Hot-Swap Server Power Supplies

Hot-swap PSUs are not interchangeable across vendors

A hot-swap server power supply is a sealed module that slides into a bay at the back of a chassis and mates with a blind connector. The connector, the module dimensions, the latch and the management signalling are all vendor-specific. A Dell supply does not go into an HPE chassis and neither goes into a Supermicro one, and the fact that they all produce twelve volts is beside the point.

Within a vendor there is more compatibility but it is not universal either. Dell's flex-slot-era supplies and HPE's Flex Slot family each cover a range of chassis, and the vendor publishes which. Where a listing states a part number we render it, because a part number can be checked against a vendor's compatibility documentation and a title claiming to fit twelve models cannot.

The practical rule for buying second-hand: find the part number of the supply currently in the machine, or the part number the machine's own documentation calls for, and buy that. Matching by wattage alone is how people end up with a module that physically fits and is rejected by the chassis management, or one that does not physically fit at all.

Wattage, and why the biggest one is usually wrong

The wattage options in the table are what the listings state, and they cover a wide range because these chassis were sold with a choice. That choice was made against a configuration - how many processors, how many drives, whether there were accelerators - and the original buyer's configuration is not necessarily yours.

Buying substantially more wattage than the machine draws is not free. Switching supplies are most efficient in a band, typically somewhere in the middle of their rating, and a 1600W supply running a 200W load sits well below that band. The efficiency certification on the label is measured at defined load points, not at ten percent.

The counterweight is redundancy. In a two-supply chassis with redundancy configured, each supply must be able to carry the whole load alone, because that is the entire point. So the sizing question is not what the pair delivers together, it is what one of them delivers by itself.

We publish no wattage recommendation per machine, because we hold no sourced power figure per configuration. Server vendors publish configuration-specific power calculators and those are the authoritative answer. What is on this page is what is on the shelf and what each listing states.

Redundancy is about the feed as much as the supply

Two supplies in a chassis protect against a supply failing. They protect against a power feed failing only if the two supplies are on different feeds, which is a wiring decision rather than a hardware one.

In a homelab both supplies are almost always plugged into the same strip on the same circuit fed by the same UPS. That configuration is genuinely useful - power supplies do fail, and they fail more often than most components in a used server - but it is worth being clear that it is protecting against one specific failure and not against the mains.

There is a related point that matters if the machine is on a UPS. If both supplies are on the same UPS, they both see whatever that UPS outputs at the same instant, so a waveform the supplies dislike takes out both halves simultaneously. Redundancy does not help with a bad transfer, which is one of the reasons the rackmount UPS page treats pure sine output as a gate rather than a preference.

The other thing two supplies buy in practice is the ability to replace one without downtime, which for a machine that is doing something useful is worth more than the redundancy itself.

Noise, which is a real specification nobody publishes

Server power supplies contain small, fast fans, and in the 1U variants those fans are very small and very fast. A machine that is otherwise tolerable in an office can be made intolerable by its supplies alone.

This is worse with a supply that is undersized for the load, because the fan speed tracks temperature and temperature tracks how hard the supply is working. It is also worse immediately after power-on, when many chassis run every fan at full speed for a self-test - a noise people frequently mistake for a fault the first time.

We publish no noise figures for any of this hardware. We have not measured any of these supplies, the vendors mostly publish acoustic figures for the whole system rather than the module, and a number we borrowed from somewhere would be worse than saying nothing. Our page on rack noise covers the general problem; the specific answer for a specific supply is not something we hold.

What is worth knowing before buying: replacing supplies is one of the few noise interventions that is straightforward, because they are modules. Replacing chassis fans on a server that monitors them is considerably less straightforward.

Where to go next

All server parts · HBAs · Used servers · Verify on arrival - a supply that was already failing is a common surprise · Rack noise - supplies are one of the easier noise interventions · Rackmount UPS - why redundant supplies do not help with a bad waveform.

Frequently asked questions