The question this piece answers: when does a home rack actually need an online (double-conversion) UPS, and what is the spec that proves it? This is the fourth piece in the home-rack power series — the 1,920 W branch, the per-outlet connector limits, the rack PDU — and it is the piece that decides what sits upstream of all of that. Every transfer-time and hold-up-time number below is quoted from vendor documentation or the specification itself, with at least two independent sources per number; where a claim rests on a single analysis, the piece says so.
What do the three UPS topologies actually do?
The names describe where the inverter sits, and the inverter's position is the whole product. A standby (offline) unit passes utility power straight to the outlets through a relay; only when the power fails does it start its inverter and switch to battery. That mechanical relay takes 10–20 ms to move (two independent sources agree on the range: a UPS-selection guide and vendor product documentation). A line-interactive unit keeps the pass-through path but adds an automatic voltage regulator (AVR) — a transformer that corrects sags and swells by tapping the windings, without going to battery. When the power does fail, the transfer to battery pauses the output briefly: 2–4 ms per Eaton's own comparison material (published through DigiKey's engineering channel), with distributor documentation citing a 2–10 ms range depending on model. An online double-conversion unit never passes mains through at all: it rectifies AC to DC, and the inverter that powers your equipment runs continuously from that DC bus, with the battery floating in parallel. Mains failure changes nothing downstream — 0 ms, a figure Eaton, distributor documentation and integrator guides all state identically, because there is no switching event to time.
| Topology | On mains | Transfer to battery | Efficiency | Built for |
|---|---|---|---|---|
| Standby (offline) | Pass-through, no regulation beyond surge clamping | 10–20 ms | ~99% (pass-through) | Non-critical loads; the outage is the only event that matters |
| Line-interactive | Pass-through + AVR corrects sags/swells without battery | 2–4 ms typical (Eaton; up to ~10 ms per vendor docs) | ~97–99% (integrator guide) | Home racks, NAS, network gear on ordinary grid power |
| Online double-conversion | Full rectify→invert regeneration; output never touches mains | 0 ms | ~90–94% (OpenZFS docs: "at most 94%" without high-efficiency mode; 90–96% per the same guide) | Bad grids, frequent generator transfers, unattended sites, medically critical loads |
One row of that table deserves emphasis because it is the one marketing skips: the AVR. Line-interactive units handle the most common power event — the sag — by regulating voltage through the transformer, without spending battery cycles. Battery cycles are the consumable that kills UPS units; lead-acid batteries last roughly three to five years in home conditions (the OpenZFS project's hardware documentation quotes the same window), and a topology that avoids discharging them for everyday fluctuations is a topology that ages slower. Double conversion, by contrast, re-regulates everything through its own inverter all the time — maximum protection, maximum wear on nothing but your electricity bill.
What is hold-up time, and why does it outrank the transfer spec?
A power supply's hold-up time is how long it keeps its DC rails inside spec after its AC input disappears entirely — the bridge built out of the primary bulk capacitor. The Intel ATX12V design guide, which every desktop PSU follows, requires 17 ms of hold-up at 100% rated load, and requires the Power Good (PWR_OK) signal to stay asserted for at least 16 ms, dropping at least 1 ms before the rails leave regulation — a warning the motherboard uses to halt the CPU cleanly rather than corrupt data. Two independent references (a homelab engineering site and the OpenZFS project's hardware documentation) state the 17 ms figure identically, and the 17 ms value has a physical explanation: one full 60 Hz cycle is 16.67 ms, so the spec guarantees survival of one complete AC cycle plus margin.
The number most buyers have not heard: the ATX 3.1 revision (September 2023) lowered the mandatory hold-up floor to 12 ms at full load, keeping 17 ms as a recommendation. Corsair's own explainer states the reduction plainly ("compared with ATX 3.0's 17 ms hold-up time, ATX 3.1 shortens it to 12 ms"), and PSU vendors and system integrators describe the change identically; the stated reason is capacitor cost and size. One engineering analysis has also measured budget consumer units failing even the published minimums under full load — we report that as a single-source finding, not a spec, but the direction is documented: hold-up time is the spec vendors don't advertise, it is inversely proportional to load (OpenZFS docs note the inverse relationship too), and it ages as the bulk capacitor dries out.
Now put the two sides of the relay together, and the purchase decision becomes arithmetic instead of ritual (ours). Worst common case: a line-interactive unit transfers in 4 ms (Eaton's comparison figure). Even a pessimistic 8–10 ms relay reading — the top of the distributor-documented range — sits inside the 12 ms ATX 3.1 mandatory floor at full load, and far inside the 17 ms ATX12V floor that most PSUs still meet at real-world loads well below rating. The PSU is spec'd to bridge the UPS's gap three to six times over. The relay is a mechanical part that moves once per outage; the "seamless" double-conversion alternative burns 24/7 watts to eliminate a pause your hardware was born unable to notice.
So when does a home rack actually need double conversion?
The honest list is short, and it matches Eaton's own guidance for unattended network closets: bad or unknown grid quality (frequent deep sags, wiring so marginal that regulation matters), generator-heavy sites where transfer events are frequent and messy, and unattended locations where nobody can reset hung gear after a dirty event. If your rack lives in a home on a municipal grid that keeps the lights on, none of those apply — and the events you actually experience are sags and momentaries, which the AVR handles without a battery cycle. The buying priority inverts: correct VA/W sizing (the piece from October covers the arithmetic), pure sine output for active-PFC supplies, AVR, and network-card graceful-shutdown all move your uptime more than the topology does.
Price the topology decision itself (ours). A 300 W home rack on a double-conversion unit running at ~93% efficiency draws ~323 W at the wall; on a line-interactive unit at ~98% it draws ~306 W. That 17 W difference runs 24/7: about 149 kWh a year, roughly $22 at $0.15/kWh — and, per the cooling piece in this series, 149 kWh a year of heat dumped into the room the rack sits in, every hour of a summer night included. Double conversion is a space heater in series with your rack that charges you for the privilege of closing a gap your PSU already closes. That is not an argument against the topology — it is the reason the topology is for sites whose grid or generator profile actually opens a gap longer than 12 ms.
One more practical bridge-check, since hold-up time is undocumented on most retail boxes (ours): run your rack at its worst-case load and pull the UPS's "on battery" test with the unit's monitoring software — if the machine rides through a transfer while saturated, your real-world margin is proven on your hardware, which beats any spec sheet. And replace the batteries on schedule; a five-year-old battery turns a 4 ms transfer gap into a different conversation entirely.
What are the caveats?
Transfer times are ranges, not constants. The 2–4 ms line-interactive figure is Eaton's own comparison; other vendors' documentation extends to ~10 ms depending on model and load, and a degraded relay changes the number over time — always check the specific unit's spec sheet. ATX 3.1's 12 ms floor is a floor, not a promise. Hold-up time is load-dependent, capacitor-dependent and age-dependent; a unit running at 90% load with dried capacitors bridges less than the spec says, which is why the saturated-transfer test above is worth doing. The efficiency figures mix sources. The 90–94% double-conversion range comes from OpenZFS's documentation plus one integrator guide; the 97–99% line-interactive figure comes from that same guide alone, so treat the gap as "several points" rather than a precise spread. This is not electrical-safety advice. The upstream rules — the 1,920 W branch limit, the per-outlet connector caps — bind no matter which UPS you buy, and a whole-rack online unit is a continuous load on the same branch as everything else.
Comments