Cold Boot Fine but Warm Reboot Fails? Cold Start vs Warm Restart, Link Retraining & Device Reset Diagnosis

Key Takeaways
When a graphics card works normally after a full power-drain cold boot but fails after a Windows Restart, the contrast itself is the diagnostic signal: Restart rebuilds the kernel yet never removes power, so a fault that only a full power drain clears points toward device state retained across a warm reset, warm-path link retraining, warm-path firmware enumeration, or retained monitor or power state. This guide defines the four boot operations precisely, sets out a four-arm test matrix, and separates those branches by observation only, without opening the card.
The Pattern, Stated Exactly
This guide covers one narrow pattern, and both halves of it are required. After a full power-drain cold boot, the graphics card works normally, including in Windows and under load. After a Windows Restart, something goes wrong: no display, a black screen at some stage of boot, low resolution, the card missing from Device Manager, a driver error status, or instability that was absent after the cold start. If both paths fail, or if the cold path also fails at times, this guide does not apply as written. That is a different diagnostic path, and forcing it into this pattern will mislead every test that follows.
The word warm in warm reboot needs defusing at the start, because it causes a predictable confusion. In this guide, warm means power was never removed. It does not mean the hardware is hot. A card can sit at an ordinary operating temperature in both paths and still show this pattern. Temperature is a separate measurement, recorded together with clock speed where it is relevant, and it is not assumed from the word warm. Keeping warm as a power-state word, rather than a temperature word, is what makes the test matrix in this guide meaningful.
A further boundary belongs here. A failure that appears only after sleep or hibernation resume is not evidence for this pattern, because sleep and hibernation resume follow their own power-state paths, with their own display-link resume behaviour. A failure that appears on the ordinary boot path, with display during early boot followed by black in Windows, is also a different pattern: in this guide, the cold boot is fully normal, including in Windows. Stage matters throughout. A reader log that says only black screen, without stating the stage at which display was lost, cannot support a conclusion. The log must state what was on screen at each stage: nothing at all, the motherboard logo, the Windows loading indicator, the login screen, the desktop, or the desktop followed by failure under load, together with the exact point at which the result diverged from normal.
Four Operations That Are Not the Same
Readers use shutdown, reboot, and restart loosely, and loose wording destroys this diagnosis, because the four operations below differ in two independent respects: whether power is removed, and whether kernel and driver state is rebuilt or restored. The table states the distinction in plain language. No result in this guide may be described as a cold boot unless it matches the fourth row.
| Reader phrase | What actually happens | Power removed? | Kernel and driver state |
|---|---|---|---|
| Windows Shut down, Fast Startup on | Apps close and the user signs out, then the kernel session is hibernated: kernel and loaded kernel-mode drivers are saved to the hibernation file and restored at next power-on | No full power drain; standby power remains while the machine is off | Restored, not rebuilt |
| Windows Shut down, Fast Startup off | Apps close, the user signs out, and the kernel session ends | Machine powers off, but standby power is still present unless the PSU is switched off or mains is unplugged | Rebuilt at next power-on |
| Windows Restart | Apps close, the user signs out, and the kernel session ends completely, followed by a full boot; Fast Startup does not apply to Restart | No; power is never removed during a restart | Rebuilt from scratch |
| Full power-drain cold boot | Machine shut down, PSU switched off and/or mains unplugged so standby power is drained, then powered on | Yes; this is the only arm in which the device loses all power | Rebuilt from scratch, and device hardware state is also reset by power loss |
Two consequences follow, and both are central to this guide. First, the statement that Restart is more thorough than Shut down is true for the kernel, because Restart rebuilds it while a Fast Startup shutdown restores it, but false for hardware power state, because neither Restart nor ordinary Shut down removes power. A fault that survives Restart but disappears after a full power drain cannot be explained by saying the kernel needed a fresh start, because the kernel already received a fresh start on Restart. Something that only power loss clears is implicated instead: device hardware state, sticky state, a link that needs retraining from power-up, or a power-rail state. Second, shutdown then power-on is not automatically a cold boot. With Fast Startup on, the Windows default, it is a hybrid resume of saved kernel and driver state. Any test that uses ordinary Shut down as its cold arm is contaminated, and its result cannot separate kernel state from hardware power state.
