Skip to content
worth noting Hardware

NVIDIA introduces the DGX Spark HPC Developer Kit – a desktop four-node cluster for research and AI models

only one source so far

NVIDIA announced the DGX Spark HPC Developer Kit at the SC26 conference – a four-node cluster with the GB10 Grace Blackwell chip, 512 GB of memory, and 20 TB of storage, enabling local execution of models with up to 700 billion parameters without the cloud.

NVIDIA announced the DGX Spark HPC Developer Kit at the SC26 conference in Chicago – a four-node cluster of DGX Spark systems intended for researchers, academics, and developers. The individual nodes are connected directly to each other in a ring via the NVIDIA ConnectX-7 network, without the need to buy and configure a network switch. The kit contains the GB10 Grace Blackwell Superchip and, in a four-node configuration, offers 512 GB of unified memory and 20 TB of NVMe storage.

According to NVIDIA, NVFP4 technology enables local deployment of models up to 700 billion parameters in size across four DGX Spark nodes, citing the DeepSeek V4.1 model as an example. The offering will include a playbook for the NVIDIA HPC SDK (currently in "coming soon" status) with GPU-accelerated compilers for C, C++, and Fortran, and MPI tools – the same tools used on NVIDIA supercomputers.

In its press release, NVIDIA cites several use cases: the company Kvantify runs tensor-network simulations of quantum-chemical circuits with up to 70 qubits on a single DGX Spark, the company CYRAN AI (a spin-out of IIT Delhi) accelerates the decoding of satellite imagery, the OpenFold project is fine-tuning the OpenFold3 model for protein-ligand structure prediction, and the Rubin Lab at UC Berkeley states that it gets results with the Evo 2 model 15 times faster than when waiting for a cluster. According to NVIDIA, the Hartwig Group at UC Berkeley performed LoRA fine-tuning of the Nemotron 3 Super 120B-A12B model directly on a single device thanks to NVFP4 quantization.

The kit is on display at SC26 and is available for purchase through OEM partners (Acer, ASUS, Dell Technologies, GIGABYTE, H3C, HP, Lenovo, MSI, xFusion) and through the NVIDIA Marketplace. Pricing was not stated in the sources. See the source article for details.

What changed

Why it matters

The kit removes one of the main obstacles to research with large models – waiting for time on a shared cluster – and makes it possible to run and train models on the order of hundreds of billions of parameters locally, without sending sensitive data or unpublished results to the cloud. For institutions with limited server capacity, according to NVIDIA, this saves hours of production system runtime, since development and debugging take place directly on the desktop.

Two audiences, two different impacts

What this means

01

For individuals

A researcher or developer can run and train large models (including LoRA fine-tuning of models on the order of 120B parameters) directly at their desk, without waiting to be allocated time on a shared cluster.

What to do Check whether the kit covers your existing workload (model size, memory, storage) before any purchase or testing.
More practical updates →
02

For a business

Companies and research institutions with their own development or research infrastructure can use the kit to reduce dependence on shared clusters and the cloud, save time spent waiting in a queue, and keep unpublished data and models on their own device.

Development
What to decide Consider acquiring the DGX Spark Developer Kit for teams that currently wait for time on a shared GPU cluster or work with sensitive data.
More business impacts →
AI inference DGX Spark HPC NVIDIA supercomputing research hardware

Check the original

Event sources

only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.

1
NVIDIA Newsroom (press releases) primary source · first detected NVIDIA DGX Spark Developer Kit Turns a Desk Into a Supercomputing Lab