GPU Crashes Only With Ray Tracing, DLSS or Frame Generation On? RT, Upscaling & Frame-Gen Feature-Gate Diagnosis Guide

2026-11-10 | GPU Repair | 18 minutes | HK GPU Fix Editorial Team

GPU Crashes Only With Ray Tracing, DLSS or Frame Generation On? RT, Upscaling & Frame-Gen Feature-Gate Diagnosis Guide

Key takeaways

When a game only crashes after ray tracing, DLSS or frame generation is switched on, the switch itself is the evidence. This guide separates the three feature paths, explains what each one asks the hardware to do, and walks through a one-variable-at-a-time diagnosis that narrows the fault to the right layer.

What this symptom actually is

The defining pattern is narrow, and it is worth stating precisely before anything else. The same game, on the same machine, with the same driver version, crashes when one specific feature — ray tracing, DLSS or FSR upscaling, or frame generation — is switched on, and runs stably when that feature is switched off. A crash here means the program ends, an error dialog appears, or the screen goes black and the game has to be forced closed. The on-off contrast itself is the most important piece of diagnostic evidence in the whole article: the problem follows the feature switch, not the hardware in general.

Three qualifying points keep this symptom distinct from neighbouring problems.

First, the comparison only counts within the same game. A crash in one game and stability in another is a different question, because the two games ask the graphics card to do different things. The test must use the same scene, the same quality settings, and the same feature switch changed in isolation.

Second, a crash is not the same as slowness. A frame-rate drop after switching ray tracing on, a softer or ghosted image after switching DLSS on, or higher input latency after switching frame generation on are performance and image-quality problems, not crashes. This article is about the program ending, not about it running badly.

Third, this is not a game that fails to launch. A game that immediately reports that it cannot initialise Direct3D has a different problem: the device was never obtained in the first place. Here the device is obtained, the game runs, and it only ends after the feature is switched on.

If the crash does not follow any feature switch — if it happens whether the features are on or off — then this is the wrong guide, and the fault belongs to a broader crash-diagnosis path. Likewise, a screen that goes black for a moment and then comes back, with Windows reporting that the driver recovered, points toward a driver timeout and recovery event rather than a feature-gate crash. A blue screen with a stop code belongs to a different diagnostic path again. An error message that names a device-removed or device-hung reason code is worth copying down in full, because the reason code is the first-hand evidence; an empty error message is also a signal, and it should be recorded rather than dismissed.

Why one switch can crash a game that otherwise runs fine

A feature switch looks like a single toggle, but behind it sit three completely different hardware and driver paths. A crash tied to one switch points at the path behind that switch. Treating all three as "turning the graphics up" hides exactly the information the diagnosis needs, so this section takes them one at a time.

Path one: ray tracing — RT cores, the BVH acceleration structure, and feature levels

Switching ray tracing on does not switch on one thing. It switches on several independent ray-tracing workloads — reflections, shadows, and global illumination are separate groups, each with its own cost and its own return. The correct tuning unit is one group at a time, and the same applies to diagnosis: switching groups off one by one narrows down which ray-tracing workload is involved.

Ray tracing runs on three things.

First, RT cores: dedicated hardware units that test rays against triangles. Once ray tracing is on, the rendering pipeline gains an entirely new hardware path that was not exercised at all when the feature was off. A fault that only appears on this path is, by definition, a fault in the ray-tracing path, not in the rasterisation path that the game was using before.

Second, the BVH acceleration structure: the scene geometry is organised into a hierarchy of bounding boxes, so a ray can start with the largest box and skip whole branches that it does not hit. Scene geometry moves and deforms constantly, so the acceleration structure has to be updated constantly. Building and updating that structure is itself work, and it is a place where a crash can happen — a failure to build or update the structure ends the program rather than slowing it down.

Third, feature levels. Hardware ray tracing support belongs to the DirectX 12 Ultimate feature set, which includes feature levels such as 12_1 and 12_2, along with mesh shaders, sampler feedback, and variable rate shading. Whether a given card supports a given feature level must be read from the dxdiag feature-level list. It must never be guessed from the model name. A card that is new in model terms does not automatically support every feature a game asks for, and a card that is older does not automatically lack them.

The diagnostic meaning follows directly. If only ray tracing crashes the game and switching it off restores stability, the problem follows the ray-tracing path. Switching the ray-tracing groups off one at a time — reflections, then shadows, then global illumination — narrows it further. If every group crashes, the fault sits in something the groups share. A ray-tracing crash does not imply a rasterisation problem, because the two paths are different.

Path two: upscaling — DLSS through NGX and Tensor cores versus FSR through generic shaders

Temporal upscaling renders at a lower internal resolution and then reconstructs a higher-resolution image, combining motion vectors, depth information, and colour data from previous frames. The sample positions shift slightly from frame to frame, and when the historical data becomes invalid the reconstruction fails. In image-quality terms that failure shows up as softness, ghosting, or flickering detail. In this article the question is different: what happens when the reconstruction pipeline itself fails and the program ends. The symptom is different, and image-quality evidence must not be treated as crash evidence.

