Official releaseLocal AI / AI PC

NVIDIA Introduces the DGX Spark HPC Developer Kit at SC26: Four Nodes in a Ring, No Switch

At SC26, NVIDIA introduced the DGX Spark HPC Developer Kit: four DGX Spark systems linked in a ring over ConnectX-7 with no switch, for 512GB of combined memory and 16TB of NVMe. The HPC SDK playbook is not out yet.

Key answer

The DGX Spark HPC Developer Kit is a four-node DGX Spark cluster NVIDIA introduced at SC26 on November 16, 2026, with each system connected directly to its two neighbors over ConnectX-7. It lets research teams run multi-node scientific and AI workloads at a desk; the accompanying HPC SDK playbook is still to come.

NVIDIA introduced the DGX Spark HPC Developer Kit on November 16, 2026, during the SC26 supercomputing conference in Chicago, aimed at researchers, academics, and developers. The kit is a cluster of four DGX Spark systems, each connected directly to its two neighbors to form a ring.

According to NVIDIA, the setup scales from one system to four, linked directly over NVIDIA ConnectX-7 networking with no network switch to buy or configure. At four nodes it provides 512GB of combined memory and 16TB of NVMe storage. NVIDIA says that with NVFP4, models of up to 700 billion parameters can be deployed locally across the four systems, naming DeepSeek V4.1 as an example. That is the vendor's statement; the post gives no performance figures for it.

On software, the installation playbook for the NVIDIA HPC SDK is described as coming soon and will bring GPU-accelerated C, C++, and Fortran compilers, libraries, and MPI tools. NVIDIA lists LAMMPS with Kokkos and NAMD as demonstrated four-system workloads, and says the community has validated GROMACS, OpenFOAM, WRF, and NIM microservices such as OpenFold3 and Boltz-2.

The post cites several user examples, with figures supplied by the users or NVIDIA. A lab at the University of California, Berkeley runs a genomics model on DGX Spark and says results arrived 15 times sooner without 11-hour cluster queue times. Another research group performs LoRA fine-tuning of a 120B-class model on a single DGX Spark using NVFP4 quantization.

On availability, NVIDIA says the kit is on display at SC26 and available for sale, with DGX Spark products and cables sold through OEM partners and the NVIDIA Marketplace. Listed partners are Acer, ASUS, Dell Technologies, GIGABYTE, H3C, HP, Lenovo, MSI, and xFusion. The post gives no pricing.

For enterprises considering GB10-class hardware for on-prem AI, two points stand out. First, direct multi-node connection is now an officially supported configuration, but combined memory does not mean any model loads across machines automatically. That depends on whether the inference engine supports tensor parallelism across nodes and on the quality of the links between them. Second, the validated workloads published so far are mostly scientific computing, so LLM serving behavior under concurrency still has to be measured.

X Cube view

A desk-side multi-node cluster lowers the cost of testing large models and scientific workloads in your own environment, keeping data and unpublished results in-house. When evaluating it, confirm node count, interconnect, and the inference engine's cross-node support together, and validate with your own workloads before deciding whether it can take over part of what a shared cluster does today.