GPU Works Sometimes, Fails Sometimes After a Drop? PCB Micro-Fractures & Solder Cracks: Intermittent Fault Mechanism & Repair Judgment

2026-10-20 ｜ GPU Repair ｜ About 12 min read ｜ HK GPU Fix Editorial

Key Takeaways

GPU faults that track temperature — working cold, failing hot, or the reverse — often point to cracked solder joints, PCB micro-fractures, or cold joints rather than dead chips. Hairline cracks open under thermal expansion and close when cool. Diagnosis needs history, flex testing, and thermal probing; reflow is temporary and X-ray cannot always see the damage.

The Signature: Faults That Follow Temperature

Some graphics cards misbehave in a way that looks almost personal. The card works fine in the morning, starts artifacting after an hour of gaming, then runs normally again after it cools down overnight. Or it refuses to boot, yet works after a few minutes of warming up. The symptoms come and go, and the common thread is temperature.

This pattern — intermittent faults modulated by temperature — is one of the most telling signatures in electronics failure analysis. Three distinct mechanisms can produce it: thermal-fatigue cracks in BGA solder joints, copper micro-cracks in the PCB itself, and cold solder joints left over from manufacturing. They feel identical to the user, but their causes, diagnostic evidence, and repair prospects are different. Understanding the difference is what separates a considered repair decision from guesswork.

How Solder Joints Crack: BGA Thermal Fatigue

Both the GPU die package and the VRAM memory chips are attached to the graphics card's circuit board through BGA (ball grid array) mounting: hundreds of tiny solder balls underneath the chip, each one acting simultaneously as the mechanical anchor and the electrical connection. During heavy load, a GPU can run at temperatures in the 70–90°C range. Every gaming session, and every shutdown, is a thermal cycle — heat up, cool down, repeat.

The solder ball, the FR-4 circuit board, and the chip's package substrate all expand and contract with temperature, but each material has a different coefficient of thermal expansion (CTE). Because they expand by different amounts, the solder ball is repeatedly sheared — stretched one way when hot, the other when cold. Published reviews of thermal fatigue in micro solder joints describe the mechanism: in the early cycles, cracks nucleate and grow along the interface between the solder and the intermetallic compound (IMC) layer; after many cycles, the solder's grain structure coarsens, and cracks propagate further through the coarsened regions. Solder operates at a high homologous temperature — it is already above half its melting point at room temperature — so creep (slow plastic deformation under stress) dominates its fatigue behavior. This is the theoretical foundation for why solder joints have a finite thermal-cycle life.

Modern graphics cards use lead-free solder (the SAC alloy family, required by RoHS regulations), which is more brittle than the older tin-lead alloy. Field practitioners widely observe that lead-free BGA joints are more prone to open-circuit failures after years of service — this is practitioner observation from repair forums, not controlled research data, and should be read as such.

Here is the point most often misunderstood: once a fatigue crack has grown inside a solder ball, simply reheating it — a "reflow" — melts and re-solidifies the same fatigued, coarsened alloy. It does not restore the original ductility and toughness of the joint. Technical guidance from board fabricator JLCPCB states this explicitly: after thermal fatigue has caused grain coarsening and micro-cracking, reballing (replacing the solder balls entirely with new ones) is the theoretically sound way to restore reliability, while reflow alone is a stopgap measure.

The Invisible Enemy: PCB Copper Micro-Cracks

Not every intermittent fault lives in a solder ball. The circuit board itself can crack, and these cracks are even harder to find. Failure-analysis literature describes three families of copper micro-crack causes: manufacturing defects (over-etched traces, poor lamination creating inherent weak points), thermal-cycling stress (CTE mismatch at layer-transition points in multilayer boards), and mechanical stress (depanelization shock, mishandling during assembly, or vibration in the field environment).

The failure is progressive. A micro-crack raises the resistance of a trace slightly, which degrades signal quality; as the crack grows, resistance rises further; eventually the trace goes fully open. The most important property is how temperature modulates this: a crack only a few micrometers wide measures nearly zero ohms at room temperature, because the two fractured faces are still pressed together. When the board heats up and expands, the faces are pulled apart — resistance jumps or the circuit opens. When it cools, they press back together and conduction returns.

This is a precise physical explanation for "works when cold, black-screens when hot, works again after cooling off." A published failure-analysis case from Spirit Electronics illustrates the point perfectly: a via-to-trace connection whose conduction degraded under thermal expansion and recovered when cool showed no suspicious damage under X-ray at all; the evidence was only found later through destructive cross-section analysis.

Standard optical inspection — even under a microscope — will not find micrometer-scale cracks. And as that case demonstrates, X-ray can miss them too. A clean visual inspection, or even a clean X-ray report, does not prove the board is free of micro-fractures.

