Octopod · Information package
2 of 2 · inside the IT hall
IT Racks — the example row inside the hall
An example of how a hall gets populated: how many racks, what type each one is — NVIDIA GPU compute, networking, storage, CPU-based management — and how they work together. Forty 52U racks in two rows of twenty, as laid out in the example rack design workbook: seven scalable units of four NVIDIA GPU racks and one leaf rack each, plus the spine, storage, management and north/south racks. The servers here are a 4U eight-GPU class, not the NVIDIA VR200 NVL72 racks SAVRN specifies at 250 kW, so the per-rack draw is lower — the population pattern is what carries over. Click any rack or any piece of equipment.
EXAMPLE ROW
40racks · 2 rows × 20
7scalable units
28NVIDIA GPU racks
224NVIDIA GPU servers · 4U
1,792NVIDIA GPUs · 8 per server
58SN5600 switches
52Urack height
Section
ROW ASU 1SU 2SU 3SU 4A1A2A3A4A5A6A7A8A9A10A11A12A13A14A15A16A17A18A19A20ROW BSU 5SU 6SU 7Spine · Storage · Mgmt · North/SouthA21A22A23A24A25A26A27A28A29A30A31A32A33A34A35A36A37A38A39A40
NVIDIA GPU serversCoolant manifolds (CDM)In-rack CDUSN5600 fabric switchesSN2201 OOB switchesCatalyst 9500 internal / externalPA-5440 firewallsISP demarcsVAST EBox storageVxRail nodesPatch panelsBlanking panels
Drag to pan · pinch or use + / − to zoom · Fit resets. Sections follow the workbook tabs.

How a scalable unit is built

The row is seven copies of the same unit, plus one set of core racks. Read it as an example of the mix — how many racks of each type a hall carries, what sits in them, and how the compute, networking, storage and management racks work together. Rack power follows whichever server class is installed; this example is not the 250 kW NVIDIA VR200 NVL72 rack. Click any card and it lights up in the row view above.

Walkthrough Runs the sequence on the row view above, about four seconds a card.

NVIDIA GPU rack — ×4 per SU, 28 in the row

8 × NVIDIA GPU servers (4U, eight GPUs each on an NVIDIA HGX baseboard, direct-to-chip liquid cooled) at U10–U45, each with a 1U coolant manifold directly above it. A 4U in-rack CDU at U6–U9. Three patch panels at the top: MPO trunks to the north/south leaf and the east/west leaf. 64 GPUs per rack in this example; SAVRN's own spec is the NVIDIA VR200 NVL72 rack at 250 kW, with the latest NVIDIA GPUs available at delivery.

Show it in the row view ↑

Leaf rack — ×1 per SU

4 × SN5600 east/west leaf switches, each between a pair of patch panels (64 MPO to trunk — spine, 64 MPO to trunk — GPUs). Below them 4 × SN2201 48-port 1G switches for out-of-band management. One leaf rack (A33 in the row view, “SU#1” on the leaf tab) also carries the network infrastructure: 2 × Catalyst 9500-32QC internal switches, 2 × PA-5440 firewalls, 2 × Catalyst 9500-24Y4C external switches and the ISP A / ISP B demarcs.

Show it in the row view ↑

Spine racks — ×2

Data Hall 1 east/west spine: 8 × SN5600 per rack in four pairs, each pair under a 3U patch panel.

Show it in the row view ↑

North/South rack — ×1

12 × SN5600 in six pairs, each pair under a 3U patch panel — the north/south tier the GPU racks trunk to.

Show it in the row view ↑

Storage rack — ×1

24 × VAST Data EBox nodes (U3–U26) with 2 × SN5600 above them and a 3U patch panel.

Show it in the row view ↑

Management rack — ×1

5 × Dell VxRail VP-760 nodes (VxRail management cluster), patch panel to north/south; vSAN switches marked TBD on the sheet.

Show it in the row view ↑

Equipment schedule — the whole row

Counted from the elevations. Blanking panels excluded.

ModelRole (per sheet)HeightWhereQty in row
NVIDIA GPU serverNVIDIA GPU server, 4U4U8 per NVIDIA GPU rack (28 racks)224
CDMCoolant distribution manifold1U8 per NVIDIA GPU rack224
In-rack CDUCoolant distribution unit, 4U4U1 per NVIDIA GPU rack28
SN5600Leaf / spine switch2U (1U in storage rack)4 per leaf rack · 8 per spine rack · 12 in North/South · 2 in storage58
SN220148-port 1G RJ45 switch (OOB)1U4 per leaf rack28
C9500-32QCInternal switch1U2 in the net-infra leaf rack2
C9500-24Y4CExternal switch1U2 in the net-infra leaf rack2
PA 5440Firewall2U2 in the net-infra leaf rack2
ISP AISP demarc A1Unet-infra leaf rack1
ISP BISP demarc B1Unet-infra leaf rack1
Vast EboxStorage node1U24 in the storage rack24
Dell VP-760VxRail management node1U5 in the mgmt rack5
Patch PanelPatch panel1U / 3Uevery rack165

Where this sits in the block

The general arrangement draws each IT building as 20 rack positions, and one IT hall as two of them — the same two-row, forty-rack shape as this example. The drawing sets the power envelope rather than the rack count: 2,200 kW per building, 8.8 MW across the block's 80 positions, an average of 110 kW a rack, with up to 250 kW available at any one position. So density follows the envelope and the server class: this example's 4U eight-GPU servers draw well under 250 kW a rack, while SAVRN's own spec — the NVIDIA VR200 NVL72 rack at 250 kW — puts 32 racks and 8 MW of IT inside the same envelope. The workbook is a layout: it fixes what goes in each rack, not what each rack draws. Open Infrastructure →

Download the workbook (.xlsx)Source: “Example Rack Design” workbook — tabs Row View, Leaf Racks, Spine Racks, North South, GPU, Mgmt Rack. Equipment identities are the sheet's part numbers; heights are as drawn.