NVIDIA Introduces DGX Spark HPC Developer Kit for Local AI and HPC Development

At the SC26 conference, NVIDIA announced a four-node DGX Spark setup for local development of distributed AI and HPC workloads without a separate network switch.

On November 16, 2026, at the SC26 conference in Chicago, NVIDIA introduced the DGX Spark HPC Developer Kit, a four-node desktop setup designed for local development of high-performance computing (HPC) and artificial intelligence workloads. The kit consists of four NVIDIA DGX Spark systems connected in a ring with direct network links.

The manufacturer describes the product as a way to test multi-node applications outside large HPC systems. The target users are said to be researchers and developers who want to prepare and validate distributed workloads locally before moving them to larger-scale infrastructure.

DGX Spark Uses a Ring Topology

The four systems are connected in a ring through direct ConnectX-7 links. According to NVIDIA, this configuration does not require a separate network switch. This arrangement lowers the technical barrier to assembling a small solution, but the manufacturer did not publish detailed data on its performance or limitations for different types of distributed workloads in the announcement.

For the four-node configuration, NVIDIA cites a combined capacity of 512 GB of memory and 20 TB of NVMe storage. However, it does not specify the exact configurations of the individual systems that OEM partners will offer.

Claimed Support for Models with Up to 700 Billion Parameters

NVIDIA claims that the setup will allow models with up to 700 billion parameters to be deployed locally when using the NVFP4 format. This is a capability claimed by the manufacturer; the performance, scaling, and practical usability of such a configuration have not yet been independently verified.

Independent measurements and technical details of the individual available configurations will therefore be important when assessing its real-world use.

Sales Through Partners, Price Not Yet Available

According to NVIDIA, DGX Spark products and the necessary cables are available through OEM partners and the NVIDIA Marketplace. However, the company did not disclose the price of the complete four-node kit, the specific models available from individual partners, or regional availability.

NVIDIA also stated that a playbook for the NVIDIA HPC SDK will arrive later. The company did not announce a specific release date or its exact feature scope.

The main open questions surrounding the new product are therefore its total price, OEM partner offerings, and the results of independent tests of the ring network topology. These will show how the four-node solution behaves with specific AI and scientific workloads.

Sources

  • NVIDIA Blog – Confirms the product announcement, four-node ring topology, stated capacities, claimed capabilities, and availability status.
  • NVIDIA at Supercomputing 2026 – Confirms that SC26 takes place November 15–20, 2026, in Chicago.

Verified and updated: November 17, 2026, 06:23

Sharing