Benchmarks Pass but Games Crash? Synthetic Load vs Game Load Diagnosis Guide

2026-11-17 | GPU Repair | 34 minutes | HK GPU Fix Editorial Team

Benchmarks Pass but Games Crash? Synthetic Load vs Game Load Diagnosis Guide

Key Summary

FurMark and 3DMark pass, yet a game crashes within minutes. The two kinds of load are not the same test: benchmarks apply steady, continuous load, while games spike, dip, compile shaders, and stream assets. This guide uses benchmark, game, feature, and driver comparisons to find which layer the crash belongs to.

The Shape of the Problem: Clean Benchmarks, Crashing Games

This guide covers one specific shape, and the shape itself is the whole diagnosis. A graphics card passes FurMark, passes 3DMark, and does so with temperatures and power draw that look normal. Then the same machine, with the same settings, is used to play a game — and within minutes it crashes. The crash may take one of several forms: the game closes suddenly back to the desktop, the screen goes black, a message appears saying the graphics driver stopped responding or that the D3D device was lost, or the whole machine restarts.

That combination is what makes this symptom different from its neighbours. A card that crashes under a benchmark as well as in games belongs with a different diagnosis: the steady-load boundary, where sustained full load itself is the trigger. A card that crashes in only one game, with no benchmark result recorded at all, belongs with a single-game diagnosis. This guide starts from a different place: the benchmarks have already passed. That starting point rules out the simplest explanations — a card that cannot sustain full load at all, or cooling that cannot cope with a long, heavy run — and moves the investigation to the things a benchmark never exercises.

The first conclusion most owners reach is that passing a benchmark proves the card is healthy, and the second is that a game crash must therefore be the game's fault. Both conclusions go too far. A benchmark proves something precise and limited: under that particular kind of load, for that length of time, the card did not cross a boundary. It proves nothing about the loads the benchmark never applied. And a game crash is evidence that some combination of load, feature, driver, and game reached a boundary — it does not, on its own, say which of those four was responsible. The question is not whether the card is healthy in general. The question is which game-only load the benchmark never tested.

One honest gap should be stated up front. No single authoritative guide isolates this exact shape — benchmarks pass, games crash — as its own diagnostic flow. The method in this guide is assembled from published technical write-ups that describe the difference between synthetic and game load, from crash-signature references for DirectX device errors, and from standard comparison testing. It is presented as a diagnostic assembly, not as official guidance from any benchmark maker, game publisher, or graphics vendor.

Why a Steady Load and a Game Load Are Not the Same Test

A benchmark and a game do not load a graphics card in the same way, and that difference is the reason one can pass while the other fails.

A synthetic benchmark applies a steady, continuous, predictable load. FurMark, for example, renders the same fur-covered object frame after frame, holding the card at a high, flat level of work. A 3DMark run is more varied, but it is still a fixed, repeatable sequence that the card works through from start to finish. The load rises, stays high, and ends. Temperatures climb to a plateau and stay there. Power draw settles at a level and holds. The clock speed settles too. Everything about the load is smooth.

A game is not smooth. A game loads a new area and the work spikes; the player stands still and the work falls away; a cutscene plays and the work changes character entirely; a menu appears with no frame-rate cap and the card suddenly renders hundreds of frames per second; a ray-traced reflection appears and a whole new set of hardware units switches on; a shader is compiled on the spot and the driver stalls for a moment. The load jumps up and down, the clock speed boosts and drops, and the power draw spikes and falls in milliseconds. The card is asked to change state constantly, and each change is a moment where something can go wrong.

That difference matters because most failure boundaries are not about the average load. They are about the transition — the instant the load changes, the clock jumps, or the power draw spikes. A card can hold a steady high load indefinitely and still trip at a sudden change, because the steady load never asks it to make that change. As one technical write-up puts it, modern graphics cards rarely crash when pegged at a steady 100% load; they usually trip up during sudden load transitions or aggressive frequency boost spikes.

There is a second difference that is easy to miss. Some cards detect FurMark and deliberately slow themselves down, treating it as a power-virus load rather than a normal workload. When that happens, the benchmark is not even testing the card at full speed — it is testing a throttled card that is, by design, staying well inside its limits. A pass under those conditions says even less about game stability than a pass at full speed would. A card can therefore pass FurMark for thirty minutes and still crash in a real game within five minutes, as another published guide describes it, without any contradiction at all.

A third difference is what the benchmark never asks the card to do. Ray tracing and path tracing, real-time shader compilation, streaming large textures and geometry into memory as a new scene loads, and the uncapped frame rates of a game menu are all game-only loads. A benchmark that renders the same scene repeatedly does not compile new shaders on the fly, does not stream in a new area, and does not switch on ray tracing halfway through. Each of those game-only features is a separate path through the driver and the hardware, and a fault in one of them is invisible to a benchmark that never takes that path.

The practical consequence is simple to state and important to keep in mind through the rest of this guide: a benchmark pass is a narrow result, not a general certificate of health. It says the card survived one kind of load. It says nothing about the others.

The Deciding Forks: Four Questions That Narrow the Field

The whole guide turns on four forks. Each fork is a comparison, and each comparison removes a whole set of possibilities. They should be done in order, before any setting is changed, because a setting changed before the fork is tested destroys the evidence the fork would have given.

Fork One: Do the Benchmarks Really Pass?

The first fork is the premise itself. A benchmark that was run once, briefly, or with different settings from the game does not establish the premise. For this guide to apply, at least two different benchmarks — for example FurMark and 3DMark — should each complete at least one full run, with temperatures and power draw recorded. If only one benchmark passes and the other crashes, the premise is not established: note which kind of load reached a boundary first, because that is evidence in itself, and do not claim that benchmarks pass in general. If both benchmarks crash, leave this guide and follow the steady-load diagnosis instead — the fault is in the sustained-load boundary, and the game crash is only a second symptom of it.

Fork Two: Every Game, or Only One?

The second fork is inside the game-load path. Try a second game that uses a different engine and comes from a different era — a different kind of game, not a sequel or a game on the same engine. If every game crashes, the fault points at a layer shared by all games: the driver, the system, or a cross-game setting. If only one game crashes and the others run normally, the fault points at that game's shaders, files, settings, or feature combination. This fork is the single most informative test in the guide, and it costs nothing more than installing or launching a second game.

Fork Three: Do the Game-Only Features Change the Result?

The third fork tests the game-only feature layer directly. Turn off ray tracing or path tracing, turn off frame generation, turn off upscaling, cap the frame rate, and clear the game's shader cache so it compiles fresh — changing one thing at a time, and retesting the same game in the same scene after each change. If the game stops crashing when one feature is off, that feature is the trigger condition. If it crashes the same way with every feature off, the fault is not in that layer, and the investigation moves on. Each change here is a comparison test, not a repair: running a game permanently with ray tracing off is avoiding the problem, not fixing it.

Fork Four: What Kind of Crash Is It?

The fourth fork is the crash signature, and it must be recorded before the crash is discussed at all. A restart, a black screen, a driver-reset message, and a silent return to the desktop are four different events that point at four different layers, and lumping them together as "the machine crashed" throws away the most specific evidence available. Note the exact wording of any error message, and note which of the four forms the crash took.

Crash formWhat it describesWhat it points toward
Silent return to desktop (CTD)The game closes without an error message and the desktop is left runningGame process, driver interaction, or a game file fault — the system itself stayed up
"D3D device lost" or DXGI errorA message naming a lost or removed DirectX device, often with a code such as 0x887A0005 or 0x887A0006The graphics device was reset or removed — see the signature table below
Driver stopped responding / driver resetA message or notification that the display driver stopped responding and recoveredA driver timeout reset (TDR) — the driver took too long and the system reset it
Whole machine restartsThe screen goes black and the machine reboots on its ownA system-level event — power, or a fault severe enough to take the whole machine down

The two DirectX codes deserve a closer look, because they are often read as a diagnosis when they are only a signature. The code 0x887A0006, labelled "HUNG", means the graphics device timed out — the driver did not respond within the timeout window, about two seconds, and the system reset it. The code 0x887A0005, labelled "REMOVED", means the device was lost entirely. The same message can come from a driver timeout, a rendering fault, shader compilation, an unstable overclock, a feature combination, or a hardware fault. The signature tells the investigator which family of causes to consider; it does not, on its own, name the cause. A "D3D device being lost" message is a graphics-device reset, not a statement that the graphics card is dead.

