On November 16, 2026, at SC26, the international supercomputing conference, NVIDIA showed off an HPC Developer Kit that links four DGX Spark desktop computers. The units connect directly to one another in a ring, letting researchers try out how computation and data are divided and how nodes communicate, all at their own desks, before moving to a larger system.
Using the four machines, NVIDIA demonstrated LAMMPS, a materials simulation code built with Kokkos, and NAMD, a molecular dynamics code. The aim is to bring more of the development process to the researcher's desk: writing distributed programs and investigating problems before waiting for time on a shared cluster.
A four-node ring lets you test communication locally
In the kit, each compute node connects directly to its two neighbors through its ConnectX-7 network interface. According to NVIDIA, the ring topology means there is no need to buy and configure a separate network switch. The DGX Spark product specifications list the ConnectX-7 speed as 200 Gbps.
Distributed computing is not just a matter of assigning work to several machines. Developers also have to think about the order in which data is handed off and about communication between nodes. NVIDIA suggests using the four machines to build and test code, then examining how to split computation and data and where communication problems arise, before moving to a large system. Algorithms that use FP32 (32-bit floating point) or mix several precisions are also supported.
Development uses the NVIDIA HPC SDK, which provides GPU-enabled C, C++ and Fortran compilers and libraries. It also includes MPI tools for exchanging data between nodes. The build and run guides for the kit are scheduled to be released soon, which is a different stage from the hardware's availability.
Beyond the demonstrations, NVIDIA cited community-verified examples that run on the setup: GROMACS for molecular dynamics, OpenFOAM for fluid analysis and WRF for weather modeling. It also presented OpenFold3 and Boltz-2, used for predicting biomolecular structures, as NIM microservices. The demonstrations and the community verifications are separate efforts, but together they show the breadth of possible targets, from materials to weather.
What 512GB and 1 PFLOP do and do not indicate
The GB10 in each DGX Spark is a processor combining a 20-core Arm CPU with a Blackwell GPU. According to the product specifications, one unit's CPU and GPU share either 64GB or 128GB of LPDDR5X memory, with a bandwidth of 273GB/s. Four 128GB units add up to 512GB.
Within each node, the CPU and GPU share memory, while computation and data are divided among the nodes. The 512GB is the combined capacity of four machines, not a sign that all four share a single address space.
NVIDIA says the four units can run a model of up to 700 billion parameters locally in NVFP4, citing DeepSeek V4.1 as an example. NVFP4 is a low-precision format that represents model values as 4-bit floating-point numbers. The size of model that can be handled and the speed at which it runs should be read separately.
The four-node configuration is intended for distributed development across nodes. The 512GB is a combined capacity, the 1 PFLOP is the theoretical FP4 performance of a single unit, and the HPC SDK setup guides are due soon. These are separate conditions. The table below matches the targets and conditions in NVIDIA's announcement and product specifications.
| Item | Figure / availability | What it tells you |
|---|---|---|
| Memory | 512GB across four units | 128GB per node, shared by the CPU and GPU within each node |
| AI compute | Up to 1 PFLOP per unit, FP4 | Theoretical performance using sparsity |
| Hardware | On sale | Purchase channels for DGX Spark and cables are listed |
| HPC SDK guides | Coming soon | Build and run guides for the four-node setup are planned |
Note: Categorized by target, precision and availability stage from the NVIDIA announcement and product specifications, retrieved November 17, 2026. These are the manufacturer's descriptions and specifications, not independent measurements.
The 1 PFLOP figure is a conditional theoretical value for 4-bit operations using sparsity. NVIDIA has not announced measured performance for scientific computing in FP32 or FP64, nor the speedup of the four-node configuration. Deciding whether to adopt it for scientific computing requires measurements with your own code and precision.
Faster research iteration, less waiting on clusters
In the division of labor NVIDIA proposes, researchers iterate on development and debugging locally and reserve shared-cluster time for large-scale production runs. The company also cites as a use case the ability to keep unpublished interim results and proprietary data within the organization.
Waiting time on a cluster determines how long it takes to get results. In a case NVIDIA described from UC Berkeley's Rubin Lab, researchers used a single DGX Spark and the genomic language model Evo 2 to predict bacterial traits from genome sequences. Comparing genome and protein language-model embeddings with conventional methods, they reached results 15 times faster under conditions that avoided an 11-hour cluster wait, according to NVIDIA. That multiple includes the waiting time and does not represent a difference in the speed of the computation itself.
For the same university's Hartwig Group, NVIDIA describes a case in which a prediction head was trained on experimental chemistry data and Nemotron 3 Super 120B-A12B was fine-tuned with LoRA. LoRA is a technique that trains small added matrices instead of updating the whole model. The group reportedly used NVFP4 to fit a 120B-parameter-class model on a single DGX Spark and ran training continuously for 8 to 24 hours, avoiding multi-day cluster waits.
Both examples are research cases using a single unit, not performance tests of the four-node kit. What the kit adds is the ability to bring development involving data distribution and communication across multiple nodes into this kind of local iteration.
NVIDIA also envisions a research workflow in which AI agents handle everything from preparing inputs to running jobs and summarizing results, while scientists verify the settings and conclusions. In an example using AI that approximates simulations to reduce wing drag, it showed information from 10 million cells being narrowed to the 50 to 500 most important dimensions. These are uses the company envisions, distinct from results demonstrated on the four-node kit.
Hardware on sale, setup guides coming soon
NVIDIA exhibited the kit at SC26 and said the hardware is on sale. It points buyers of DGX Spark and the cables to participating OEMs and the NVIDIA Marketplace. The HPC SDK guides for building and running the four-node setup, however, are scheduled for release soon.
On storage, the announcement blog says 20TB across the four units, while the product specifications say up to 4TB per unit, so the kit's actual capacity awaits confirmation. The materials obtained also do not settle the price, regional inventory, whether cables are included, or the release date of the guides.
For adoption decisions, the question is whether you can split your own code and data across the four units and test it, communication included. Once the guides are out, being able to reproduce that configuration and confirm the steps before moving to a large system should make it easier to devote allocated shared-cluster time to production runs.
