VRAM Fault or GPU Core Fault? How to Tell from Artifact Patterns and Test Logic
Key Takeaways
The GPU core does the computing and VRAM stores picture data, so each tends to fail differently. VRAM faults tend to corrupt the picture itself, while core faults tend to freeze, crash, or kill output entirely. This guide explains the artifact patterns, the test logic that narrows the fault down, and why pattern alone never proves a diagnosis.
Introduction: The Screen Is Corrupted — But Which Part Is Guilty?
A game starts throwing coloured blocks across the screen. Text turns to garbage. The display freezes, the driver crashes, or the machine dies the moment a 3D workload begins. The reader's first question is rarely "how do I fix this?" — it is "which part is failing?" That question matters, because a VRAM fault and a GPU core fault point in different repair directions, with different difficulty and different considerations.
This guide explains the division of labour between the core and the memory, the artifact patterns each tends to produce, and the test logic that narrows the fault down. One warning belongs up front: patterns are clues, not proof. No picture pattern by itself can definitively diagnose a card. The patterns guide the investigation; the workbench confirms it.
What the Core and the VRAM Each Do
| Part | Role | Physical layout |
|---|---|---|
| GPU core | Performs the computation: geometry, shading, and general calculation | One large chip at the centre of the board |
| VRAM (GDDR memory) | Stores picture data: frame buffer, textures, and other data the core works on | Many smaller chips arranged around the core |
This division explains why their failures look different. A fault in storage corrupts the stored data — the picture itself comes out wrong. A fault in computation stops the computation — the picture freezes, the driver gives up, or output ceases altogether. That is the core intuition behind everything that follows.
What a VRAM Fault Tends to Look Like
A VRAM fault tends to corrupt the contents of the picture. Typical tendencies include:
- Coloured blocks or squares appearing in the picture, sometimes in a regular grid-like arrangement.
- Horizontal or vertical lines that track the picture rather than the screen — they move with the content because the corruption lives in the stored data.
- Checkerboard patterns over parts of the image.
- Garbled text — characters replaced by wrong glyphs or random symbols.
- Corrupted textures — surfaces in a game rendered with wrong colours or patterns while the geometry stays correct.
Two observations sharpen a VRAM suspicion:
- The corruption appears in screenshots. A screenshot captures the frame buffer — the stored picture data. If the corruption is visible in a screenshot viewed on another device, the corruption is in the stored data, not in the monitor or the output path. This is one of the most useful distinctions in artifact diagnosis.
- The corruption changes with the picture. Because the fault sits in stored data, what is corrupted depends on what is stored. Different scenes, windows, or textures show different damage.
These are tendencies, not rules. A VRAM fault does not always produce textbook patterns, and a fault elsewhere can occasionally imitate them.
What a GPU Core Fault Tends to Look Like
A core fault tends to stop or distort the computation itself:
- Whole-picture freezing — the image locks and the system stops responding.
- Driver crashes — the display driver stops and recovers, or the system restarts, particularly when a 3D load begins.
- Immediate failure under 3D load — the desktop is fine, but starting any 3D workload kills the output.
- Distorted geometry — shapes stretched, collapsed, or twisted, where the structure of the scene itself is wrong rather than its surface detail.
- Black screen from boot — no output at all from the first instant, with no picture to corrupt.
Core power delivery faults and core solder-joint faults can produce similar behaviour, which is why "core-like" symptoms are not automatically "the core chip is dead". A core that is starved of stable power, or whose solder connection to the board is failing, can freeze and crash just as a damaged core can. The symptom points at the core's neighbourhood; the workbench determines which part of that neighbourhood is responsible.
Comparison Table
| Symptom | Tends toward VRAM | Tends toward core | How to confirm further |
|---|---|---|---|
| Coloured blocks, lines, checkerboard in the picture | Yes | Less typical | Check whether the corruption appears in a screenshot |
| Garbled text or corrupted textures | Yes | Less typical | Check whether the geometry of the scene stays correct |
| Whole picture freezes, driver crashes | Less typical | Yes | Check whether the failure tracks 3D load rather than picture content |
| Fails the moment any 3D load starts, desktop fine | Possible | Yes | Observe whether load, rather than content, is the trigger |
| Distorted or twisted geometry | Less typical | Yes | Check whether the scene structure itself is wrong |
| No output at all from boot | Less typical | Possible, or power / VBIOS | Check whether the card initialises at all |
| Corruption visible in a screenshot | Yes — corruption is in stored data | Less typical | View the screenshot on another device |
| Corruption visible only on screen, not in a screenshot | Less typical | Less typical — points toward the output path or monitor | Test another cable, port, or monitor |
Test Logic: How to Narrow the Fault Down
No single test proves a diagnosis, but a combination of observations narrows the range considerably:
Idle versus load. If the card is stable at idle and fails only under 3D load, the fault is load-related. That implicates the core, the power delivery, the cooling, or the memory timings — all of which are stressed by load. It does not single out one of them, but it rules out faults that would also appear at idle.
