Games Run Fine but Benchmarks / Stress Tests Crash, Go Black or Reboot? Game Load vs Synthetic Full Load & Power Margin Diagnosis Guide

2026-11-15 | GPU Repair | 28 minutes | HK GPU Fix Editorial Team

Games Run Fine but Benchmarks / Stress Tests Crash, Go Black or Reboot? Game Load vs Synthetic Full Load & Power Margin Diagnosis Guide

Key Summary

A game that runs fine proves only that ordinary play never reached the power or cooling boundary. A benchmark can hold full load long enough to reach it, so a crash under test and stability in games can both be true. Compare crash form and timing, return to stock settings, and test power, temperature, and combined load in order.

The Opposite Shape: Games Fine, Benchmarks Fail

A computer plays games for hours without a problem. Then a benchmark or stress test is started, and within minutes the screen goes black, the driver resets, the picture freezes, or the whole machine reboots. The first reaction is usually to conclude that the test has proved the graphics card is broken. That conclusion is too fast, and it is also the wrong shape of question.

This guide is about the opposite shape from the familiar one in which a game and a benchmark both crash. In that familiar shape, the same trigger appears in both places, and the fault is likely to show up under any heavy load. Here the trigger appears in only one place: ordinary games are stable, and only a synthetic test fails. Both facts can be true at the same time, because a game and a synthetic test do not place the same load on the machine. A game that runs fine proves only that ordinary play never reached a particular boundary. A benchmark can hold the machine at full load long enough to reach that boundary for the first time.

This guide covers desktop computers where games run normally, including long sessions, while a benchmark or stress test crashes, goes black, resets the driver, freezes, or reboots the machine. It does not cover the case where games and benchmarks both crash, which is a different shape with a wider range of suspects. It does not cover a score-reading question — whether a benchmark number is high or low — which is a different subject altogether. The core claim is simple: the test is not the game, and a crash under test is evidence about a boundary, not a verdict that the card has failed. Record the form of the crash and the time at which it happens, note which test failed, return to stock settings, and then work through power, temperature, and combined load in order, instead of deciding at the start that the card must be replaced.

Why a Benchmark Is Not a Game

The central mechanism of this guide is that game load and synthetic load are different kinds of load, not different amounts of the same thing.

A real game varies its load constantly. Scenes change, the camera moves, some frames are light and some are heavy, and the graphics card's power draw rises and falls with what is on screen. Even a demanding game with ray tracing spends much of its time below the card's maximum power draw, and it rarely holds that maximum for long without a break. That variation matters, because a marginal power supply, a warm voltage regulator, or a borderline setting may never be pushed far enough, for long enough, during ordinary play to show itself.

A synthetic test is designed to remove that variation. FurMark, in particular, is a power virus: it renders a fur-covered donut using extremely heavy shaders and pushes the graphics card to its maximum power draw, drawing far more power and generating far more heat than any real game ever would. That is not a flaw in FurMark; it is its purpose. It is a deliberate worst-case load, and it should be read as one. A machine that passes FurMark has survived an unusually severe sustained load. A machine that fails FurMark has failed that severe load, which is not the same as failing in ordinary use.

Two consequences follow, and both are easy to get wrong. First, a graphics card stable under FurMark may crash in games, and vice versa. FurMark's load is single-minded: sustained power and heat. A game adds other stresses — changing shader work, heavy memory access, ray tracing, and rapid changes in load — that FurMark does not reproduce. Passing one does not prove the other. Second, a single pass of any test is useful evidence, not proof of long-term health. An intermittent fault can be entirely absent from a short test and still be present in the machine, and a test that passes once has shown only that this test, at this time, did not trigger the fault.

NVIDIA's own warning belongs here, because it sets a hard boundary on what testing may involve. FurMark can push power demand far beyond real applications, which can trigger the card's over-temperature or over-current protection. Those protections must never be disabled in order to run a test. Disabling a protection to get a test to run can permanently damage the card and affect its warranty. A test that only passes with a protection switched off has proved nothing except that the protection was doing its job.

Four Crash Forms Point at Different Layers

People say "it crashed" as though that were one event, but this guide's crashes take four different forms, and each form points at a different layer. Naming the form correctly is the first diagnostic act, because the four forms are not interchangeable.

