Power supply (PSU): no power, random shutdowns and the failure signs of a dying supply

First reading: is it "dead" or "standby"?

Look at the case before touching anything. A standby LED lit (or USB ports that charge a phone) means the +5V standby rail is alive — the supply is at least partially working, the AC chain is intact, and the problem sits downstream: the main rails, the power button path, the 24-pin connection, or the motherboard's power sequence. No light at all means the break could be anywhere from the wall to that small standby circuit. That single glance halves the search space, and it's free.

One caution that saves a wrong conclusion: a lit standby LED does not exonerate the PSU. The standby rail is a small, separate circuit, and a supply can run a healthy +5Vsb while its main section is dead — "standby but no start" is a classic partially-failed PSU presentation. The LED tells you where to start looking, not where the problem is.

The wall, the cable, and the switch: the checks that end most cases

Three checks, in order, that clear the majority of "PC dead" reports. The outlet: plug something with a known working charger into the same socket (or try the cable in a different one) — dead multi-outlet strips and tripped RCDs/GFCIs are common and invisible. The cable: seated at both ends, not cracked at the PSU end where the strain relief lives; a second cable is a ten-minute test. The PSU's own switch: the O/⏻ rocker on the back is off more often than you'd think — by a previous owner, a cleaner, or a cable yank that flicked it. Then the last free check: the front-panel power connector itself, a small header that works loose and makes a healthy machine look dead.

Why this order: each check is cheaper than the next, and each one, if it's the answer, would have made every later test meaningless. People who start with "the PSU must be dead" are usually skipping the two cheapest links in the chain.

Starts for a second, then turns off

When the fans spin up and everything dies a moment later, something is either tripping a protection or telling the board to stop. The three usual stories: a short circuit somewhere (the PSU's short-circuit protection doing exactly its job), a failing PSU whose rails can't hold, or a motherboard fault. The test that separates them is minimal configuration: power the system with only the CPU, one stick of RAM in the known-good slot, and nothing else — no GPU (if the CPU has integrated graphics), no storage, no peripherals, front panel power only.

If the minimal config stays on, the fault is in one of the things you removed — reconnect them one at a time, retesting after each, until the failure returns with a specific component in the box. If the minimal config still dies, the field is down to the PSU or the motherboard, and the known-good supply swap (Section 09) is what splits them: swap in a good PSU, retry — if it's still dead, the board path; if it lives, the verdict is in.

Power is on, but no display (no POST)

Fans spinning steadily with a black screen is the most common "no power" complaint that isn't a power problem. A board that fails to POST with the fans alive is, in the large majority of cases, a RAM seating issue, a GPU that isn't fully seated or that's failed, a monitor/cable problem, or a stuck BIOS state — the supply is delivering, and the machine just isn't completing its boot sequence. Work the cheap sequence: monitor and cable first (a dead display mimics a dead PC), then RAM reseat and single-stick test, then the GPU in a different slot, then a BIOS reset (the small jumper or button on the board, per the manual).

Where the PSU legitimately re-enters this story: a supply can deliver enough power to spin fans and still be too weak or too degraded to complete POST under the board's full draw. You can't fully clear the supply from a no-POST machine — the honest statement is "the PSU can't be excluded until the system is running, and if it then fails under load, come back to the load branch." The no-POST sequence above is where you start, though, because it's where the answer usually is.

The connectors: 24-pin, CPU power, PCIe

Three connections matter for a powered system, and all three fail in the same subtle way: latched, but not fully seated. The 24-pin ATX connector (the big one) can sit a millimeter off with its latch clicked — enough to lose contact on the pins that carry the power-good signal, which produces "standby but no start" or "spins but no POST." The 4+4 CPU/EPS connector at the top of the board is the one that gets forgotten when a board is swapped, and a machine without it often powers on with fans and dies at exactly the moment the CPU is asked to do something. The 6+2 PCIe connectors feed the GPU directly from the supply (see the diagram) — a GPU with an unseated or missing PCIe cable can boot to a black screen or die the moment a game starts.