When a Drop Is the Trigger: Impact Damage

A card that worked perfectly, then fell off a desk, survived a bumpy courier journey, or took a knock — and started acting intermittently afterwards — deserves special attention. Academic research on BGA failure under drop impact (published in ScienceDirect) identifies the root cause as the combined action of mechanical shock and PCB bending vibration. Two findings are particularly relevant: the crack propagation rate increases with the amplitude of the vibration power spectral density, and the failure location migrates as loading intensifies — from the solder bulk near the PCB side, to the joint's neck, to the near-package interface — while the failure mode shifts from fatigue cracking to brittle fracture.

The practical meaning: a drop can flex the board for a fraction of a second and leave invisible cracks in corner BGA joints or inner-layer copper. The card's surface can look completely unblemished — no scratches, no dents — yet hairline damage is already there, and it presents as an intermittent fault from day one. Structural reasoning also suggests that a long, heavy graphics card mounted in a PCIe slot concentrates bending stress near the PCIe gold fingers and the far-end solder joints when shaken during transport, because of the lever effect — this is a structural inference, and quantitative confirmation is unverified.

Cold Joint vs. Solder Crack: Not the Same Thing

The terms get thrown around interchangeably, but they describe different defects with different implications:

Because their symptoms overlap (both produce intermittent faults), confusing them leads to wrong repair expectations — which is why the diagnostic workflow below matters.

Diagnosing Intermittent Faults: A Technician's Workflow

Before any of this applies, the usual suspects must be ruled out first: cables, monitor, power supply, motherboard slot, and drivers. The standard approach is cross-testing — different slot, different machine, different power supply, and the integrated GPU as a control. Only then does board-level diagnosis begin.

Step 0 — Take a history (often the most valuable step)

• Has the card ever been dropped, knocked, or shipped? When did the problem start — did it begin right after an impact? This points toward impact damage.
• Is the fault linked to temperature? Failing when hot but fine when cold (or the reverse)? This points toward solder cracks, micro-fractures, or cold joints.
• Are the symptoms visual artifacts, black screen, or a complete failure to boot? Artifacts are more often associated with VRAM data channels; black screens or undetected cards more often with the GPU itself, power delivery, or BIOS — these are tendencies, not conclusions.

Step 1 — Visual inspection under magnification

Check for visible PCB bending, impact marks, missing components, cracked solder on large joints (power inductors, capacitor leads), and scratched or oxidized PCIe gold fingers from transport or repeated insertion. Note the limits: BGA solder balls are hidden beneath the chip, and inner-layer copper micro-cracks are invisible. A clean visual inspection does not rule out a hairline fracture.

Step 2 — Flex testing

With the card powered and at idle (or under light load), gently press different points with an insulated tool — the four corners of the GPU package, individual VRAM chips, different board areas — and watch whether the fault appears or disappears instantly. If pressing a particular chip or spot immediately triggers artifacts or a black screen, that location's solder joints or the traces beneath it are strongly suspect. Pressing the GPU package itself with a reaction points at BGA joints; pressing a VRAM chip points at that chip's joints.

This is a diagnostic aid, not a treatment, and there is a real risk: excessive flexing can widen a hairline crack or tear a pad off the board (pad lift), turning a repairable fault into an unrepairable one. Press gently and briefly. Folk remedies such as padding material under the heatsink to keep pressure on the chip only mask the symptom and should not be presented as legitimate repair.

Step 3 — Thermal probing

Use a hot-air station at low temperature (around 100–120°C — enough to expand the board, not to melt solder) to heat the suspect area evenly and observe whether the fault is induced or temporarily disappears. Thermal expansion can pull a crack open (inducing the fault) but can also push contact faces together (temporarily improving it) — both directions are diagnostically meaningful. Conversely, freeze spray on the suspect chip or area tests the cold direction. Repair guides list freeze spray plus controlled heating as a standard technique for isolating intermittent solder-joint and marginal-chip faults.

Interpretation: a fault that clearly tracks temperature suggests a mechanical connection problem (solder crack, micro-fracture, cold joint); a fault completely unaffected by temperature leans toward internal chip or power-logic issues.

Step 4 — Electrical measurement

Measure the power rails with a multimeter — 12V, GPU core voltage, VRAM voltage, 3.3V — checking for correct voltages and for shorts to ground. A shorted rail points toward a MOSFET or internal chip failure. Measuring the resistance or diode drop of GPU/VRAM rails to ground can rule out short-circuit-type faults, but it cannot distinguish a solder crack from an internal chip failure. Professional shops also use PCIe graphics-card diagnostic fixtures that test the data bus from the gold fingers to the GPU for opens and shorts with indicator lights.