The Five Suspicion Layers

A crash that happens in games but not in benchmarks can fail at five layers, and they must be kept separate. Changing the driver, the settings, and the overclock all at once and finding the game stable proves nothing about which change did the work, and the game will crash again the next time only one of those conditions is present. The layers below run from the most game-specific to the most system-wide, and each one has its own deciding evidence.

LayerWhat it coversDeciding evidence
1. Game-only feature layerRay tracing, path tracing, live shader compilation, asset streaming, uncapped menu frame ratesTurning one feature off stops the crash
2. Transient power and load-change layerSudden load changes, boost clock spikes, millisecond power peaksThe crash happens at load transitions, and ordinary monitoring is too slow to see the spike
3. Video memory allocation layerSudden, burst-style memory allocation when a new scene loads or ray tracing switches onThe crash happens at scene loads or when a memory-hungry feature switches on
4. Driver, game file, and DirectX path layerDriver interaction with one game, DirectX 11 against DirectX 12 paths, driver and game version combinationsA clean driver reinstall, or a different DirectX path, changes the result
5. Settings and overlay layerOverclock or undervolt stable under steady load but not under changing load; overlay, recording, and anti-cheat software injected into the game processReturning to factory defaults, or closing all injected software, stops the crash

Layer One: The Game-Only Feature Layer

The first layer is the set of things only a game does. Ray tracing and path tracing switch on dedicated hardware units that a rasterised benchmark never touches. Live shader compilation asks the driver to build rendering programs on the spot, in the middle of play, and a stall or failure during that compilation can take the game down. Asset streaming pulls large textures and geometry into memory the moment a new area loads, creating a sudden burst of work that a benchmark's repeated scene never creates. And an uncapped menu frame rate can ask the card to render hundreds of frames per second of a nearly static image — a strange, intense load that no benchmark reproduces, and a common place for device-lost crashes to appear.

The evidence for this layer is the feature-by-feature comparison in fork three. If the game stops crashing when ray tracing is off, or when the frame rate is capped, or after the shader cache is cleared and rebuilt, the fault points at the feature layer. Note the precise trigger: which feature, switched off on its own, changed the result. A game that crashes only with ray tracing on, in every other respect unchanged, is pointing at the ray-tracing path — the combination of that feature, the driver version, and the game version — and not at the card in general.

One point about shader compilation deserves care. When a game updates its engine, the way shaders are compiled, the lifetime of resources, and the way memory is used can all change. The shader cache can also be a leftover from an older engine, an older driver, or an older set of settings. Clearing the cache and letting the game rebuild it is a comparison test: if the rebuilt cache behaves differently, the fault pointed at stale compiled shaders. It is not a repair, and it does not say the card was at fault.

Layer Two: The Transient Power and Load-Change Layer

The second layer is the one benchmarks are worst at testing: the change itself. A game's load jumps up and down constantly, and with each jump the boost clock rises or falls and the power draw spikes or drops. The spikes can be far above the card's rated power draw for a few milliseconds — brief, but real — and they arrive at the exact moment the card is also being asked to do something new.

The evidence for this layer is timing, not a number on a monitoring screen. If the crash happens at a load transition — the moment a new scene loads, the moment the player enters a busy area, the moment a menu appears or disappears, the moment the frame rate suddenly jumps — the fault points at the transition layer. And here is the trap: ordinary monitoring software samples too slowly to catch a millisecond-scale power spike. Most monitoring tools read power draw a few times per second at best, and a spike that lasts a few milliseconds passes between two samples. Seeing no spike on a monitoring graph does not mean there was no spike. It means the monitoring did not catch one. Absence of evidence on a slow sampler is not evidence of absence.

This layer is also the reason a benchmark pass is weak evidence about power. A benchmark's power draw is flat and predictable, so the power delivery never faces a sudden demand. A game's power draw is a series of sudden demands, and the boundary it tests is the delivery system's ability to answer a sudden demand, not its ability to sustain a flat one.

Layer Three: The Video Memory Allocation Layer

The third layer is how the game asks for video memory. A benchmark's memory use is steady: it allocates what it needs at the start and holds it. A game's memory use is bursty. Loading a new scene, switching on ray tracing, or turning up texture quality can ask for a large amount of memory all at once, in a single moment. The question is not whether the card has enough memory in total over a long session. The question is whether it can get the amount it needs at that one moment.

The evidence for this layer is when the crash happens. A crash that consistently arrives at a scene load, at a fast-travel, at the moment a high-resolution texture pack engages, or at the moment a memory-hungry feature switches on points at the allocation layer. A crash that happens at random moments in the middle of steady play, with no scene change, points elsewhere.

One important adjustment follows from the premise of this guide. In a diagnosis where the benchmarks have already passed, video memory exhaustion should be treated as a lower-priority suspect than it would be in a general game-crash guide. Sustained full load has already shown that the card can hold its memory under steady pressure. What is left is the burst — the single moment when a large allocation is requested and cannot be satisfied — and that is a narrower, more specific claim than "the card does not have enough memory."

Layer Four: The Driver, Game File, and DirectX Path Layer

The fourth layer is the interaction between the driver and one particular game. The same driver that passes a benchmark can crash in one game, because the benchmark and the game take different paths through the driver. A game may use DirectX 12 where the benchmark used DirectX 11, or the reverse, and the two paths exercise different driver code. A driver version and a game version are a combination, and a combination that is stable in one pairing can be unstable in another — a crash can even appear after a driver update, when the new driver first enables a feature that turns out to be the one that fails.

The evidence for this layer is the driver comparison and the DirectX path comparison. A clean driver reinstall — removing the old driver completely with a tool such as Display Driver Uninstaller, then installing the current official release — that changes the symptom points at the driver layer. A game that crashes on DirectX 12 but not on DirectX 11, or the reverse, points at the path layer. Neither result should be written up as a verdict on a particular driver version: versions change, and the current official release is the reference. No driver version and no game version should be treated as generally good or generally bad.

Game files belong in this layer too. A damaged or incomplete game file, or a game file written for an older engine version, can crash the game in the same way a driver fault would. Verifying the game files through the game platform's own verification tool is a comparison test in this layer, and it belongs before any conclusion about the hardware is drawn.

Layer Five: The Settings and Overlay Layer

The fifth layer is the settings around the card and the software injected into the game. An overclock or an undervolt can pass a benchmark and fail in a game, because the benchmark's steady load never pushes the card through the clock and voltage changes that a game's jumping load demands. As one published guide puts it, a graphics card overclock or undervolt that flies through a traditional rasterised benchmark might instantly crash the moment ray tracing enters the equation. The profile looks stable under the benchmark and is not stable under the game, because the two loads test different boundaries.

The evidence for this layer is the return-to-defaults comparison. Restoring the overclock, the undervolt, and the power limit to factory defaults and replaying the same game in the same scene is a comparison test, not a repair. If the game stops crashing at factory defaults, the direction is clear: the settings were unstable under changing load, not that the hardware is broken. If it crashes the same way at factory defaults, the settings layer is cleared and the investigation moves on.

Overlay and injected software belongs in the same layer. Overlay software, recording software, and anti-cheat software inject themselves into the game process, and a benchmark never runs through that injection path. If the game stops crashing when every overlay and recording tool is closed, the fault points at the injection layer. If it crashes the same way with everything closed, that layer is cleared.

The Seven Diagnostic Steps

The steps below are comparison tests. Each one changes one variable and records the result. Changing two things at once and finding the game stable proves nothing about which change mattered, and the crash will return the next time only one of those conditions is present. Write down each result before moving on.

Step 1: Record the Baseline

Before anything is changed, record the baseline. Which benchmark was run, how long it ran, whether it completed, and the temperatures and the power draw at the time of the crash. Then the game side: which game, which scene, how long from launching the game to the crash, which crash form it took, the full text of any error message, and whether the driver or the game was recently updated. Without this record, later comparisons have nothing to compare against, and a crash that changes form halfway through the investigation will look like a new problem when it is the same one wearing a different shape.

Step 2: Confirm the Benchmark Premise

Run the benchmark comparison. FurMark passing and 3DMark passing, each at least once through a full run, is what establishes the premise of this guide. If only one passes and the other crashes, record which kind of load reached a boundary first — that is evidence, not a failed test — and do not claim that benchmarks pass in general. If both crash, stop here: this guide no longer applies, and the steady-load diagnosis is the right path. The premise is the foundation of everything that follows, and a premise that was never actually established will mislead every step after it.

