Skip to main content
A supercomputing lab on any desk: Nvidia unveils DGX Spark HPC Dev Kit with 512 GB of memory - Notebookcheck

A supercomputing lab on any desk: Nvidia unveils DGX Spark HPC Dev Kit with 512 GB of memory - Notebookcheck

NVIDIA GPUs
17 lis 2026, 09:07Hardware News
Oryginał
3 min read

Nvidia has introduced the DGX Spark HPC Developer Kit, a four-node cluster that turns four compact AI systems into a desktop supercomputer with 512 GB of unified memory.

The DGX Spark HPC Developer Kit is designed to bring supercomputing capabilities to the desktops of researchers, academics, and developers. It combines four DGX Spark systems into a single cluster with 512 GB of unified memory and 20 TB of NVMe storage. This setup allows users to run scientific simulations, AI models, and agent-based workloads locally without relying on cloud servers.

Announced at the SC26 supercomputing conference, the developer kit connects four DGX Spark systems in a ring topology using Nvidia ConnectX-7 networking. Each node connects directly to its two neighbors, eliminating the need for a separate network switch. This design simplifies the setup and improves communication between nodes.

Each DGX Spark is powered by the Nvidia GB10 Grace Blackwell Superchip. This chip combines a Blackwell GPU with a 20-core Arm CPU. The CPU includes 10 Cortex-X925 performance cores running at up to 4 GHz and 10 Cortex-A725 efficiency cores at up to 2.8 GHz. The Blackwell GPU features 6,144 CUDA cores and supports up to 1 PFLOP of FP4 AI performance.

The system uses 128 GB of LPDDR5X coherent unified memory with a 256-bit interface and 273 GB/s of memory bandwidth. The GB10 has a thermal design power (TDP) of 140 W. Nvidia says the four-node configuration can run models with up to 700 billion parameters using NVFP4 precision. The combined memory and storage capacity allow researchers to distribute larger models and scientific workloads across the cluster.

Software and applications

Nvidia is preparing an installation playbook for its HPC SDK, which will bring GPU-accelerated C, C++, and Fortran compilers, libraries, and Message Passing Interface tools to the DGX Spark cluster. These are the same development tools used on Nvidia-powered supercomputers. The software will allow researchers to build and test scientific applications across all four nodes, examine how workloads and data are distributed, and troubleshoot communication issues before moving to larger systems.

The kit has already been demonstrated running LAMMPS with Kokkos for materials simulation and NAMD for molecular dynamics. Nvidia also notes that the broader DGX Spark community has validated several other applications, including GROMACS for molecular dynamics, OpenFOAM for fluid dynamics, and WRF for weather modeling.

Real-world impact

Nvidia highlights several research projects already using DGX Spark. The Rubin Lab at the University of California, Berkeley, uses the system to analyze genomic data and predict bacterial traits. Nvidia says the team received results 15 times sooner than with its previous workflow, which involved waiting in an 11-hour cluster queue. Meanwhile, researchers at UC Berkeley's Hartwig Group used a single DGX Spark system with NVFP4 quantization to fine-tune a 120-billion-parameter-class model for chemistry research, avoiding multi-day cluster queues and enabling continuous training runs.

Availability and partners

The DGX Spark HPC Developer Kit is being demonstrated at SC26 and is already available for purchase. Nvidia lists Acer, Asus, Dell, Gigabyte, H3C, HP, Lenovo, MSI, and xFusion among its hardware partners. Individual DGX Spark systems and cables are also available through the Nvidia Marketplace. The company has not disclosed a price for the four-node developer kit.

Z.ai
Article written by AI
Z.ai GLM-4.7-FlashFree
Free AI articles150 of 150
No limit after signing in with Telegram