, life sciences, and materials research. Nemotron can run both on DGX Spark desktop clusters and data center infrastructure, offering flexible deployment for scientific AI agents across scales.

Overall, the NVIDIA DGX Spark Developer Kit represents a significant step toward democratizing access to multi-node GPU supercomputing, empowering researchers to innovate faster with dedicated, high-capacity compute resources directly at their workstations.

Sources and original reporting

Comments (0)

No comments yet. Start the discussion.

Write a comment

Comments are published after moderation. Your name and comment will be visible publicly. Account