A reboot — the whole machine restarting by itself — is the most serious form. A shutdown is more serious than an application crash. It may point to power supply protection, poor power delivery, overheating, graphics card failure, cabling, motherboard problems, or instability elsewhere in the machine. A reboot is a system-level event, not a game-level one, and it should be treated as such.

A black screen is a different form. The picture disappears, the monitor may report no signal or go to sleep, and the machine may or may not still be running underneath. A black screen that persists belongs with the guide on a persistent black screen; only the shape is used here: a picture that vanishes under load, rather than a picture that never appears.

A driver reset is a third form. The screen flickers or goes black for a moment, the picture returns, and an error message reports that the display driver stopped responding and recovered, or the game closes with a driver error. That form points at the driver layer, including the timeout detection and recovery mechanism; only that mechanism is used here, without repeating its full diagnosis.

A freeze is a fourth form. The picture stops moving, the machine stops responding to input, and nothing recovers without a forced restart. A freeze under load belongs with the guide on freezing; only the shape is used here: a machine that locks up rather than closing a program cleanly.

Write down which of the four forms occurred before changing anything. A reader who records "it crashed" has recorded almost nothing. A reader who records "the machine rebooted three minutes into FurMark, at stock settings, with the driver version noted" has recorded something a later step can actually use.

The Five Layers of Suspicion

Once the crash form is fixed, the fault can be placed in one of five layers. Each layer has its own decisive check, and the order runs from the easiest to confirm to the deepest.

LayerWhat it coversDecisive check
1. Power boundary layerThe sustained full load of a benchmark reaches the limit of the power supply, the cables, or the connectors — a limit ordinary games never reachedLower the power limit by about a fifth and re-run the same test; if the crash disappears, the power boundary is implicated
2. Cooling boundary layerCore temperature, hotspot temperature, or memory junction temperature climbs under sustained load until a limit is reached, or the voltage regulators overheatRecord all three temperatures at the moment of the crash and see which one reached its limit first
3. Driver reset layerThe crash takes the form of a driver stopping and recovering, or a driver error closing the test, rather than a reboot or a freezeNote the exact driver error and whether the picture returns by itself; a driver reset points at the driver layer
4. Settings layerAn overclock, an undervolt, or a raised power limit is stable in games but unstable under sustained full loadReturn every setting to the factory default and re-run the same test; if the crash disappears, the setting was the cause
5. Test itself layerThe test's own load, or the combination of that test with this driver and this card, is the suspect — FurMark's load is not a game loadChange the test: if FurMark fails but a game-like benchmark passes, the extreme power load is the distinguishing factor

The order is itself the diagnosis. A reader who jumps to layer one and buys a larger power supply has skipped the settings check, which costs nothing and often answers the question. A reader who works through the layers in order will either find the fault early or arrive at the deeper layers with the cheap explanations already ruled out.

Crash Timing Is Evidence, Not Decoration

When a crash happens matters as much as how it happens. The timing table below is drawn from published symptom analysis of stress testing, and it should be read as a guide to where to look next, not as a guarantee about any single machine.

Crash timingWhat it most often suggestsNext step
Within seconds of starting the testVoltage too low, or a setting that is incompatible with the loadReturn to stock settings first, before suspecting hardware
After 5 to 30 minutesMarginal instability at the edge of what the machine can sustainCompare temperatures and the power-limit probe; the fault is at a boundary, not an immediate failure
After 1 to 4 hoursCooling throttling or voltage regulator overheating under sustained loadRecord all three temperatures and check airflow and the cooling path
Only when idle, never under loadIdle-state instability, such as an unstable low-power stateThat is a different shape from this guide; do not apply the full-load steps to it

The timing must be recorded against the same test and the same settings, or it means nothing. A crash after twenty minutes in FurMark and a crash after three minutes in a combined CPU and graphics card test are two different pieces of evidence, and they should not be averaged into a single vague statement such as "it crashes after a while". Note the test, the settings, the elapsed time, the crash form, and the three temperatures together, as one record.

There is a related point about temperature that is easy to miss. Core temperature, hotspot temperature, and memory junction temperature can tell different stories. A card can show an acceptable core temperature while its hotspot or its memory is already at its limit, and the crash that follows will look sudden and unexplained if only the core temperature was watched. Only the three-temperature point is needed here, and their full diagnosis is not repeated. The rule here is only this: record all three numbers at the moment of the crash, and see which one reached its limit first. A machine that slides gradually hotter and then crashes points towards a cooling boundary. A machine that crashes suddenly while all three temperatures are still moderate points towards the power or driver layer instead.

Three Tests, Three Different Meanings

FurMark, 3DMark, and OCCT Power are not three versions of the same test. They ask the machine three different questions, and the pattern of which ones pass and which ones fail is itself diagnostic.

FurMark asks the power question in its most extreme form. It holds the graphics card at maximum power draw with heavy shaders, generating more heat and drawing more power than any real game. Published guidance treats fifteen to thirty minutes of FurMark as enough for its purpose; it is not a suitable long-term stability test, and it should not be run for hours in the hope of a stronger result. A FurMark failure says the machine cannot sustain that extreme power load. It does not, by itself, say the machine cannot play games. FurMark's unusual sustained power demand can expose a marginal power supply or a marginal connection that ordinary games never trigger — which is exactly why the test is useful, and exactly why its result must not be read as a game result.

3DMark asks a game-like question. Time Spy is a Direct3D 12 benchmark, and its stress-test mode runs more than twenty loops; Speed Way is a Direct3D 12 Ultimate benchmark with ray tracing, closer to a modern game. A machine that fails FurMark but passes a 3DMark stress test has drawn a clear line: the extreme power load is the distinguishing factor, and the game-like load is within what the machine can sustain. A machine that fails both has a wider range of suspects, and the investigation should not stop at the power boundary. Unigine Superposition sits between the two in character: it stresses the graphics card and its memory together, and it needs a longer run to catch temperature-related instability, which is a reminder that a short run can miss a fault that only appears once heat has built up.

OCCT Power asks the whole-machine question. It pushes the processor and the graphics card at the same time, to reproduce power instability that appears only when both are loaded together. Combined load raises the total power draw of the whole machine, and it can expose a weak power supply, a motherboard voltage regulator limit, or a case airflow problem that separate tests never catch. The order of OCCT testing matters, and it is the reason this guide insists on a sequence: run the processor-only test and the graphics-card-only test separately first, and only then run the combined Power test, so that the results can be compared. Published technician practice is to compare processor-only, graphics-card-only, and combined Power runs to isolate load-related failures. If only the combined run fails, the fault points at the whole machine's power delivery — power supply, voltage regulators, or airflow. If a separate run already fails, return to the range of that separate test before blaming the combination.

OCCT also carries its own safety rules, and they must be followed, not treated as optional. If the 12-volt rail drifts by more than plus or minus five percent, stop the test. If there is a smell of burnt plastic, a repeated electrical noise, or any smoke, stop the test at once. Never open a power supply: its capacitors can remain charged after it is disconnected from power. The OCCT Power test must not be run unattended, and it must not be run for a long time on a card already known to have a cooling or power problem. A test that is run while nobody is watching cannot be stopped when the warning signs appear.

The Seven Steps, In Order

The sequence below is the diagnostic method of this guide. The order is deliberate: each step rules out one layer before the next layer is touched, and no step is skipped because its answer seems obvious.

