NVIDIA Unveils Multi-Node HPC Developer Kit

The new desk-side cluster links four DGX Spark systems with Grace Blackwell Superchips to enable local testing of complex HPC workloads.

NVIDIA has unveiled a new HPC Developer Kit designed to bring the power of a multi-node cluster directly to the desks of researchers and software developers. By linking four DGX Spark systems into a single desk-side unit, the kit allows for local testing of complex multi-node workloads without the need for a full-scale data center environment. Each of the four compute nodes is powered by the GB10 Grace Blackwell Superchip, providing a high-performance foundation for experimentation and code development.

Seamless Networking Without the Switch

One of the most notable features of the new kit is its networking efficiency. The four systems are linked in a ring configuration using NVIDIA ConnectX-7 networking, a setup that utilizes a Direct ConnectX-7 ring interconnect to bypass the need for a separate network switch. Despite its compact footprint, the combined cluster packs a significant punch, offering 512 GB of memory and 20 TB of NVMe storage. NVIDIA claims this hardware profile is capable of distributing models with up to 700 billion parameters when utilizing NVFP4.

From Weather Modeling to Molecular Research

While the kit is intended for development rather than replacing production-grade supercomputers, it is built to handle demanding scientific workloads. NVIDIA highlighted applications such as LAMMPS and NAMD, as well as OpenFOAM, weather modeling, and biomolecular structure prediction as ideal use cases for the platform. To support these efforts, a forthcoming NVIDIA HPC SDK playbook will provide the necessary compilers, libraries, and message-passing tools specifically optimized for the Spark cluster.

Availability and Ecosystem Support

Showcased recently at SC26, the HPC Developer Kit is now available through the NVIDIA Marketplace and a wide range of OEM partners, including Acer, ASUS, Dell, Gigabyte, HP, Lenovo, and MSI. While NVIDIA has not disclosed a specific standalone price for the four-node bundle, the DGX Spark hardware continues to be sold through both NVIDIA and its global partner network, targeting academics and developers looking to bridge the gap between local coding and massive-scale deployment.