NVIDIA DGX Spark Developer Kit Turns a Desk Into a Supercomputing Lab
Reported by NVIDIA, Unite.AI, Mezha and 2 more outlets5 sources ↓
NVIDIA introduced a four-node DGX Spark developer kit at SC26 in Chicago for researchers, academics and developers. The company says the kit is available through OEM partners and the NVIDIA Marketplace.
Key points
- The four-node cluster has 512GB of combined memory and 20TB of NVMe storage
- NVIDIA says NVFP4 supports local deployment of models with up to 700 billion parameters across the four systems
- The cluster uses ConnectX-7 links in a ring and does not require a network switch
- The kit is available through NVIDIA Marketplace and OEM partners
- The NVIDIA HPC SDK installation playbook is forthcoming
Sources
- NVIDIARead ↗NVIDIA DGX Spark Developer Kit Turns a Desk Into a Supercomputing Lab
- Unite.AIRead ↗NVIDIA Introduces DGX Spark HPC Developer Kit for Scientific Computing
- MezhaRead ↗NVIDIA Unveils Four-Node DGX Spark Developer Kit at SC26
- GIGAZINERead ↗NVIDIA has announced the 'DGX Spark Developer Kit,' which connects four DGX Spark mini AI supercomputers to achieve a memory capacity of 512GB, enabling the execution of models up to 700B.
- Blockchain NewsRead ↗NVIDIA Unveils DGX Spark HPC Developer Kit at SC26
Details
The four-node configuration links DGX Spark systems directly to their neighbors in a ring using ConnectX-7 networking. NVIDIA says this setup avoids the need to buy or configure a network switch; the kit can also scale from one node to four.
NVIDIA says the systems use GB10 Grace Blackwell Superchips and provide 512GB of combined memory and 20TB of NVMe storage. The company says NVFP4 quantization supports local deployment of models with up to 700 billion parameters across the four systems, including DeepSeek V4.1.
The NVIDIA HPC SDK installation playbook is not yet available. NVIDIA says it will provide GPU-accelerated C, C++ and Fortran development tools and message-passing interface tools for building and testing multi-node scientific applications.
NVIDIA says demonstrated workloads include LAMMPS with Kokkos for materials simulation and NAMD for molecular dynamics. The company also reports that UC Berkeley’s Rubin Lab produced bacterial-trait prediction results 15 times sooner than with its traditional cluster-based method; its Hartwig Group used the kit for local chemistry-model fine-tuning.
Written from coverage as of
Images
- NVIDIA
- Unite.AI
- Blockchain News
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh