Topology
Power and topology: One plan, not two
Power enters a fabric at defined entry points, usually a small number of connectors or busbars, and from there it is distributed along rails to every device in the tree.
Power enters a fabric at defined entry points, usually a small number of connectors or busbars, and from there it is distributed along rails to every device in the tree. When a rail drops, every device downstream of that rail loses power at once, regardless of how the topology is drawn. The power plan and the topology plan are read together because a rail is not an abstract supply line: it is a physical boundary that cuts the fabric into powered and unpowered regions.
Where does power enter a fabric?
Power enters a fabric at the boundary between the enclosure and the outside world. In a rack, that boundary is typically a power distribution unit feeding one or more bulk supplies, which in turn feed the board or backplane. On a board, the entry points are the connector pins or busbars that carry the incoming rails. Everything past that point is downstream.
The entry point matters because it sets the number of independent power domains. A single feed creates one domain. Two feeds with separate conversion stages create two domains, and a device can be placed in either. The topology plan records where each device sits in the tree; the power plan records which domain that device belongs to. Neither document is complete without the other.
A useful way to read the entry point is to ask what is upstream of it. If a rail is generated on the board from a wider input, the entry point for that rail is the regulator, not the connector. If the rail arrives already regulated from an external supply, the entry point is the connector or the cable. The distinction changes what a failure at the entry point takes down.
Readers who work on the energy side of infrastructure will recognize the same pattern in other domains: a supply point defines a boundary, and everything behind it shares a fate. A site such as hydrogen system planning treats production, delivery, and end use as one chain for the same reason, because a break at one point propagates along the chain rather than staying local.
What happens when a rail drops?
A rail drop removes power from every device that draws from that rail. The effect is not gradual. Devices with their own holdup capacitance may ride through a short interruption, but once the holdup is exhausted, the device resets or shuts down. Devices without holdup stop immediately.
The scope of the outage is set by the rail map, not by the tree depth. Two devices at very different depths in the topology can fail together if they share a rail. Two devices at the same depth can fail separately if they sit on different rails. This is why a topology diagram alone is a poor predictor of failure scope.
A rail drop also changes the state of the fabric itself. Links that were trained go down. Upstream ports see receivers disappear. Enumeration state held by the operating system becomes stale. When power returns, the fabric does not resume where it stopped: it re-trains, re-enumerates, and re-allocates resources. The recovery path is a cold start for the affected region.
Partial rail loss is the harder case. If a rail feeds several devices and only one of them has an independent path to another rail, that device may survive while its neighbors do not. The surviving device can be left in a state where its links point at devices that are no longer present. Firmware and drivers have to tolerate this, which is one reason hot-plug and surprise-removal handling is specified in such detail.
Why are the two plans read together?
The power plan answers which devices share a fate. The topology plan answers which devices can talk to each other. Read separately, each gives a partial picture. Read together, they show which communication paths survive a given power failure.
Consider a switch with two upstream ports on different rails and several downstream devices on a single rail. The topology plan shows redundant upstream paths. The power plan shows that the downstream devices have no redundancy at all. The combined reading shows that the redundancy is real for the switch but not for the endpoints behind it.
The combined reading also exposes single points of failure that neither plan shows alone. A regulator that feeds two branches of the tree is invisible in a topology diagram, because the diagram draws links, not supplies. It is visible in the power plan, but only when the power plan is overlaid on the tree.
A third reason is change control. Adding a device to the topology can move it onto a rail that is already close to its current limit. Removing a device can leave a rail underloaded in a way that changes regulator behavior. Neither change is visible if the two plans are maintained in separate documents with separate revision histories.
Reading the two plans together: a checklist
- Confirm the entry points: list every connector, busbar, and on-board regulator that introduces a rail.
- Map each rail to the devices it feeds, including devices that draw from more than one rail.
- Mark shared rails on the topology diagram, so that devices with a common supply are visually grouped.
- Identify devices with a single rail and no alternate path; these are the endpoints that a rail drop removes entirely.
- Check holdup: note which devices can ride through a brief interruption and for how long.
- Trace the recovery path: after a rail returns, which links re-train and which enumeration state is rebuilt.
- Compare revision histories: a topology change and a power change should appear in the same revision note.
- Verify current limits against the device list, not against the rail label alone.
Rail map
| Rail | What it feeds | What a loss takes down |
|---|---|---|
| Main input | Bulk supply, then all on-board regulators | The entire fabric in the enclosure |
| Regulator A output | Root complex, upstream switch ports | All links and all downstream devices |
| Regulator B output | Downstream switch ports, endpoints | The endpoints on that branch only |
| Auxiliary rail | Management controller, sideband signals | Out-of-band access, telemetry, wake logic |
| Standby rail | Holdup capacitors, clock generation | Ride-through capability, then all clocked devices |
Common mistakes
- Treating the topology diagram as a complete failure map. It shows links, not supplies.
- Assuming that depth in the tree equals failure scope. Rail membership decides scope.
- Ignoring the auxiliary rail. Losing it removes management access while the data path still looks healthy.
- Reading holdup as unlimited. It buys time, not immunity.
- Updating one plan and not the other. A rail change that is not reflected in the topology diagram produces a diagram that predicts the wrong failure scope.
- Forgetting that recovery is a cold start for the affected region, not a resume.
- The practical rule is simple: any document that shows where devices sit should also show where their power comes from. When the two are drawn on the same page, the failure scope of a rail drop becomes readable at a glance, and the redundancy that exists in the topology can be checked against the redundancy that actually exists in the supply.
The page behind these facts
The figures on this page are read from the source the text names, energy.gov, consulted on September 6, 2026. How this desk reads a document, and what it does with a figure it cannot source, is set out in sources and method.