Placeholder content

This package content is written, not the real redacted document. This is written to demonstrate the flow end to end. It is not sourced from NVIDIA's published corpus and must be replaced before anyone outside the company sees it.

The reference package

A 20,736 GPU AI factory, compiled and traceable.

GB300 NVL72 on Quantum-3 InfiniBand with WEKA storage. Redacted, and otherwise exactly what lands on day seven.

5,184compute trays
288racks
36scaling units
18deployment zones

01

Executive summary

This package describes a 20,736 GPU AI factory built on the GB300 NVL72 platform with a Quantum-3 InfiniBand compute fabric and WEKA storage. It was compiled deterministically. Every quantity below traces to the reference architecture rule that produced it, and the trace is included in the evidence manifest.

The recommended architecture is a rail optimised fat tree across 36 scaling units, distributed over 18 deployment zones so that the loss of any one zone removes at most one eighteenth of capacity. Storage is served over a fabric separated from compute, which keeps a storage incident from becoming a training incident.

Release status: bounded baseline approved. Procurement and installation remain locked pending the customer's own gate review. This document is a solution architecture, not an engineering release, and the boundary between the two is stated explicitly in the handoff section.

02

Compute

Compute specification
PlatformGB300 NVL72
Total GPUs20,736
Compute trays5,184
Racks288
Scaling units36
Deployment zones18

03

Fabric

Fabric specification
Compute fabricQuantum-3 InfiniBand
TopologyRail optimised fat tree
Storage fabricSeparated from compute
Out of bandIsolated management network

04

The rest of the package

Everything above renders for anyone. The depth below is the exchange, because the full document is the thing we would rather discuss than post.

  • Full topology Architectural, systems and conceptual views, with copyable diagram source.
  • Bill of materials Decomposed to what you buy, every line traced to the rule that produced it.
  • Port and connectivity budget Port counts, media, and cable lengths reconciled against the topology.
  • Power and rack envelope Per zone and per rack, with the assumptions stated.
  • Requirements traceability matrix Where each requirement came from and what satisfies it, with criticality flags.
  • Architecture decision record The choices made, the alternatives, and why each was rejected.
  • Evidence manifest How the design was arrived at, for due diligence.