  1. Record the crash form and timing. Which form was it — reboot, black screen, driver reset, or freeze? How long after starting the test did it happen — within seconds, after five to thirty minutes, or after an hour or more? Which test failed — FurMark, 3DMark, or OCCT — and do games really run without any problem, including long sessions? Do not move to the next step until this record exists, because every later step is judged against it.
  2. Return to stock settings for comparison. Set any overclock, undervolt, and power limit back to the factory default, and re-run the same test that crashed. If the crash disappears at stock settings, the fault was setting instability, and the investigation ends here — that is a settings problem, not a hardware failure. Only the stock-settings comparison is needed here, and it must be run before any hardware is suspected.
  3. Change the test for comparison. Run a different test under the same settings. If FurMark fails but 3DMark passes, the fault leans towards the pure power boundary, because FurMark's load is the most extreme of the three. If both fail, the range of suspects is wider. If only the OCCT Power combined test fails while the separate graphics card test passes, the fault leans towards the whole machine's power combination. A single test result is one data point; the pattern across tests is the diagnosis.
  4. Run the lower power limit comparison. Lower the power limit by about a fifth, and re-run the test that previously crashed. If the crash disappears at the lower limit, the power boundary is implicated, and the investigation should stay in that layer — power supply, cables, connectors, and the card's own power delivery. If the crash continues unchanged at the lower limit, the chance that the fault lies outside the power boundary rises, and the investigation should move to temperature and driver layers. Only the lower-power-limit comparison is needed here, and it is used as a probe, not as a permanent setting.
  5. Compare the three temperatures. At the moment of the crash, record the core temperature, the hotspot temperature, and the memory junction temperature, and see which one reached its limit first. If temperatures climb gradually and then the machine crashes, the fault leans towards a cooling boundary. If temperatures are moderate and the crash is sudden, the fault leans towards the power or driver layer. Do not rely on a single temperature number: the three numbers can tell different stories, and that point is all that is needed here.
  6. Run the OCCT Power combined test, in the right order. First confirm that the processor-only and graphics-card-only runs are each stable for a short run, and that temperatures fall back afterwards. Then run the combined Power test for a short time, with someone watching it throughout. If only the combined run fails, the fault points at the power supply, the motherboard voltage regulators, or case airflow. If a separate run already fails, return to the range identified in step three. Follow the OCCT safety rules in the previous section: stop if the 12-volt rail drifts beyond plus or minus five percent, stop at any smell of burning, repeated electrical noise, or smoke, and never run this test unattended.
  7. Cross-test or bring the machine in. Cross-test with a different power supply, or move the graphics card into another machine with ample power, and see whether the fault follows the card or stays with the original machine. Only the cross-testing idea is needed here. Bring the machine in with the full record: which test, which settings, the crash form, the crash timing, the three temperatures and the power draw at the moment of the crash, and the driver version.

That record should be complete enough that an inspection can start at the right layer instead of repeating the whole sequence from the beginning. A record that says only "it crashes in benchmarks" will be re-tested from scratch. A record that says which test, at which settings, after how long, in which form, and at which temperatures, can be acted on at once.

Diagnostic Logic: Reading the Symptom

SymptomMost likely causeNext step
Games fine for hours, FurMark crashes, 3DMark passesPure power boundary: FurMark's extreme sustained load is the distinguishing factorLower power limit probe, then check power supply, cables, and connectors
Games fine, FurMark and 3DMark both crashWider range: power, cooling, driver, or settingsReturn to stock settings, then compare the three temperatures
Crash within seconds of starting a testVoltage too low, or a setting incompatible with the loadReturn to stock settings first
Crash after 5 to 30 minutesMarginal instability at a boundaryTemperature and power-limit comparisons
Crash after 1 to 4 hoursCooling throttling or voltage regulator overheatingRecord all three temperatures and check airflow
Only the OCCT Power combined test crashes, separate tests passWhole-machine power delivery: power supply, voltage regulators, or airflowDo not blame the card alone; check the combined load path
Crash disappears at stock settingsSetting instability, not hardware failureKeep stock settings and re-test; do not replace hardware on this evidence
Crash disappears when the power limit is lowered by about a fifthPower boundary implicatedCheck power supply, cables, and connectors before buying a larger unit
Temperatures climb gradually, then the machine crashesCooling boundaryCheck the cooling path and airflow
Temperatures moderate, crash is suddenPower or driver layerPower-limit probe and driver error record
Machine reboots under testSystem-level event: power protection, power delivery, overheating, or instability elsewhereTreat as more serious than an application crash; stop and record
A test passes onceUseful evidence only, not proof of long-term healthDo not declare the machine healthy on one pass

Edge Cases That Are Not This Problem

Games and benchmarks both crash. That is not this guide. This guide is the opposite shape, in which games are stable and only a benchmark fails. Only the paired-trigger description is used here as the contrast case — one game and one benchmark crashing together — and that diagnosis is not repeated.

The crash happens only under an overclock or undervolt, and disappears at stock settings. That is a settings problem, not a hardware failure. Only the stock-settings comparison is used here, and the investigation ends at step two when the crash disappears at stock settings.

Artefacts appear in only one test, while other tests and games are normal. Suspect the test software, the driver, or that particular load combination first, and cross-check with a different test. Only the cross-test comparison is used here, and its full score-reading diagnosis is not repeated.

The first run from cold and a later consecutive run give different results. Cooling state differs between the two runs, which is not itself a fault. Only that point is used here: a cold first run and a warm later run are not directly comparable.

An intermittent fault does not appear in a short test at all. A test that passes has shown only that this test, at this time, did not trigger the fault. Only that point is used here. A single pass is useful evidence, not proof of long-term health.

Artefacts appear, there is a smell, the fan makes an unusual noise, or temperatures climb out of control during a test. Stop the test at once and do not run it to completion. Only that stop rule is used here.

The machine crashes only when idle, never under load. That is a different shape from this guide, which is about crashes under sustained full load. Idle-only crashes point at idle-state instability, and the full-load steps here do not apply to them.

The question is whether a benchmark score is high or low. That is not this guide. Score reading is a separate subject, and a crash under test is not a score problem. Do not use a score difference as a pass or fail standard for a crash diagnosis.

Six Common Mistakes

  1. Assuming that a benchmark crash means the card is broken. Wrong. A crash under test says only that a boundary was reached under that particular load. It does not prove the card will fail in ordinary use, and it does not prove the card has failed at all.
  2. Assuming that games running fine means the card is completely healthy. Wrong. Game load may never have pushed the power or cooling boundary, and the benchmark may be the first thing that ever reached it. Stability in games and a fault at the boundary can both be true.
  3. Assuming that passing FurMark means every game will be stable. Wrong. FurMark's load is single-minded — sustained power and heat — and passing it does not prove stability under ray tracing, heavy memory access, or rapidly changing game load. A card stable under FurMark may crash in games, and vice versa.
  4. Assuming that one temperature number is enough. Wrong. Core, hotspot, and memory junction temperatures can tell different stories, and a card can crash while its core temperature alone looks acceptable. Record all three numbers at the moment of the crash.
  5. Assuming that no power spike visible in monitoring software means there is no power problem. Wrong. Ordinary monitoring software samples too slowly to catch spikes that last microseconds or milliseconds. A clean-looking monitoring graph does not rule out a power boundary problem. Only that point is used here.
  6. Assuming that a larger power supply will certainly fix the problem. Wrong. A poor connector contact, a split cable, or ageing power delivery on the card itself will not necessarily be fixed by more watts alone. Check the cables, the connectors, and the card's own power path before buying a larger unit. Only that point is used here.

When to Stop and Bring the Machine In

Stop running tests when any warning sign appears: artefacts on screen, a smell of burning, an unusual fan noise, temperatures climbing out of control, smoke, or a repeated electrical noise. Stop sooner if the machine reboots repeatedly under test, because a reboot is a system-level event and a test that keeps forcing it is not gathering useful information.

Stop and bring the machine in when all seven steps have been completed and the pattern still points at hardware — for example, when the crash continues at stock settings, at a lowered power limit, and across more than one test, with temperatures recorded — or when the cross-test shows the fault following the card or the power supply. Bring the machine in with the full record: which test failed, which settings were used, the crash form, the crash timing, the three temperatures and the power draw at the moment of the crash, and the driver version.

Inspection takes place at Room 1209, Dali Plaza, Shek Mun. You can call 2186 8010, or reach the shop on WhatsApp at 6558 6806. Inspection is charged at the site's published three-tier fees of HK$400–2,800, with the actual quote following inspection. There is no home-visit service and no Mong Kok shop; inspection is carried out in the shop.

Frequently Asked Questions

My games run fine for hours, but FurMark crashes — is my graphics card broken?

Not necessarily, and that conclusion should not be drawn from one test. FurMark is a power virus: it pushes the card to maximum power draw with extremely heavy shaders, drawing far more power and generating far more heat than any real game ever would. A crash under FurMark says the machine cannot sustain that extreme load. It does not, by itself, say the card will fail in games. Record the crash form and timing, return to stock settings, and try a game-like benchmark such as 3DMark for comparison before deciding what is at fault.

If my card passes FurMark, does that mean every game will be stable?

No. A graphics card stable under FurMark may crash in games, and vice versa. FurMark's load is single-minded — sustained power and heat — while a game adds changing shader work, heavy memory access, ray tracing, and rapid changes in load that FurMark does not reproduce. Passing FurMark is useful evidence about sustained power and heat, not a guarantee about every game. Each game and each test must be judged on its own evidence.

What is the difference between a reboot, a black screen, a driver reset, and a freeze?

They point at different layers. A reboot is a system-level event and the most serious form: it may point to power supply protection, poor power delivery, overheating, or instability elsewhere, and a shutdown is more serious than an application crash. A black screen means the picture has vanished under load. A driver reset means the driver stopped and recovered, or closed the test with a driver error, and the picture returns. A freeze means the picture and the machine stop responding altogether. Write down which form occurred, because the four forms lead to different next steps.

How long should I run FurMark, and is a longer run better?

Published guidance treats fifteen to thirty minutes of FurMark as enough for its purpose, and FurMark is not suitable as a long-term stability test. A longer run does not produce a stronger result, and it should not be run for hours, or on a card already known to have a cooling or power problem. Different tests have different suggested durations, so no single number of minutes should be used as a universal pass or fail standard for all tests. Record the test, the duration, and the result together.

What does 3DMark tell me that FurMark does not?

3DMark is a game-like load. Time Spy is a Direct3D 12 benchmark whose stress-test mode runs more than twenty loops, and Speed Way is a Direct3D 12 Ultimate benchmark with ray tracing, closer to a modern game. If FurMark fails but 3DMark passes, the extreme power load is the distinguishing factor, and the game-like load is within what the machine can sustain. If both fail, the range of suspects is wider, and the investigation should not stop at the power boundary.

How do I use OCCT Power to separate the processor, the graphics card, and the whole machine?

Run the tests separately first, then combined. Confirm that the processor-only run and the graphics-card-only run are each stable for a short run, with temperatures falling back afterwards, and only then run the combined Power test for a short time, with someone watching throughout. If only the combined run fails, the fault points at the whole machine's power delivery — power supply, motherboard voltage regulators, or case airflow. If a separate run already fails, return to the range of that separate test. Technicians compare processor-only, graphics-card-only, and combined Power runs to isolate load-related failures.

When should I stop a stress test immediately?

Stop at once if artefacts appear, if there is a smell of burning or a repeated electrical noise, if there is any smoke, if the fan makes an unusual noise, or if temperatures climb out of control. In OCCT, stop if the 12-volt rail drifts by more than plus or minus five percent. Never disable the card's over-temperature or over-current protection to run a test, and never open a power supply, because its capacitors can remain charged after it is disconnected. The OCCT Power test must not be run unattended.

My crash happens after five to thirty minutes, not immediately — what does that mean?

Timing is evidence. A crash within seconds of starting a test most often suggests voltage too low or a setting incompatible with the load, so the first step is to return to stock settings. A crash after five to thirty minutes most often suggests marginal instability at the edge of what the machine can sustain. A crash after one to four hours most often suggests cooling throttling or voltage regulator overheating under sustained load. These are guides to where to look next, not guarantees about any single machine, and they must be recorded against the same test and the same settings.

The crash disappears when I lower the power limit — does that prove the power supply is faulty?

It implicates the power boundary, but it does not by itself prove which part of that boundary is at fault. Lowering the power limit by about a fifth and finding the crash disappears points at power supply, cables, connectors, or the card's own power delivery. A larger power supply alone will not necessarily fix a poor connector contact, a split cable, or ageing power delivery on the card. Check those parts before buying a larger unit, and do not treat the lowered limit as a permanent setting — it is a probe.

I have run all seven steps and it still crashes — what should I do next?

Bring the machine in with the full record: which test failed, which settings were used, the crash form, the crash timing, the three temperatures and the power draw at the moment of the crash, and the driver version. Inspection takes place at Room 1209, Dali Plaza, Shek Mun. You can call 2186 8010, or reach the shop on WhatsApp at 6558 6806. Inspection is charged at the site's published three-tier fees of HK$400–2,800, with the actual quote following inspection.

Related articles

Related services

If you need Power connector repair, see these service pages: