Question

A research organization is designing an on-premises AI cluster that must start with 40 GPUs but scale to over 500 GPUs within two years without redesigning the network fabric or storage architecture each time capacity is added. They want a validated, modular building-block approach that NVIDIA supports for predictable expansion. Which infrastructure design principle best fits this requirement?

Accepted Answer

Adopt a DGX SuperPOD reference architecture built from repeatable scalable units (SUs), so compute, networking, and storage expand as pre-validated increments

Answer

Deploy DGX systems as independent islands, each with its own isolated switch, and interconnect them ad hoc with Ethernet as new nodes are purchased

Answer

Buy the largest single HGX chassis available today and vertically scale by upgrading GPUs inside that one chassis over time

Answer

Start with a workstation-class DGX Station and later migrate the workload to a public cloud region when GPU demand grows