GPU Keeps Disappearing from Device Manager Until Rescan? Enumeration, Link Retraining & Power State Diagnosis Guide

Key Summary
A graphics card that vanishes from Device Manager and returns after a rescan is not necessarily dead: a rescan that brings it back proves enumeration can still succeed. This guide uses that fork, with Kernel-PnP records, a power-management comparison, and a link reading under load, to narrow five suspect layers through seven steps.
The Shape of This Problem: Four Conditions, All Required
This guide is about one narrow shape of failure, and the shape has four conditions that must all be present before anything in it applies. First, the graphics card entry in Device Manager has completely disappeared — no entry, no error code, nothing to click on. Second, the card was working normally before, with no hardware changed and no motherboard changed. Third, the card comes back after a scan for hardware changes or after a restart. Fourth, the disappearance repeats intermittently rather than happening once.
Each condition matters, because each one excludes a different neighbouring problem. A card that is present with a yellow exclamation mark and an error code is a different state: it is an error-code problem, not a disappearance problem. A card that stopped being recognised after a hardware or motherboard change is a different trigger: something changed, and the investigation starts from that change. A card that never comes back is a different outcome: it is a persistent failure, not an intermittent one. And a disappearance that happened once and never repeated is not a pattern at all — it is a single event, and no diagnosis built on patterns can be applied to it.
The shape is worth defining this precisely because it is the gap between four neighbouring problems. The guide on a card not being recognised at all covers a persistent failure after a hardware change. The guide on a card not being recognised after a motherboard change covers a different trigger. The guide on error codes covers a card that is present with a code attached. The guide on a monitor disappearing intermittently covers a display that vanishes while the graphics card itself is still in Device Manager — and its own stop line says plainly that a graphics card disappearing from Device Manager is outside its scope. This guide is the space between those four: the card itself, gone, intermittently, and back again after a rescan.
The third condition — the card comes back after a rescan — is not a repair. It is a fork. Rescanning for hardware changes asks Windows to ask the bus again whether the card is there. If the answer comes back yes, enumeration can still succeed sometimes, and the fault sits in one of the layers that fail intermittently: link training, power state, or slot contact. If the answer comes back no, the card is not in this guide's territory at all, and the persistent-failure path is the one to follow. The rescan is a diagnostic question, not a fix, and treating it as a fix is one of the six misconceptions this guide corrects.
Why a Disappearance Is Not an Error Message
When enumeration fails, the system's behaviour is that the device is not there — not that the device is there with an error message. That sentence, from the neighbouring guide on a monitor disappearing, is the key to understanding this whole problem. A missing entry is not a missing diagnosis; it is the diagnosis. Windows could not enumerate the device at that moment, so there is nothing to attach an error code to.
That distinction changes what evidence to look for. The evidence is not in Device Manager, because the device is not there to be looked at. The evidence is in the system event log, recorded at the moment the device disappeared. Windows logs Plug and Play events under the Kernel-PnP source, and the timestamps on those records can be compared against the time the card disappeared. That comparison turns "I think the card dropped out" into "the system recorded an enumeration event at this time" — a checkable fact rather than an impression.
Two user reports, published as Microsoft Q&A questions with community answers rather than as official Microsoft documentation, describe what those records look like. One report describes an RX 5700 XT crashing and then showing Event 411 under Kernel-PnP, with Problem Status 0xC00000E5 and the device driver listed as amduw23g. Another report, dated 26 January 2025, describes an RX 7900 XT whose three monitors went black at random and never came back; disabling and re-enabling the graphics card in Device Manager restored it temporarily, until the next load made it crash again; Task Manager could not find the graphics card; and Kernel-PnP events 400, 411, 420, and 430 were all recorded under the graphics card entry. Both reports are individual user accounts, and they cannot be generalised into a universal rule — but they establish what to look for: whether a Kernel-PnP record appears at the moment of disappearance, which Event ID it carries, and whether one appears every time the card disappears.
An important caution belongs with Event 411 specifically. A community answer in the first report suggests that Kernel-PnP errors point to a driver or hardware problem and recommends a clean driver removal and an older driver version; the same discussion also mentions DXGI_ERROR_DEVICE_REMOVED, the error raised when a device is removed. Event 411 on its own records that a device initialisation or driver load went wrong. It is not proof of a cause. It must be combined with the disappearance pattern — whether the card vanishes at idle, under load, or after waking — and with the rescan comparison, before it can point at any layer. Jumping from an Event ID straight to "the hardware is broken" is exactly the kind of shortcut this guide is built to prevent.
The Five Suspect Layers
Everything that can make a graphics card disappear intermittently and return after a rescan sits in one of five layers. The layers are listed here so that each later step can be read as testing one of them, and so that a result can be mapped to a layer without guessing.
Layer 1: The Enumeration and Driver Layer
This layer is enumeration itself, the driver loading, and the question of whether the device is gone or present with an error code. Its evidence is the Kernel-PnP record: whether one appears at the moment of disappearance, which Event ID it carries, and whether it appears every time. If the card disappears and no Kernel-PnP record appears at that moment, the enumeration layer is not the one speaking — the disappearance may have been recorded elsewhere, or the timing may be wrong, and the record needs to be checked again rather than assumed. If a record appears every time, with the same Event ID, that regularity is itself evidence: the same event, repeating, points at a repeatable mechanism rather than a random one.
This layer also holds the distinction between the two states that are most often confused. A card that is present with a yellow exclamation mark and Code 45 is not the same as a card that is completely gone. Code 45 means Windows knows the device exists but believes it is not connected. A complete disappearance means Windows has nothing to report on at all. The two states follow different flows, and the guide on error codes covers the first while this guide covers the second. Intermittency itself is a signal here, in the words of that neighbouring guide: intermittent Code 45 — sometimes recognised, sometimes Code 45 — points to the contact, slot, and power layers more than to the driver layer. The same reasoning applies to a complete disappearance: a fault that comes and goes points at something that fails sometimes, not at a setting that is wrong all the time.
Layer 2: The Link Training Layer
This layer is the PCIe link itself, and it needs explaining because a link is not a fixed wire. A PCIe link's speed and width are not fixed in advance; the two ends settle them step by step when the machine starts up. The training logic begins at the slowest rate and works its way up, settling lane width, speed, and equalisation as it goes. When the link is unsteady, that logic falls back to re-aligning the clock in its Recovery state, or it slows the link down to keep the connection alive. That process is called link training, and the account of it here follows the neighbouring guides on link speed and on cold-boot behaviour, restated for this guide.
The relevance to a disappearing card is direct: if training fails at the margin, the device may not be enumerated at all, and it simply is not there. A riser cable, a marginal slot contact, or a link that trains successfully most of the time and fails occasionally will produce exactly this symptom — the card is there, then it is not, then a rescan asks again and this time training succeeds. The distinguishing evidence for this layer comes from Step 5: the link speed and width read under load, compared against what the card and the slot are designed to support.
Layer 3: The Power State Layer
This layer is ASPM — Active State Power Management — and the power states a PCIe link moves through. L0 is the normal operating state. L0s is a standby state that is logically still L0 but with reduced physical signalling. L1 saves more power but has a larger wake latency, and disables the link and transaction layers; it can have L1.1 and L1.2 sub-states. A link enters L1 when it is idle, and either end can start the transition by sending an Electrical Idle Ordered Set. These facts come from NXP application note AN14913, "PCIe Power Management using ASPM", revision 2.0, dated 1 April 2026.
The relevance is that waking from a low-power state requires the link to retrain, and if retraining fails, the signal is lost — and with it, the device. A card that disappears after waking from sleep points at this layer. A card that disappears at idle points at the transitions between L0s and L1. Windows exposes one control over this behaviour in Power Options, under PCI Express and Link State Power Management, with three settings: Off, Moderate power savings, and Maximum power savings. Changing that setting to Off is an isolation test, following the neighbouring guide on idle and light-load crashes: if the disappearances stop, the evidence points at the ASPM and wake-training layer; if they continue unchanged, power management is not the variable. The test is a diagnostic probe, not a recommendation to turn power saving off permanently.
Layer 4: The Slot, Contact, and Power Layer
This layer is the physical connection: the gold fingers, the slot, the power connectors, and the 12V supply. Intermittent contact here is read by the system as the device being removed — which is, from the system's point of view, exactly what happened. The connection was there, then it was not. A riser cable, a card in a lower slot rather than the top slot connected directly to the processor, a power connector that is not fully seated, or a machine that was recently moved are all facts that belong in the record for this layer.
The trap in this layer is the reseat. Pulling the card out and seating it again alters four things in one move: the power is cut, a stretch of time goes by, the slot contacts mate afresh, and the link negotiates itself again. If the card then works, it is impossible to tell which of the four mattered — a point from the neighbouring guide on cold-boot behaviour. A reseat that "fixes" the problem proves nothing about the cause, and the next disappearance can still happen. This layer is therefore a recording layer in this guide, not a repair step: note what is there, and do not change several things at once.
Layer 5: The Firmware and Settings Layer
This layer is the firmware: the BIOS Hot Plug option, Above 4G decoding, Resizable BAR, Compatibility Support Module, the forced PCIe generation, and the firmware version. Its most distinctive check is Hot Plug. If PCIe Hot Plug is enabled in BIOS, Windows can treat a normal internal graphics card as a hot-pluggable device, listing it under "Safely Remove Hardware". That behaviour is documented in a WindowsForum article citing Microsoft Learn driver documentation, and in a MakeTechEasier article on stopping Windows from offering to eject the GPU. The underlying mechanism is Windows reading the read-only device property DEVPKEY_Device_SafeRemovalRequired, which depends on whether the surprise-removal capability bit CM_DEVCAP_SURPRISEREMOVALOK is unset and whether the device sits under a removable parent device.
The important distinction is that being listed as removable is a classification problem, not a disappearance. The device is still there. Hot Plug being enabled does not, on its own, explain a card vanishing — but it is one check in the firmware layer, and if the machine does not use an external GPU, Hot Plug should not be enabled. Above 4G decoding, Resizable BAR, Compatibility Support Module, and a forced PCIe generation are the other settings in this layer, taken as concepts from the neighbouring guide on a card not being recognised: any of them changed recently is a fact worth recording, because the enumeration path itself runs through them.
| Layer | What it covers | Strongest clue | What would weaken it |
|---|---|---|---|
| 1. Enumeration and driver | Kernel-PnP records, driver loading, gone vs error code | A Kernel-PnP record at the exact moment of disappearance, every time | No Kernel-PnP record at the disappearance time |
| 2. Link training | LTSSM, Recovery state, speed or width drop, riser cable, slot contact | Link speed or width degraded under load | Link reads normally under load, every time |
| 3. Power state | ASPM L0s and L1, Link State Power Management, sleep wake, D3 | Disappearances stop when Link State Power Management is set to Off | Disappearances continue unchanged with it Off |
| 4. Slot, contact, and power | Gold fingers, slot, power connectors, 12V supply | Disappearances follow movement, a riser cable, or a loose connector | No physical change, no riser, connectors fully seated |
| 5. Firmware and settings | BIOS Hot Plug, Above 4G, Resizable BAR, CSM, forced PCIe generation, firmware version | Hot Plug enabled without an external GPU, or a setting recently changed | All settings unchanged and Hot Plug disabled |
The First Decisive Fork: Does a Rescan Bring It Back?
The most valuable single observation in this diagnosis is also the simplest to obtain. Open Device Manager, go to Action, and choose Scan for hardware changes. Then watch what happens.
If the card comes back after the scan, stay in this guide. The scan asked Windows to ask again, and this time enumeration succeeded. That proves enumeration can succeed sometimes, and the fault sits in one of the layers that fail intermittently — link training, power state, or slot contact — rather than in a dead card. A dead card does not come back intermittently.
If the card does not come back after the scan, this guide no longer applies. Follow the persistent-failure path covered by the neighbouring guide on a card not being recognised, because the fault is not an intermittent enumeration failure. There is a third outcome worth recording separately: the card comes back after a restart but not after a scan. That difference points at different enumeration paths — a restart runs a full enumeration, while a scan only asks again — and it should be written down, because it is evidence about which path fails.
A note of caution applies to the fork. The rescan is a question, and like any question it can be asked badly. If the card is rescanned while it is still in the middle of disappearing, or if settings were changed between the disappearance and the scan, the answer is weakened. Record the conditions: what was done, in what order, and what the result was. "It came back after a rescan" is evidence. "It came back after I rescanned and restarted and reseated it" is not, because three things changed at once.
The Second Decisive Test: The Power-Management Comparison
The second decisive test is the power-management comparison, and it should be run as soon as the rescan fork has been observed. In Power Options, under PCI Express and then Link State Power Management, change the setting to Off, and then observe for several days. Change nothing else at the same time.
There are two possible outcomes, and each one narrows the field sharply. If the disappearances stop after the setting is changed to Off, the evidence points at the ASPM and wake-training layer: with the link no longer dropping into L0s or L1, the wake retraining that was failing no longer happens, and the card stays enumerated. If the disappearances continue unchanged, power management is not the variable, and the investigation moves on to the link and slot layers. Either outcome is useful, and either outcome is only useful if the comparison ran for several days — a single day without a disappearance proves nothing when the disappearances were already intermittent.
Two warnings belong with this test. The first is that it is a diagnostic probe, not a permanent setting. Turning power saving off to see whether the symptom stops is a legitimate test; leaving it off forever without understanding why it worked is not a repair, and it leaves the original question unanswered. The second warning is about changing more than one thing at once. If Link State Power Management is changed, the driver is also updated, and the card is also reseated, and the disappearances then stop, nothing has been learned about which of the three changes mattered. One change per attempt, with the result recorded, is slower in the moment and far faster overall.
The Seven-Step Diagnosis
The seven steps below are ordered so that the cheapest and most decisive comparisons come first, and so that each step's result decides which step follows. They should be worked through in order, with the record kept as they go, rather than jumped between.
Step 1: Characterise the Disappearance — What Does Device Manager Show, and What Does Event Viewer Say?
At the moment the card disappears, open Device Manager and look at the Display adapters section. A completely disappeared entry — no entry, no error code — is this guide's shape. An entry present with a yellow exclamation mark and Code 45 belongs to the error-code guide. An entry present with Code 10, 12, or 31 also belongs to the error-code guide. A screen that still shows a picture while the graphics card is missing from Device Manager is a combination worth writing down, and the investigation continues. At the same time, open Event Viewer and record whether a Kernel-PnP record appeared at the moment of disappearance, which Event ID it carries, and whether its timestamp matches the disappearance. That record is the evidence that turns an impression into a checkable fact.
Step 2: Run the Rescan Comparison
In Device Manager, go to Action and choose Scan for hardware changes. If the card comes back after the scan, continue in this guide — enumeration can succeed sometimes. If it does not come back after the scan, this guide no longer applies, and the persistent-failure path is the one to follow. If the card comes back only after a restart and not after a scan, record that difference: it points at different enumeration paths, with a restart running a full enumeration and a scan only asking again.
Step 3: Record the Pattern
Over several consecutive days, record every disappearance: did it happen at idle, under load, after waking from sleep, or at a completely random time? How often? Did every disappearance end with a rescan bringing the card back, or did some need a restart? A pattern cannot be claimed from a single event. If the card has disappeared only once and never again, the pattern has not been established, and this guide's flow should not be used. The pattern is the evidence that the later steps test against, and without it the later steps are guesses.
Step 4: Run the Power-Management Comparison
In Power Options, under PCI Express and Link State Power Management, change the setting to Off, and observe for several days. If the disappearances stop, the evidence points at the ASPM and wake-training layer. If they continue unchanged, power management is not the variable, and the investigation moves on. This step is only a diagnostic probe, not a permanent change to power saving.
Step 5: Read the Link Status Under Load
Once the card is back, open GPU-Z and read the link generation and width while the card is under load, and compare the reading against what the card and the slot are designed to support. The reading must be taken under load, because at idle the PCIe link drops speed to save power, and an idle reading that looks low is expected behaviour that proves nothing. A link that still runs degraded under load — a lower generation or a narrower width than designed — or one accompanied by crashes or black screens, makes the link-training layer suspect. A link that reads normally under load does not clear the layer entirely, because training can fail sometimes and succeed at other times, but it moves the investigation on to the next step.
Step 6: Record the Slot, Contact, and Power Facts
Record whether a riser cable is in use, which slot the card is in — the top slot connected directly or a lower slot — whether the power connectors are fully seated, and whether the machine has been moved recently. Do not change several things at once. Pulling the card out and seating it again alters four things at the same time — the power is cut, a stretch of time goes by, the slot contacts mate afresh, and the link negotiates itself again — and which one mattered cannot then be separated. If the reader is not familiar with removing and installing a card, stop here: this step is only a record.
Step 7: Check Firmware and Driver Settings, Then Seek Inspection
Check whether PCIe Hot Plug is enabled in BIOS — if the machine does not use an external GPU, it should not be — and whether Above 4G decoding, Resizable BAR, or Compatibility Support Module settings have changed recently, and whether the PCIe generation has been forced to an older generation. Perform one clean driver installation, with the version chosen against current official guidance rather than on the assumption that the newest version is correct. If the card still disappears intermittently after all seven steps, bring the full record for inspection: the time of every disappearance, the Kernel-PnP Event IDs, the rescan-versus-restart comparison results, the power-management comparison result, the GPU-Z readings taken under load, and the list of BIOS settings.
| Step | What to do | Result that narrows the fault | Next step |
|---|---|---|---|
| 1 | Check Device Manager state at disappearance; check Event Viewer for Kernel-PnP records | Completely gone with no error code — this guide; present with a code — error-code guide | Step 2 |
| 2 | Scan for hardware changes | Card returns — enumeration can succeed sometimes; does not return — persistent-failure path | Step 3, or the persistent-failure guide |
| 3 | Record every disappearance for several days: idle, load, after sleep, or random | A pattern forms, or it does not | Step 4 |
| 4 | Set PCI Express Link State Power Management to Off; observe for several days | Disappearances stop — ASPM and wake-training layer; continue — power management is not the variable | Step 5 |
| 5 | Read link generation and width in GPU-Z under load | Degraded under load — link-training layer suspect; normal — move on | Step 6 |
| 6 | Record riser cable, slot position, power connectors, recent movement; change nothing | Physical facts recorded for the slot and power layer | Step 7 |
| 7 | Check BIOS Hot Plug, Above 4G, Resizable BAR, CSM, forced PCIe generation; one clean driver installation | Firmware or driver layer isolated, or fault persists | Inspection with full record |
Prevention: Keeping the Next Disappearance From Being a Mystery
Prevention for an intermittent disappearance is less about any single setting and more about keeping a record that makes the next diagnosis shorter. A card that disappears once a week and is never written down is a mystery every time it happens; a card whose disappearances are logged with times, Kernel-PnP Event IDs, and rescan results is a pattern with a history.
Keep a short log of the facts that matter: the BIOS settings as they are now, the driver version in use, whether a riser cable is installed, and which slot the card sits in. When any of those changes — a BIOS update, a driver update, a card moved to another slot — note the date, and then watch for several days before concluding that the change helped or harmed. When a disappearance happens, write down the time, what the machine was doing, and whether a rescan or a restart brought the card back, before doing anything else. None of these habits prevents a marginal link or a failing contact. What they do is make the next diagnosis start from evidence rather than from memory, and the seven steps above have less ground to cover.
Boundaries: When the Problem Belongs to Another Guide
A disappearing graphics card shares symptoms with several neighbouring problems, and the boundaries are worth stating plainly, because working through the wrong guide wastes the time that the right guide would have used.
If the card is present in Device Manager with a yellow exclamation mark and Code 45, that is the error-code guide — a different state with a different flow. If the card stopped being recognised after a hardware or motherboard change and stays unrecognised, that is the guide on a card not being recognised, or the guide on a card not being recognised after a motherboard change. If the monitor disappears while the graphics card itself is still in Device Manager, that is the guide on a monitor disappearing intermittently — and its own stop line says that a monitor disappearing and never coming back, a graphics card disappearing from Device Manager, or the machine freezing are all signs beyond the scope of self-diagnosis. If the problem is a first cold boot with no picture, fixed by restarting, that is the guide on cold-boot behaviour, where the cold and warm paths are different. If the link runs at a lower speed or narrower width but the device is still present, that is the guide on link speed dropping. If the machine crashes or shows a black screen at idle or under light load but the device is still present, that is the guide on idle and light-load crashes. If the driver stops responding and the device is still present, that is the guide on driver timeouts, where the driver is reset but the device remains. And if a rescan does not bring the card back at all, that is the persistent-failure path, not this guide.
The boundary that most often causes confusion is the one with the error-code guide, because Code 45 and a complete disappearance feel like the same problem — a card that is not working. The test that separates them is Step 1: an entry present with a code belongs to the error-code guide; no entry at all belongs here. If both states appear at different times, record each one separately, because they are two different states and they may not have the same cause.
Six Misconceptions
1. A disappeared card means the card is dead. It does not. A rescan bringing the card back proves that enumeration can still succeed sometimes, and a dead card does not come back intermittently. Intermittency itself points at the contact, slot, power, link-training, or power-state layers more than at a dead piece of hardware.
2. A rescan bringing the card back means it is fixed. It does not. A rescan only asks Windows to ask again; the conditions that made the first attempt fail have not reappeared, which is not the same as those conditions having gone away. The card can disappear again at the next idle period, the next wake from sleep, or the next load.
3. Code 45 and a complete disappearance are the same thing. They are not. Code 45 is an entry present with an error code; a complete disappearance is no entry and no error code. The two states follow different flows and should not be folded together.
4. Updating to the latest driver will definitely fix it. It will not. The driver is only one of five layers. Enumeration, link training, power state, and firmware can all make a device disappear intermittently, and the version should be chosen against current official guidance rather than on the assumption that the newest version is correct.
5. A disappearance at idle must be ASPM. It is not necessarily ASPM. ASPM is only one layer, and the power-management comparison in Step 4 has to be run before that claim can be made. Attributing an idle disappearance to ASPM without running the comparison is an unproven leap.
6. Reseating the card and having it work means the cause was found. It does not. Pulling the card out changes four things at once — the power is cut, a stretch of time goes by, the slot contacts mate afresh, and the link negotiates itself again — and which one mattered cannot be separated. Working again is not the same as knowing why, and the problem can return.
Sources
External sources are cited here by source and date only, as published material consulted for this guide, not as testing carried out by this site. Microsoft Q&A, "Event 411, Kernel-PnP error" (user report with community answers, not official documentation). Microsoft Q&A, 26 January 2025, Kernel-PnP errors 400, 411, 420, and 430 with DXGI_ERROR_DEVICE_REMOVED (user report with community answers, not official documentation). NXP application note AN14913, "PCIe Power Management using ASPM", revision 2.0, 1 April 2026. WindowsForum article on Windows 11 showing a GPU as removable, citing Microsoft Learn driver documentation. MakeTechEasier article on stopping Windows from offering to eject the GPU. The link-training, cold-and-warm-path, power-management, and driver-timeout concepts used here come from neighbouring guides on this site, restated for this guide; neighbouring guides are referred to by subject for boundary purposes only, and their mechanisms are not reproduced here.
FAQ
My graphics card has completely disappeared from Device Manager — does that mean the card is dead?
Not necessarily. A rescan for hardware changes bringing the card back proves that enumeration can still succeed sometimes; a dead card does not come back intermittently. Intermittency itself points at the contact, slot, power, link-training, or power-state layers more than at dead hardware.
How do I tell a complete disappearance apart from Code 45?
Code 45 is an entry present with a yellow exclamation mark and an error code; a complete disappearance is no entry and no error code — the device entry itself cannot be found. The two states follow different flows: Code 45 follows the guide on error codes (#350), while a complete disappearance follows this guide.
If rescanning brings the card back, is it fixed?
No. Rescanning only asks Windows to ask again; the conditions that made the first attempt fail have not reappeared, which is not the same as those conditions having gone away. The card can disappear again at the next idle period, after the next wake from sleep, or under the next load. A rescan is an isolation comparison, not a repair.
When the card disappears, what should I look for in Event Viewer?
Look for Kernel-PnP records: whether one appeared at the moment of disappearance, which Event ID it carries (user reports have shown 400, 411, 420, and 430), and whether its timestamp matches the time the card disappeared. That comparison turns "I think the card dropped out" into "the system recorded an enumeration event at this time".
How do I use the power management comparison to tell whether it is an ASPM problem?
Go to Power Options, then PCI Express, then Link State Power Management, and change it to Off, then observe for several days. If the disappearances stop, the evidence points at the ASPM and wake-training layer; if they continue unchanged, power management is not the variable. This step is only a diagnostic probe, not a recommendation to turn power saving off permanently.
Why must I read the link status in GPU-Z under load?
Because at idle the PCIe link drops speed to save power, so a low idle reading is expected behaviour and proves nothing. The link generation and width have to be checked again while a rendering test is actually running, before any judgement can be made about whether they are normal.
What is the difference between disappearing after sleep and disappearing at idle?
Disappearing after waking from sleep points at a retraining failure on wake (the power-state layer); disappearing at idle points at a failure in switching between ASPM L0s and L1. The two triggers point at different layers, so they must be recorded separately and not folded together.
What does the Hot Plug option in BIOS have to do with the card disappearing?
Enabling PCIe Hot Plug in BIOS can make Windows treat an internal graphics card as a hot-pluggable device and list it under "Safely Remove Hardware". That is itself a classification problem — the device is still there — and it is not the same as the card disappearing on its own. But it is one check in the firmware layer, and if the machine does not use an external GPU, Hot Plug should not be enabled.
If reseating the card fixes it, does that mean I found the cause?
No. Pulling the card out changes four things at once — the power is cut, a stretch of time goes by, the slot contacts mate afresh, and the link negotiates itself again — and which one mattered cannot be separated. Working again is not the same as finding the cause, and the problem can happen again.
If all seven steps still leave the card disappearing intermittently, what should I bring for inspection?
The time of every disappearance, the Kernel-PnP Event IDs, the rescan-versus-restart comparison results, the power-management comparison results, the GPU-Z link readings taken under load, and the BIOS settings list (Hot Plug, Above 4G decoding, Resizable BAR, Compatibility Support Module, and PCIe generation). Inspection is charged under the website's published three-tier pricing of HK$400–2,800, with the actual price confirmed after inspection. The shop is at Room 1209, Dali Plaza, Shek Mun; phone 2186 8010, WhatsApp 6558 6806.