Why a server needs two power supplies
A redundant power supply (redundant PSU) protects not against the server breaking as a whole, but against one specific point of failure: if one unit burns out or shuts down due to a voltage spike, the server keeps running from the other one without a reboot. In 1U servers, two 400-550 W units are typically installed; in 2U, 750-1200 W units, with each unit on its own capable of fully powering the server under maximum load.
Without power redundancy, all other fault-tolerance measures — RAID, redundant network ports, a rack UPS — lose their point: a single power supply remains the point that brings the whole service down.
Redundancy scheme: N+1 and separate feeds
The standard scheme for a server is N+1: one power supply carries the load, the second is standby, both are physically identical and powered on at the same time, splitting the load in half. A more reliable scheme adds power separation at the rack level:
- Each server power supply is plugged into its own PDU (Power Distribution Unit) in the rack — if one PDU fails or is accidentally switched off, the other power supply keeps working.
- Both PDUs are connected to different electrical feeds of the data center, and ideally to different UPS units and different diesel generators.
- Cables from the server's two power supplies are routed along different sides of the rack, so one accidental cable snag does not switch off both units at once.
This scheme protects not only against the power supply itself failing, but also against a failure of an entire data center power line — an incoming feed fault, a breaker trip, or scheduled maintenance on one of the UPS units.
How to test redundancy in practice
Power redundancy cannot be considered working until it has been verified with a physical disconnection. Test procedure:
- Confirm via IPMI or the management panel that both power supplies are healthy and sharing the load.
- Unplug the cable from one power supply (not the switch on the unit itself — actually pull the cable, simulating a real outage).
- Check that the server keeps responding to ping and services have not restarted.
- After 2-3 minutes, plug the cable back in and confirm the second unit returns to standby rather than staying as the only active one.
This test should be done once during server acceptance — the redundancy stated in the specification does not always work as described if one of the units is faulty or misconfigured.
Power draw and headroom: how to calculate the load
Power supply sizing is done not from the rated power of components but from real power draw under peak load with a 20-30% margin:
| Component | Typical draw |
|---|---|
| CPU (2 units, TDP 150-200 W) | 300-400 W |
| RAM (16 DDR5 modules) | 80-150 W |
| NVMe drives (8 units) | 60-100 W |
| Fans and motherboard | 50-80 W |
Total load for such a server is roughly 490-730 W, so a 750-800 W power supply with N+1 redundancy leaves a comfortable margin. Do not undersize the margin: under peak load (server boot, RAID rebuild, full CPU load), power draw can briefly exceed the average by 15-25%.
Power monitoring via IPMI and BMC
The baseboard management controller (BMC) reports the state of each power supply separately, and this is worth checking regularly, not only during an outage:
ipmitool sdr type "Power Supply"
ipmitool sel list | grep -i "power supply"The first command shows the current state of both units (Presence detected, Failure detected, Predictive failure), the second shows the power-related event history from the server event log (SEL). An entry like Power Supply | Predictive Failure asserted means the unit still works but is already degrading, and it should be replaced at the next planned maintenance window, without waiting for a full failure.
Power redundancy checklist
- The server physically has two power supplies, both reporting OK status via IPMI.
- The units are connected to different PDUs and, where possible, to different electrical feeds of the rack.
- Redundancy has been tested manually — by unplugging one cable while checking service availability.
- Power sizing was done with a 20-30% margin above peak load, not from average figures.
- Alerts are configured for SEL events about a predictive power supply failure, so the unit is replaced on a schedule rather than after the server goes down.
A redundant power supply earns its cost only when all of these points are met together: a second unit in the chassis by itself does not save you from an outage if both are plugged into the same outlet on the same PDU.