GPU VRM Failure Explained: MOSFET Shorts, No-Power Symptoms & Repair Diagnosis
Key Takeaways
A graphics card's VRM steps 12V down to the low voltage the GPU core and memory need. When a MOSFET shorts, a phase drops out, or a controller fails, symptoms range from instant shutdown to crashes under load. If a short is suspected, stop powering the card on repeatedly, since each attempt can spread the damage.
Introduction: When the Card Itself Is the Power Problem
A graphics card is plugged in, the power button is pressed, and the whole machine instantly goes dead. Or the machine boots, the desktop looks fine, and the system crashes the moment a game or a render starts. Or there is a smell of burning and a scorch mark near the power connector. The usual suspects get blamed: the power supply, the motherboard, the cable, the drivers. Sometimes one of those is indeed at fault. But there is another possibility that sits between the power connector and the GPU itself, and it is one that is easily misdiagnosed: the card's own on-board power delivery circuit, the voltage regulator module, or VRM.
The VRM is the part of the graphics card that turns the 12V arriving from the power supply into the low voltage, high current supply that the GPU core and the memory actually run on. When this circuit fails, the symptoms can look exactly like a dead power supply, a dead motherboard, or a dead GPU — which is why so many cards are written off, or so many power supplies replaced, for a fault that was on the card's own power stage all along. This guide explains how the VRM works, what happens when each of its parts fails, how to tell a VRM fault apart from the look-alike faults, and how a repairer decides whether the card can be saved.
What the VRM Is: A Step-Down Converter on the Card
Think of the VRM as a small, fast, tightly regulated substation sitting on the graphics card. The power supply delivers roughly 12V. The GPU core and the memory need something around 1V, delivered at high current and held stable even as the load changes in microseconds. Feeding 12V straight into a GPU core would destroy it. The VRM's job is to step that voltage down, smooth it, and hold it steady.
Each phase of the VRM is a synchronous buck converter, built from four basic parts:
| Part | What it does |
|---|---|
| PWM controller | The brain: it watches load demand, switches the MOSFETs on and off at high speed, and adjusts the output to hold the target voltage. The GPU tells it what voltage is wanted through signals such as VID (voltage identification). |
| High-side and low-side MOSFETs | The switches: they chop the 12V input into pulses. The high-side MOSFET connects the output toward the 12V input; the low-side MOSFET connects it toward ground. They take turns, never both on at once. |
| Choke (inductor) | The energy store and smoother: it resists sudden changes in current and, together with the capacitors, turns the switched pulses into steady direct current. |
| Capacitors | The reservoir and filter: they store charge and filter out voltage ripple, so the output stays smooth under rapidly changing load. |
The PWM controller monitors what the load needs and adjusts the switching continuously. When the GPU suddenly demands more current — a scene change in a game, the start of a render — the VRM has to respond immediately. That fast, closed-loop regulation is what separates a working VRM from a crude voltage dropper.
Multi-Phase Power: Sharing the Load
Modern graphics cards do not rely on a single phase. They use a multi-phase VRM: several identical phases working in turn, each carrying a share of the total current. Splitting the current across phases means each phase carries less, runs cooler, and the combined output is smoother than any single phase could deliver on its own.
Card specifications often describe the arrangement as something like "8+2". The first number is the number of phases feeding the GPU core; the second covers the memory and other sections. That notation describes how the phases are divided, not a quality score. More phases do not automatically mean a better power stage: the quality of the components, the current each phase can actually carry, the cooling of the power stage, and the design of the printed circuit board all matter just as much. Comparing two cards by phase count alone is a mistake.
Some cards use DrMOS (Driver-MOSFET) packages, which integrate the driver together with the high-side and low-side MOSFETs in a single package. Integration shortens the signal path, allows higher switching frequency, and raises power density. From a repair point of view, the difference matters: a failed DrMOS device is replaced as a whole package, rather than as individual discrete MOSFETs. It also looks like nothing more than a small square block on the board, so someone who does not recognise it may not realise it is the switching element of the power stage.
More Than One Rail, and a Start-Up Order
A graphics card does not have one single supply voltage. It has several power rails, each feeding a different part of the card: the core voltage rail (VCore / NVDD), the memory voltage rail (VMem / FBVDD), PCIe-related rails (PEX), and auxiliary rails such as 5V, 1.8V, and 3.3V.
These rails do not all switch on at once. They start up in a sequence, one stage at a time. A diagnostic guide for NVIDIA Pascal (GTX 1000 series) cards published on Repair Wiki records a typical order of 5V, then 1.8V, then VCore, then VMem / PEX. If one stage fails to start, the later stages never start either.