Lowering clocks as an observation. Reducing the memory clock and observing the result, then separately reducing the core clock and observing the result, can provide narrowing clues. If lowering one side clearly eases the symptoms, that side is implicated — but this is a clue, not a confirmation. Clock changes alter power draw, heat, and timing together, so an improvement does not prove which of those was the cause.
Memory testing, understood correctly. Dedicated tools exist that write patterns to video memory and read them back, looking for chips that return wrong data. In principle, such a test can identify a faulty memory chip. In practice, interpreting the results needs expertise: a test can miss an intermittent fault, a fault that only appears when the memory is hot, or a fault in the memory controller rather than in a memory chip. A test that reports no error does not prove the memory is healthy, and a test that reports errors does not by itself identify the repair.
Temperature correlation. If symptoms worsen as temperature rises and ease when the card is cool, the fault is temperature-linked — cooling, solder joints, or power delivery are in the frame. If symptoms appear identically when the card is cold, the fault is less likely to be thermal. Temperature is a modifier of the evidence, not a verdict.
Consistency. Intermittent symptoms — sometimes fine, sometimes failing, with no pattern — are among the hardest to interpret. Intermittence points toward a marginal connection, a temperature-dependent fault, or a fault that depends on specific data patterns. It argues against a simple "this chip is dead" reading, because a wholly dead chip usually fails consistently.
After the Diagnosis: Why the Distinction Changes the Repair Direction
The reason the VRAM-versus-core question matters is that the two faults point in different repair directions:
- Individual VRAM chip damage points toward replacing or reworking the affected memory chip — a component-level repair on one of several identical chips.
- Core damage or a core solder-joint fault points toward BGA rework of the core, or core replacement — work on a single large chip with a far more demanding thermal and alignment profile.
Those are differences in repair direction only. This guide deliberately does not discuss pricing or repair outcomes. Whether a repair makes sense depends on the card's model, its remaining value, and the extent of the damage — a judgement for the owner and the repair shop, made after a workbench diagnosis, not from a symptom table.
FAQ
Q: The picture is corrupted but screenshots look fine. Does that rule out VRAM?
A: It points away from VRAM and toward the output path: the cable, the port, the monitor, or the output stage of the card. A screenshot captures the stored frame; if the stored frame is clean, the corruption is being introduced after storage. Test another cable, port, and monitor before suspecting the card's internals.
Q: Lowering the memory clock made the artifacts go away. Does that prove the VRAM is faulty?
A: No. It is a useful narrowing clue, but lowering the memory clock also lowers power draw, heat, and timing stress. The improvement could reflect any of those. It suggests the memory subsystem deserves closer investigation — it does not confirm a faulty chip.
Q: Why does the card sometimes work and sometimes fail?
A: Intermittent behaviour usually means the fault depends on conditions: temperature, specific data patterns, a marginal solder connection, or a power-delivery component operating at its edge. A wholly failed chip tends to fail consistently. Intermittence is a reason for careful workbench testing rather than a quick verdict.
Q: Can a VRAM fault cause a driver crash?
A: It can contribute: corrupted data can crash a workload or a driver. But a driver crash is more characteristic of core-side or power-delivery problems, and a crash by itself does not identify the faulty part. Look at the pattern of what happens before the crash — picture corruption beforehand leans toward memory; a sudden freeze under load leans toward the core side.
Q: If the core is faulty, is the card unrepairable?
A: Not necessarily, and not automatically. Core-side faults include solder-joint failures that can be reworked, and power-delivery faults in the core's supply circuit. Whether a given card can be repaired depends on the actual fault found on the workbench, not on the symptom category. Do not write a card off from a symptom description alone.