NVIDIA Introduces DGX Spark HPC Developer Kit for Local Multi-Node Scientific Computing

·2026.11.17 08:00

Key point

The four-node cluster offers 512GB of combined memory and 20TB of NVMe storage, allowing researchers to run multi-node workloads without shared cluster queues.

1 / 2

Details

  • NVIDIA introduces the DGX Spark HPC Developer Kit at SC26, a four-node cluster of DGX Spark systems connected via NVIDIA ConnectX-7 networking without a central switch.
  • The kit provides 512GB of combined memory and 20TB of NVMe storage, supporting local deployment of models with up to 700B parameters using NVFP4, such as DeepSeek V4.1.
  • An NVIDIA HPC SDK playbook is coming soon, providing GPU-accelerated C, C++, and Fortran compilers and MPI tools for multi-node development and debugging.
  • Demonstrated workloads include LAMMPS with Kokkos, NAMD, GROMACS, OpenFOAM, and WRF, alongside NVIDIA NIM microservices like OpenFold3 and Boltz-2.
  • Use cases include agent-assisted simulation workflows, training AI models to approximate slow simulations, and quantum-chemistry circuit simulations with up to 70 qubits and 975 gates using NVIDIA cuQuantum.
  • The Rubin Lab at UC Berkeley used the DGX Spark and NVIDIA Evo 2 model to predict bacterial traits, reporting that results arrived 15x sooner by avoiding 11-hour cluster queue times. The Hartwig Group fit a 120B-class Nemotron 3 Super 120B-A12B model on a single device using NVFP4 quantization.
  • The kit is on display at SC26 and available for sale through OEM partners including Acer, ASUS, Dell Technologies, GIGABYTE, H3C, HP, Lenovo, MSI, and xFusion.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.