Ray Tracing Killing Your FPS? RT Reflections, Shadows & Global Illumination Cost, Denoising & Combination Guide

2026-10-26 ｜ GPU Repair ｜ About 15 minutes ｜ HK GPU Fix Editorial Team

重點摘要

Ray tracing can transform reflections, shadows and indirect light, but the three do not cost the same, and low ray counts leave noise that denoisers must hide. This guide separates reflections, shadows and global illumination by cost, explains how denoising works and where it fails, and shows how to combine ray tracing with upscaling sensibly.

The Short Answer First

Ray tracing is often discussed as if it were a single switch: on, the game looks better and runs slower; off, it looks worse and runs faster. That picture hides the important structure. A game rarely enables "ray tracing" as one indivisible thing. It enables ray-traced reflections, or ray-traced shadows, or ray-traced global illumination, or some combination, each at its own quality level, and each with its own cost and its own failure mode. Turning on the cheapest class while leaving the most expensive class off can give a visible improvement at a manageable cost. Turning on the most expensive class at its highest quality can overwhelm the machine while adding detail that is hard to see in motion.

The three classes also fail differently when they are starved of rays. Reflections become noisy or smeared when too few rays are traced. Shadows become noisy or blocky when their ray budget is small. Global illumination, which describes light bouncing off surfaces to light other surfaces, is the most sensitive of all, because it depends on many diffuse bounces that are individually faint and hard to sample. Understanding which class is active, and how many rays it is allowed, is the key to tuning ray tracing sensibly rather than by trial and error.

A second structure matters just as much: ray tracing in games almost never works alone. The raw ray-traced result at a playable ray budget is noisy, and a denoiser — a filtering system that uses information from neighbouring pixels and previous frames — cleans that noise up. The final image is therefore a partnership between ray tracing and denoising, and many ray-tracing artefacts are denoising artefacts, not ray-tracing errors. A smeared reflection or a ghosting shadow may be the denoiser struggling, not the ray tracer failing.

This guide separates the cost classes, explains denoising and its limits, and then sets out sensible combinations, including ray tracing with upscaling. It does not repeat the full upscaling or frame-generation guides: upscaling is covered by article #162 and frame generation by article #175, and this guide points past them where they touch ray tracing. Its companion in this batch, article #211, covers per-setting tuning and the controlled comparison method in general; the method there applies directly to ray tracing, and this guide uses it without repeating its full detail.

What Ray Tracing Actually Does, in Plain Terms

Traditional rasterisation draws a scene by projecting triangles onto the screen and shading each visible pixel from lights, textures, and approximations. It is fast because it works forward from geometry to pixels, but it must approximate anything that depends on light travelling between surfaces: reflections, indirect light, and accurate shadows. Ray tracing works in the opposite direction: for each pixel, it casts rays into the scene to discover what light actually reaches that point, following rays as they bounce off surfaces, pass through transparent materials, or are blocked by occluders.

Following rays naively would be impossibly expensive, because light bounces in every direction and a single pixel could require an enormous number of rays to sample adequately. Games therefore use acceleration structures — organised hierarchies of bounding volumes, commonly called bounding volume hierarchies — that let a ray quickly skip empty space and large groups of geometry and test only the surfaces it might actually hit. Even with that acceleration, games trace far fewer rays per pixel than a film renderer would, often on the order of a few rays per pixel for a given effect, and rely on denoising to fill in the rest. That compromise — few rays plus aggressive denoising — is the central fact of real-time ray tracing, and almost every artefact and cost in this guide follows from it.

Ray tracing cost also depends on scene complexity in a way rasterisation cost does not. A scene with many small, detailed objects forces rays to traverse deeper hierarchies and test more surfaces. A scene with large, simple surfaces is cheaper to trace. A scene with many reflective or transparent surfaces forces more secondary rays — rays spawned when a first ray hits a reflective or refractive surface — which multiply the work. That is why ray tracing cost varies so much between scenes in the same game, and why a ray-tracing setting that feels affordable in one level can feel expensive in another.

Cost Class One: Ray-Traced Reflections

Ray-traced reflections replace approximate reflections with reflections computed by tracing rays from reflective surfaces to discover what they should show. The improvement is most visible where traditional approximations fail: reflections of objects that are not currently visible on screen, reflections on curved or rough surfaces, and reflections that move correctly as the camera moves. A puddle that shows the building behind the camera, or a polished floor that reflects a character who has just left the frame, is the kind of detail ray-traced reflections add.

The cost of reflections depends on several factors. Reflection resolution — the resolution at which the reflected image is computed — is a direct multiplier: a full-resolution reflection costs far more than a half-resolution one, and the difference is often hard to see on a rough or distant surface. Surface roughness matters too: a mirror-smooth surface needs a coherent set of rays in one direction, while a rough surface scatters rays in many directions and needs more rays to sample adequately, which is why rough reflections are often noisier or more heavily denoised. The number of reflective surfaces in view matters as well: a scene with one polished floor is cheaper to reflect than a scene full of glass, water, and chrome.

Reflections are usually a middle-cost class: more expensive than ray-traced shadows, less expensive than full global illumination. They are also the class whose benefit is most scene-dependent. In a game built around water, glass, and polished interiors, reflections transform the image. In a game set mostly in rough, matte environments — dirt roads, rock, cloth — reflections add little, because there is little to reflect. Tuning reflections therefore starts with a scene question: are reflective surfaces prominent in the scenes you actually play, or are you paying for reflections you rarely see?

Cost Class Two: Ray-Traced Shadows

Ray-traced shadows compute shadowing by tracing rays from each surface point toward each light source and checking whether anything blocks the way. The benefit over shadow maps — the traditional method described in article #211 — is accuracy and softness: shadows that soften correctly with distance from the occluder, shadows from small or thin objects that shadow maps miss or render blockily, and shadows that remain stable as the camera moves rather than shimmering between map resolutions.

Shadow cost depends on the number of lights, the number of rays traced per light, and the complexity of the geometry between surface and light. A scene with one dominant light, such as the sun, is cheaper to shadow than a scene with many local lights, because each light requires its own shadow rays. Soft shadows are more expensive than hard shadows, because a soft shadow edge must be sampled at multiple points across the light's area rather than tested once against a point. Long, detailed shadows in an open scene are more expensive than short shadows in an enclosed one, because the rays must traverse more geometry.

Ray-traced shadows are usually the cheapest ray-tracing class, and they are often the best value when only one class can be afforded. Their improvement is subtle rather than dramatic — more accurate contact and edge behaviour rather than a transformed image — but it is visible in the places where shadow maps traditionally struggle: thin objects, distant shadows, and soft penumbra edges. If a game's ray-tracing menu offers shadows separately, shadows are the natural first class to try, and the natural class to keep when other classes must be dropped.

Cost Class Three: Ray-Traced Global Illumination

Global illumination is the most ambitious and most expensive class. It describes indirect light: light that bounces off one surface and lights another. In the real world, a red wall tints the floor beside it, a sunlit courtyard lights the room beyond its doorway through bouncing light, and a dark corner is never truly black because light arrives from many indirect paths. Traditional rasterisation approximates indirect light with baked lightmaps, light probes, or screen-space approximations, all of which are limited to what was precomputed or what is visible on screen. Ray-traced global illumination computes indirect light live, by tracing rays that bounce between surfaces, so it responds correctly to moving lights, moving objects, and changing scenes.

That ambition is exactly why it is expensive. Indirect light arrives from every direction, not from a few known lights, so it must be sampled broadly. Each bounce scatters light further and weakens it, so many bounces contribute a little each, and sampling them all adequately requires many rays. Diffuse surfaces scatter light in all directions, which means the rays from a diffuse surface fan out unpredictably and converge slowly on a stable answer. The result is that global illumination needs far more rays per pixel than reflections or shadows to reach the same noise level, and at a playable ray budget it relies most heavily on denoising.

Global illumination is therefore the class to approach last, not first. Its benefit is real — scenes feel more grounded, colours bleed believably between surfaces, and dark areas gain believable indirect detail — but it is also the class most likely to dominate the ray-tracing budget and most likely to show denoising artefacts when starved. A sensible tuning order tries reflections and shadows first, judges their combined cost, and only then considers whether global illumination can be added at a modest quality level without pushing the machine past its limit.

