
What Is Ray Tracing? How Rays, Bounces and Samples Become Render Time
Overview
Introduction
What is ray tracing? It is a way of working out what a 3D image should look like by following rays of light through a virtual scene. For each pixel, the renderer sends a ray from the camera, finds the first surface it hits, and then sends more rays to learn what light reaches that point: one toward each light to check for shadow, others for reflections and refractions, and others for light that bounced off nearby surfaces. The pixel's color is what those rays report back.
The term covers two very different working worlds. In a game, ray tracing means far fewer rays per pixel, cleaned up by a denoiser, inside a frame budget of roughly 16.7 milliseconds at 60 frames per second. In production rendering with Arnold, V-Ray, Corona or Cycles, it means many more rays per pixel and minutes or hours per frame. The underlying operation is the same: trace a ray, find what it hits.
This guide explains the idea once, then shows how rays per pixel and bounces turn into render time, how real-time and offline ray tracing differ, and why offline work still spreads across many machines. One boundary keeps it short: the side-by-side comparison of ray tracing and path tracing, including noise and sampling, is in ray tracing vs path tracing, so we link to it instead of repeating it. For the broader picture of what rendering is, see what rendering is in computer graphics.
Ray Tracing in One Paragraph
Wikipedia's article on ray tracing (read 19 November 2026) describes it as a technique for modeling light transport that traces light-like rays through a scene to approximate pixel colors. It credits Arthur Appel with the first published computer method for ray-traced shaded images, in 1968, which used rays for visibility and further rays toward the lights to find shadows. Turner Whitted then showed the recursive version in 1979 and 1980, where a ray keeps going after a hit to produce mirror reflections, refraction through transparent objects, shadows and anti-aliasing. NVIDIA's ray tracing overview (read 19 November 2026) adds the practical point that renderers trace backward, from the camera through each pixel, because that wastes far less work than tracing every ray a light source emits.
The idea to hold onto is that ray tracing is a way of asking a scene questions. What is the first surface visible through this pixel? Can this point see the light? What does this mirror reflect? Every extra question is another ray, and every ray costs time.
How a Ray Tracer Builds One Pixel
Here is what happens for a single pixel, in order:
- Camera ray. A ray leaves the camera and passes through the pixel on the image plane, the virtual screen between the camera and the scene.
- Closest hit. The renderer tests the ray against the scene geometry and keeps the nearest surface, together with its surface normal and material.
- Shadow ray. From the hit point, a ray goes toward a light. If something opaque sits in between, that light does not reach the point and the point is in shadow for that light.
- Reflection and refraction rays. If the material is shiny or transparent, the ray continues in the mirror direction or through the material. Each of those rays lands on another surface and the process repeats.
- Combine. The results are weighted by the material and added into the pixel's color.
Diagram of the rays behind one pixel: a camera ray through the image plane hits an object, a shadow ray goes from the hit point to the light, and a reflection ray continues to a second surface
Testing every ray against every triangle would be hopeless, so renderers build an acceleration structure, usually a bounding volume hierarchy (BVH): boxes inside boxes that let a ray skip large parts of the scene without testing them. Wikipedia describes BVHs as bringing intersection cost down from linear in the number of objects to somewhere between linear and logarithmic, and NVIDIA calls BVH traversal the most time-consuming part of ray tracing, which is the work its RT cores accelerate.
Ambient occlusion is one place where the ray idea and a shortcut meet. A ray tracer can measure how enclosed a point is by casting short rays around it, while real-time engines approximate the same effect from screen data. Our ambient occlusion explainer covers those shortcuts.
Rays Per Pixel and Bounces: Where Render Time Comes From
Two multipliers decide how much work a frame needs: how many rays each pixel gets, and how many times a ray is allowed to bounce.
Rays per pixel. One camera ray per pixel gives jagged edges and one sample of every effect. Renderers send many and average them. Arnold's sampling documentation (read 19 November 2026) shows how quickly this grows. The Camera (AA) value is squared, so a setting of 3 means 9 camera samples. The diffuse rate applies per camera sample, and it is squared too, so Diffuse Samples set to 2 gives 4. Together that is 9 x 4 = 36 diffuse samples per pixel.
Apply that to a 1920 x 1080 frame, which has 2,073,600 pixels, and one frame carries 74,649,600 diffuse samples, before counting shadow rays, glossy and transmission samples or any bounce beyond the first. These settings are hypothetical and chosen to show scale, not to recommend values.
Chain of arithmetic showing 2,073,600 pixels, 9 camera samples per pixel, 36 diffuse samples per pixel and 74,649,600 diffuse samples per frame under Arnold's squared sampling rule
Because the values are squared, a small slider change is a large change in rays. Raising Camera (AA) from 3 to 4 takes 9 camera samples to 16, about 78% more, and with the diffuse rate held at 2 the per-pixel diffuse total goes from 36 to 64.
Bounces. Each time a ray hits a surface and carries on, that is a bounce. The Blender manual's Cycles light paths page (Blender 5.2 LTS, read 19 November 2026) lets you cap total bounces and also cap diffuse, glossy, transmission and volume bounces separately. It states that higher values give better quality and lower values render faster, and that a limit of 0 gives direct lighting only.
That last case shows what bounces are for. With direct lighting only, a surface that no light reaches directly renders dark. Each extra bounce lets light travel one surface further: the soft fill from a wall, or color picked up from a red sofa on a white floor.
Three panels showing the same light path with bounce limits of 0, 1 and 2: limit 0 traces only direct light, limit 1 adds one indirect bounce, limit 2 adds two, and anything beyond the limit is not traced
Both multipliers grow together. More rays per pixel and more bounces per ray each mean more work, which is why scenes with small or hidden lights, glass and glossy surfaces cost more than a simple sunlit exterior. The noise and sampling side of that story is covered in ray tracing vs path tracing.
Which Settings Multiply the Ray Count
| Setting or scene feature | What it changes | Effect on ray count |
|---|---|---|
| Image resolution | Number of pixels | Rays scale with pixel count. A 3840 x 2160 frame has 4 times the pixels of 1080p |
| Camera (AA) or pixel samples | Rays per pixel | In Arnold the value is squared, so it grows faster than the number on the slider |
| Diffuse, glossy and transmission samples | Secondary rays per hit | Applied per camera sample, so they multiply the camera samples |
| Bounce limits | How many hits a path may have | Deeper paths trace more rays; a limit of 0 means direct lighting only |
| Number and type of lights | Shadow rays per hit | More lights mean more shadow checks, although renderers can sample a subset |
| Motion blur and depth of field | Spread of rays over time and lens | Need more samples to settle, so they add rays |
Ray Tracing Versus Rasterization
Real-time graphics were built on rasterization, not ray tracing. NVIDIA's overview (read 19 November 2026) describes rasterization as converting triangle meshes into pixels, which is less computationally intensive than ray tracing, while ray tracing is better suited to physically accurate reflections, refractions and shadows. Rasterization answers "which pixels does this triangle cover?" and approximates lighting with shortcuts. Ray tracing answers "what does this ray hit?" and asks again for every effect.
Hybrids are common: rasterization handles much of the scene and ray tracing handles reflections, refractions and shadows. Our real-time vs offline rendering guide goes through how each approach draws a frame and which one fits which job.
Real-Time Ray Tracing and Offline Ray Tracing
Real-time ray tracing is not a different technique. It is the same ray operation with hardware help and a much smaller ray budget. Wikipedia records that NVIDIA's RTX GPUs, built on the Turing architecture with dedicated RT cores, launched in September 2018, and that Battlefield V was the first game to use DirectX Raytracing (DXR), in November 2018. NVIDIA explains the division of labor: RT cores accelerate BVH traversal, and Tensor Cores run AI-based denoising, which reduces the number of rays that need to be cast.
The deadline is what changes everything. At 60 frames per second a frame has about 16.7 milliseconds. A ten-minute offline frame has roughly 36,000 times that, and an offline renderer spends it on many more rays per pixel and more bounces, aiming for a converged, controllable result instead of a denoised approximation. The history of ray tracing in games, from Quake II RTX onward, is in real-time ray tracing in games.
Why Production Ray Tracing Still Needs Many Machines
NVIDIA's overview notes that film studios traditionally relied on CPU render farms, with complex effects sometimes taking multiple days to render. Hardware has sped up since, but production scenes grow to match: more geometry, more complex lighting and higher resolution.
The arithmetic is what makes a farm worthwhile. A hypothetical example, for illustration only: a 30-second shot at 25 frames per second is 750 frames. If each frame takes 45 minutes on one machine, that is 562.5 machine-hours, about 23 days on a single machine. Frames are independent, so spread across 100 machines at once the same shot finishes in about six hours of waiting, if the machines are available and the scene loads cleanly on each. The total compute does not shrink; the wait does.
Two practical notes follow from how rays work:
- The whole scene has to be reachable. A ray can go anywhere, so the renderer generally needs the geometry and its acceleration structure available at once. On GPU renderers that means the scene has to fit in the card's memory; NVIDIA's overview notes that onboard memory once capped the scene complexity GPUs could handle.
- Test the hardest frame first. Render the frame with the most glass, hair or lights before committing a sequence. If it takes ten minutes more than expected, multiply that by the frame count.
For ways to cut the per-frame cost, our render time optimization guide covers sampling, denoising and scene preparation.
Ray-Traced Scenes on Super Renders Farm
Ray tracing is how these renderers work, so any scene that renders offline is a ray-traced scene, and it renders on a farm like any other: frames are processed in parallel as batch jobs.
Super Renders Farm is a fully managed farm: you upload the scene and do not set up machines, install software or manage render licenses yourself. Cycles and EEVEE in Blender, Arnold, V-Ray and Corona in 3ds Max, and Redshift in Maya and Cinema 4D are ready without a setup request. Other engine and host pairings vary: some are ready without a request, some are provisioned on request, and some do not exist as products (Corona, for example, is made only for 3ds Max and Cinema 4D). The supported software page shows the engines for each host, and for any pairing not named here, contact us to confirm before you upload. Real-time viewers such as Enscape, D5 Render or Lumion are not rendered on the farm; it renders offline frames.
Corona, and Arnold or V-Ray in their CPU modes, render on our CPU nodes. Redshift renders on our GPU cloud render farm, where each node carries two RTX 5090 cards with 32 GB of VRAM per card; memory is per card and is not pooled, so a GPU scene has to fit in 32 GB. New accounts receive $25 free render credits, which is enough to render test frames of your heaviest shot before committing a sequence, and our team is on 24/7 live chat for scene checks. Current rates are on the pricing page.
Common Misunderstandings
"Ray tracing needs an RTX card." Ray tracing is an algorithm that has existed since the 1960s and runs on CPUs as well as GPUs. RTX is hardware that accelerates parts of it. Offline renderers such as Arnold, V-Ray and Corona do their ray tracing on CPUs routinely.
"Ray tracing means a realistic image." Realism still depends on materials, lighting and how many bounces and samples you allow. A ray tracer limited to direct light renders areas without a direct light source dark.
"Real-time and offline ray tracing are the same thing." They share the operation but not the budget: real-time hits a frame deadline with few rays, hardware and denoising, while offline spends minutes per frame to converge.
Summary: What to Remember
| If you are asking | Short answer |
|---|---|
| What is ray tracing? | Following rays from the camera through each pixel to find what is visible and what light reaches it |
| Where does render time come from? | Rays per pixel, bounce limits, resolution, lights and effects such as glass and motion blur |
| How is it different in a game? | The same operation with RT cores, far fewer rays and a denoiser, inside about 16.7 ms per frame |
| Do I need an RTX card? | Not for offline CPU renderers; RTX accelerates parts of the work on GPU |
| Why does production rendering need many machines? | Frames are independent, so machines split the sequence; the total compute stays, the wait shrinks |
FAQ
Q: What is ray tracing in simple terms? A: Ray tracing is a rendering method that follows rays of light through a 3D scene to work out each pixel's color. A ray leaves the camera, finds the first surface it hits, and more rays check for shadows, reflections and light bounced from other surfaces. The pixel takes its color from what those rays find.
Q: What is ray tracing in games? A: In games, ray tracing is a hardware-assisted way of adding accurate reflections, shadows and lighting to a mostly rasterized frame. RTX GPUs have dedicated RT cores that speed up the ray work, and AI denoising cleans up the small number of rays that can be cast in roughly 16.7 milliseconds per frame at 60 frames per second.
Q: What is the difference between ray tracing and ray casting? A: Ray casting follows a single ray per pixel to find which surface is visible, and sometimes a ray toward a light to check for shadow. Ray tracing adds recursion: a ray that hits a reflective or transparent surface spawns further rays, which gives reflections, refraction and richer lighting. Appel's 1968 method is the ray-casting form and Whitted's 1979 to 1980 work is the recursive form.
Q: How does ray tracing work? A: A ray tracer casts a camera ray through each pixel, tests it against the scene using an acceleration structure such as a BVH, and keeps the closest hit. From there it casts shadow rays toward lights and reflection or refraction rays for shiny or transparent materials, then combines the results by the material's properties. Many rays per pixel are averaged.
Q: Do you need an RTX graphics card for ray tracing? A: No. Ray tracing is an algorithm that also runs on CPUs, and offline renderers such as Arnold, V-Ray and Corona routinely use CPU nodes. RTX cards include RT cores that accelerate BVH traversal, which speeds up ray tracing on the GPU in software that uses them, such as real-time games.
Q: Is path tracing better than ray tracing? A: Path tracing is a form of ray tracing, not a rival to it. It follows many random bounce paths per pixel and averages them, which handles indirect light and soft shadows in one process at the cost of noise. The comparison is covered in our ray tracing vs path tracing article.
Q: Why do ray-traced renders take so long? A: Render time grows with the number of rays per pixel, the number of bounces, the resolution and the complexity of lights and materials. Arnold squares its sample settings, so a small increase on a slider can add a large share of rays. A test render of the hardest frame shows the real cost before you commit a sequence.
Q: Can ray-traced scenes be rendered on a render farm? A: Yes. Offline renderers are ray tracers, and frames render in parallel as batch jobs, so a farm shortens the wait without reducing the total compute. On Super Renders Farm, Cycles and EEVEE in Blender, Arnold, V-Ray and Corona in 3ds Max, and Redshift in Maya and Cinema 4D are ready without a setup request. Other engine and host pairings vary, and a few do not exist as products, so for any pairing not named here, contact us to confirm before you upload.
About Thierry Marc
3D Rendering Expert with over 10 years of experience in the industry. Specialized in Maya, Arnold, and high-end technical workflows for film and advertising.