Step 3: Compare All Games Against One Game

Try a second game that uses a different engine and comes from a different era. Play it in the same way, for a similar length of time, and record whether it crashes. If every game crashes, follow the driver and system layer. If only one game crashes, follow that game's layer — its shaders, its files, its settings, and its feature combination. Do not skip this step by assuming a second game would crash too; the assumption is exactly what the step exists to test.

Step 4: Return to Factory Defaults

Restore the overclock, the undervolt, and the power limit to factory defaults, and replay the same game in the same scene. Restore all three settings at once, not one by one: restoring them one at a time leaves the result unreadable — if the game stops crashing, there is no way to tell which setting mattered, or whether it took the combination. If the game stops crashing at factory defaults, the direction is that the settings were unstable under the game's changing load — not that the hardware is broken. If it crashes the same way, the settings layer is cleared. This step is a comparison test, not a repair: running at factory defaults permanently is a choice the owner makes afterwards, and it is not the same thing as having found and fixed a fault.

Step 5: Switch the Game-Only Features Off, One at a Time

Turn off ray tracing or path tracing, frame generation, and upscaling — one at a time, retesting the same game in the same scene after each change. Cap the frame rate, including in menus. Clear the game's shader cache and let it compile fresh once, then retest. Note which single change, if any, stops the crash: that change names the trigger condition. If the game crashes the same way with every feature off, the feature layer is cleared. Do not change three features at once and then try to work out which one mattered — the evidence is gone the moment the changes are combined.

Step 6: Close the Overlays and Injected Software

Close all overlay software, recording software, and any injected tool other than the game's own anti-cheat, and launch the game again. If the game stops crashing, the fault points at the injection layer. If it crashes the same way, that layer is cleared. This step is often skipped because overlay software feels harmless, but the injection path is a real path through the game process, and a benchmark never exercises it.

Step 7: Clean Driver Reinstall, Then Prepare for Inspection

Do a clean driver reinstall as a comparison test: remove the old driver completely with a clean-removal tool such as Display Driver Uninstaller, then install the current official release — the version current at the time, checked against the vendor's official release, not a version number written down in any guide. Retest the same game in the same scene. Then assemble the three-sentence conclusion: what the symptom is, which variable made the symptom appear or disappear, and which possibilities are still not ruled out. Bring that record, with the baseline from step one and the results of every comparison, to inspection. A record that says "the game crashes with ray tracing on and does not crash with it off, at factory defaults, on the current driver" is worth far more than a record that says "the game crashes."

Six Misconceptions That Mislead This Diagnosis

Misconception 1: Passing a Benchmark Means the Card Is Fully Healthy

A benchmark pass proves the card survived that benchmark's load, for that length of time, under those settings. It does not prove the card is healthy in general, because the game's jumping load, ray tracing, shader compilation, and burst memory allocation were never tested. A card can be stable under a steady load and unstable under a changing one, and both results can be true at the same time. The benchmark is a narrow test, and treating it as a general certificate of health is the single most common error in this diagnosis.

Misconception 2: A Benchmark Crash Means the Card Is Broken, and a Game Crash Must Be the Game's Fault

Both halves of this are wrong. A crash under either kind of load only says that some load reached a boundary — it does not name the boundary. A game crash can come from the driver, the settings, the power delivery, or the hardware, just as a benchmark crash can. The two kinds of load test different boundaries, and a fault can live in either one. The crash tells the investigator where to look, not what is broken.

Misconception 3: If Only One Game Crashes, the Hardware Cannot Be Involved

A single-game crash does point toward that game, but it does not clear the hardware. If that game is the only one that uses a large amount of video memory, or the only one that uses a particular graphics feature, a memory or driver fault can show up in that game alone and in no other. The single-game result narrows the investigation to the game's layer; it does not close the hardware question. The game-to-game comparison is a fork, not a verdict.

