Your PC shuts off forty seconds into a game, or it refuses to spin a fan at all, and the side panel comes off. Out comes the multimeter, the yellow wire reads 12.1 V, and the power supply gets declared healthy. Then the same crash happens on the next boot.

A voltage number by itself says very little. It means something only when you know the load it was taken under, the accuracy of the instrument, the input voltage at the wall, and what the rail does when demand spikes. Learning how to test power supply correctly means measuring the same rail several times under different conditions and writing all of it down, not once on a bench with nothing connected.

The order matters. Cheapest test first, and escalate only when the symptom survives a check that should have caught it.

Expected Voltage Ranges for Each PC Power Rail

Every rail in a working desktop has a window it has to stay inside. Keep this table in view while you probe: it covers the four rails a modern ATX supply produces, plus the legacy negative rail you will still find on some units.

RailNominalWhat it powers
+12 V12 V (11.40 V to 12.60 V)CPU, GPU and most high-draw components
+5 V5 V (4.75 V to 5.25 V)Older drives and some motherboard circuits
+3.3 V3.3 V (3.135 V to 3.465 V)Motherboard logic and memory
+5VSB5 V (4.75 V to 5.25 V)Standby power, live when the PC is off
-12 V-12 V (-10.8 V to -13.2 V)Legacy designs, often absent on newer units

+12 V does the heavy lifting. On any modern build, the CPU voltage regulators and the graphics card draw from +12 V, so it gets the first probe and the closest attention. The standard working tolerance for the main ATX rails is plus or minus 5% of nominal, which puts a healthy +12 V anywhere from 11.40 V to 12.60 V. Getting a trustworthy reading follows a fixed order, and the Fluke procedure for testing a DC power supply with a digital multimeter lays it out: disconnect or power down the load where practical, inspect the meter and leads, verify the meter on a known voltage source, select a range above the expected value, then compare the reading against the supply's specification.

The minor rails fail less often, but they still fail. +5 V feeds older drives and some board circuits, +3.3 V runs memory and motherboard logic, and both sit inside their own 5% windows in the table. The purple +5VSB wire deserves the most respect of the group: it stays live whenever the cord is plugged in, so a board can wake and a probe can bite even with the machine shut down. The -12 V rail is nearly vestigial and absent on many current units. Don't chase it unless a specific card calls for it.

Idle voltage tells you almost nothing

A rail reading 12.1 V at the desktop is behaving normally. That same rail dropping to 11.2 V when a game loads the GPU is a fault, because it has fallen through the 11.40 V floor while doing the job it exists to do. That drop is what poor regulation looks like from the outside, and the usual causes are an overloaded unit, aged capacitors, a damaged switching stage, or a cable and connector carrying more current than they are rated for.

A sag inside the window can be a warning too. A rail that slides from the top of its range down to the bottom edge as load climbs has passed the table check and still shown weak regulation. That is why the Fluke procedure has you repeat the measurement at idle and under a controlled load while recording input voltage, output voltage, load current and temperature. A reading taken with nothing attached proves the rail exists, and that is all it proves.

Write down the conditions, not just the number

A voltage is only as trustworthy as the instrument behind it. The National Institute of Standards and Technology power and energy measurements program reports that its power-measurement bridge holds a basic uncertainty of about 15 parts per million at 120 V, 5 A, 60 Hz across power factors from 0 to 1, and for practical calibration work it gives typical uncertainty values of 50 to 500 parts per million, depending on power level and instrument type. A national lab publishes its uncertainty to six digits. Your handheld meter carries a spec of its own, printed in the manual, and on a 12 V rail it can eat a visible slice of the window you are checking against. That spec belongs next to every number you record.

Practical rule: log four things beside every reading. The meter and its range, the load, the input voltage, and the temperature. A number missing those four can't be compared with anything you measure next week.

A rail outside its window is a fault to investigate, full stop. A rail inside its window is a rail that passed one check.

Choosing the Right Test for Each Fault

Six tests are recommended for power supplies, and they don't answer the same question. Each one can tell you something different.

Test methodWhat it catchesWhat it missesBest for
Multimeter on DC outputsMissing or grossly wrong railsRipple, transient response and load sagChecking rail voltages at idle
Paperclip testWhether the PSU starts and the fan spinsRail accuracy and load behaviorConfirming a dead or silent unit
PSU tester (LED type)Rail presence and some connector faultsRegulation under load and rippleQuick presence check before buying a replacement
Built-in self-test (Dell systems)Internal PSU failure with other parts isolatedLoad behavior and non-Dell systemsConfirming a suspected PSU fault on Dell hardware
OscilloscopeRipple, noise and switching artifactsRequires skilled probing and setupRandom crashes under load with good rail readings
Electronic load with multimeterRegulation across the full load rangeRequires equipment and heat managementVoltage that sags only at high draw

Read that table by fault, not by tool. A machine that won't turn on needs a jumper and a fan check, not an oscilloscope. A machine that turns off under load needs a load and a ripple measurement, not another LED tester.

The escalation order is cost and setup, in that sequence. Multimeter first, paperclip second, LED tester third if you want a connector-by-connector sweep. The oscilloscope and the electronic load come last because they need the most preparation and, in the load's case, somewhere to dump several hundred watts of heat. A passing LED tester proves rails exist. It says nothing about whether the unit holds regulation or stays quiet when a GPU pulls 300 W for a second, so a green light does not clear a supply that drops out mid-render.

Improvised loads built from resistors or hair dryers are worth avoiding outright. They give unrepeatable results, they get hot enough to burn a bench, and they tell you nothing you can compare against a spec. Use a purpose-built electronic load, or accept that you're doing a symptom test rather than a measurement. And whatever tool is in your hand, the mains side of the unit stays closed.

For a picture of what full validation looks like, the Cybenetics SAFE test protocol evaluates fuse and inrush protection, startup at 50% rated load at the minimum declared input voltage, over-voltage shutdown thresholds, and rail regulation across tested loads. That last item is the one a multimeter never covers.

USB-C supplies add a complication the old fixed-rail tutorials miss entirely. The USB Implementers Forum charger page states that USB Power Delivery 3.1 supports fixed 28 V, 36 V and 48 V modes reaching up to 240 W, while Adjustable Voltage Supply mode can request intermediate voltages. A USB-C meter showing 5 V at idle hasn't shown you that the charger can negotiate the profile your laptop asks for, hold its rated power, or read the cable's rating. Negotiation is a separate test from voltage presence, and confusing the two sends people out to buy replacement hardware that was never broken.

Safety Steps and Multimeter Test Procedure

OSHA's rules on de-energizing equipment in 29 CFR 1910.333 require circuits and equipment to be disconnected from electrical energy sources before work, and a qualified person must use test equipment to verify that exposed circuit elements are deenergized. That test also has to catch induced voltage and backfeed, energy arriving from a source other than the one you disconnected.

The field version of that requirement is the "test before touch" sequence. Deenergize, lock out or otherwise prevent restart, verify the tester on a known source, test the equipment, then recheck the tester to confirm it survived the measurement. IEC 62368-1, the safety standard for audio, video, information and communications equipment, applies the same logic: it classifies energy sources by the hazard they present and prescribes a safeguard for each class.

For a home user the line is clear. Low-voltage output checks on the DC side are within reach with basic precautions. Primary-side, mains-input and capacitor measurements need competent training, category-rated equipment, insulated probes, and a procedure that assumes hazardous voltage may still be present. In a desktop, everything you need to measure is on the output side of the supply, so there's no reason to go further.

The steps below are the ones I'd follow on a switched-off machine, in this order.

  1. Shut the PC down, switch off the PSU and unplug its cord. The unit has to be dead while your hands are inside the PC, and the supply's own casing stays shut for the whole procedure.
  2. Check the leads and verify the meter on a known source. Look for cracks in the insulation, then read an AA battery. About 1.5 V means the meter and leads are working; anything else means the numbers you're about to collect aren't trustworthy.
  3. Set the range above the expected voltage. The 20 V DC setting suits a +12 V rail. If the meter autoranges mid-measurement, the display flickers and you can misread a digit at exactly the wrong moment.
  4. Touch the black probe to a black COM wire, then the red probe to the rail. Black is ground, orange is +3.3 V, red is +5 V, yellow is +12 V, purple is +5VSB. Every reading needs the same reference point.
  5. Plug the cord back in, switch the PSU on and read with the connector fully seated. The outputs only exist while the unit is energized, and a half-seated connector produces a false low reading that looks like a fault.
  6. Record the reading with the load and input voltage in use. A number without its conditions is a number you'll have to take again.

Three slips that produce bad readings

The current jack instead of the voltage jack. Clipping the red lead into the 10 A socket puts a near-short across the rail the moment you touch a live pin. Best case the meter's fuse blows. Worst case you've added a fault to a circuit you were trying to diagnose. Check which socket the red lead is in before anything gets powered.

Shorting neighbouring pins. The +12 V and ground pins on a 24-pin connector sit millimetres apart, and a probe tip is wide enough to bridge them. Wrap electrical tape around the tip or use a fine needle probe, keep one hand behind your back, and look at the connector rather than at the meter while you place the tip.

Treating a no-load reading as a full diagnosis. Nothing attached means nothing drawn, and a supply with weak regulation or dried capacitors will often produce a perfect 12.00 V into an open circuit. Even with a load attached, the meter reports a smoothed average and cannot show the ripple riding on the rail. It's the single most misleading number in this whole process.

Mains-side parts, capacitors that hold charge after shutdown, and everything behind the supply's casing are work for a qualified technician with the right equipment.

Paperclip, Standby and Self-Test Checks

The paperclip test answers one question: does this unit start on its own, with the motherboard and everything else out of the picture. Unplug the 24-pin, CPU and GPU cables from the board, bend a paperclip into a U, and bridge the green PS_ON wire to any black ground wire. On a 24-pin connector that's pin 14 to pin 15, counted along the clip-side row. On modular cables where every wire is black, count pins rather than hunt for green.

how-to-test-power-supply-psu-jump.jpg

Plug the cord in, switch the rocker to on, and watch the fan. Turn off any Eco Mode, Quiet Mode or Zero RPM switch first, because a fan-stop supply looks dead at zero load. When you're done, switch off, unplug the cord, then remove the jumper. Never leave an unloaded supply running unattended, and stop after a few seconds.

The paperclip test answers one question. Don't let it answer three.

A spinning fan proves the unit switches on. It doesn't prove regulation, it doesn't prove the unit survives a load, and it doesn't prove the supply is the reason your PC won't POST.

Here's the same procedure walked through on a bench.

Standby power is the next quick check, and it explains a specific symptom: a PC with a lit motherboard LED that won't wake from the power button. With the cord plugged in and the switch on, the purple +5VSB wire should read about 5 V against a black ground. No standby voltage means the supply's always-on circuit has failed, and no amount of button-pressing will fix that.

The built-in self-test on Dell systems

Dell ships a BIST button on many desktop supplies, and it isolates the PSU from a suspect motherboard or a disconnected internal cable. According to Dell's power supply troubleshooting guide, you power the machine off, press the button, and watch both the status LED and the fan.

  • Pass: a solid, non-flickering LED with a fan spinning normally.
  • Fail: no light at all, or a fan that doesn't spin correctly.
  • Fail even with the LED lit: treat a non-spinning fan as a failure, because the LED alone doesn't clear the unit.
  • Not available: non-Dell systems don't have this button, so fall back to the paperclip test.

A plug-in PSU tester is the last of the quick tools. Connect the unit's 24-pin, 4-pin and 8-pin cables to the matching sockets, switch on, and confirm each rail LED lights and holds steady. A dark or flickering LED marks a rail to investigate. Read the tester's labels to see which connectors it covers, because not every model checks every rail, and a missing socket isn't a passing result.

Ripple and Load Testing Under Real Demand

A rail can hold a textbook 12.0 V on the meter and still crash the PC the moment the GPU pulls hard. The DC volts setting averages, while ripple is a peak-to-peak artifact riding on top of that average. Switching ripple is usually nonsinusoidal, and many handheld meters are specified mainly for sinusoidal measurements, so a clean-looking reading proves nothing about noise on the rail. You need a scope, a defined load, and the discipline to repeat the measurement under more than one condition.

how-to-test-power-supply-oscilloscope-measurement.jpg

The measurement method matters as much as the instrument:

  1. Verify probe compensation. Touch the probe to the scope's calibration terminal and adjust the trimmer until the square wave has square corners. An uncompensated probe inflates or softens every edge you measure.
  2. Set AC coupling and a time base showing several cycles. AC coupling removes the 12 V DC offset so the ripple fills the screen, and several repetitive cycles make the peak-to-peak reading meaningful rather than accidental.
  3. Probe at the output with a short ground. Use the manufacturer's fixture or a spring-tip ground, not the six-inch alligator clip, because a long ground lead picks up loop noise and invents ripple that isn't there.
  4. Apply a defined load near the rated maximum. A stress test or an electronic load works. Ripple measured at idle is a comfort blanket, not a test.
  5. Read peak-to-peak on each rail and compare it with the ATX spec limit for that rail. A healthy supply lands well under the limit, while an unfiltered or failing unit can show noise far above it. Crossing the spec is a real fault indicator, not a cosmetic wobble.
  6. Repeat at low and high input voltage and at several load levels. Ripple that only appears at 80% load and 100 V input is still a fault, and it's the kind a single test misses.

What a passing result looks like

Take that same rail at 80% load: 12.0 V on the meter, and the scope shows a peak-to-peak figure past the limit. The meter said fine; the scope says out of spec. That gap between the two instruments is the whole reason this test exists.

For scale, KitGuru measured the Corsair RM850x at full 850 W load with ripple of 17 mV peak-to-peak on +3.3 V, 16 mV on +5 V and 23 mV on +12 V in its 2024 RM850x ATX 3.1 review. Igor'sLAB recorded 24.2 mV of +12 V ripple at 100% load on the same model and marked it a pass (Igor'sLAB RM850x test). Two labs, two slightly different numbers, both far inside spec. That spread is normal, and it's why the conditions travel with the result: a ripple figure only means something next to the load and input voltage where it was taken.

The recurring mistakes are a long ground lead, a probe with less bandwidth than the switching frequency, skipping the load entirely, and confusing peak-to-peak ripple with RMS noise. Each one distorts the figure enough to flip a passing verdict into a failure, or hide a real fault behind a clean-looking trace.

Diagnosing Common Symptoms and Their Likely Causes

The most common complaint is a PC that shuts down under gaming or rendering load and runs perfectly the rest of the time. That pattern points at the +12 V rail sagging as the GPU draws, or at a protection circuit tripping earlier than it should. It's also the symptom a no-load test is guaranteed to miss.

SymptomLikely causeNext test
Shuts down under gaming or rendering loadSagging +12 V rail or overcurrent protectionLoad test with multimeter on +12 V, then ripple check
No power and no fan spinDead PSU, failed switch or cordPaperclip test, then check the outlet and power switch
Random reboots during heavy useRipple above 120 mV or failing capacitorsOscilloscope ripple test against the 120 mV limit
Standby light on but no startFaulty PS_ON control or motherboard faultPaperclip test, then swap in a known-good PSU
Good rail readings but unstable operationLoad-dependent regulation faultElectronic load test across the full load range
Burning smell or crackling soundInternal damageStop, unplug and replace the unit without probing it

A burning smell or a crackle from the unit ends the diagnosis. Unplug it, don't open it, don't probe it, replace it. The same goes for a supply that has visibly vented a capacitor or shows scorching around a connector.

When the power supply isn't the culprit

Several of those symptoms belong to other parts just as often. Random reboots fit failing memory, a loose GPU power cable or an unstable overclock. No-start conditions fit a dead motherboard, a shorted front-panel switch or a standoff touching the case. Before you spend money, run MemTest86 on the memory and reseat the GPU and 24-pin connectors, because those cost nothing and produce identical behaviour.

Rule out the cheap parts first: memory and a half-seated cable mimic a dying supply more often than people expect.

A known-good PSU is the cleanest ruling-out tool there is. If a borrowed unit of adequate wattage runs the same workload without a shutdown, the original is the problem. If the symptom survives the swap, the fault is elsewhere and you've saved the price of a replacement.

Intermittent faults deserve their own paragraph, because they beat short tests. A supply with a marginal capacitor or a cracked solder joint can pass every check on a cold bench and fail twenty minutes into a session. Repeat the load test after warm-up, then repeat it again after the unit has been under load for half an hour. A fault that only appears hot is still a fault, and it's the one most likely to be declared fixed prematurely.

Decide Whether to Repair or Replace the Unit

The replacement rule is short. Replace the unit if it fails the paperclip test, if the fan doesn't spin or the bearing screeches, if any rail falls outside its window, if ripple exceeds the ceiling that applies to your platform, or if it smells burnt. Those are all terminal for a home user, and none of them are worth a repair attempt on a part that costs less than the labour to diagnose it.

Internal repair is work for a qualified technician with the schematic and the equipment. A home user should not open the case, and the storage charge in the primary capacitors stays lethal long after the cord comes out.

Before buying anything, swap the modular cable and try a different connector on the supply. A damaged cable or a melted connector housing mimics a dead unit convincingly, and it's a fraction of the price to fix. Check both ends of every modular lead for discoloration at the same time.

What to check before you buy

  • Wattage with headroom: add up the CPU and GPU draw, then leave margin for transient spikes, which is where an undersized unit fails.
  • Connector match: confirm the new unit has the EPS and PCIe connectors your board and card need, and the right number of them. Adapters on a 12VHPWR card are a bad place to save money.
  • Published efficiency: a rated efficiency figure is a proxy for build quality. Tom's Hardware measured the Corsair RM750e at an average of 89.5% efficiency with 115 VAC input and 92.4% with 230 VAC input across loads from 20% to 100% (Tom's Hardware RM750e review), which is a useful reminder that the same unit performs differently on a different wall supply.
  • Warranty length: a ten-year warranty on a mid-range unit tells you what the maker expects its capacitors to survive.
  • Guides for the rebuild: NeoTeo's PC building and troubleshooting guides cover replacement selection and the checks that follow a swap, if you want a second reference while the case is open.

Then repeat the basic checks on the new unit before it goes into a full build. Multimeter on the 24-pin at idle, paperclip and fan check, and a standby reading. Five minutes now beats discovering a dead replacement after the cable management is finished.


If you want to keep going past this guide, NeoTeo publishes hands-on tutorials that take a repair from symptom to fix, including the memory and stress-test procedures that rule out everything a power supply gets blamed for. Its hardware section covers parts selection, BIOS-level voltage monitoring and the software tools that log rail behaviour under load, so you can pair a multimeter reading with what the board reports. Head to NeoTeo and work through the checks in the order this article lays out, cheapest test first.