GB10 Cluster Guide DGX Spark and every GB10 workstation

Sanitization notes

This guide was written from Petronella Technology Group, Inc. internal build notes, session handoffs and benchmark logs. Before publication the following classes of detail were generalized. This file lists what was changed and how, so a reader knows which identifiers are placeholders; it deliberately does not contain any of the original values.

Addressing

  • All management-network addresses were replaced with the RFC 5737 documentation range 192.0.2.0/24 (switches at 192.0.2.11 and 192.0.2.12).
  • The RDMA fabric's two planes (one per PCIe half of the ConnectX-7 port) were replaced with 198.51.100.0/24 (plane A, half enp1s0f0np0) and 203.0.113.0/24 (plane B, half enP2p1s0f0np0). Host numbers follow the node number.
  • The point-to-point /30 links from the earlier direct-attach ring were re-expressed in the same two documentation ranges.
  • Overlay-network (mesh VPN) addresses that appeared in the source notes were removed; they play no part in the fabric.

Hostnames

  • The four MSI EdgeXpert MS-C931 units became gb10-01 through gb10-04.
  • The two NVIDIA DGX Spark Founders Edition units became gb10-05 and gb10-06.
  • The two switches became crs812-a (MSI pod) and crs812-b (DGX Spark pod).
  • Every other host on our network that appeared in the source notes (workstations, other GPU servers, storage, management VMs, the campus core switches) was removed or replaced with a role description such as "the campus network" or "the management switch". Their names, addresses, ports and VLANs are not needed to reproduce the fabric.

Switch identifiers

  • Switch serial numbers, the switch MAC OUI, and the RouterOS login are omitted.
  • References to the file that holds the switch credentials are omitted. The guide says "credentials" where the notes named a file.
  • Physical port numbers on our campus switches (which uplink, which patch) are omitted; the RSTP discussion refers to "the campus uplink" and "an RJ45 patch".
  • The campus network controller, the firewall, and their vendors, versions and addresses are omitted.

Container images and repositories

  • Locally built image tags (the names we gave the patched vLLM builds) are omitted, because they are not on any registry and would not resolve. The guide describes the patch chain generically and credits the public GitHub repositories by name.
  • One community-published Docker Hub image is referred to as "the image published by the recipe author; see the repo" rather than by account name.
  • Internal script names, launcher environment-variable interfaces, results-log paths and memory-file names were replaced with the commands and flags they wrap.

People, clients and business detail

  • Fleet-wide credentials, the fact that some hosts share credentials, and which do not, are omitted.
  • Names of individuals other than the firm's principal are omitted, including a colleague's launcher scripts and address expectations.
  • No client or engagement detail existed in the source material for this topic; none appears here. The firm's compliance work is described only in the "About us" paragraph of the README.
  • Purchase-decision discussion (which unit to buy next, inventory, margins, order velocity, supplier costs) is omitted except for the list prices of GB10 units observed in June 2026 and the breakout DAC price, both dated in the bill of materials.
  • Operational incidents unrelated to the fabric (voice systems, chat bridges, unrelated outages, download throttling on one site's uplink) are omitted; the last of these survives only as the generic advice to pull images where bandwidth is good and fan them out over the fabric.

What was not changed

  • Every measured number (bandwidths, tokens per second, KV-cache sizes, boot times, memory floors, counter deltas) is exactly as logged.
  • Interface and RDMA device names (enp1s0f0np0, enP2p1s0f0np0, rocep1s0f0, roceP2p1s0f0) are NVIDIA-documented and appear unchanged.
  • RouterOS interface names (qsfp56-dd-1-1, qsfp56-2-1, ether1, ether2) are the switch's own names and appear unchanged; which physical uplink is ether1 versus ether2 on each switch is arbitrary and was kept for readability.
  • Community repository names, Hugging Face checkpoint names, part numbers, product prices and NVIDIA documentation URLs are unchanged.

Grep used before publication

The tree was searched for internal address prefixes, internal hostnames, switch serial prefixes, the switch MAC OUI, and the word for a login secret. The search returned no hits at commit time. Re-run it after any edit that pastes from internal notes.

View or correct this page on GitHub