- Four nodes interconnect directly via ConnectX-7, eliminating a separate switch purchase
- 512GB memory plus 20TB NVMe supports local models up to 700B parameters
- Bundled HPC SDK offers GPU-accelerated C/C++/Fortran compilers and MPI tools
AI CURATED HOT STORY · 11/17/2026
NVIDIA Unveils Four-Node Desktop Supercomputer Kit with 512GB Memory
NVIDIA has launched the DGX Spark HPC Developer Kit at SC26, comprising four DGX Spark units connected in a ring via ConnectX-7 without an external switch, delivering 512GB of aggregate memory and 20TB of NVMe storage. The kit enables researchers to run multi-node scientific computing, AI models, and agent workloads right on their desktop, reducing reliance on shared GPU cluster queues.
Technically, the four nodes are directly interconnected through ConnectX-7, eliminating the need for a separate network switch. The aggregated 512GB memory and 20TB NVMe storage support distributed deployment of large models, with the ability to run models up to 700 billion parameters locally. The accompanying HPC SDK provides GPU-accelerated C/C++/Fortran compilers and MPI tools, targeting multi-node GPU application development, FP32 mixed-precision algorithms, and LAMMPS molecular dynamics simulations.
The kit lowers the barrier to multi-node HPC development, allowing research teams and AI developers to perform distributed training and simulation tasks locally that previously required large clusters. It is poised to accelerate innovation in scientific computing and agent applications, complementing traditional shared GPU cluster usage models.
Hourly heat curve
5 fully observed hoursThe line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.
Coverage timeline
Total 2 reports · Latest first- IT之家 · 智能时代T2·62 pts
- NVIDIA BlogT1·62 pts
- Four nodes ring-connected over ConnectX-7, eliminating switch purchase and setup
- 512GB combined memory and 20TB NVMe support distributed model and scientific workloads
- Targets multi-node GPU app development, FP32 mixed-precision algorithms and LAMMPS molecular dynamics