

NVIDIA announced the DGX Spark HPC Developer Kit on November 16 at the SC26 supercomputing conference in Chicago, USA, allowing researchers and developers to form a computing cluster with 4 desktop AI computers, the DGX Spark. According to NVIDIA's official announcement, the four machines provide a total of 512GB of memory and 20TB of NVMe storage space, and with the NVFP4 numerical format, can deploy AI models with up to 700 billion parameters (700B) locally.
4 DGX Sparklines Connected in a Ring, No Additional Network Switch Required
The DGX Spark HPC Developer Kit uses 4 DGX Sparks, each equipped with the NVIDIA GB10 Grace Blackwell Superchip and 128GB of unified memory. Existing product specifications from NVIDIA indicate that a single machine can be used for inference and testing models with up to 200 billion parameters; this new kit allows larger models to be distributed across multiple compute nodes for execution.
The 4 computers are connected via NVIDIA ConnectX-7 networking, each directly linked with two adjacent machines in the ring architecture, eliminating the need to purchase an additional network switch. Research teams can also gradually expand from a single machine to a 4-machine setup as needed. It is important to note that the 512GB refers to the total memory across the 4 machines and not that a single computer has 512GB of memory.
NVIDIA announced that the entire configuration includes 20TB of NVMe storage, but the current specifications listed for a single DGX Spark machine show a capacity of 4TB. The new announcement did not clarify how the kit is configured to total 20TB, and it is advisable to verify each machine’s storage specifications at the time of purchase.
NVFP4 Supports 700B Models, Officially No Inference Speed Released
NVIDIA stated that with the four-machine configuration using the NVFP4 low-precision numerical format, AI models up to 700 billion parameters, like DeepSeek V4.1, can be deployed locally. NVFP4 can reduce the memory requirement for model weights; however, running large models still requires distributing the workload across different nodes and handling data transfer between machines.
The official announcement details the parameter scale of the deployed models but does not specify the tokens per second output, inference latency, or cross-node performance testing for the 700B model. This specification does not equate to the entire system being able to fully train the 700B model under the same conditions. For developers looking to establish local AI systems in their offices or labs, actual speeds will still need to be tested based on the models and workloads used.
Scientific Simulations and AI Agents Can Be Tested on the Desktop
NVIDIA has demonstrated the four-machine configuration running materials simulation software LAMMPS (with Kokkos) and molecular dynamics software NAMD. The company also mentioned that the DGX Spark development community has validated applications like GROMACS, OpenFOAM, and the weather simulation tool WRF, which can assist research teams in testing programs and computational workflows before using large shared supercomputers.
NVIDIA plans to provide HPC SDK installation guides later, covering GPU-accelerated C, C++, Fortran compilers, libraries, and MPI messaging tools. Researchers will be able to troubleshoot data allocation and communication issues for multi-node programs in a desktop environment before transferring work to large computing systems. The company also proposed using AI Agents to assist in preparing simulation input data, initiating jobs, and organizing results, with research conclusions still needing to be verified by personnel.
Purchase Channels Provided, Kit Price Not Yet Announced
The DGX Spark HPC Developer Kit is currently being showcased from November 15 to 20 at SC26. NVIDIA stated that the required DGX Spark computers and connection cables are now available for purchase through NVIDIA Marketplace and partners such as Acer, ASUS, Dell Technologies, GIGABYTE, H3C, HP, Lenovo, MSI, and xFusion.
NVIDIA has not yet announced a unified price for the four-machine kit, sales arrangements in Hong Kong, or the exact launch date for the HPC SDK installation guide. Users looking to build their own cluster need to confirm the number of hosts, cabling, storage configuration, and software deployment conditions with suppliers.
This article is reproduced from unwire.hk, with copyright belonging to the original author and original media.
免责声明:本文章仅代表作者个人观点,不代表本平台的立场和观点。本文章仅供信息分享,不构成对任何人的任何投资建议。用户与作者之间的任何争议,与本平台无关。如网页中刊载的文章或图片涉及侵权,请提供相关的权利证明和身份证明发送邮件到[email protected],本平台相关工作人员将会进行核查。
