Image: NVIDIA.
NVIDIA is expanding DGX Spark for locally hosted AI. Its October 21 announcement set October 23 as the start of the partner rollout of 64GB systems.
The company also announced support for manually configured clusters of up to eight devices. A Model Launcher intended to simplify downloading and running models is planned for later in October.
Small teams could use this approach for some experiments and internal requests on their own hardware. Local deployment does not mean that every large model will fit into memory or deliver acceptable speed.
The economics depend on the workload: equipment, electricity, maintenance and utilisation all matter. Comparing a cloud request fee with the absence of that particular charge on a local machine would leave out substantial parts of the cost.