PCIe Reset Types: Why Reset Is Not One Thing
The PCIe specification defines four reset mechanisms, and the distinctions among them do diagnostic work in this guide. A cold reset is a Fundamental Reset that occurs when main power is applied to the device: system power-on, a full power cycle, or card reinsertion with power removed. The whole device starts from a clean hardware state. A warm reset is also a Fundamental Reset, triggered by hardware without removing and reapplying power, through the sideband PERST# signal. Internal logic is reset, but the device was never unpowered. The specification does not prescribe one single mechanism for generating a warm reset; platform design decides, which is itself a reason warm-path behaviour can differ between motherboards.
A hot reset is an in-band reset, sent through the PCIe link itself rather than through the sideband signal. It resets the downstream device and restarts link training, and software can request it through the Secondary Bus Reset bit in a bridge control register. A Function Level Reset, or FLR, resets a single function of a multi-function device, initiated by software writing the Initiate FLR bit. The function must complete FLR and its own initialisation within a documented interval of 100 ms in the specification summaries. FLR does not reset sticky bits, hardware-initialised registers, or the PCIe link itself: the physical link stays up.
For this guide, three distinctions matter. Cold versus warm asks whether power was removed at all. Fundamental versus hot versus FLR asks whether the reset reaches the whole device, travels over the link, or touches only one function while the link stays up. Sticky registers ask what survives: configuration state that is excluded from Fundamental Resets while auxiliary power remains present. Described qualitatively, some state survives a warm reset while standby power is present, and is only cleared when power is fully removed. This guide does not invent a list of which registers, on which card, behave this way. That is model-specific and unverified here, and naming registers without a model-specific source would turn a mechanism into a false precision.
Link Training and Retraining: The Handshake That Can Fail on the Second Attempt
A PCIe link is not a fixed wire; it is negotiated. The Link Training and Status State Machine moves through Detect, Polling, and Configuration into L0, the active state, with a Recovery state used for retraining. During training, the two ends agree on lane width, on data rate or generation class, on lane polarity and numbering, and on equalisation settings that compensate for signal loss across the physical channel, including transmitter presets and receiver equalisation in phases at higher speeds.
Training starts low and steps up: negotiation begins at the lowest rate and attempts higher rates only if the lower-rate handshake succeeds. A link that trains at a lower generation or narrower width than the card and slot are capable of has still trained successfully. There may be no error message at all, only less bandwidth. A generation drop at idle can also be normal power management, so a low reading at idle proves nothing on its own. The comparison that matters is under load, in both paths, on the same workload.
Retraining is not identical to first training. On a warm reboot, the link goes through Recovery and renegotiates, potentially under different conditions from power-up: a different equalisation starting point, a different power-management state, a different BIOS enumeration path, and a device whose internal state was reset but never unpowered. A marginal signal path, involving slot contact, a riser, a retimer, or a lane-sharing configuration, can therefore pass a from-power-up negotiation and fail, or downgrade, on renegotiation, or the reverse. This is presented as a candidate mechanism to test, not as a proven cause in any individual machine.
Width downgrade and generation downgrade are different evidence and must not be collapsed into a single phrase such as the link is slow. Generation can legitimately drop at idle and recover under load. Width does not dynamically scale the same way, so a persistently narrower width under load points more toward the physical link path or lane allocation, such as BIOS bifurcation or lane sharing, or a slot wired for fewer lanes by design. Capability must be checked before any width reading is called a downgrade: what the card and slot are designed for is the baseline, not an assumption.
Display link training is a second, separate handshake, and this guide keeps it labelled separately at every point. DisplayPort or HDMI link training between GPU and monitor, including monitor power state and the EDID handshake, is redone on reboot and can fail independently of the PCIe link. A warm-reboot failure that changes when the monitor, cable, or port changes belongs in the display-link branch, not the PCIe branch. The phrase link retraining must never be left ambiguous between PCIe link retraining and display link retraining.
Device Reset and Driver Reload: What Restart Rebuilds, and What It Does Not
On Restart, Windows ends the kernel session and loads drivers fresh. That clears a hung or corrupted driver instance in memory, user-session and application state, and pending driver-install completion steps that require a full boot. It does not, by itself, clear a driver file problem, because corrupted or mismatched driver files are simply reloaded; a driver version problem, because the same incompatible version loads again; firmware or BIOS enumeration decisions made during the warm path, such as BAR allocation, Resizable BAR or Above 4G Decoding state, PCIe generation setting, or display-output priority; device hardware state that survives a warm reset while standby power is present; a marginal PCIe link that retrains to a degraded state; or a monitor or dock state that was never power-cycled.