Other Ray-Traced Classes, Briefly

Two further classes appear in some games and deserve a brief mention so that readers are not confused by menu labels. Ray-traced ambient occlusion computes contact darkening by tracing short rays around each surface point to test how much of the surroundings occlude it. It is conceptually similar to the screen-space ambient occlusion described in article #211, but it can see occluders that are off-screen, which screen-space methods cannot. Its cost is moderate and its benefit is subtle, concentrated in corners and crevices.

Ray-traced caustics and transparency effects — light focused through water or glass, or accurate refraction through transparent materials — appear in a smaller number of games and are usually the most specialised class of all. Their cost can be high and their benefit is confined to scenes with prominent water or glass. Where they appear, they should be treated as a luxury class: tried last, kept only if the scenes that showcase them are scenes the reader actually plays, and dropped first when cost must be cut.

Denoising: How Few Rays Become a Clean Image

Denoising is the system that makes real-time ray tracing possible at all. A raw ray-traced image at a few rays per pixel is speckled with noise: neighbouring pixels receive different random samples and therefore different brightness values, producing a grainy, flickering image. A denoiser cleans that noise by combining information across space and time. Spatial denoising averages each pixel with its neighbours, using surface and lighting information to avoid blurring across genuine edges. Temporal denoising reuses information from previous frames, tracking how surfaces move so that a pixel can borrow samples from where it was in the last frame.

Temporal denoising is powerful because it effectively multiplies the ray budget: a pixel that borrows samples from several previous frames behaves as if many more rays had been traced for it. It is also the source of the most visible ray-tracing artefacts, because it depends on correctly tracking motion. When tracking fails — on fast-moving objects, on disoccluded surfaces that were hidden in the previous frame and are newly visible, or on thin objects that move between frames — the denoiser borrows the wrong samples, and the result is ghosting, smearing, or lagging detail that trails behind moving objects.

Spatial denoising has its own failure mode: over-smoothing. When the denoiser averages too aggressively, fine detail is lost and reflections look waxy or smeared, shadows lose their edge definition, and indirect light looks blotchy. Under-smoothing leaves noise visible, especially in dark areas and on rough surfaces where samples vary most. Denoiser quality settings, where a game exposes them, trade between these two failures, and the right setting depends on the scene: a slow, static scene tolerates heavy temporal reuse, while a fast, chaotic scene needs the denoiser to trust current-frame samples more and previous-frame samples less.

A practical consequence is that ray-tracing artefacts should be diagnosed as a partnership problem. A noisy reflection may mean the ray budget is too low, or it may mean the denoiser is under-smoothing for that surface. A smeared reflection may mean the denoiser is over-smoothing, or that temporal tracking is failing on a moving object. The test that separates the two is to raise the ray-tracing quality one step in the same scene: if the artefact improves, the ray budget was the limit; if it stays the same, the denoiser or the tracking is the limit, and a higher ray budget will not fix it.

The Ray-Tracing Cost Table

As with the per-setting table in article #211, this table is a map, not a measurement. The same class costs different amounts in different games, because engines choose different ray budgets, different denoiser strengths, and different quality presets, and because the scene determines how much work the class actually triggers. The table tells you where to look first and what failure to expect; the controlled comparison method from article #211 — one change at a time, same scene, same route — tells you whether the guess was right on this machine.

Tuning Ray Tracing, Step by Step

Start by identifying which classes the game actually offers. Open the graphics menu and list the ray-tracing options separately: reflections, shadows, global illumination, ambient occlusion, and any others. Do not treat a single "ray tracing" preset as one thing; open its details or its individual toggles and note which classes it enables. A preset labelled "high" in one game may enable all classes, while the same label in another may enable only reflections and shadows. The menu label is not a reliable guide to what is actually running.

Next, establish a baseline with ray tracing off, using the controlled method from article #211: a representative scene, a repeatable route, and a note of how the baseline feels. Then enable one class at a time, starting with shadows, then reflections, then global illumination, judging each addition in the same scene before adding the next. This order is deliberate: it reveals the cost of each class separately, so that when the total becomes too expensive, the reader knows exactly which class to drop. Enabling everything at once and then dropping classes blindly teaches nothing about which class was responsible.

Within a class, tune quality before adding the next class. A class at medium quality that is affordable is better than the same class at ultra quality that forces everything else off. Pay particular attention to reflection resolution and to global illumination quality, because those two quality levers have the largest cost range within their classes. Where a game offers a denoiser quality setting, try it only after the ray budget is set, because denoiser changes alter artefacts without altering the ray cost, and changing both at once confuses the diagnosis.

Judge ray tracing in motion, not in screenshots. Denoising artefacts — ghosting, smearing, flicker — appear and disappear as the camera moves, and a still image can hide them completely. A scene that looks clean in a screenshot but smears during play is a denoising problem, not a success. Conversely, a scene that looks slightly noisy in a screenshot but stable in motion may be perfectly acceptable in play. The test route should include camera movement, moving objects, and transitions between light and shadow, because those are the moments that stress tracking and denoising.

Finally, consider the scene mix of the whole game, not only the showcase scene. A game may have one spectacular reflective level and many matte ones; tuning for the showcase level alone can leave the rest of the game paying for reflections it rarely shows. Tune in the scenes you play most, and treat showcase scenes as a check rather than the standard.

Combining Ray Tracing with Upscaling

Ray tracing and upscaling are natural partners, and the combination is the most common way to make ray tracing affordable. Upscaling renders the scene at a reduced internal resolution and reconstructs a higher-resolution image, which reduces the shading work of the whole frame, including the ray-tracing work that scales with resolution. The saving is real, but the partnership has its own interactions, and they deserve attention rather than assumption.

The first interaction is that upscaling and denoising both reconstruct images from limited information, and they can compound each other's artefacts. An upscaler that reuses previous frames can add its own ghosting to a denoiser's ghosting, and a denoiser that smooths aggressively can soften detail that the upscaler is trying to reconstruct. The combination should therefore be judged in motion, in scenes with moving objects and fine detail, not only by average frame rate. If ghosting appears only when both systems are active, the combination is the suspect, and the remedy may be a different upscaler quality preset or a lower ray-tracing quality, rather than disabling either system entirely.

The second interaction is that upscaling quality presets change the ray-tracing budget indirectly. A higher-quality upscaling preset renders at a higher internal resolution, which raises the ray-tracing cost; a lower-quality preset renders lower and costs less. Tuning ray tracing and upscaling together therefore means tuning two linked variables, and the controlled method must be respected: change one, judge, then change the other. Changing both at once and judging the combined result hides which change mattered.

The third interaction is frame generation, which is sometimes offered alongside ray tracing and upscaling as a further speed lever. Frame generation inserts synthesised frames between rendered frames and raises displayed frame rate, but it does not raise the rate at which ray tracing is computed, and it adds its own latency and artefact considerations. Article #175 covers frame generation in full, and this guide does not repeat that analysis. The brief guidance here is to settle ray tracing and upscaling first, judge that combination in motion, and only then consider whether frame generation adds anything the reader actually wants, rather than adding all three systems at once and debugging their combined artefacts together.

A sensible combination order therefore runs: first, ray-tracing classes one at a time at modest quality; second, an upscaling preset that restores comfortable speed without visibly degrading the ray-traced detail; third, a final motion check for compounded ghosting and smearing; and only then, if desired, frame generation as a separate decision. Each step is judged before the next is taken, so that any artefact can be attributed to the step that introduced it.

Common Mistakes and Misconceptions

"Ray tracing is one switch." It is a set of classes with different costs and different benefits. Treating it as one switch leads to all-or-nothing choices that are worse than a selective combination: either everything on at unaffordable cost, or everything off with no benefit at all.

"More rays always fix artefacts." More rays fix noise, but they do not fix denoiser tracking failures or over-smoothing. An artefact that survives a quality increase is a denoising or tracking problem, and raising quality further wastes cost without fixing it.

"Ray-traced shadows are the most dramatic improvement." Shadows are usually the best value, but not the most dramatic. Reflections and global illumination transform the image more visibly in the right scenes; shadows improve accuracy and stability more quietly. Value and drama are different judgements, and both matter.

