Skip to content

Models & products

Nvidia links four DGX Sparks in a 512 GB developer kit

Edited and reviewed by AI Ping Editorial Team

3-line summary

  1. Nvidia unveiled its research kit with four DGX Spark computers and four direct-connect cables at SC26 in Chicago.
  2. The switch-free ring links 512 GB of total memory. Nvidia says it can run models of up to 700 billion parameters with NVFP4.
  3. Nvidia’s U.S. listing directs buyers to sales for a price. The launch announcement lists 20 TB of storage, while the per-unit specs on the sales page add up to 16 TB.
PracticalDevelopers
Four imagined compact computers in two rows on a wooden desk
AI-generated conceptual image: Four imagined compact computers show the scale of a DGX Spark kit that fits on a desk. · AI Ping

Nvidia announced a kit that links four DGX Spark computers at the SC26 supercomputing conference in Chicago on November 16. In the DGX Spark HPC Developer Kit, each machine connects directly to its two neighbors. The four connections form a loop, eliminating the need for a separate network switch in that configuration.

The machines give a research team its own hardware for testing AI models and scientific code that have to run across multiple computers. Together the four Sparks have 512 GB of memory. Nvidia says the cluster can run models of up to 700 billion parameters locally when their weights use NVFP4, a four-bit format that reduces the memory needed to store them.

Four computers, four cables

Nvidia's U.S. Marketplace lists four computers and four connecting cables for the kit. Each computer has a GB10 chip, which pairs a Grace CPU with a Blackwell GPU, as well as 128 GB of unified memory and a ConnectX-7 network adapter. The 512 GB is the total across the four machines. Software has to distribute a model across the machines to use memory beyond the capacity of one Spark.

People could already connect multiple Sparks. Nvidia documented a four-machine ring without a switch in an October 22 technical article. The new bundle brings the machines and cables for that layout together as a product for researchers. Nvidia's U.S. listing also describes a switch-based option with a separately sold switch. In its October testing notes, Nvidia used a ring to distribute one large model across the machines and pointed to a switch for splitting a cluster into groups or for jobs that need direct connections between any pair.

Individual Sparks come in 64 GB and 128 GB versions. Nvidia lists the 64 GB version through hardware partners and shows a 256 GB total for a cluster of four of those units. The developer kit's 512 GB comes from four 128 GB machines.

Nvidia gives two different storage specifications. Its November 16 announcement lists 20 TB of NVMe storage for the four machines. The U.S. Marketplace entry for the kit lists 4 TB in each Spark, which would add up to 16 TB. The documents give no explanation for the difference or details of any additional storage.

A diagram showing four machines and four cables in a switch-free ring
Illustration: Each machine in the four-node ring connects to two neighbors. · AI Ping

What happened when testers added machines

Nvidia's October article gives one measure of how a four-node ring can perform. Using the NVFP4 version of a 27-billion-parameter Qwen3.8 model with one request at a time, Nvidia measured 32.82 output tokens per second on the ring, up from 11.86 on one Spark. Input processing was about 1.69 times as fast in the same test. The figures describe that model and workload. Nvidia's 700-billion-parameter figure concerns model size.

A separate preprint dated November 15 reports tests conducted at the University of Delaware with one or two Sparks. The researchers evaluated 17 sets of public model weights and precision settings, including versions of the same model stored at different precisions. Across 13 sets with complete comparisons, splitting a model's calculations across two Sparks yielded a median peak output throughput 1.59 times the single-machine result. The study tested configurations of up to two machines.

For models that fit on one Spark, the authors also estimated what would happen if each machine ran its own copy of the model. With 64 requests in progress and an even split between the replicas, that arrangement had higher estimated throughput on 11 of the 13 comparisons than splitting one model across both computers. The estimate uses single-machine measurements and assumes that routing requests adds no overhead.

Two imagined compact computers linked by one short cable on a green workbench
AI-generated conceptual image: Two imagined computers and a cable depict the two-machine setup examined in the University of Delaware study. · AI Ping

Nvidia says it is displaying the kit at SC26 and has made it available for sale. The U.S. Marketplace lists a price for an individual Spark. For the four-machine kit it provides a sales contact instead of a public price, so the cost of the complete kit is still unspecified there.

The launch announcement describes demonstrations on four machines, including a materials simulation with LAMMPS. A separate guide for using Nvidia's high-performance computing toolkit to build and test scientific software across the cluster is still in preparation. On November 17, Nvidia's developer site listed it as coming soon.

Sources