DLSS and FSR are not the same thing, and that difference is itself a diagnostic tool. DLSS runs through NVIDIA's NGX driver path, uses Tensor cores, and relies on model files distributed with the driver — the DLSS model version follows the driver, not the game. FSR and XeSS run through generic compute shaders, without NGX and without Tensor cores. That gives a clean cross-check: if DLSS crashes the game but switching to FSR at the same quality concept leaves it stable, suspicion falls on the NGX, Tensor-core, and model path. If DLSS and FSR both crash, suspicion does not fall on NGX; it falls on something the upscaling paths share — motion vectors, the history buffer, memory allocation, or the driver layer.

The quality preset is evidence too. If Quality is stable and only Performance crashes, or the reverse, that contrast should be recorded. The internal resolution is lower at Performance, and the reconstruction workload is structured differently, so the preset that crashes and the preset that does not are both part of the record. Writing only "DLSS crashes" throws that information away.

Path three: frame generation — interpolated frames, Reflex, and the base frame-rate threshold

Frame generation inserts a computed intermediate frame between two genuinely rendered frames. The generated frame contains no new user input, and a higher displayed frame rate does not mean the amount of genuine rendering has fallen. In image-quality terms the known problems are higher latency and artefacts; here the question is what happens when the generation process itself crashes.

Frame-generation crashes have three distinctive trigger points, and they should be kept separate.

First, the base frame-rate threshold. When the base frame rate is low, the change between two base frames is large and the interpolation is harder. A crash that only appears in scenes with a low base frame rate is threshold evidence, not hardware evidence.

Second, the Reflex and frame-pacing dependency. Frame generation usually runs together with a low-latency mode, so the frame-scheduling chain gains extra links. Testing Reflex on and off is a separate test line of its own.

Third, interaction with overlays and filters. Public reporting has documented a case in which driver version 621.01 fixed a crash that occurred when DLSS frame generation and a Freestyle filter were enabled together in one game. That is evidence only that "frame generation plus another image-processing layer" can be an independent trigger combination. It must not be generalised into a conclusion about other games or other filters, and it must not be turned into a list of games or versions to avoid.

The diagnostic meaning: if frame generation crashes the game while DLSS upscaling on its own is stable, the problem follows the interpolation path, not the upscaling path. The two are often switched on together, so testing must switch them on one at a time.

The diagnostic tree: change one variable at a time

Every step below changes one thing and records the result before moving on. Skipping ahead, or changing several settings at once, destroys the evidence.

StepWhat to doHow to read the result
0Write down the full error message, the state of every feature switch, the scene, how often it happens, and what changed before it first happened (driver or game version)An error message that names D3D12, a feature, or a reason code is first-hand evidence; a message with nothing in it is also a signal
1Feature-switch isolation: same scene, same quality settings, change only one switch at a time — ray tracing, DLSS, frame generation, each tested separately, repeated several timesWhichever switch crashes the game when it is on points at that path; stability with everything off means isolation is complete, not that the fault is fixed
2Ray-tracing branch: switch the ray-tracing groups off one at a time — reflections, shadows, global illuminationThe group that crashes when it is on is the group involved; if every group crashes, the fault is in something the ray-tracing path shares
3Upscaling branch: compare DLSS against FSR at the same quality concept, then try different presets (Quality, Performance)DLSS crashes and FSR is stable points at the NGX and Tensor-core path; both crashing points at shared upscaling pipeline, memory, or driver layers
4Frame-generation branch: frame generation on and off, Reflex on and off, and record the base frame rateA crash only with frame generation plus Reflex points at the scheduling chain; a crash only at a low base frame rate is threshold evidence
5Compare the dxdiag feature-level list against what the game requiresA required level missing from the list is a feature-level problem; a level present in the list rules the feature level out; never guess from the model name
6Clear the game shader cache and the driver shader cache, then let the paths rebuildStability after clearing points at the cache layer; stuttering on the first entry into a scene afterwards is normal rebuilding, not proof the fault remains
7Monitor dedicated video-memory usage over time while the feature is onA crash only as usage approaches the top points at the allocation layer; a crash at low usage is not a memory problem
8Compare the timeline: driver version and game version, and which update preceded the first crash — record only, never downgrade a driver on this basisA version is a time marker, not a quality label; one driver release has been publicly reported to fix a frame-generation crash in one game while also drawing public reports of instability elsewhere
9Still crashing: record the error message, the switch-isolation results, the dxdiag list, the driver and game versions, and which steps were done, then seek inspectionDo not keep changing settings at random; if artefacts or whole-system symptoms appear as well, the direction changes

How this relates to API crashes, driver timeouts, DXGI errors, and memory

An API-path crash and a feature-gate crash are different layers. If a game crashes under DirectX 12 but is stable under DirectX 11, that is an API-path problem. If it crashes only when ray tracing or DLSS is switched on, that is a feature layer problem. Ray tracing and DLSS only appear on the DirectX 12 or Vulkan paths, so the two belong to the same API family — but the direction of investigation is the feature switch, not the whole API. Switching back to DirectX 11 is a diagnostic isolation step, not a fix, and it should never be presented as one.