Step 5 — Putting the evidence together

These are tendencies, not a diagnostic formula. True confirmation usually requires desoldering for inspection (X-ray, cross-section) or validating through a successful reball.

What Tests Can and Cannot Prove: The Hard Limits

Honesty about diagnostic limits is part of the service:

1. Crack location decides everything. A crack in a corner BGA ball of the GPU (reballable) versus a crack in an inner-layer copper trace (invisible, hard to locate) versus a crack at a via barrel (the Spirit Electronics scenario) are worlds apart in repair difficulty. Inner-layer micro-cracks can be so hard to locate that destructive cross-sectioning is needed just to find them.
2. Multiple faults can stack. A dropped card can simultaneously have BGA micro-cracks, dislodged components, and a warped PCB. Fixing one spot leaves the others, producing the frustrating "repaired but still unstable" outcome. Any repair judgment should leave room for this possibility.
3. Chip-internal fracture is indistinguishable from the outside. Experienced repair practitioners note that many faults are not in the BGA solder at all but in the connection between the silicon die and the package substrate inside the chip; heating merely "revives" it for days to months. No external test can reliably separate an internal fracture from a solder-ball crack — this is a hard boundary of diagnosis.
4. X-ray has blind spots. X-ray reveals solder-ball shape, bridging, and large voids, but fine fatigue cracks may be invisible — the Spirit Electronics case found nothing on X-ray. A clean X-ray image is not a certificate of health.
5. Diode-mode measurements only exclude. Ground-resistance checks can rule out shorted rails but say nothing about cracks versus internal chip failure.
6. Reflow recovery is a clue, not a cure. If heating restores function, it corroborates a mechanical connection problem — but it does not prove a specific joint was the root cause, and it does not mean the repair is permanent.
7. Within warranty, use the warranty. Do not open, heat, or attempt anything yourself — self-repair voids the warranty. This is a general principle.

Reflow, Reball, or Retire: The Repair Judgment

The framework below is a decision guide, not a price promise. Actual repair charges follow the publicly listed three-tier pricing (HK$400–2,800); no specific repair fee is stated or implied here.

When repair is worth considering

• The card has meaningful residual value (mid-range or better, still worth something on the secondhand market) and the fault is reasonably localized — pressure or thermal testing points at a specific chip or region.
• A simple BGA fatigue crack is suspected and the chip itself is likely healthy — reballing is then the theoretically sounder option.
• Power delivery, BIOS, and missing-component issues have already been excluded.

Reflow: diagnostic and temporary

Reflow is a fast, low-cost diagnostic tool: if heating restores function, the evidence for a mechanical connection fault strengthens. It should not be packaged as a permanent fix, and the risk of recurrence should be stated plainly. Reflow has its own temperature-window risks: too cool produces cold joints; too hot softens the laminate, lifts pads, and damages innocent neighboring components. Reference figures from soldering guides place preheat around 150–180°C and reflow around 220–245°C. Large GPU dies are also reported by practitioners to be sensitive to overheating — careless operation can write the card off entirely (a forum-practitioner claim, not manufacturer data).

The home-oven approach deserves a direct word. Forum threads contain anecdotes of oven-bake "successes," but from a professional standpoint: the temperature profile is uncontrolled, there is no preheat, plastic parts and capacitors can crack, and solder fumes are toxic. This article does not present oven reflow as a recommended practice — folk attempts exist, but the risks are high, and it is not endorsed.

Reball: the sounder option when the chip is healthy

Reballing replaces the old solder balls with entirely new ones, theoretically restoring joint reliability. But it has strict limits: it cannot fix internal chip fractures, inner-layer copper micro-cracks, or substrate problems such as Black Pad. If a reball does not help, the suspicion shifts to the chip body or board interior. It is also labor- and equipment-intensive. And per JLCPCB's guidance, reballing is appropriate where the chip itself is healthy and only the joints are bad; if diagnosis shows the chip is internally dead, reballing wastes time — the chip must be replaced, and sourcing new chips for older cards is a separate difficulty.

When retirement is the honest answer

• Low-end or old cards where the repair cost approaches or exceeds the card's residual value.
• A power rail shorted to ground that points at the GPU's internals (replacement chip unavailable or prohibitively priced).
• A visibly warped board, delamination, or large-scale burn damage.
• An inner-layer micro-crack that cannot be located, or cross-section evidence of via / inner-trace fracture — restoration difficulty is extreme and reliability low.
• A board already subjected to repeated improper heating (lifted pads, blistered laminate) — further heating only makes things worse.
• The card is under warranty — use the warranty, do not attempt repair.

The governing principle: complete the diagnosis first, then discuss repair. The line between "worth repairing" and "not worth repairing" rests on three things — how clearly the fault is localized, repair cost versus the card's residual value, and whether the owner accepts the recurrence risk. If any one of the three does not hold, the honest recommendation is to not repair.