The check is physical and takes two minutes: unplug and reseat all three, watching the latch click and the connector go in straight, then look for bent pins in the 24-pin — a single bent pin can knock out one rail's worth of function and produce a symptom list that looks like a failing supply.

Not every shutdown is the PSU — the honest base rate

Say it plainly: most random shutdowns are not caused by the power supply. Thermal protection (a GPU or CPU hitting its limit and cutting power to itself), the OS (a crash that looks like a shutdown, a power plan, a driver), the GPU (its own VRMs and memory fail under load exactly like a PSU does), and memory faults all produce "the PC just turned off." The PSU's share of random-shutdown cases is real but minority — which is why this page keeps checking the alternatives before the supply, and why the swap test, not the symptom, is the verdict.

The practical consequence: a shutdown problem is a pattern problem. Before buying anything, spend a few days (or a few sessions) recording when it happens — idle, in a game, during a render, at a specific app — because "load only" and "random" are two different diagnoses with different suspect lists, and the recording is what tells them apart.

The pattern tells you: idle, load, or random

Three patterns, three suspect orders. Load only (gaming, rendering, stress tests — never at the desktop): thermals first, then unstable overclock/XMP, then PSU capacity or health. Idle or after sleep (it dies while you're away and you come back to a dark machine): USB power to a peripheral, sleep/hibernate settings, the standby rail, or a power plan — the supply's main rails are barely working in this scenario, which makes a main-rail PSU fault less likely, not impossible. Genuinely random (no reproducible trigger): the broadest list — OS, drivers, RAM, GPU, thermals that only build up over time, and PSU — which is why the event log becomes the first tool here.

A simple log earns its keep: date, what you were doing, temperatures if visible, and what exactly happened (off, restart, or freeze-then-off). Two or three recorded failures usually collapse "random" into one of the three patterns — and a pattern is a diagnosis you can act on, while "random" is not.

What Windows' own record says (and doesn't)

After an unexpected power loss, Windows writes a Kernel-Power 41 event to the System log — "the system has rebooted without cleanly shutting down first." Read it correctly and it's useful; misread it and it's noise. The event is present in almost every hard power loss — PSU, thermal, someone pulling the plug — so its mere existence proves only that the power died uncleanly. What separates the causes is what's around it: a bug check (a blue screen code) logged just before means the OS caught a fault first, which points at drivers, memory, or a GPU fault; a bare 41 with nothing around it is a clean hard loss, which points at power delivery or thermal protection. This read-only query pulls the last ten:

The limit to keep in mind: a 41 entry is the OS describing a corpse, not performing an autopsy. It tells you the power died without warning and whether software saw anything first — it cannot tell you what cut the power. That's still the pattern + temperatures + swap test doing the work.

Shutdown vs restart: which one is your PC doing?

The distinction is diagnostic gold, because the two mechanisms are different. A restart — the machine powers off and, seconds later, POSTs again on its own — means something ordered a reboot: a driver crash, a Windows Update action, a BIOS setting, a thermal watchdog, or a brief voltage collapse (brownout) long enough to reset but not to stay down. A clean shutdown to a dark machine means the power was cut and nothing restarted it — thermal cutoff, a protection trip, or a supply that simply stopped. And a blue-screen-then-off means the OS caught a fault it could not survive, which is a different suspect list again (memory, drivers, GPU).

So before any testing: watch the next failure, or ask yourself what the last one looked like. "It came right back" and "it stayed dead" are two different first answers, and they route you to different branches of the decision tree below. People who describe their problem as "it just shuts down" are usually holding one of these three and don't know which yet.

Transients: why gaming is a different kind of load

A modern gaming workload doesn't draw power steadily — it swings. The GPU and CPU can each step their demand up within a few milliseconds, and the combined spike (a "transient") can be meaningfully higher than the average draw of the same game. The PSU's 12 V rail has to absorb those swings without sagging, and that's where two different failures show up: a supply that's undersized for the system hits its limit on the biggest transients, and a supply that's aging — capacitors dried out, regulation degraded — hits a limit it never had before. From the user's chair they're identical: the machine is fine, a heavy scene starts, and the power cuts. The label and the age of the unit are what tell them apart, and the swap test confirms either.

Capacity vs health: two "not enough" problems

Both cause load-only shutdowns, and both end in a new PSU — but they're different purchases. A capacity problem has a clean story: you upgraded the GPU (or added drives), the system now draws more than the supply was sized for, and the failures started around the upgrade, at the same high draw, every time. The fix is sizing up honestly (Section 11). A health problem has no clean story: the same system that ran this workload fine for two years now fails at draws it used to handle, the failures creep in over weeks, and the unit is several years old. The fix is replacement at whatever wattage the system actually needs. "It used to be enough and now it isn't" is the sentence that separates health from capacity — and it's the question to ask before looking at any wattage number.

Check thermals first — the load failure that isn't electrical

Before the supply enters the conversation, the failing session needs temperatures. Run the game or workload that kills the machine with a monitor visible (the taskbar overlay, or a second screen), and watch: a GPU in the mid-80s °C and climbing, or a CPU at its 100 °C limit, is about to execute a thermal shutdown — which cuts power to the offender and, depending on the board, can take the whole system with it. That looks exactly like a load failure, costs nothing to test, and is the most common answer in hot rooms and dusty cases. If the shutdown tracks the temperature curve — happens as the temp crosses a line, gets more likely as the room warms, or disappears on a fan-cleaning anniversary — it's cooling, and no amount of PSU testing will find it.

Overclocks and XMP: the free things that mimic a dying supply

An unstable CPU overclock, GPU undervolt/overvolt, or memory profile (XMP/EXPO) produces load failures that are indistinguishable from a PSU problem — same timing, same clean cutoff, same "works at the desktop" behavior — because an unstable voltage or timing is a power-delivery problem, just one your BIOS created instead of the supply. The test is free and takes one reboot: reset to stock (default clocks, memory at JEDEC or the profile disabled), and run the failing workload. If the failures stop, you had a tuning problem wearing a PSU costume, and the supply was innocent. If they continue, the overclock is cleared and the suspicion list tightens back toward thermals and the supply.

Smells and sounds: which ones end the investigation

Smell is the unambiguous one. A burning-plastic, hot-electronics, or sharp ozone smell coming from the PSU's area means stop immediately: unplug, don't power on again to "listen for it," and replace the unit. If the smell is faint and the cause is ambiguous (the GPU's coils also get hot-smelling under sustained load), identify the source with the case open and everything off before deciding — but when in doubt, the supply retires. Sounds are the nuanced one: a steady, tone-varying coil whine from the GPU or board is usually harmless (vibrating inductors, not failure); a buzz, crackle, or click from the PSU's own enclosure is not normal and is a reason to replace rather than investigate further; and a fan that's grinding rather than whirring is a bearing ending its life, which on a PSU is a replacement signal even before the electronics care.

The fan question: silent at idle is not a failure

The most-misread PSU symptom is a fan that doesn't spin. A large share of modern supplies use a zero-RPM (hybrid) fan curve: at low load and low temperature the fan is deliberately off, and it wakes as the unit warms. So "PSU fan not spinning" at the desktop, with everything working, is frequently by design — and blowing into the vents to check, or concluding the unit is dead from a quiet idle, are both mistakes. The useful observations are the opposite edges: a fan that never spins even under sustained load (the unit may be running hot with no cooling), a fan that spins erratically or stutters, or a fan that recently started making new noise. Judge the fan across a range of loads — idle, a game, a stress test — and only the erratic end of that range is a signal.

