GPU Faults Only When Cold vs Only When Hot? Temperature-Dependent Failure Diagnosis Guide

Key Takeaways
A graphics card that fails only when cold and works fine once warm, or works fine when cold and fails only once hot, is giving you a temperature clue. Comparing symptoms across temperature states is how you separate which layer is actually affected.
What This Article Covers, and What It Does Not
This article covers one specific pattern: the same graphics card behaves differently at different temperatures. It fails during the first few minutes after a cold start and then recovers as the machine warms up, or it works perfectly when cold and only starts failing after sustained use has raised temperatures. The key idea is temperature comparison: the same system, the same card, different results at different temperature states.
Several neighbouring topics need to be marked off first so they are not mixed into this one. First, faults that appear all the time regardless of temperature belong to a general fault guide; this article mentions them only in passing. Second, fan noise or dust build-up as a cleaning topic on its own is not the focus here; cleaning is mentioned only where the cooling layer is being checked. Third, problems that began immediately after a driver update belong to a driver topic and are mentioned only as one exclusion step. Fourth, power-supply insufficiency under load can look very similar to a hot-state fault, but it sits on a different layer; this article uses comparison tests to separate the two rather than treating them as the same thing.
Keeping these boundaries clear matters because temperature is the easiest factor to misread. Temperature can be the trigger, it can be an amplifier, or it can be a coincidence: as a machine warms up, load, fan speed, and power demand all change at the same time. If you do not separate those variables one at a time, you may blame temperature for what is really a load problem, or blame load for what is really a component that only misbehaves within a certain temperature range. Every diagnostic step below follows one rule: change one condition at a time, and write down the result before and after.
Why Temperature Changes Fault Behaviour
To understand temperature-related faults, it helps to know which parts of a graphics card actually respond to temperature. The first is thermal expansion and contraction of metals and circuit boards. A graphics card carries many solder joints connecting the processor, memory, and other components to the board. Metals expand when heated and contract when cooled, and different parts expand by different amounts. If a solder joint already has a microscopic crack or a weakened contact, expansion at one temperature may temporarily improve the contact, while contraction at another temperature may temporarily worsen it. This mechanism can explain faults that appear only when cold and vanish once warm, and equally faults that appear only once hot.
The second factor is the temperature behaviour of electronic components themselves. Capacitors, resistors, and semiconductor parts change their electrical characteristics with temperature. Within the normal design range these changes do not affect operation, but when a component has aged or is already near a critical threshold, a temperature shift can push it past that threshold. A component that behaves normally at room temperature may drift out of specification at high temperature, disturbing power delivery or signal timing; conversely, a component that behaves normally when hot may fail to start properly when cold because its characteristics have shifted in the other direction. These are component-layer possibilities that cannot be judged from appearance alone.
The third factor is the cooling system itself. Fan speed, the thermal interface between the processor and the heatsink, and airflow inside the case all determine the actual temperature the card experiences. Poor cooling makes temperature rise faster and higher than normal, so a problem that would otherwise only appear under extreme conditions can show up during ordinary use. In the other direction, a cooling system that has not yet reached its normal operating state during a cold start, for example because fan speed strategy differs at low temperature, can produce brief early anomalies. Cooling-layer symptoms and component-layer symptoms can look very similar, which is why they must be checked separately.
The fourth factor is the tight coupling between load and temperature. A machine is usually hot because it is under load, and high load also means higher power demand, more frequent memory access, and higher clock speeds. When several factors change together, the fault you observe may be caused by load rather than by temperature itself. Diagnosis must separate load from temperature: for example, observing behaviour under low load while temperature rises slowly, or under high load while improved cooling keeps temperature lower, and watching which variable the fault follows. A fault that follows temperature is temperature-related; a fault that follows load belongs to a load-layer topic.
Classifying Symptoms First: Cold-Only, Hot-Only, and Temperature-Unrelated
The first category is cold-only: the fault appears in the first few minutes after switching on, then disappears as the machine warms up, and repeats on every cold start. Typical presentations include no display at start-up, visual artefacts during start-up, or a crash shortly after start-up, followed by normal operation for the rest of the day. The most diagnostically useful feature is the time window: symptoms fade gradually between cold and warm states. Record how long after start-up the symptoms disappear, and whether the improvement is gradual or sudden. Gradual improvement points to a process that changes steadily with temperature; sudden recovery may involve a fan-speed strategy change or a threshold being crossed.
The second category is hot-only: the machine is completely normal when cold and only fails after sustained use has raised temperatures. Typical presentations include visual artefacts after prolonged use, crashes, black screens, automatic restarts, or a sudden sharp drop in performance. The key comparison is the state when symptoms appear and what happens when load is reduced or cooling is improved. If symptoms clear immediately when load drops, the load layer is more likely; if symptoms persist until temperature has clearly fallen, the temperature layer is more likely.
The third category looks temperature-related but is not. A common case is a fault whose timing happens to coincide with warm-up, but whose real cause is a scheduled task, a program that only starts some time after boot, or a driver problem that only appears under a particular operation. The signature of this category is a fixed timing that does not move with temperature: the fault appears at the same point after start-up regardless of room temperature or cooling conditions. The test is to change temperature conditions and see whether the fault timing moves. If it does not move at all, temperature is not the main cause.
After classifying, record four things: whether the machine was cold or hot when the fault appeared, how long the machine had been running before the fault, roughly what the load was at the time, and whether the fault faded gradually before disappearing. These four items form the baseline for every comparison test that follows.
The Diagnostic Core: Using Temperature as a Variable
The principle is simple: hold everything else fixed, change one temperature-related factor, and watch whether the fault follows. The four comparison tests below are ordered from safest and easiest to more involved.
Test one: cold-start reproduction. After the machine has fully cooled, for example overnight without use, switch it on and immediately perform the same light operation you would normally do, recording whether the fault appears, how long after start-up it appears, and whether it clears on its own. Then leave the machine running until it is fully warm and repeat the same operation, comparing results. A fault that appears only when cold and never when warm is evidence for the cold-only pattern; a fault that appears in both states belongs to a persistent-fault workflow instead.
Test two: hot-state load comparison. Once the machine is warm, increase load gradually and record the load level and approximate temperature at which the fault appears. Then reduce load and observe whether the fault eases. The purpose is not to push the machine to its limit but to find the boundary: at what load level the fault begins, and below what load level it never appears. A clear boundary suggests a definite trigger condition; a vague, inconsistent boundary suggests other variables such as case ventilation or room temperature are also involved.
Test three: cooling-condition comparison. Without opening any hardware, change the ventilation conditions around the case, for example by making sure vents are not blocked and room temperature is normal, and observe whether the fault is delayed or appears earlier. If improved ventilation clearly delays the fault, the cooling layer becomes more likely; if fault timing does not change at all, cooling is probably not the main cause. Do not direct an external cold source at the inside of the card, and do not reach inside a running machine.
Test four: immediate-restart comparison. After a fault appears, restart immediately and observe whether it reproduces straight away. A hot-state fault will usually reproduce more easily on immediate restart because the machine is still hot; a cold-only fault will usually not reproduce on immediate restart because the machine is already warm, and will only return after a full cool-down. This test helps separate whether the fault follows temperature state or follows the number of start-ups or a start-up procedure.
After all four tests, you should have a simple comparison table recording temperature state, load, fault presence, and timing for each test. That table is the evidence for judging layers.
How to Read the Test Results
| Test combination | Result | Most likely layer | Next step |
|---|---|---|---|
| Cold-state light operation | Fault appears, clears once warm | Cold-only temperature-related | Record disappearance timing and improvement pattern, prepare inspection notes |
| Cold-state light operation | No fault; fault appears only under hot high load | Hot-state temperature or load-related | Continue load and cooling comparisons to separate temperature from load |
| Hot high-load fault, load reduced | Symptoms clear immediately | Load layer more likely | Move to a load-related workflow rather than blaming temperature alone |
| Hot high-load fault, load reduced | Symptoms persist until temperature falls | Temperature layer more likely | Check cooling conditions, then prepare inspection notes |
| Ventilation improved | Fault clearly delayed | Cooling layer more likely | Check case ventilation and fan operation |
| Ventilation improved | Fault timing unchanged | Temperature may not be the main cause | Check scheduled tasks, drivers, and other time-related factors |
| Immediate restart | Fault reproduces straight away | Fault follows the hot state | Strengthens hot-state evidence |
| Immediate restart | Fault clears, returns only after cooling | Fault follows the cold state | Strengthens cold-only evidence |
The key point of this table is that no single test can directly prove which component has failed. The endpoint of temperature comparison testing is to narrow the range to a temperature-related layer, not to identify a specific part yourself. From outside, solder-joint, component-characteristic, and cooling layers cannot be told apart reliably; that distinction requires professional inspection.
Separating the Layers: Cooling, Fans, Power, and Components
The cooling layer is the easiest to observe from outside. Whether fans spin, whether fan speed changes with temperature, whether case vents are blocked, and whether airflow inside the case looks normal can all be checked without opening hardware. A fan that does not spin at all, or whose speed behaves erratically, is a clear warning sign; but a fan that spins normally does not guarantee cooling is fine, because an aged thermal interface or a contact problem between heatsink and processor is invisible from outside. Cooling-layer judgement must rest on comparison results, not on fan movement alone.
The fan layer deserves one specific caution: fan noise and fan performance are two different things. A loud fan is not necessarily a poor cooler, and a quiet fan is not necessarily a good one. Judge the fan layer by whether speed changes with temperature and load, and whether fault timing changes with fan behaviour. If the fan already behaves abnormally before the fault appears, for example speed suddenly dropping to zero or fluctuating wildly, the fan layer becomes more suspect; if the fan runs normally throughout and the fault still follows temperature, the fan is probably not the main cause.
The power layer is easily confused with the temperature layer because high load raises both temperature and power demand. The way to separate them is comparison: if the fault appears even under low load at lower temperature, the power layer is less likely; if improved cooling keeps temperature lower under high load but the fault still appears, the power or load layer becomes more likely. This comparison is not an invitation to modify power delivery yourself; it is a way to keep your conclusion honest, leaving a question mark rather than writing temperature down as a certainty when load and temperature are bundled together.
The component layer is what remains when cooling conditions, fan behaviour, and load comparisons have all been done and the fault still follows temperature reliably. When that happens, the component layer becomes more likely. It must be stressed, though, that within the component layer, finer distinctions such as solder joints, capacitors, or memory parts cannot be made from outside. The purpose of reaching this point is not to name a part but to arrive at a complete diagnostic conclusion: a temperature-related component-layer range, ready to hand over with your notes.
Edge Cases That Are Easy to Misread
The first edge case is a fault that appears only in colder seasons or at lower room temperatures and vanishes in summer. This is easily mistaken for a cold-only fault, but it may involve the gap between room temperature and internal temperature, or a component characteristic at low ambient temperature. Judge it by cold-versus-warm comparisons within the same environment, not by impressions across different seasons.
The second edge case is a fault with no fixed pattern, sometimes cold-only and sometimes hot-only. This is no longer a classic temperature-related fault; it may involve several overlapping factors, or an intermittent problem that merely coincides with temperature changes. Do not force it into the cold-only or hot-only category. Record temperature state, load, and timing for every occurrence and look for a common factor behind them.
The third edge case is normal temperature readings combined with a fault that still follows usage time. Monitoring software usually reports one sensor point, which does not represent the temperature at every location on the card. A localised hot spot can exist while the reported figure looks normal. A normal reading therefore cannot overturn comparison-test results; readings are reference data, and comparison results are the evidence.
The fourth edge case is a fault that appears or disappears after thermal paste replacement or cleaning. The time link tempts people to credit or blame the cooling work, but the disassembly process itself may have changed contact states, or cleaning may have changed case airflow. A time link on its own is not evidence; temperature comparison results are.
Common Misconceptions
Misconception one: a fault that appears only when hot must be insufficient cooling. Cooling is one possibility, not the only one. Component characteristics, power delivery, and load can all first show themselves in the hot state. Do not equate a hot-state fault with a cooling problem before doing a cooling comparison.
Misconception two: a cold-only fault that clears once warm means the problem has fixed itself. Recovery at high temperature only means that contact or component behaviour has temporarily returned to a workable range; the underlying cause remains. The next cold start is very likely to reproduce it. Treating temporary disappearance as a cure is the easiest way to delay proper handling.
Misconception three: if temperature readings are not high, temperature cannot be involved. Reported figures cover sensor points only, not every location, and not the rate or magnitude of temperature change. A component can fail because of how fast temperature changes rather than because absolute temperature is extreme.
Misconception four: pointing a fan at the machine counts as diagnosis. An external fan changes more than temperature: airflow direction and humidity distribution inside the case also change, and the variable is hard to control. Diagnosis should use comparison under normal ventilation conditions, not an extreme external cooling state that would never occur in ordinary use.
Misconception five: opening the heatsink yourself counts as deeper diagnosis. Removing a heatsink involves the thermal interface, screw pressure, and warranty considerations, and even after opening, the faulty layer cannot be identified by eye. Temperature comparison testing already narrows the range; opening hardware does not make the conclusion more accurate and only adds risk.
Misconception six: a temperature-related fault must get progressively worse. The rate of progression depends on the cause: some worsen gradually, others stay at the same trigger condition for a long time. Do not infer the cause from the rate of worsening, and do not assume a fault that has not worsened will never return.
Stop Points: When Not to Test Further
Stop if any of the following occur. First, any unusual smell, smoke, abnormal heat felt from the machine, or scorch marks. Second, the fault changes from occasional to frequent, or spreads from one symptom to several. Third, you are starting to consider opening the graphics card, removing the heatsink, replacing the thermal interface yourself, or prying at components. Fourth, comparison testing is complete and consistently points to a temperature-related layer, but you are unsure whether it is cooling or component layer. Fifth, the machine develops a new problem during testing, such as failing to shut down or start normally.
Stopping is not giving up; it is protecting evidence. Your temperature comparison table is the most valuable material you hold. Continuing, especially by opening hardware or applying extreme cooling, will only blur results that are currently clear.
Information to Record Before Seeking Inspection
First, whether the fault is cold-only or hot-only, and how long the machine had been running when it appeared. Second, the timing pattern of appearance and disappearance, including roughly how long disappearance takes and whether reducing load helps. Third, the results of each of the four comparison tests: cold-start reproduction, load boundary, ventilation comparison, and restart comparison. Fourth, fan behaviour, case ventilation, and room conditions. Fifth, anything that changed before the fault first appeared, such as moving the machine, cleaning, part changes, or driver updates. Sixth, temperature readings if available, recorded together with the load state at the time, never as a lone figure. Seventh, the concrete presentation: no display, visual artefacts, crash, black screen, or automatic restart. Describe facts only, without inferring the cause yourself.
Write these down as text, not as a verbal summary from memory. A written record lets inspection start directly at the temperature-related layer instead of repeating the whole comparison sequence.
Frequently Asked Questions
Q: A fault that appears only when cold and clears once warm, does that mean the card has fixed itself?
A: No. Recovery at high temperature only means that some contact or component behaviour has temporarily returned to a workable range. The underlying cause remains, and the next cold start is very likely to reproduce the fault. Temporary disappearance should never be treated as a cure.
Q: A fault that appears only when hot, is it always insufficient cooling?
A: Not necessarily. Cooling is one possibility, but component characteristics, power delivery, and load can all first show themselves in the hot state. Cooling and load comparisons are needed to separate whether the fault follows temperature or follows load.
Q: Temperature readings look normal, can temperature-related causes be ruled out?
A: No. Readings cover sensor points only, not the temperature at every location on the card, and not the rate of temperature change. Comparison-test results are the evidence; readings are only reference data.
Q: How can I tell whether a fault follows temperature or follows load?
A: Change one variable while holding the other steady. If symptoms clear immediately when load drops, the load layer is more likely; if symptoms persist until temperature has clearly fallen, the temperature layer is more likely. A test that changes both at once has no diagnostic value.
Q: Improved case ventilation delays the fault, does that prove a cooling problem?
A: It raises the likelihood of the cooling layer, but it does not rule out the component layer. Better cooling can delay the trigger condition while the underlying cause remains at component level. Ventilation improvement is a reasonable everyday measure, but it is not proof the problem is solved.
Q: What does an immediate restart result tell me?
A: For a cold-only fault, an immediate restart should clear the fault because the machine is already warm, which strengthens cold-only evidence. For a hot-state fault, an immediate restart should reproduce the fault more easily because the machine is still hot. Restart results must be read together with cold-versus-warm comparisons.
Q: Is opening the heatsink myself helpful?
A: No, and it carries risk. Removing a heatsink involves the thermal interface, screw pressure, and warranty considerations, and the faulty layer still cannot be identified by eye afterwards. Temperature comparison testing already narrows the range sufficiently; opening hardware does not improve the conclusion.
Q: Will a temperature-related fault necessarily get worse over time?
A: Progression depends on the cause: some worsen gradually, others stay at the same trigger condition for a long time. Do not infer the cause from the rate of worsening, and do not assume a fault that has not worsened will never return. Recording the conditions of each occurrence is more useful than estimating a worsening rate.
Conclusion: Temperature Is a Variable, Not a Verdict
Temperature-related faults test discipline more than technical skill. Change one condition at a time, record temperature state, load, and result each time, and follow whichever variable the fault follows. Temperature itself is not a conclusion; the result of temperature comparison is. Write the comparison table clearly, stop at the stop points, and every step afterwards becomes much clearer.
If comparison testing consistently points to a temperature-related layer, or if warning signs such as unusual smell or smoke appear, bring the notes listed above when seeking inspection. That is far more useful than simply saying the fault appears when hot. HK GPU Fix is located at Room 1209, Dali Plaza, Shek Mun. Phone 2186 8010, WhatsApp 6558 6806.