That sequence must not be treated as a universal rule. Power-up order varies by generation, manufacturer, and board design; the Pascal order is an example from one documented design, not a truth that applies to every card. Any diagnosis that assumes a fixed order for an unknown card is building on sand. What is general is the principle: rails start in an order, and a missing rail means the diagnosis stops at the stage that failed.
Behind each rail there are also enable signals and power-good signals. A controller only starts working when it receives its enable signal, and once its output is stable it issues a power-good signal telling the next stage it may start. A missing voltage can therefore mean the output components have failed, but it can equally mean the controller has failed, the enable signal never arrived, or an earlier rail never came up. A missing rail is a symptom with several possible causes, not a diagnosis on its own.
Why a VRM Fails
VRM components carry real electrical and thermal stress, and several mechanisms can push them past their limits. These are possible causes, described as mechanisms — not a ranking, and not a claim about how often each one occurs.
- Heat and poor cooling. Switching and conduction losses in the MOSFETs generate heat. If the thermal pads are not making proper contact, if there is no airflow over the VRM area, or if heat builds up inside the case, sustained high temperature accelerates ageing of every component in the stage.
- Current surges and abrupt load changes. Heavy load, overclocking, and supply fluctuations all add stress to the switching components.
- Capacitor ageing. Electrolytic capacitors are sensitive to heat. Ripple current flowing through a capacitor's internal resistance generates heat inside it, and over time that can dry the capacitor out or make it fail. As filtering degrades, output ripple grows and stability under load gets worse.
- Component defects or batch problems, and physical damage. A cracked or knocked-off component, or corrosion from liquid exposure, can disable a phase or create a short.
- Abnormal input power. A poor-quality power supply with large ripple, or a connector with bad contact that overheats locally, can stress the on-board power stage.
No single cause explains every VRM failure. Treating any one of these — overclocking, for example — as the cause of all VRM faults would be wrong.
VRM Failure Modes: Symptom Against Mechanism
Different VRM faults produce different symptoms. Learning the mapping between them is what turns "the card is dead" into a useful diagnosis.
| Failure mode | Mechanism | Typical symptoms |
|---|---|---|
| High-side MOSFET short (drain-source short) | 12V connects directly to the output | The most dangerous failure: the core or memory may be exposed directly to 12V and destroyed. The power supply protection trips instantly. There may be a burning smell or scorch marks. Even if the MOSFET is replaced, the core downstream may already be dead. |
| Low-side MOSFET short | Output rail shorted to ground | Instant shutdown or power supply protection trip on power-on; the affected rail measures extremely low resistance to ground. |
| MOSFET open circuit / one phase dead | That phase stops delivering power | The remaining phases must carry the whole load. The card may still boot under light load, then become unstable, throttle, or crash under heavy load. Over time the surviving phases overheat, and the failure can cascade from one phase to the next. |
| PWM controller failure | No drive signals, or wrong timing | The whole group of phases served by that controller stops working, so the associated rail has no voltage at all. The card will not start, but there may be no short — resistance measurements can look normal. |
| Capacitor failure (dried out, shorted, burst) | Filtering fails, or a direct short | Large ripple leads to instability and crashes under load. If the capacitor itself shorts, the rail is shorted to ground and the machine trips on power-on. A failed capacitor may bulge, leak, or burst, but many failed capacitors look completely normal from the outside. |
| Choke (inductor) open circuit | That phase is disconnected | Equivalent to losing a phase: reduced power capability and instability under heavy load. Chokes are comparatively durable components and tend to fail less often than MOSFETs or capacitors. |
| Fuse blown | Upstream overcurrent protection has operated | The card shows no response at all. A blown fuse is a result, not a cause: before replacing it, the short or overcurrent that blew it must be found and removed, or the replacement will blow immediately. |
| VRM overheating (not yet destroyed, but unable to cope) | Over-temperature or overcurrent protection operates | The GPU core temperature can look normal, yet the card throttles or crashes as soon as it is loaded. VRM thermal pads not seated properly, or no airflow over the VRM area, are common triggers. This is a different condition from core overheating. |
The practical point of the table is that three broad symptom patterns — trips instantly on power-on, shows no response at all, and boots but dies under load — point toward different failure modes. Noting which pattern a card shows, before anything is taken apart, saves a great deal of diagnostic time.
First: Stop Powering the Card On
If a short is suspected — the machine trips the moment the card is installed, there is smoke, a burning smell, or scorch marks — do not keep switching the machine on to try again. This is the single most important safety message in this guide, and it belongs before any diagnostic step.
Every power-on attempt can widen the damage. With a high-side short, each attempt feeds 12V into components designed for roughly 1V. With any short, each attempt dumps current into the shorted point. The damage done in the first second can be the difference between a repairable board and a destroyed GPU core. Stop, unplug, and diagnose without power first.
What a Reader Can Safely Check
The steps below are divided into two levels: checks a reader can safely do without taking the card apart, and board-level diagnosis that belongs to a repairer with the right equipment. The boundary matters. Nothing in the reader-level list requires opening the card's power stage, desoldering components, or probing a live board.
Visual inspection, with no power applied
Under bright light, ideally with magnification, look for scorching, melted plastic, capacitors that are bulging or leaking, MOSFET or DrMOS packages that are cracked or burst, components that are missing or knocked off, signs of liquid corrosion, and discolouration of the board. If visual inspection finds clear damage, that already establishes a hardware fault, and there is no need to continue with software troubleshooting.
Isolation testing: is it the card, or something else?
- Remove the graphics card and boot with integrated graphics, or with another known-good card. If the machine boots, the problem very likely lies with the original card or its power cable.
- Swap in a known-good power cable and try a different PCIe slot, to rule out the cable and the slot.
- Install the suspect card in a different, known-good computer. If the same fault follows the card, that confirms the fault is in the card itself. This cross-test is the most reliable way to separate "the card is faulty" from "the motherboard or power supply is faulty".
Understanding the power supply trip signal
When a card has a hard short, the power supply's overcurrent and short-circuit protection (OCP / SCP) cuts the output immediately, and the machine looks completely dead. That is a protection action, not evidence that the power supply has failed. The machine returning to normal once the card is removed only shows that the source of the short is in the card. It does not prove the power supply is healthy, and it certainly does not justify fitting a larger power supply to a shorted card — a larger supply can only deliver more current into the short and burn the fault worse. The correct order is to confirm whether the card is shorted first.
Other safe reader checks
- If there is a burning smell, smoke, or scorching, cut the power immediately and do not try to switch on again.
- Check that the power cables are fully seated, and look for melting or discolouration at the connectors. If a connector itself has melted, that is a separate connector fault, covered by the guides on 12VHPWR and 8-pin connector melting.
- Check airflow and VRM cooling as routine maintenance: confirm there is airflow over the VRM area and that thermal pads are making contact. This covers cleaning and routine thermal maintenance only; it does not mean dismantling power components.
- Record the symptoms precisely: does the machine trip the instant the card is installed, does it boot and then die under load, or does it show no response at all? Those three patterns point toward different failure modes, and describing them accurately when arranging a repair saves diagnostic time.
Board-Level Diagnosis: Principles for Professional Repair
The rest of this section describes how a repairer approaches a VRM fault. It explains the reasoning and the traps, so a reader can understand what a repair involves. It is not a step-by-step repair tutorial, and board-level work should not be attempted without the proper equipment.
Resistance measurement with power off
With power removed, a multimeter is used to measure each power rail's resistance to ground. A diagnostic guide for Pascal cards on Repair Wiki uses 50 ohms or less on the 12V and 3.3V base rails as a reference threshold for suspecting a short. That figure is specific to that guide and to those cards. It is not a universal standard, and applying it to other generations or designs would be wrong.
The VCore measurement trap
The core voltage rail is a trap for the unwary. Its resistance to ground is naturally extremely low, because the GPU core itself is a very low-resistance load. Measuring VCore to ground and concluding "low resistance means short" will condemn healthy cards. The same Pascal diagnostic guide points out that directly measuring VCore resistance to ground is of limited help on those cards, and suggests measuring VCore against 12V instead. The general lesson holds beyond Pascal: a low reading on a low-voltage, high-current rail is not, on its own, proof of a short.
Powered voltage measurement, in sequence
Only after a short has been ruled out is the card powered up, and the rails are then measured one at a time in start-up order. The diagnosis stops at the first missing rail rather than skipping ahead. A missing voltage leads back to that stage's controller and to its enable and power-good signals: is the controller working, did the enable signal arrive, did the previous stage come up?
Hotspot location
Once a short is confirmed, the repairer needs to find where on the board it sits. Options include a thermal camera or infrared thermometer, and, in a professional environment with a laboratory power supply and current limiting, the alcohol evaporation method — alcohol applied to the board evaporates fastest on the component that is heating up. This technique belongs to a controlled bench with current-limited power, not to a home power supply, and it must never be attempted by feeding a shorted card from a standard PC power supply.
Low-voltage, current-limited injection
A related professional technique is voltage injection: feeding the shorted rail from a laboratory power supply at low voltage with a current limit, and using the voltage distribution and hotspot location to find the shorted point. Like the other bench techniques, it is described here as a concept, so readers understand what a repair involves. It is not an instruction to attempt it.
The in-circuit MOSFET test trap
A MOSFET contains an internal parasitic body diode, and the board has many parallel paths around any one component. Measuring a MOSFET while it is still soldered into the circuit therefore very easily produces misleading readings: a reading that looks like a short may only be the response of other parallel components or the body diode, and does not mean that particular MOSFET has failed. Reliable identification usually requires removing the suspect component and measuring it off the board, or cross-confirming with several independent tests — resistance, powered voltage, and hotspot location together. A multimeter beep on an in-circuit MOSFET is not, on its own, a diagnosis.
There is also a repair principle specific to multi-phase VRMs: if one phase has burned out, the remaining phases in the same group have often already carried excess stress. Replacing only the visibly burned component without checking the rest of the group invites a repeat failure. That is a matter of repair judgement, not a guarantee about any particular card.
Telling a VRM Fault Apart from the Look-Alikes
Several other faults produce symptoms that resemble VRM failure. Each has a distinguishing test.
| Other fault | How it differs from a VRM fault |
|---|---|
| Power supply too weak or failing | The external supply cannot deliver enough power. The distinguishing test is isolation: a known-good supply, or the suspect card in a known-good machine. A VRM fault persists even with a healthy power supply. |
| Melted 12VHPWR / 8-pin connector | The damage is at the connector or plug itself — melting, discolouration, a loose or heat-damaged plug. The VRM can fail with the connector looking perfectly normal, and a melted connector is a connector fault even if the VRM is fine. |
| VBIOS damage or a wrong BIOS flash | A firmware problem. The VRM can be faulty with the firmware intact, and a firmware fault can exist with a healthy VRM. Firmware troubleshooting does not fix a shorted power stage. |
| VRAM versus GPU core fault | Those faults typically show visual artifacts — patterned corruption, coloured blocks, garbled output — and are diagnosed with artifact patterns and tests. A VRM fault typically shows no power, instant shutdown, or crashes under load, without artifact patterns, and is diagnosed with multimeter and voltage measurement rather than artifact analysis. |
| Card not detected (PCIe / motherboard / BIOS) | A system-level detection problem. A VRM fault can make a card completely undetectable, because a missing power rail means the card never starts. The root cause in that case is the missing rail, not the slot or the settings. |
| Fan not spinning / fan PWM fault | The fan PWM control circuit is a different circuit from the VRM PWM controller. Both are called PWM, but they are entirely separate circuits. A fan fault does not mean the power stage has failed, and a VRM fault does not necessarily stop the fan. |
| Coil whine | A noise problem: the inductor vibrates under changing load and produces an audible whine. Coil whine on its own generally does not indicate VRM failure, and it is not evidence that the power stage is broken. |
Where monitoring software fits in is also worth stating. A guide on reading GPU-Z and HWiNFO sensor data covers what those tools can show. But a card with a severe VRM fault may never power up or be detected at all, in which case monitoring software reads nothing. That gap — what to do when the software cannot see the card — is exactly what this guide covers.
Common Mistakes
- "Bake it in the oven / blast it with a hot-air gun." Wrong. A VRM fault is an electrical failure of a component, not a cracked solder joint. Heating will not repair a shorted MOSFET, and it can widen the damage. The oven myth is covered in detail by the existing guides on reflow myths; it does not apply to VRM failure.
- "It whines, so the power stage is broken." Coil whine is an inductor vibration noise and generally does not indicate VRM failure. Noise alone is not evidence of an electrical fault.
- "VCore resistance is low, so the rail is shorted." VCore is naturally a very low-resistance rail. A low reading to ground is expected and proves nothing on its own.
- "The multimeter beeped, so that MOSFET is the culprit." In-circuit measurement is full of false positives from parallel paths and the body diode. Cross-confirm before blaming a component.
- "A bigger power supply will fix it." If the card is shorted, a larger supply only feeds more current into the short. Confirm whether the card is shorted before changing the supply.
- "More phases means a better power stage." Phase count is one design choice among several. Component quality, per-phase current capability, cooling, and board design matter as much.
- "The fuse blew, so replacing the fuse fixes it." The fuse is the result, not the cause. Without finding the short or overcurrent that blew it, the replacement blows again.
- "The core temperature is normal, so the card is not overheating." VRM overheating can trigger protection, throttling, or crashes while the core temperature looks normal. Watching only the GPU temperature misses this possibility.
- "The fan PWM and the VRM PWM are the same thing." They share a name and nothing else. They are separate circuits with separate jobs.
- "The Pascal start-up order applies to every card." It is one documented example. Start-up order varies by design, and the card in front of you is the authority.
Repair or Retire: The Decision Framework
The decision between repair and retirement does not turn on whether a component can be replaced. It turns on whether the downstream components survived the fault.
Repairable cases tend to look like this: a single component has shorted — a MOSFET, a capacitor, a fuse — and the core and memory were not exposed to the fault; or the PWM controller has failed without any rail being driven to a damaging voltage. In those cases, replacing the failed component and checking the rest of the affected group can restore the card.
Cases that are likely beyond repair tend to look like this: a high-side short has already fed 12V into the core and destroyed it internally; several rails are shorted at once; or the board itself has burned through or its copper traces have lifted. In those cases, replacing the visibly failed component cannot bring back what the fault already destroyed.
The key question, then, is not "can this component be swapped?" but "did the downstream parts survive the fault while it was happening?" That is the question a repairer answers with resistance measurements, voltage checks, and inspection — and it is the question a card owner needs answered before deciding whether to send a card in or let it go. No figure for cost or repair time is given here, because neither can be honestly stated without seeing the specific card.
Prevention: Reducing VRM Stress
No maintenance routine can make a VRM immune to failure, but a few habits reduce the stress it operates under. Keep airflow moving over the VRM area rather than letting the card bake in still, hot air inside the case. Treat thermal pads and paste as maintenance items: pads that no longer make contact leave the power stage to cool itself by radiation alone. Avoid running a card with a power supply whose output is suspect, since abnormal input power stresses the on-board stage. And when a card shows early warning signs — instability that appears only under load, a new burning smell, a connector that runs hot — stop and investigate rather than pushing through. Early attention does not guarantee a save, but repeated power-on attempts after a fault has declared itself can turn a repairable fault into a destroyed core.
FAQ
Q: Can a VRM fault damage the motherboard?
A: A short on the card can trip the power supply's protection and shut the machine down, which is the protection working as designed. Damage beyond the card depends on the fault and on how many times the machine was powered on after the fault appeared. That is one more reason to stop powering on a card that trips the supply, and to isolate-test it rather than repeatedly retrying it in the same machine.
Q: Will replacing the power supply help?
A: Only if the power supply itself is the fault. If the card's VRM is shorted, a new or larger power supply does not fix anything — it can make the damage worse by delivering more current into the short. The correct order is to establish whether the card is shorted, using the isolation tests in this guide, before spending anything on a new supply.
Q: Why does the card crash in games when the GPU temperature looks normal?
A: Core temperature is only one of the card's thermal stories. If the VRM area is overheating — because its thermal pads are not making contact, or because there is no airflow over the power stage — the VRM's own over-temperature or overcurrent protection can throttle the card or shut it down while the core temperature reads normally. A fault in a phase can produce the same pattern: the card copes at light load and fails as soon as the load rises.
Q: Is coil whine a sign of VRM failure?
A: Generally no. Coil whine is the inductor vibrating under changing load and producing an audible sound. It is a noise phenomenon, not evidence that the power stage has electrically failed. A card can whine and be electrically healthy, and a card can have a serious VRM fault without any whine at all.
Q: The card shows no signs of life at all. Is it definitely a VRM fault?
A: Not necessarily. A completely dead card can also mean a blown fuse, a failed controller, a fault in an auxiliary rail, or a problem outside the card entirely. No response is a symptom with several possible causes. The isolation tests — the card in another machine, a known-good cable and slot — narrow the possibilities before any board-level conclusion is drawn.
Q: Can I test the MOSFETs with a multimeter at home?
A: Not reliably. In-circuit measurement is distorted by the body diode inside each MOSFET and by the many parallel paths on the board, so a reading that looks like a short may be entirely misleading. Reliable identification normally requires removing the suspect component and measuring it off the board, combined with resistance, voltage, and hotspot evidence. A single in-circuit beep is not a diagnosis.
Q: My card has a high phase count. Does that mean its VRM cannot fail?
A: No. Phase count describes how the load is divided. Component quality, per-phase current capability, cooling, and board design all affect whether a power stage survives its working life. A high phase count does not make a card immune to any of the failure mechanisms in this guide.
Q: The fuse blew. Can I just replace it?
A: Replacing the fuse without finding what blew it is not a repair. The fuse blew because of a short or an overcurrent event somewhere upstream of it. Until that cause is found and removed, a replacement fuse will blow the same way.