NVIDIA DGX SuperPOD

That depends on the workload rather than a catalogue minimum, and it is worth being honest that many teams asking about SuperPOD are better served by a handful of nodes with a properly designed fabric. We will size it against what you actually intend to run, and say so if the architecture is heavier than your requirement.

Because it is the commonest way one of these projects goes wrong. The fabric determines how the nodes are laid out in the rack, what optics and cable lengths you need, and how the cluster scales; retrofitting it means re-cabling and frequently re-racking. On a multi-node training build the fabric is a substantial share of the cost, and leaving it out of the first budget is what causes the overrun.

As a configured system rather than a boxed part. We confirm the exact platform revision, the networking and storage around it, rack power and cooling requirements, and the support term, then issue an itemised quote. Current availability and lead time are stated on that quote — accelerator allocation moves, and any figure published on a web page would be wrong for most buyers by the time they read it.

The GPUs and the NVLink fabric are the same silicon either way. DGX gives you one vendor accountable for the whole stack including the software environment, and is the designed unit if you are heading toward SuperPOD. An HGX system from a server vendor gives you your existing management tooling, support relationship and configuration choices. If your operations are already standardised on Dell or HPE, that consistency is frequently worth more than it looks on a spreadsheet. We supply both and will tell you which fits.