A practical integration ladder
GPU
The processor or an accelerator product built around it. Exact form factor and memory still matter.
HGX
A multi-GPU server platform used by system builders; it is not a complete end-user deployment by itself.
DGX
An NVIDIA-branded integrated system with a defined hardware and software configuration.
NVL
A tightly integrated multi-system configuration whose claim must name the exact generation and topology.
Four common comparison errors
- Using total rack memory beside per-GPU memory without labeling units.
- Comparing an announced rack configuration with a currently shipping accelerator.
- Treating an architecture family as a single product specification.
- Ignoring networking, power, cooling, software, and provider constraints.
NVLink support is an exact-product question
NVLink appears in NVIDIA’s data-center system designs, but the label does not mean that every GeForce or professional board has an NVLink connector or supports a bridge. NVIDIA’s GeForce RTX 4090 and RTX 5090 specifications both say No for NVLink (SLI-Ready). For another GPU, use its own official specification and the intended system topology; do not carry an answer across generations.
For cluster networking rather than card-to-card support, use the platform network map to distinguish NVLink, Spectrum-X Ethernet, and Quantum InfiniBand.
[99][100][96]A claim template that survives scrutiny
As of [observation date], [source owner] describes [exact product or system] at the [component / module / server / rack / cloud] level as [announced / sampling / shipping / generally available] for [named market or provider].
If any bracket is unknown, keep it unknown. Do not fill the gap with a broader marketing label.
Rubin examples preserve the system boundary
NVIDIA’s Tile IR table names HGX Rubin NVL8 and Vera Rubin NVL72 as Rubin targets. These system labels do not identify an individual retail graphics card. Its preview qualification applies to that software path; consult system documentation separately for configuration and availability.
[118][114]