NVIDIA introduced the DGX Spark HPC Developer Kit at SC26 as a four-system cluster for developing and debugging scientific and AI workloads locally before moving them to larger systems. The nodes connect in a direct ConnectX-7 ring; NVIDIA says a separate network switch is not required. (NVIDIA’s announcement)

What the four-node setup provides

NVIDIA states that the cluster has 512 GB of combined memory and 20 TB of NVMe storage across its four nodes. The memory figure is an aggregate, not one shared-memory pool.

NVIDIA also says its NVFP4 format enables deployment of models with up to 700 billion parameters across the four nodes. That is a deployment claim for the cluster.

When sizing a workload, keep the cluster-wide totals separate from the resources of any one node.

Which workloads NVIDIA cites

For four-system demonstrations, NVIDIA names LAMMPS with Kokkos and NAMD. It separately says the broader DGX Spark community has validated GROMACS, OpenFOAM, WRF, OpenFold3 and Boltz-2; those examples are not all identified as four-node demonstrations. (NVIDIA’s workload examples)

NVIDIA also describes project uses of DGX Spark beyond the four-node demonstrations. It says Kvantify reported running tensor-network quantum-chemistry simulations of circuits up to 70 qubits and 975 gates on a single DGX Spark. NVIDIA says CYRAN AI uses DGX Spark for JPEG 2000 decoding of large satellite images. The company also describes agent-assisted workflows that prepare simulation inputs, launch runs and summarize results for scientists to review.

In separate case studies, NVIDIA says the Rubin Lab at IGI used DGX Spark and Evo 2 for bacterial-trait prediction, reporting results 15 times sooner than a previous workflow involving an 11-hour cluster queue. NVIDIA describes the Hartwig Group at UC Berkeley using one DGX Spark with NVFP4 for a 120-billion-parameter-class model and avoiding multi-day cluster queues. These are NVIDIA-reported project examples, distinct from the four-system demonstrations.

Software and sales

NVIDIA says an installation playbook for the NVIDIA HPC SDK is forthcoming. The playbook is intended to cover GPU-accelerated C, C++ and Fortran compilers, libraries and message-passing tools.

The kit combines four DGX Spark systems; NeoTeo’s earlier coverage of DGX Spark’s local-AI role provides background on the individual platform. NVIDIA says the HPC kit is available through NVIDIA Marketplace and hardware partners.