Partial failure: the state that fools every cheap test

The most common condition of a dying PSU is not "dead" — it's "intermittently wrong." The unit passes the paperclip start, spins up fine, delivers correct voltages at rest to a multimeter, and sits on the bench passing a tester — and then collapses the 12 V rail the moment a game asks for a transient spike. This happens because the degradation (dried capacitors, weak solder, aging rectifiers) shows up under the specific electrical stress of load, which is exactly the condition none of the cheap tests apply. Two consequences follow. First, "I tested it and it's fine" is not a conclusion about a PSU that fails under load — it's a conclusion about a PSU at rest. Second, the only tests that see this state are the load test and the known-good swap run through the failing scenario, and that's why they anchor Section 09.

Aging: what actually happens inside over the years

The dominant aging mechanism is the electrolytic capacitors: they dry out over years of heat and duty, their capacitance drops, and the filtering and regulation they provide degrade — which is precisely the "works at rest, sags under transients" behavior of the partial failure above. Secondary mechanisms add up too: solder joints cycling with heat, fan bearings wearing, and dust insulating the cooling. The practical numbers: quality supplies in moderate conditions commonly last five to ten years; a unit in a hot case feeding a power-hungry GPU ages at the top of that range or worse. Age is not a fault — a nine-year-old quiet, stable supply that passes load is fine to run — but age changes how you weigh every other sign: the same new noise or the same load failure on a two-year-old unit and a nine-year-old unit are different diagnoses.

Overheating vs a failing supply

The two produce the same visible event — a clean power-off under load — and the separator is temperature. Thermal shutdowns are caused by the GPU or CPU cutting their own power at a temperature limit (or a board-level cutoff), and they correlate with the temp curve: the failure arrives as the temperature crosses a line, recurs in hot conditions, and improves when cooling is cleaned or improved. A supply's load failure doesn't care about the temperature curve — it arrives at a specific electrical draw, in any room temperature, and persists after a fan-cleaning. The test that splits them: watch temperatures during a failing session. If the failure tracks the temps, fix the cooling and retest before touching the power story; if the temps are comfortably in range at the moment of failure, thermal is cleared and the electrical suspects (overclock, then supply) take over.

The GPU's turn: when the "PSU" is the graphics card

A failing GPU — its VRMs, its memory, its power stage — shuts the system down under exactly the same conditions a failing PSU does: heavy load, clean cutoff, desktop unaffected. The two are separable in three ways. First, the event log and crash artifacts: a GPU fault often leaves a bug check or a display-driver timeout record; a power loss leaves a bare Kernel-Power 41. Second, the load that triggers it: a GPU failure is often tied to specific work (a particular game, ray tracing on, a render) in a way a capacity-limited supply is not — the supply trips at whatever draw crosses its limit. Third, and conclusively, the known-good PSU swap: if the failure follows the old supply, it's the supply; if it persists with a healthy supply in the box, the GPU (and its own power path — the PCIe cables and the card's connectors) becomes the primary suspect. A dying GPU is also the reason "new PSU, same problem" happens more often than people expect.

Motherboard and RAM: the other half of "no power"

For the dead-machine symptoms — no power, standby-only, spins-then-off — the motherboard is the PSU's chief rival. A short on the board trips the supply's protection (the classic "spins for a second" machine), a failing VRM section mimics a weak supply, and a dead power sequence on the board makes a healthy PSU look inert. The minimal-configuration test is the separator: a board that stays on with CPU and one RAM stick but dies when the rest returns is pointing at the removed components one by one; a board that dies even at minimal config, with a known-good PSU in it, is the diagnosis. RAM enters the same symptoms through the same test — a bad stick or a bad slot produces "spins, beeps or doesn't, no POST" and is cleared by the single-stick, different-slot sequence. Neither component can be cleared from the outside by any test that doesn't involve isolating it.

