英伟达 launches DGX Spark HPC development kit: four-node cluster, 512GB memory, aimed at research and AI development
IT之家 November 17 news, 英伟达 today at the SC26 conference announced the launch of the DGX Spark HPC development kit, aimed at researchers, academics, AI developers and others. The kit consists of four 英伟达 DGX Spark units, powered by the GB10 Grace Blackwell superchip.

IT之家 has learned that the DGX Spark HPC development kit has 4 nodes, with each node directly interconnected via 英伟达 ConnectX-7 network adapters, with no need to purchase an additional network switch. The product can provide 512GB of memory and 20TB of NVMe storage, allowing researchers to distribute models and scientific computing tasks across the entire cluster, and also to deploy AI models with up to 700 billion parameters, such as DeepSeek V4.1, locally.
With this device, researchers with limited server quotas no longer need to queue up for each task, but can complete development and debugging themselves, leaving server resources for formal production tasks.
At the same time, 英伟达 will soon release the NVIDIA HPC SDK installation guide, providing GPU-accelerated C, C++ and Fortran compilers, libraries and Message Passing Interface (MPI) tools for the DGX Spark cluster. Teams can build and test scientific applications across the four nodes, check the distribution of computing tasks and data, and troubleshoot communication issues before scaling to larger systems.
The device is currently on display at the SC26 conference. Users can purchase DGX Spark devices and accompanying cables through 英伟达 OEM partners (Acer, ASUS, Dell, Gigabyte, H3C, HP, Lenovo, MSI and xFusion), or buy the reference machine directly.
Related articles
- 英伟达 releases Rubin HPC GPU: over 50,000 CUDA cores per card sets a record, expected to launch in 2028
- 英伟达 DGX Spark supports 8-unit clusters: 128GB version unified memory up to 1TB, RTX PRO 6000 Blackwell-related tasks speed up by 40%
- China's strongest DeepSeek V4.1 Flash-1017 tops the Agent Arena open-source leaderboard, No. 11 overall
- DeepSeek V4.1 Flash-1017 new version released: improved performance, shorter runtime, wish-based programming made real with one click
- OpenCode announces DeepSeek V4.1 Flash is provided free of charge, but only supports 200K context
- A new choice for desktop AI supercomputing: 英伟达 NVIDIA DGX Spark 64GB memory version officially released, $4999