Myths That Refuse to Die

"Bake it and it's fixed." A card that works again after reheating is not necessarily repaired. The fatigued alloy has not been replaced, and if the root cause is inside the chip or in an inner copper layer, the heat merely masked it. Recurrence is the common outcome — a consistent observation across practitioner forums.

"The home oven method is fine — people online did it." Anecdotes exist; control does not. Uncontrolled temperature profiles, no preheat, toxic fumes, and the risk of cracking plastic parts and capacitors make this a high-risk gamble, not a recommended practice.

"The flex test is a repair." Pressing only diagnoses. Excessive force damages more, and folk pressure-padding tricks mask symptoms rather than fixing anything. Do not present them as legitimate repair.

"One symptom equals a diagnosis." Hot-artifacts can also be VRAM thermal failure or ageing power delivery; a black screen after a drop can also be a dislodged component or a shaken-loose BIOS chip. History and the full test set are needed — never judge from a single symptom.

"Reballing fixes everything." It fixes solder joints. It does not fix internal chip fractures, inner-layer copper cracks, or substrate defects. The word should be used precisely: before confirmation, "suspected solder-joint or trace contact issue" is safer than declaring "solder crack," since the true failure point is often not the solder ball at all.

FAQ

Q1: Why does my graphics card work when cold but fail when hot?

Heat expands the circuit board and the materials on it. If a solder joint or copper trace has a hairline crack, the crack's two faces are pressed together at room temperature and conduct normally; under thermal expansion the faces are pulled apart, resistance jumps, and the circuit opens — producing artifacts, a black screen, or a driver crash. When the card cools, the faces close again and it works. This temperature modulation is the hallmark of cracked joints and micro-fractures, though the same symptom can also come from internal chip thermal failure, so temperature-tracking alone is not a confirmed diagnosis.

Q2: The X-ray showed nothing. Does that prove there is no crack?

No. X-ray can show solder-ball shape, bridging, and large voids, but fine fatigue cracks are frequently invisible to it. In a documented PCB failure-analysis case, a via-to-trace connection with temperature-dependent intermittent conduction showed no suspicious damage on X-ray; the evidence only emerged through destructive cross-sectioning. A clean X-ray does not exclude micro-cracks in solder joints or inner-layer copper.

Q3: Reflow made my card work again — does that confirm it was a solder crack?

Not conclusively. Recovery after heating corroborates a mechanical connection problem, which strengthens the suspicion, but it does not identify the exact failure point — and it does not mean the repair is permanent. Practitioners consistently report that reflow revivals often recur within weeks or months, either because the fatigued alloy was never replaced or because the real fault (for example, an internal chip fracture or an inner-layer trace crack) was never in a solder ball at all.

Q4: Is the home-oven reflow method recommended?

No. While forum anecdotes of oven-bake "successes" exist, the temperature profile is uncontrolled, there is no proper preheat, plastic components and capacitors can crack, and solder fumes are toxic. It is a high-risk gamble rather than a recommended practice, and it can leave the board worse off (lifted pads, blistered laminate) than before. Professional judgment treats oven reflow as a folk attempt, not a repair method.

Q5: Between reflow and reball, which should I choose?

They serve different roles. Reflow is best understood as a quick, low-cost diagnostic: if heating restores function, the fault is likely a mechanical connection problem — but it should not be sold as a permanent fix, and the recurrence risk should be disclosed. Reballing replaces the solder balls entirely and is the theoretically sounder option for restoring joint reliability — but only when the chip itself is healthy and the fault is localized to its joints. It cannot resolve internal chip fractures, inner-layer copper micro-cracks, or substrate defects, and if a reball does not help, the problem likely lies elsewhere.

Q6: My card was dropped and looks fine on the outside. Can the damage really be invisible?

Yes. Drop-impact research shows that mechanical shock combined with PCB bending vibration can crack BGA joints or inner-layer copper while leaving the card's surface completely unmarked. Corner solder joints and inner traces can develop hairline fractures that present as intermittent faults immediately — or surface later under thermal cycling. Neither visual inspection nor X-ray can reliably rule this out, which is why the history ("it started after the drop") carries so much diagnostic weight.

Q7: How do I decide whether repair is worth it?

Three tests: can the fault be reasonably localized to a specific chip or region? Does the repair cost stay well below the card's residual value? And can you accept the recurrence risk, given that reflow is temporary and diagnosis has hard limits (internal chip fractures are indistinguishable from the outside)? If any of the three fails — especially with under-warranty cards, shorted power rails pointing at the GPU, or unlocatable inner-layer cracks — the honest answer is to use the warranty or retire the card.