A related Windows mechanism deserves careful mention. Timeout Detection and Recovery, or TDR, applies when the GPU stops responding for longer than the documented default timeout: Windows resets the display driver. A successful TDR appears as a brief black flash and a message that the display driver stopped responding and has successfully recovered, logged as Event ID 4101 from source Display. A failed or too-frequent recovery can end in a VIDEO_TDR_FAILURE stop code. TDR is relevant to this guide only when the warm-reboot symptom is a driver recovery event after reaching Windows. It is not an explanation for no display before Windows loads, and a 4101 event identifies a reset that happened, not which component caused the stall. Adjusting TDR timeout values is not presented as a fix in this guide: it is out of scope for an observation guide and can mask a real fault.
Power-State Retention and Firmware Enumeration on the Warm Path
While a PC is off but still connected to mains, standby power is present. While a PC restarts, power is never interrupted at all. The GPU, motherboard, monitor, and any dock or hub may therefore retain some powered state across a warm reboot that a full power drain clears. Power-management states are part of the warm path: PCIe ASPM low-power link states and device sleep states can leave a device that fails to exit a low-power state cleanly on warm re-enumeration appearing missing, degraded, or unstable, while a from-power-up enumeration succeeds.
This is also why a report that removing the card and putting it back made the problem go away is weak evidence on its own. That action combines power drain, time passing, slot re-seating, and fresh link training into one uncontrolled change. The test matrix below exists to separate those confounded factors as far as observation allows. Handling of internal hardware is a professional-inspection matter, not a reader step in this guide.
During boot, BIOS or UEFI firmware enumerates PCIe devices and assigns address space. Settings that interact with this enumeration, and that can therefore behave differently between a cold enumeration and a warm re-enumeration, include Above 4G Decoding, which allows address space above 4 GB to be assigned to large PCIe devices, and Resizable BAR, which lets the CPU address the GPU's full frame buffer through one large BAR. These are a pair: Resizable BAR requires Above 4G Decoding, and a toggle that is on without its prerequisite silently does nothing. A mismatch here can show as degraded behaviour or a detection problem, though a mismatch on its own is not a warm-only pattern. Also relevant are the PCIe generation setting, Auto versus a forced generation, slot bifurcation or lane-sharing configuration, boot mode, UEFI versus CSM, Secure Boot state, and display-output priority, meaning which output is tried first. A BIOS update can restore defaults and change several of these at once, which is why a pattern that started after a BIOS update must be investigated as a settings-change event, not as proof of a hardware fault. The rule for settings in this guide is that they are evidence to record, one at a time, with the before-value noted. No sequence of BIOS toggles is presented as a recipe, and no BIOS update is presented as a first step. If a setting is changed as a test, exactly one setting changes, the result is logged, and the setting is restored if it does not help.
The Four-Arm Test Matrix
Principle: one variable changes at a time, every run is logged with the same fields, and a conclusion requires the pattern to repeat, not a single success or failure. Run each arm more than once, in varying order where practical, and do not draw a conclusion from a single pass. Arms B and C exist to expose the Fast Startup contamination. Arm D exists to separate Restart from any warm path.
| Arm | Operation | What it tests |
|---|---|---|
| A: Warm reboot | Windows Restart, power never removed | The reported fault path itself |
| B: Hybrid power-on | Windows Shut down with Fast Startup on, the default, then the power button | Whether saved kernel and driver state changes the outcome versus Restart |
| C: True cold boot | Shut down, then PSU switched off and/or mains unplugged so standby power is drained, then power on | Whether full power loss clears what Restart does not |
| D: Restart after a cold boot, without any intervening load | Cold boot by arm C, reach the desktop, then immediately Restart with no game or heavy workload in between | Whether the warm failure needs a preceding workload, heat, or driver session, or follows any warm reboot regardless |
What to log on every run: date and time, and which arm was used; what was on screen at each stage, and the exact point of divergence from normal; whether the card appears in Device Manager, and its status text quoted exactly; whether resolution and refresh rate are normal or locked low; the GPU link reading under load, not at idle, meaning negotiated generation and width as reported by the reader's monitoring tool, compared with the card and slot capability, with idle readings recorded separately and never used as the load comparison; GPU temperature and clock together, at the same moment, under the same workload, because a temperature alone, or a clock alone, proves nothing about throttling; recent changes before the pattern started, covering driver update, Windows update, BIOS update, new monitor, cable, or dock, new software, a moved PC, or an added card or device, with a timeline of changes treated as evidence and nothing changed recorded as a logged answer rather than an assumption; and event evidence for that run, as described below.
Reading the matrix: if arm A fails, arm C passes, and the pattern repeats, the fault lives in something Restart does not clear: device hardware state retained across warm reset, a warm-path link retraining result, warm-path firmware enumeration, or retained power or monitor state. The next step is to split those branches, not to declare a failed card. If arm A fails, arm B also fails, and arm C passes, the conclusion is the same, with the added information that Fast Startup state is not the differentiator, because both the restored-kernel path and the rebuilt-kernel path fail and only power loss helps. If arm A passes at times and fails at times while arm C passes consistently, intermittency is itself information: log the conditions of the passing and failing warm runs separately, including preceding workload, uptime, monitor state, and whether a second display was connected, rather than averaging the results away. If arm A fails and arm C also fails, this is no longer the pattern of this guide. Reclassify toward a general detection or black-screen path and stop forcing the warm-only explanation. If arm D passes but arm A after a heavy session fails, the trigger is not any warm reboot but something accumulated during the preceding session: driver session state, a workload left in a particular state, a power-management state entered under load, or a display state. That narrows the investigation to session-dependent branches. If arm D fails exactly like arm A, the trigger is in the warm path itself, independent of preceding load, and link retraining, firmware enumeration, and retained device or monitor state take priority.
Event and Status Evidence After a Failed Warm Run
In Device Manager, under Display adapters, open the GPU, then Properties, then Device status. Code 43 means Windows has stopped this device because it has reported problems. It does not by itself prove a dead card: the framing for Code 43 lists both driver problems and hardware problems as possible causes, and the status must be read together with the matrix result and the change timeline. A status reading of This device is working properly, quoted exactly, is also evidence, and it points the investigation away from a driver-reported device fault and toward link, enumeration, or display branches.
In Event Viewer, under Windows Logs and then System, look around the exact time of the failure. Display source Event ID 4101 indicates a driver recovery event happened. Kernel-Power Event 41 indicates the whole PC rebooted unexpectedly and points away from a pure GPU-driver story. WHEA-Logger entries point toward hardware error reporting in CPU, PCIe, or memory class, and should be quoted exactly and taken to professional inspection rather than interpreted by the reader as a named failed part. Reliability Monitor for the same window places driver, application, and hardware-error entries on one timeline. Absence of an event is also information, but weak information: no 4101 does not rule out a display-link or firmware-enumeration failure, because those may leave no driver event at all. A reader who treats no error in Event Viewer as no fault will miss exactly the branches this guide is built to separate.
One Change at a Time After the Matrix
After the matrix, each branch test changes exactly one thing, logs the result, and restores the change if it does not help. First, the display branch, because it is external and reversible: same PC, different monitor port or cable type, single monitor only, and the monitor power-cycled independently. This tests display-link retraining and monitor retained state, not the PCIe link. Second, the driver branch: note the exact driver version and date, and whether the pattern started with a driver or Windows update. A clean driver reinstall is a test, not a conclusion: if the pattern survives a clean reinstall of a driver version that previously worked, the driver-install hypothesis weakens and the investigation moves on. Repeated reinstalls of the same version as a loop are not instructed. Third, the firmware-settings branch: record the settings listed above before changing anything, change one setting, test, and restore if unhelpful. Never present a batch of BIOS changes. Fourth, the link branch: compare negotiated generation and width under load after a cold boot versus after a warm reboot, on the same workload. A warm-only downgrade, generation or width, that repeats is link-branch evidence. A downgrade present in both paths is not evidence for this guide's pattern. Fifth, the session branch: arm D versus arm A after load, as described in the matrix reading.
Branch-by-Branch Differentiation
Driver and operating-system state: this branch is favoured when failure appears only after reaching Windows, at login, on the desktop, or at first application load; when Device Manager status or a 4101 event is present; and when the pattern started with a driver or Windows update. It is disfavoured when failure occurs before any Windows display, with no logo and no loading indicator, because the driver is not yet the actor at that stage. In that case, move to the firmware and link branches rather than lingering on drivers.
PCIe link retraining: this branch is favoured when failure or degradation appears at or before Windows load, or as a repeatable warm-only link downgrade under load; and when behaviour changes with slot-related variables, such as which slot is used or whether a riser or extension is in the path, recorded as a configuration fact rather than as an instruction to rebuild the machine. It is disfavoured when full generation and width under load are identical in cold and warm paths and the display itself is normal until an application loads, a shape that belongs more to the driver branch or the power-state branch.
Firmware enumeration: this branch is favoured when the pattern started with a BIOS update or a settings change; when the card is missing, resolution is low, or performance is degraded from the first Windows display after a warm reboot but not after a cold boot; and when recorded settings show a prerequisite mismatch, for example Resizable BAR shown as enabled without Above 4G Decoding, or a forced generation that differs from Auto behaviour. The test is recording and single-setting comparison, not a settings recipe. If the reader cannot confidently navigate firmware settings, that is a stop-line case, not a reason to guess.
Power-state and retained state: this branch is favoured when only a full power drain, arm C, restores normal behaviour, while both Restart and hybrid power-on fail; when behaviour depends on how the machine was powered off, PSU switch or unplug versus on-screen Shut down; and when behaviour depends on preceding power-management states, for example a device that entered a low-power state and does not re-enumerate cleanly. This branch cannot be confirmed by observation alone at home beyond the matrix result. It is a referral conclusion, meaning the pattern points to retained device or power-rail state and professional inspection is needed to identify which, not a named failed component.
Display link and monitor state: this branch is favoured when the PC itself reaches Windows, suggested by sounds, keyboard response, or remote behaviour, but the screen shows no signal; when changing the monitor, cable, port, or power-cycling the monitor alone changes the outcome; when the fault follows a particular monitor or port rather than the PC's reset path; and when the fault appears only with a second display connected. DisplayPort link training, monitor power state, and the EDID handshake are the mechanisms here, and they must be named as display-link mechanisms, not folded into PCIe link language. This branch is disfavoured when the card is missing from Device Manager or shows a driver error even on a known-good display, because that is not a monitor handshake story.
Session accumulation: this branch is favoured when arm D, an immediate restart after a cold boot, passes while arm A after a long or heavy session fails. Candidate session factors, to be logged rather than assumed, include a particular application or overlay left running, a driver session that has performed a TDR recovery earlier in the session, a second display connected mid-session, or a power-plan or display-sleep transition during the session. The test is to reproduce the session factor in isolation, same preceding workload followed by restart, versus a clean session, one factor at a time.
What a Link Reading Does and Does Not Prove
A low generation reading at idle does not prove a fault: idle downshift and ASPM power saving legitimately lower the generation at idle, and the comparison must be under load, in both paths. A low width reading is stronger evidence than a low generation reading, because width does not scale down at idle the same way, but a slot wired for fewer lanes by design, or a BIOS lane-sharing setting, can also produce a narrower width legitimately. Capability must be checked before calling a width reading a downgrade. The statement that the link retrained is not a diagnosis: retraining happens on every boot and on every Recovery event. The diagnostic question is always to what did it train, in which path, under load, and does that result repeat. The statement that reseating fixed it is not a diagnosis either: reseating combines power drain, time, contact renewal, and fresh training. Reporting reseating as the fix without separating those factors misleads the next reader whose fault is in a different branch.
Boundary Cases
Intermittent warm failure: log passing and failing warm runs separately. Intermittency does not invalidate the pattern, but a single warm failure after a single cold success does not establish it. Warn against changing drivers, firmware, and hardware in response to one event. Warm failure only after heavy load, not after idle use: use arm D to separate session accumulation from the warm path itself, and do not slide into temperature-fault language without temperature-plus-clock evidence recorded together. Warm failure only with a second monitor, dock, or hub: reclassify toward the display-link branch and toward the display-path articles, and do not proceed with a PCIe diagnosis as if the second display were irrelevant. Warm failure only after a driver or BIOS update: the change timeline is the lead, the branch order becomes driver first or firmware first, and hardware failure must not be presented as the default. A laptop with switchable or mux graphics adds internal display routing, a mux switch, and integrated-versus-discrete selection that desktop PCIe reasoning does not cover: the matrix still applies, but a laptop-specific routing conclusion must not be drawn from desktop PCIe logic, and the case goes to professional inspection rather than improvised laptop steps. No display at any stage versus display that appears then disappears are different branches even though readers describe both as black screen, and the article forces the reader to state the stage. A card that works after a warm reboot but degrades only under load shows a load-path pattern, not a warm-only reset pattern, unless the same load after a cold boot is normal: the matrix comparison is cold-load versus warm-load on the same workload, not warm-load versus idle. Virtualisation and passthrough setups are out of scope: FLR and device reset behaviour in a passthrough or multi-function context is a specialist topic, and desktop reasoning from this guide must not be applied to it.
What Not to Do, and the Stop Line
The reader's job in this guide is to log arms, stages, statuses, events, and link readings, and to change at most one reversible software or firmware setting at a time with the original value recorded. Stop and seek professional inspection if the pattern survives the matrix and the branch tests; if Device Manager shows a persistent error status or the card is repeatedly missing; if WHEA-Logger hardware-error entries accompany the failure; if the machine needs internal handling; if the card is in warranty; or if the reader is not confident navigating firmware settings. Do not open the card, remove its cooler, tighten screws, or replace pads or paste. Do not present any of those as reader steps. If the tests in this guide narrow the fault to retained device state, a link path, or firmware enumeration but do not identify a reversible cause, that narrowing is itself the useful result to bring to an inspection. It is not a failed diagnosis.
Do not ignore warranty. A card still under warranty should not be opened for any reason arising from this guide. The correct response to a suspected reset-path or enumeration problem on a warranted card is to contact the warranty provider with the observation record and follow their instructions, rather than attempting any internal step. Even for a card out of warranty, internal steps belong to a professional with the right materials and the experience to assess the link path and enumeration state, because the difference between a correct assessment and a guessed fix is not visible from outside until the card is tested again in both paths.
Prevention, in this context, is record-keeping and restraint. Keep the change timeline: driver versions and dates, Windows update dates, BIOS update dates, and any monitor, cable, dock, or slot configuration change, noted at the time they happen rather than reconstructed later from memory. Keep power-on order constant across test arms, so that a monitor-order effect is not mistaken for a reset-path effect. When a setting is tested, restore it if it does not help, so that the machine does not accumulate unrecorded changes that confound the next diagnosis. A professional inspection, when it is needed, will start from the matrix record and the branch evidence: which arms passed and failed, at which stage, with which Device Manager status and event entries, and with which under-load link readings in the cold path versus the warm path. That record lets the inspection start from the narrowed branches rather than repeating the whole matrix from the beginning.
If you live nearby, you are welcome to bring your observation record to HK GPU Fix for advice, at Shatin Dali Plaza Room 1209, phone 2186 8010, WhatsApp 6558 6806. Bring the arm-by-arm log, the stage at which each warm run diverged, the Device Manager status text, the event entries, and the under-load link readings for cold and warm paths. That record is enough for the other party to judge the next step quickly, without needing to repeat the whole sequence from the start.
Misconceptions
Misconception one: shutdown then power-on is a cold boot. False with Fast Startup on, because the kernel and drivers are restored and standby power remains. Only arm C qualifies as a cold boot in this guide.
Misconception two: Restart is the thorough reset. True for the kernel, false for hardware power state. Restart never removes power, so it cannot clear state that only power loss clears.
Misconception three: warm reboot means the card is hot. False. Warm means power-retained. Temperature is a separate measurement, recorded together with clock speed.
Misconception four: fans spinning means the card is fine, or fans not spinning means the card is dead. Fan behaviour is not a diagnostic verdict: some cards do not spin fans at low temperature by design, and a powered card with a failed link or driver can still spin fans. Record fan behaviour as an observation, never as a conclusion.
Misconception five: Code 43 means the card is dead. False. Code 43 reports that Windows stopped the device because the driver reported a problem. Driver, firmware, power, and hardware causes are all still open until the matrix narrows them.
Misconception six: no error in Event Viewer means no fault. False. Firmware enumeration and display-link failures may leave no driver event at all.
Misconception seven: the newest driver or newest BIOS will fix it. Unproven in any individual case, and a BIOS or driver change is itself a new variable that can confound the matrix. Record versions, change one thing at a time, and do not present updating as a guaranteed fix.
Misconception eight: forcing the highest PCIe generation fixes a link problem. Forcing a generation removes the link's ability to negotiate down to a stable rate and can make a marginal link worse. Generation settings are recorded evidence, and any change is a single-variable test with a restore step.
Misconception nine: retraining failed, therefore the slot or card is faulty. Retraining to a lower state is evidence about the link path, which includes slot, riser, retimer, lane allocation, firmware settings, and the card. Naming one of those as the culprit from a downgrade alone is over-claiming.
Misconception ten: if it works after unplugging it, the power supply is faulty. Power drain clears retained state in several branches, not only power-supply faults. The matrix result points to retained state, not to a named component.
Frequently Asked Questions
Why does my card work after I switch the machine off at the wall, but fail after a Restart?
Because Restart and a full power drain clear different things. Restart rebuilds the kernel and reloads drivers, but it never removes power, so device hardware state, sticky state, warm-path link retraining results, and warm-path firmware enumeration decisions survive it. A full power drain removes standby power as well, resetting device hardware state from power-up. When only the power drain helps, and the pattern repeats, the fault lives in something Restart does not clear, and the branch tests in this guide split those candidates.
Is shutdown then power-on the same as a cold boot?
Not with Fast Startup on, which is the Windows default. With Fast Startup on, shutdown hibernates the kernel session and restores it at next power-on, and standby power remains. A cold boot in this guide means shutdown, then PSU switched off and/or mains unplugged so standby power is drained, then power on. Only that arm may be called a cold boot when testing this pattern.
What does warm reboot mean, if not that the card is hot?
Warm means power was never removed during the reboot. It says nothing about temperature. Temperature must be measured separately, together with clock speed, under the same workload, before any thermal conclusion is drawn. A warm-reboot fault at an ordinary temperature is still a warm-reboot fault, and its explanation lies in reset path, link retraining, enumeration, or retained state, not in heat assumed from the word warm.
What does Code 43 actually mean?
Code 43 means Windows has stopped this device because it has reported problems. It does not, by itself, prove a dead card, because both driver problems and hardware problems remain possible causes. Read the status together with the test matrix result and the change timeline, and quote the status text exactly rather than paraphrasing it into a verdict.
Why do I need to run each test arm more than once?
Because a single pass cannot establish a pattern. A warm run can pass or fail for a session reason, a monitor state, or a one-off job, and a single cold success can coincide with a changed condition. Repetition in varying order, with the same logged fields each time, is what separates a repeating reset-path pattern from a single event. Changing drivers, firmware, or hardware in response to one event destroys the evidence before the pattern is established.
My card is missing from Device Manager after a warm reboot but present after a cold boot. What does that point to?
That shape points toward the warm-path branches: firmware enumeration, link retraining, or retained device state, rather than a driver that fails after reaching Windows. The next step is the branch evidence: the stage at which display was lost, the under-load link readings in both paths, the recorded firmware settings, and whether a display change alters the outcome. If those do not identify a reversible cause, the narrowed result goes to professional inspection.
Should I force the PCIe generation to the highest setting?
No. Forcing a generation is not presented as a fix in this guide. It removes the link's ability to negotiate down to a stable rate and can make a marginal link worse. Record the generation setting and the negotiated generation under load in both paths. Any change to the setting is a single-variable test with the original value recorded and restored if it does not help.
What if my machine is a laptop?
The test matrix still applies, but a laptop's internal display routing, mux switch, and integrated-versus-discrete selection add a layer that desktop PCIe reasoning does not cover. Do not draw a laptop-specific routing conclusion from desktop PCIe logic. Log the arms, stages, and statuses as described, and take the record to professional inspection rather than improvising laptop steps.
What should I bring when seeking professional help?
Bring the arm-by-arm log with dates and times, the stage at which each warm run diverged, the Device Manager status text quoted exactly, the Event Viewer and Reliability Monitor entries for the failure window, the under-load link readings for the cold path versus the warm path, temperature and clock readings taken together where relevant, and the change timeline covering driver, Windows, and BIOS updates and any monitor, cable, or dock change. That record lets an inspection start from the narrowed branches.
Does reseating the card prove what was wrong?
No. Reseating combines power drain, time passing, slot contact renewal, and fresh link training into one uncontrolled change. It may restore normal behaviour without identifying which of those factors mattered. The matrix exists to separate them as far as observation allows, and a reseat result should be logged as a combined change, not reported as a diagnosis of one component.