Misconception 4: Turning the Settings Down and Seeing No Crash Means the Problem Is Fixed

Turning settings down is a comparison test. It proves the problem lives in a particular feature or load combination, and that is valuable evidence. It does not repair anything. Running the game permanently at lower settings is avoiding the problem, not fixing it, and the underlying boundary is still there — it will be reached again the next time a game asks for the same feature or the same load. A comparison result should be recorded as a comparison result, not reported as a fix.

Misconception 5: Updating the Driver Will Always Fix a Game Crash

A crash can appear after a driver update, not only before one. A new driver can be the first version to enable a particular feature, or the first version to interact with a game in a particular way, and the crash can start with the update. A crash can also be a game-version problem that no driver change will touch. Driver versions should be compared — the version before the crash against the version after it — and no version should be written down as a generally good or generally bad version. Versions change, and the current official release is the only reference that stays valid.

Misconception 6: If Monitoring Software Shows No Power Spike, There Is No Power Problem

Ordinary monitoring software samples too slowly to catch a millisecond-scale power spike. A spike that lasts a few milliseconds passes between two samples and leaves no trace on the graph. Seeing no spike means the monitoring did not catch one. It does not mean the spike did not happen, and it does not clear the power-delivery layer. The absence of a spike on a slow sampler is not evidence, and it should never be used to close the investigation.

What the Sources Disagree About

The published sources do not agree with each other, and both sides are given below.

Most of the sources support this guide's position. WCCFTech's guide to stress-testing a graphics card states that a graphics card overclock or undervolt that passes a traditional rasterised benchmark can crash immediately when ray tracing enters the picture, that modern cards rarely crash under a steady full load and usually trip during sudden load transitions, and that a profile can look flawless in 3DMark and crash five minutes into an actual game. Digital Citizen's guide states that a card can pass FurMark for thirty minutes and still crash in a real game within five minutes, that many cards detect FurMark and slow themselves down, and that a profile passing every benchmark but crashing in the one game played every night is not stable. Both sources agree with the position taken here, and with the position taken in the companion guide on the opposite shape, that passing FurMark does not mean a game will be stable.

One source disagrees directly. A FurMark guide from mypcbottleneck states that if a card can handle FurMark without overheating or crashing, it can handle basically anything a game throws at it. That claim contradicts WCCFTech, Digital Citizen, and the companion guide's position. This guide follows the majority position and the published evidence for load transitions: FurMark passing does not guarantee game stability, and the contradiction is set out here so that the reader can see both sides rather than only the side that agrees with the guide.

There is a second disagreement, about how to describe the comparison steps. Third-party tutorials often describe returning to factory defaults, switching off ray tracing, capping the frame rate, clearing the shader cache, and reinstalling the driver as repairs. This site's position is that all of them are comparison tests. They produce evidence about which layer the fault belongs to, and they are not repairs. This guide follows that position throughout: no comparison step is presented as a fix.

A third point concerns unsourced claims in the companion guide on the opposite shape. That guide contains statements attributed to unnamed technical analysis, without a source listed. Those statements are not used as evidence here. Every claim in this guide rests on the sources listed at the end, or on the comparison logic of the steps themselves.

When to Stop Testing

Benchmarks and crash-reproduction testing carry risk, and the testing should stop before the risk outweighs the evidence. Do not disable any protection to run a test, and do not repeatedly re-run a test that makes the machine restart — each restart is another stress event, and repeating it does not add information after the first few results. Stop immediately if the screen shows artefacts, if there is any unusual smell or noise, or if temperatures run out of control. A test that has already given a clear result does not need to be run again, and a test that crashes the machine is a result in itself — record it and move to the next comparison instead of running the same crash again.

Do not edit the registry to change the driver timeout. Some third-party tutorials describe changing the timeout that governs how long the driver is allowed before the system resets it. That change reaches into the system's low level, and it is not a step to take on one's own. This guide does not give instructions for it, and the timeout itself is only one possible cause among several — changing it would hide the evidence rather than produce it.

What to Bring for Inspection

If the comparisons point toward a hardware layer, or if they point nowhere conclusive, the next step is a professional inspection. Bring the full record: the benchmark results, the baseline temperatures and power draw, the game and scene, the time from launch to crash, the exact crash form and error message, which driver and game versions were in use, and the result of every comparison — which feature, switched off on its own, changed the result, and which did not. The three-sentence conclusion from step seven is the summary of that record.

Inspection and repair pricing follows the site's published three-tier scale of HK$400 to HK$2,800, with the actual quote given after inspection. No price in this guide should be read as a quote for a particular fault, and no fee is stated here for any specific repair.

The workshop is at Room 1209, Dali Plaza, Shek Mun. The telephone number is 2186 8010, and the WhatsApp number is 6558 6806. There is no Mong Kok shop, and there is no on-site service — inspection is carried out at the workshop.

Frequently Asked Questions

If the benchmarks pass, does that mean my graphics card is fine?

No. Passing a benchmark only proves nothing went wrong under that kind of sustained load. A game's changing load, ray tracing, live shader compilation, and burst memory allocation are all things a benchmark does not catch. Passing a benchmark only says the card was fine under the benchmark's load — it does not mean it is fine under a game's load.

If FurMark passes, does that mean my graphics card is fine?

No. FurMark passing means the card survived FurMark's steady, continuous load for the length of the run. It does not test the sudden load changes, ray tracing, live shader compilation, or burst memory allocation that a game applies. A card can pass FurMark for thirty minutes and still crash in a real game within five minutes, and both results can be true at the same time.

My game crashes but the benchmark does not. Is it definitely the game's fault?

No. A game crash says that some combination of load, feature, driver, and game reached a boundary — it does not say which one. A driver fault, a settings fault, a power-delivery fault, or a hardware fault can all show up in a game first, because the game exercises paths the benchmark never takes. The game-to-game comparison and the feature-by-feature comparison are what narrow it down.

Only one game crashes. Does that rule out a hardware problem?

No. If that game is the only one that uses a large amount of video memory, or the only one that uses a particular graphics feature, a memory or driver fault can appear in that game alone. A single-game crash points the investigation at that game's layer, but it does not close the hardware question.

I turned ray tracing off and the game stopped crashing. Is it fixed?

Not necessarily. Turning a feature off is a comparison test: it proves the problem lives in that feature's load combination, and that is valuable evidence. Running the game permanently with the feature off is avoiding the problem, not fixing it. The underlying boundary is still there, and it can be reached again by another game or another feature.

I updated the driver and the crashes started. Can a new driver cause a crash?

Yes. A crash can appear after a driver update, when the new driver first enables a feature or first interacts with a game in a particular way. A crash can also be a game-version problem that no driver change will touch. Compare the driver version from before the crash with the version after it, and use the current official release as the reference rather than treating any version as generally good or bad.

My monitoring software shows no power spike. Does that rule out a power problem?

No. Ordinary monitoring software samples too slowly to catch a millisecond-scale power spike. A spike that lasts a few milliseconds passes between two samples and leaves no trace on the graph. Seeing no spike means the monitoring did not catch one — it does not mean the spike did not happen.

After clearing the shader cache, the first time I load the game it stutters badly. Is something wrong?

No. After the cache is cleared, the game has to compile its shaders again, so stuttering or slowness the first time a scene loads is normal. Wait for the compilation to finish once before judging the result — do not call it a crash while it is still compiling. Only a crash after compilation has finished is a result worth reading.

What does "D3D device lost" actually mean?

It means the graphics device was reset or removed, not that the graphics card is dead. The same message can come from a driver timeout, a rendering fault, shader compilation, an unstable overclock, a feature combination, or a hardware fault. The error code narrows the family of causes — 0x887A0006 means the device timed out, 0x887A0005 means it was lost entirely — but the signature on its own does not name the cause.

How should I record a crash so the record is actually useful?

Record six things: which benchmark was run, how long it ran, and whether it passed; which game, which scene, and how long from launching the game to the crash; which crash form it took; the full text of the error message; whether the driver or the game was recently updated; and what each comparison step did and what it found. Then assemble a three-sentence conclusion: what the symptom is, which variable made the symptom appear or disappear, and which possibilities are still not ruled out.

Should I keep re-running the game until it crashes again to be sure?

No. Once a test has given a clear result, running it again adds risk without adding information, and repeatedly re-running a test that restarts the machine is a stress event each time. Record the result, move to the next comparison, and stop immediately if there are artefacts, unusual smells or noises, or temperatures running out of control.

Sources

Related articles

Related services

If you need Chip-level GPU repair, see these service pages: