Power Circuit Headroom
Back to Data Center Management · Rack Unit Tracking · Cooling Capacity · Growth Forecasting · Service Offerings
This is confirming new hardware won't push a circuit past its safe capacity before it's racked. Note the tense. Power draw per circuit is about measuring what is being drawn now and noticing when it creeps; this page is the decision made in advance, about equipment that is not yet in the building.
The two depend on each other. Without measurement there is no credible starting figure, and the check degenerates into adding nameplate ratings — which produces a number so conservative that people learn to ignore it.
1. The Headroom Is Half What You Think
Three ceilings apply in order, and only the first is widely observed.
- 80 per cent of the breaker rating, because everything here is a continuous load. A 32 A circuit offers 25.6 A.
- Half of that again, if the rack is dual-fed and expected to survive a feed failure. Each feed must be able to carry the whole load alone, so steady state sits near 40 per cent of rating per side.
- Whatever the weakest link downstream allows — see section 4.
The illustration above is the case that catches people. Existing draw is 38 per cent, the proposed addition takes it to 52, the breaker is nowhere near its limit, and every dashboard is green. Then feed B is lost, feed A is asked for 104 per cent, and the rack that was built for redundancy goes dark because it was dual-fed.
If the answer you want is "yes" and the redundancy arithmetic says no, the honest options are to rebalance across circuits, to accept explicitly that the rack is no longer feed-redundant, or to put the hardware somewhere else. Quietly proceeding is the option that looks fine for a year.
2. Estimating Draw for Hardware You Do Not Have
You cannot measure what has not arrived, and nameplate is not an estimate — real draw is commonly 30 to 50 per cent of it. Better sources, in order of usefulness:
- Measure an identical machine you already run, under its real workload. By far the best evidence, and frequently available.
- The vendor's power calculator, configured to the actual specification — the processor, the memory population, the drive count. These are reasonably good and much better than nameplate.
- Nameplate as an upper bound only, used to confirm the worst case does not trip anything.
Then add margin for what the machine will become. Servers gain memory and disks over their life, and the draw you plan around should be the configured maximum rather than the as-delivered figure.
3. Peak, Not Just Steady State
- Inrush at power-on can be several times running draw, and it arrives for every machine at once after an outage. A circuit sized on steady state can trip on recovery.
- Workload peaks. A machine at idle during the survey and at full load in production is a planning error waiting to happen. Size against busy, not against now.
- Staggered start is the mitigation, by outlet group or BIOS delay, and it needs to be configured when the hardware is installed rather than discovered during the first recovery.
4. The Breaker Is Not the Only Limit
Headroom at the circuit means nothing if something upstream or downstream runs out first. Check the whole chain:
| Link | What to confirm |
|---|---|
| PDU outlets | Enough of them, of the right connector, on both feeds, reachable from the chosen units |
| Circuit and phase | 80 per cent, and the halved figure where redundancy is required — per phase, not per PDU |
| UPS | Capacity in kVA, and runtime at the new load. Runtime falls as load rises, and the figure that matters is the one after the addition |
| Generator | Capacity, and whether it is still sized for the room as it now is |
| Cooling | Every watt added is a watt to remove. See cooling capacity |
UPS runtime is the one most often missed. Adding load to a UPS that is not near its capacity limit still shortens the ride-through for everything behind it, which changes the assumptions in your recovery plan.
5. Reserve What You Have Already Promised
Capacity committed to an approved project is not available, even though it is not yet drawn. A headroom figure that ignores pending installs will approve two things that cannot both happen.
Track committed capacity the same way reserved rack units are tracked, with an owner and an expiry, and subtract it from available headroom in every answer.
How We Approach It
- Start from measured draw, per circuit and per phase, rather than from nameplate sums.
- Establish the real ceiling per circuit — 80 per cent, halved where feed redundancy is required — and record which racks are genuinely dual-fed.
- Estimate the new load properly, from an identical machine where one exists and from the vendor's calculator otherwise, at its configured maximum.
- Check the whole chain: outlets, phase, circuit, UPS capacity and runtime, generator, and the cooling consequence.
- Subtract committed capacity for approved but uninstalled work.
- Give an answer with its reasoning — yes, no, or yes-with-a-condition such as rebalancing first — and record it so the next request starts from it.
What You Get
- A per-circuit headroom figure based on measurement, with the redundancy-halved number shown alongside the breaker number.
- A list of racks that are nominally dual-fed but would not survive losing a feed today.
- A repeatable pre-install check covering outlets, phase, circuit, UPS runtime, generator and cooling.
- Committed-but-uninstalled capacity tracked and deducted, so two projects cannot be approved into the same headroom.
- A staggered power-on order, so recovery does not trip what normal running never would.
The deliverable is a sentence someone can act on before the hardware is ordered: this fits here, on these outlets, and here is what happens to it when a feed fails.