When it's the OS: the shutdowns Windows causes itself

A meaningful share of "random shutdown/restart" reports end in the operating system: a driver crash that reboots the machine, a Windows Update applying and restarting, a power plan with aggressive sleep, USB selective suspend dropping a device that the machine then reacts to, or a watchdog service configured to reboot on error. The tell is in the record — a restart ordered by the OS leaves a shutdown/restart event with a reason code in the log, which a hard power loss never does; a blue screen before the off is the OS catching a fault, not power dying. The check is the same event-log read from Section 05: if the entries around the failures show software ordering the event, the power hardware is out of the case entirely, and the fix is a driver, a setting, or an update — free, and often boring, which is the best ending this page has.

The known-good swap: the only test that's a verdict

The method: remove the suspect, install a known-good supply of adequate wattage and the right connectors, and run the system through the scenario that fails — the game, the render, the idle overnight, whatever the pattern is. If the failure follows the old unit (gone with the swap, back when it returns), the verdict is in and the rest of the machine is exonerated. If the failure persists with the good supply in the box, the PSU is cleared — a genuinely useful negative result, because it sends the investigation to the GPU, thermals, or board instead of a second PSU purchase. Two requirements make it valid: the donor must be known-good (not "another old unit") and adequate for the system, and the test must reproduce the failing condition — a swap that only idles the machine has tested idle, which the suspect PSU was probably fine at.

PSU testers: what a pass and a fail actually mean

A bench or plug-in PSU tester checks the resting rail voltages and, on better models, the protection circuits' response — useful, and the right first tool when you have one. Read the results correctly: a fail is strong evidence the unit is bad (though confirm which rail or protection failed — a weak 5 V standby on an otherwise fine unit is a different story than a dead 12 V). A pass is weaker than it feels: it means the unit presents correct voltages at rest, which is exactly the condition a partially failing supply fakes. So the honest reading of "tester says fine, PC still dies under load" is "the supply remains a suspect" — the tester has narrowed the field, not closed it, and the load test or the swap is what closes it.

Multimeter testing: resting rails and their blind spot

With the unit powered standalone (started per the paperclip method below) or in the system, a multimeter on the 24-pin's rail pins reads the DC voltages: 12 V, 5 V, and 3.3 V, each acceptable within roughly ±5% of nominal. A reading well outside the window (say 11.2 V on the 12 V rail) is real evidence of a bad unit. The blind spot is the same one the tester has, magnified: at rest, the rails are barely loaded, and a supply whose capacitors have dried out will hold correct voltage at near-zero current and sag the moment the GPU draws. "My multimeter says 12.01 V" and "my PC dies in games" are both true, and only the second one is the diagnosis. Treat a static reading as a health check with a narrow field of view, not a certificate.

The paperclip test: what it proves, and the three things it can't

The "paperclip test" (properly, a jumper start) powers a PSU on without a motherboard by bridging the green PS_ON wire to a ground wire on the 24-pin connector — the fan spins, and the unit is "working." What that proves is precisely one thing: the unit will start. What it cannot prove, the three things people wrongly conclude from it: that the rails are correct (a started unit can still read wrong voltages), that it holds them under load (the partial-failure state passes this test every time), and that the protections are alive (a unit can spin with dead over-voltage protection). Use it for what it is — a quick "is this unit completely dead" check — with the cautions: unplug the unit first, bridge only PS_ON to ground (a wrong pair of pins can damage the unit), work only on the low-voltage DC connector, and understand that a failed paperclip start on a unit with no other symptoms is often a bad test, not a bad PSU.

Load testing: the only test that reproduces the failure

Load testing applies real draw to the rails and watches what happens — the condition where partial failures actually fail. In-system, it means running the failing scenario (the game, the render, a CPU+GPU stress tool) with the monitoring visible: temperatures, and the moment of failure. On the bench, a load box applies controlled draw per rail — more precise, and a service task rather than a home one. The honest limitation of home load testing: consumer stress tools load the CPU and GPU through the system, which is exactly the path a failing supply lives on, so it's a good reproduction — but if the unit only fails on a specific rail the workload never stresses, the test passes and the mystery continues. Which is why the load test and the swap are companions: the load test reproduces, and the swap adjudicates.

Keep a test log — the pattern is half the diagnosis

PSU problems are pattern problems, and the pattern lives in your notes, not in the hardware. Before and during any testing, log each session with the template below — the scenario, the temperatures, what happened, whether a swap was in, and the result. Two or three logged sessions usually convert "it's random" into "it's the 12 V rail at roughly this draw," which is a statement every test above can act on:

When replacement is the right call

Five situations, in descending clarity. The known-good swap gave the verdict — replace, no further discussion. The unit has burned smells, visible damage, or failed its protections repeatedly — replace, and don't run it in the meantime. It survived a lightning strike or a surge — retire it or verify it properly; a supply that absorbed a spike can be quietly wrong in ways no resting test sees. It's aging (roughly five-plus years of hard duty) and showing any warning sign — the sign plus the age is the diagnosis. And you're sizing up for a new GPU and the current unit is at or under the recommended draw — replace as part of the upgrade, because "at the edge" is where transient failures live. The economics settle any hesitation: a power supply is the cheapest major component in the machine, and the one whose job is to keep the expensive ones alive — a cheap PSU that fails its job is the most expensive part in the box.

Sizing honestly: the wattage math without the guesswork

The sound method anchors on the components, not on a percentage folklore. Find the GPU's power recommendation — the manufacturer states the recommended PSU wattage for that card, and it already includes the system overhead around it; that number is your floor for a typical build. For a CPU that's a desktop flagship, a water-cooled overclock, or a very drive-heavy system, add headroom on top. Then check the shape, not just the total: the 12 V rail's combined amp rating is the rail your CPU and GPU actually draw from, and a unit whose headline wattage looks adequate but whose 12 V rail is thin is a unit that's inadequate in the one place it matters. Two anti-patterns to avoid: sizing for idle (idle draw tells you nothing about load), and buying 1000 W "to be safe" for a mid-range system (you'd be paying for capacity the system will never touch, in a unit that runs at its least efficient loads).

Quality: what the label tells you — and the two things it doesn't

The 80 Plus badge is an efficiency rating: how much of the drawn power becomes DC instead of heat, measured at defined load points. It says nothing about protection quality, capacitor grade, or how the unit behaves at the edge of its rating — a unit can earn a badge and still be a poor power source, which is why "it's 80 Plus Gold" is not an answer to "is it good?" What actually separates supplies: the protection circuitry (OVP, OVP per rail, OCP, OPP, SCP, OTP — present and correctly set), the component quality (capacitor makers and grades, topology), and the reputation track record — the tier lists the community maintains from teardowns and long-term failures encode years of exactly this information, and consulting one is the cheapest quality test available. Read the label's fine print alongside the badge: the per-rail amp table is the unit's real specification, and the 12 V row is the one your system lives on.

Modular cables: the compatibility trap

Modular PSU cables look interchangeable and are not. The pinouts on the PSU-end connector differ between brands, between models, and in some cases between production runs of the same model — because those pins carry sensing lines (power management, fan control, per-rail identification) that aren't visible in the plastic. Plugging a cable from one unit into another can read fine at the connector and still mis-sense a rail, which in documented cases has damaged the GPU or the board, not just the cable. The rules: use the cables that ship with the unit; if a cable is lost, buy the replacement for that exact model; never mix cables between different units "because they fit." The one related trap for newer builds: the 12VHPWR/12V-2×6 GPU connector — use the proper native cable or the manufacturer-approved adapter, and seat it fully, because that connector carries the entire GPU load on contacts that are also where the well-publicized melting failures happened when adapters or partial seating were involved.