A driver timeout is a separate mechanism. Windows watches the driver for roughly two seconds, and if the driver does not respond it resets the driver. A feature-gate crash is sometimes what happens after that reset, when the game cannot survive the device being rebuilt and ends with an error. The timeout mechanism itself is a different subject.

A DXGI device-removed or device-hung error carries a reason code, and the code matters. A hung code leans toward the game sending an invalid instruction; a driver-internal-error code leans toward the driver itself; a message with no reason code at all is most often associated with a driver upgrade, a switch between two graphics devices, or a driver timeout reset. Copy the full error message and the reason code.

Video memory is involved in a specific way. Ray tracing has to store the BVH structure, and upscaling has to store the history buffer, so both consume video memory. A crash that only happens as usage approaches the top is an allocation-layer problem, not damage. Error correction and retransmission in video memory is a separate layer and should not be mixed into this one.

Overclocking deserves a brief note. An unstable overclock usually follows high load in general; it does not politely target one feature switch. Returning the card to factory defaults for one isolation run is an exclusion step, not a repair.

Boundary cases

Common misconceptions

MisconceptionWhat is actually true
"A ray-tracing crash means the card is broken"No. A hardware fault does not usually target one feature switch so neatly. If the same load is completely stable with the feature off, the hardware should be suspected last.
"Switching the feature off fixed it"No. Switching a feature off is diagnostic isolation. It proves the problem follows the feature path, but the underlying cause — a driver-path defect, a game implementation issue, a damaged cache, or an insufficient feature level — is still unresolved, only bypassed.
"The newest driver is always the best one for this game"No. Driver versions should be judged by timeline. Public reporting shows that one release can fix a frame-generation crash in one game while also drawing reports of instability elsewhere — a version is not a quality label.
"DLSS and FSR are the same thing, so either one will do"No. DLSS runs through NGX and Tensor cores with driver-distributed models; FSR runs through generic shaders. Comparing the two is itself a diagnostic tool.
"Frame generation raised the frame-rate number, so the card got faster"No. Generated frames contain no new input and do not represent a reduction in genuine rendering work.
"A new model guarantees support for every feature"No. Feature levels must be read from the dxdiag list, never guessed from the model name.
"Clearing the shader cache always helps"No. The cache is only one layer. Stuttering on the first entry into a scene after clearing is normal rebuilding.

Before seeking inspection checklist

Bringing that record to an inspection is what turns a vague "it crashes when I turn things on" into a diagnosis that can actually be worked on. The published inspection and repair pricing for this service is HK$400–2,800 across three tiers, with the actual figure confirmed after inspection. The shop is at Room 1209, Dali Plaza, Shek Mun. The phone number is 2186 8010 and WhatsApp is 6558 6806. There is no Mong Kok shop and no on-site service.

FAQ

Why does my game only crash when ray tracing is on?

Because ray tracing exercises a hardware and driver path — RT cores and the BVH acceleration structure — that the game does not use at all when the feature is off. Switching the ray-tracing groups off one at a time (reflections, shadows, global illumination) narrows down which workload is involved. It does not by itself prove the card is faulty.

DLSS crashes my game but FSR does not. What does that mean?

It points at the path DLSS uses and FSR does not: the NGX driver path, Tensor cores, and the driver-distributed model. If both DLSS and FSR crash, the fault is more likely in something they share, such as the motion-vector and history-buffer pipeline, memory allocation, or the driver layer.

Frame generation crashes the game but upscaling alone is fine. Why?

Frame generation adds its own layer — interpolating frames between genuinely rendered ones, usually tied to a low-latency mode and a frame-scheduling chain. If upscaling is stable on its own, the problem follows the interpolation path, not the upscaling path. Record the base frame rate and whether Reflex was on, because both are part of the evidence.

Is switching the feature off a fix?

No. It is isolation. It proves the crash follows the feature path, but the underlying cause is still there and still unresolved. The next steps are the diagnostic tree above: the cache, the feature-level list, the memory picture, and the timeline.

The crash only started after a driver update. Should I roll the driver back?

No. A version change is a time marker, not a verdict. Record the driver and game versions and which update preceded the first crash. Public reporting shows that a single driver release can fix a crash in one game while drawing instability reports elsewhere, so versions should not be labelled good or bad.

After clearing the shader cache, the first entry into a scene stutters badly. Does that mean it is not fixed?

No. Stuttering on the first entry into a scene after clearing the cache is normal rebuilding. Run through the scene a few more times and let the paths rebuild fully before judging whether it is stable again.

The game asks for a feature level my card does not have. Is the card broken?

Not necessarily. A missing feature level is a capability question, answered by the dxdiag list — never by guessing from the model name. The direction is the driver version or the hardware capability, not a repair of a card that is working as designed.

The error message says the device was removed. What do I do with that?

Copy the full message, including the reason code, before doing anything else. The code separates a game-side invalid instruction from a driver-side internal error and from the common no-code cases, and it is the single most useful line in the whole record.

Related articles

Related services

If you need Chip-level GPU repair, see these service pages: