Kedios B300 Server Farm — Planning Summary

Date: March 30, 2026

Scope: 32-node deployed B300 compute farm with 72-node-standard-aligned network baseline

Purpose: High-level planning and review summary for the current source-of-truth markdown set

---

1. Executive Position

The Kedios package remains a 32-node NVIDIA B300 training farm. The change in this revision is not a move to 72 deployed servers. Instead, the network, management, border, and BOM basis now aligns to the 72-node B300 standard reference pattern, while the live compute population remains 32 nodes / 256 GPUs.

This summary therefore follows a two-layer rule:

1. current deployment = 32 compute nodes, 256 GPUs
2. standard-aligned network baseline = retained Spectrum-4 plus SN5610, SN4700, SN2201, and dual UFM

---

2. Locked Planning Figures

---

3. Power and Rack Framing

Validated Compute-Side Anchors

Allocation Rule

Use these two facts together:

• 1.5 MW = facility-allocation envelope
• 680 kW = current rack-constrained hard ceiling

Because the refreshed network baseline adds SN5610, SN4700, SN2201, and a second UFM node, this planning summary does not publish a new all-in network watt total until the refreshed procurement BOM carries the missing per-model power entries.

---

4. Network Architecture Summary

Live Compute Fabric

Non-Compute Network Layers

Management Rule

The old single 96-port OOB switch story is withdrawn. The base management model is now the SN2201-based management layer.

The server-side port mapping remains:

• X710 Port 0 = OS management
• X710 Port 1 = BMC / IPMI

---

5. Split Baseline vs Current Deployment

The practical meaning is simple:

• the compute population differs between baseline and deployment
• the network-side switch counts do not in the current working design

---

6. Placement and Cabling Rules

Locked Server-Side Cable Counts

Placement Rule

The old assumption that only N1 and N2 are occupied is no longer authoritative. Treat N1–N6 as the logical network/services placement envelope until the refreshed draw.io diagrams lock the final physical placement of:

• Spectrum-4
• SN5610
• SN4700
• SN2201
• dual UFM nodes

---

7. Downstream Reporting Rules

Every downstream summary, questionnaire answer, or viewer card must preserve these distinctions:

1. Never imply that 72 compute nodes are deployed today.
2. Present 32 deployed nodes and 72-node standard baseline as separate labeled facts.
3. Do not reuse the old 1 MW, 1 UFM, 1 generic OOB switch, or N3–N6 empty wording.
4. Do not publish a refreshed all-network watt total until the final BOM provides the missing switch power entries.