"Upscaling makes ray tracing free." Upscaling reduces cost, but it does not remove it, and it adds its own artefacts that can compound denoising artefacts. The combination must be judged in motion, not assumed to be free.

"A screenshot proves ray tracing works." Screenshots hide temporal artefacts. Ghosting, smearing, and flicker are motion phenomena, and only a moving test reveals them. Judge ray tracing by playing, not by photographing.

"If ray tracing is slow, the graphics card is failing." Ray tracing is expensive by design, and a slow result at high quality in a complex scene is expected behaviour, not evidence of a fault. Hardware becomes a suspect only when the symptom survives across settings, games, and scenes — for example, when ray tracing at its lowest quality in a simple scene behaves worse than rasterised rendering in the same scene in a way that settings cannot explain — and even then, drivers and cooling should be checked before hardware is blamed.

Frequently Asked Questions

Question: Which ray-traced class should I enable first?

Answer: Ray-traced shadows, because they are usually the cheapest class and the best value, improving shadow accuracy and stability without dominating the budget. Add reflections next if your scenes have prominent reflective surfaces, and consider global illumination last.

Question: Why do my ray-traced reflections look smeared when I move?

Answer: That is usually a denoising or temporal tracking effect, not a lack of rays. When the camera or an object moves quickly, the denoiser may borrow samples from the wrong previous-frame position, producing smear or ghosting. Try a different denoiser quality if the game offers one, and judge in motion rather than in screenshots.

Question: Why does global illumination cost so much more than shadows?

Answer: Because indirect light arrives from every direction and bounces many times, with each bounce scattering light further and weakening it. Sampling that diffuse, multi-bounce light adequately needs far more rays than testing whether a point is blocked from a known light, which is all shadows need.

Question: Should I use upscaling with ray tracing?

Answer: Yes, in most cases, because upscaling reduces the internal resolution at which ray tracing is computed and can restore comfortable speed. Judge the combination in motion, because upscaling and denoising artefacts can compound, and settle ray-tracing classes before choosing an upscaling preset. Article #162 covers upscaling in full.

Question: Why does ray tracing look good in one level and noisy in another?

Answer: Because ray-tracing cost and noise depend on the scene: reflective surfaces, light count, geometry complexity, and dark areas all change the workload. A level with many lights and reflective surfaces needs more rays than a simple matte level, and dark areas show noise more readily because their samples vary proportionally more.

Question: Does frame generation help ray tracing?

Answer: Frame generation raises displayed frame rate by inserting synthesised frames, but it does not increase the rate at which ray tracing is computed, and it adds its own latency and artefact considerations. Settle ray tracing and upscaling first, then judge frame generation separately. Article #175 covers it in full.

Question: My reflections show objects that are not on screen. Is that a bug?

Answer: No, that is exactly what ray-traced reflections are for. Traditional screen-space reflections can only show what is already visible on screen, while ray-traced reflections trace rays into the full scene and can therefore reflect objects behind or beside the camera. It is a benefit, not an error.

Question: When should I turn ray tracing off entirely?

Answer: When no selective combination — shadows alone, or shadows plus reflections at modest quality, with upscaling — gives a visible benefit worth its cost in the scenes you actually play, or when denoising artefacts in motion outweigh the image improvement. Turning everything off after testing selectively is a decision, not a failure, and it preserves speed for the rest of the game.

Getting Help in Hong Kong

If ray tracing behaves in a way that settings cannot explain — for example, artefacts that appear in every game regardless of class or quality, or a game that crashes whenever any ray-tracing class is enabled while rasterised rendering works normally — the cause may sit below the settings layer, in drivers, cooling, or hardware. Note which classes were tested, in which scenes, and which artefacts survived quality changes, and bring the machine in for inspection. HK GPU Fix is at Shek Mun, Dali Plaza, Room 1209. The phone number is 2186 8010 and WhatsApp is 6558 6806. There are no home visits and no Mong Kok shop. Inspection pricing follows the published three-tier range of HK$400 to HK$2,800, with a final quote after inspection.