Skip to content
nodeza.learn

Reference architectures and positioning: Dell AI Factory with NVIDIA

The design layer above the NIC. Read the published NVIDIA reference architectures — DGX SuperPOD, Enterprise RA (HGX AI Factory, NVL72 AI Factory), and the per-platform validated stacks — and know which one governs a given Dell deal. Decode node nomenclature (2-8-9-400), separate NVLink scale-up from the RDMA fabric, and size a cluster end to end: GPUs to SuperNICs to planes to leaf/spine ports to optics to BOM, with every ratio cited to the page it came from. Then map the RA onto what Dell actually sells (PowerEdge XE9680/XE9780/XE9785, PowerSwitch SN-series and Q-series, PowerScale F710), learn the three real support paths, and run the Ethernet-vs-InfiniBand positioning conversation on published claims only. Ends with timed FAE scenarios and an NCP-AIN gap plan. Interactive sizing tools, two-variant labs (containerlab/NVIDIA Air/Cumulus VX with no NVIDIA switch, or the Dell-lab BlueField-3 and ConnectX hosts).

4 modules. Dots: grey not started, blue in progress, green complete with quiz ≥ 80%, amber complete but below 80%. Each module ends in a checkpoint that unlocks the next.

Reference architectures and positioning: Dell AI Factory with NVIDIA

0%

0 of 20 cases closed

RA landscape

Sizing

Dell AI Factory

Positioning

The reference-architecture landscape

Pick the governing reference architecture for a deal and speak its vocabulary: scalable units, node nomenclature, and the line between NVLink and the fabric.

  1. 1.1 Which reference architecture governs this dealTwo leaf counts for the same server
  2. 1.2 Reading node nomenclature: 2-8-9-400Nine adapters and a cable nobody ordered
  3. 1.3 Scale-up is not scale-out: NVLink vs the fabricThe commodity fabric
  4. 1.4 The networks in an AI factory, and what the RA refuses to specifyThe chapter that is not missing
  5. 1.5 The generation map: which SuperNIC is on the baseboardBlueField-3 is a family, not a speed
  6. Checkpoint

Sizing a fabric: from GPU count to BOM

Turn a GPU count into leaf, spine, optics and BOM arithmetic you can do out loud in front of a customer and defend line by line.

  1. 2.1 Rail-optimized design and what breaks itThe pod that passed every test
  2. 2.2 Dual plane: two fabrics, not a bondThe bond with no partner
  3. 2.3 The sizing chain: GPUs, NICs, planes, leaves, spinesThe quote that fit in half a chassis
  4. 2.4 The other three networks: converged, storage, out-of-bandThe call at 02:10
  5. 2.5 Optics, cables and the power nobody budgetedWhere are the optics?
  6. Checkpoint

Dell AI Factory with NVIDIA: what Dell actually sells

Map the NVIDIA reference architecture onto Dell's real catalog, a real solution BOM, and the three real support paths.

  1. 3.1 The PowerEdge AI server lineNine adapters, four slots
  2. 3.2 Dell's switch catalog: SN, Q and Z in one price listThe pallet that was four times too big
  3. 3.3 Reading a real Dell AI Factory BOMAn architecture with the reasoning deleted
  4. 3.4 PowerScale back-end networking and the version-pin trapThe two switches procurement wanted back
  5. 3.5 Who do I call: three support paths, one asset tagDell says NVIDIA, NVIDIA says Dell
  6. Checkpoint

Positioning, competition, and the FAE bench

Win the Ethernet-versus-InfiniBand conversation on published facts, stay neutral on competitors, and close the NCP-AIN gaps.

  1. 4.1 The Ethernet vs InfiniBand decision frameworkTwo quotes, one server
  2. 4.2 Seven objections and the published answersSix months of PFC storms
  3. 4.3 Competitors, stated neutrallyThe sentence in yellow
  4. 4.4 Scenario: a 256-GPU Dell AI Factory design reviewFriday's quote, Tuesday's review
  5. 4.5 NCP-AIN: mapping the blueprint and closing the gapsWho answers at 02:10
  6. Checkpoint