NVIDIA announced the DGX Spark Developer Kit, a four-node cluster for researchers, academics, and developers to run multi-node scientific workloads and AI models on their desks. The kit scales from one to four nodes, providing 512GB of combined memory and 16TB of NVMe storage.
What happened
At SC26, NVIDIA introduced the DGX Spark Developer Kit, a four-node cluster designed for researchers, academics, and developers to run multi-node scientific workloads and AI models directly from their desks. The kit scales from one to four nodes, connected directly over NVIDIA ConnectX-7 networking without the need for a network switch. Each node is equipped with the GB10 Grace Blackwell Superchip, offering 512GB of combined memory and 16TB of NVMe storage. This setup enables researchers to run large-scale simulations and AI models locally, bypassing the need for shared GPU clusters and reducing wait times for server hours.
What to weigh
- The kit scales from one to four nodes without a network switch.
- The kit supports up to 700B parameters for models like DeepSeek V4.1.
FAQ
What is the maximum number of nodes?
The kit scales from one system, or node, to four — linked directly to one another over NVIDIA ConnectX-7 networking, with no network switch to buy or set up.
What kind of storage does it provide?
At four nodes, it provides 512GB of combined memory and 16TB of NVMe storage that provides researchers more capacity to distribute models and scientific workloads across the cluster.
What models can be run on the kit?
With that much power at the desk, researchers with a fixed compute and server allocation no longer have to wait in a queue for every job. Teams can develop and debug on their own, saving server hours for production runs and keeping unpublished results, and proprietary datasets on premises.
Source: NVIDIA


