Path tracing. The heaviest load a browser can put on a GPU.
A real-time global illumination load that measures GPU stability and sample throughput, in your browser.
Illustrative progress. Not a measurement.
Pacing
About 150 samples
Load type
Extreme compute
Render mode
Path tracing + GI
Reports
Samples/s + consistency
It stops at a sample count, not at a clock.
Every GPU does the same quantity of work, so the cost of each sample is the measurement rather than how many a fast card squeezed into a fixed window.
One path per pixel per frame
A fragment shader raymarches a small signed-distance-field scene, two spheres above a reflective plane under two coloured area lights, taking one jittered path per pixel and accumulating a running mean into a half-float texture.
A fixed time, timed per sample
The stage runs for a set time, paced to about 150 samples, so the run is the same length on every card. Each sample is timed around a blocking draw, so the figure is the card's real cost per sample.
Timed around the draw
Each sample is timed around its draw with a blocking call, so the figure reflects real GPU cost. The first sample is excluded; samples per second is the reciprocal of the mean per-sample cost.
Capped at 1000 by 560
Render resolution is capped at 1000 by 560 pixels with the device pixel ratio capped at 1.25, so a high-density display does not quietly change the workload.
Two numbers, and what each one means.
The on-screen resolve from noise to a clean image is paced to about 14 seconds so a fast GPU does not converge in a second and sit idle. That pacing is presentation. The measurement is separate.
Samples per second
The reciprocal of the mean per-sample GPU cost across the run, with the first sample excluded. Higher is faster. It is comparable between your own runs at the same backend and measurement revision, and directional against anything else.
Consistency
One minus the coefficient of variation across the per-sample times, expressed out of 100. A card that delivers samples unevenly scores lower than one that is merely slow, which is the distinction a single throughput figure hides.
Validity gate
Without a valid path-tracing result, including one taken with the tab in the background, path tracing is left out of the score and the other components carry it. The report marks the score as partial, and a partial run cannot grade above B. Skipping the stage, or a fallback, cannot inflate a score.
What it does not read
No temperature, no power draw, no VRAM. The browser exposes none of them. If the tracer stalls or floating-point render targets are unavailable, the stage substitutes a flagged placeholder that scoring refuses to trust.
Illustrative progress. Not a measurement.
Hardware
What it needs. Three tiers, no fine print.
Minimum
GTX 1050 / RX 560
Completes the run. Expect a long resolve and a low samples-per-second figure.
Recommended
GTX 1070 / RX 580
Resolves comfortably inside the paced window with a consistency figure worth reading.
Optimal
RTX 2060+ / RX 5700+
Keeps the pace easily. The figure is the cost of each sample, not how soon the image looks done.
Questions
Path tracing, answered
- How does path tracing work?
- For every pixel, a ray is fired from the camera into the scene. Where it hits a surface, it bounces in a direction chosen at random according to how that material scatters light, and it keeps bouncing until it reaches a light or is stopped. The light it collects on the way is that pixel's colour for one sample. One sample is noisy, so many are taken and averaged, and the image converges from grain to a clean result. Because every bounce is followed, soft shadows, reflections and light bleeding colour from one surface onto another all emerge on their own rather than being faked.
- What rules does path tracing follow?
- It follows the physics of light transport, usually written as the rendering equation: the light leaving a point is the light it emits plus all the light arriving from every direction, weighted by how the surface scatters it. In practice that means a few rules. Energy is conserved, so a surface never reflects more light than reaches it. Each material has a scattering function that decides where a bounce goes. Directions are chosen at random but in proportion to how likely light is to travel that way, which is called importance sampling. And paths are ended early at random, with the survivors weighted up to compensate, so the average stays correct without tracing forever.
- What is the difference between ray tracing and path tracing?
- Ray tracing is the general technique of following rays through a scene, and games usually use it for one effect at a time, such as reflections or shadows, on top of a rasterised image. Path tracing is ray tracing taken all the way: every pixel's light is computed by following full random bounce paths, so all lighting comes from one physically based calculation. It looks more correct and costs far more, which is why it makes a demanding benchmark.
- Does a browser benchmark actually use the GPU for path tracing?
- Yes. This stage runs as a fragment shader through WebGL2, which executes on the graphics card, not the processor. Each sample is timed around a blocking draw so the figure reflects the card's real cost, and the backend in use is recorded with the result.
- What is path tracing in GPU benchmarking?
- Path tracing simulates how light realistically bounces through a scene. In benchmarking it measures GPU performance under an extreme, physically-based lighting workload, far more demanding than rasterization.
- Why is path tracing harder to run than rasterization?
- Rasterization approximates lighting cheaply. Path tracing traces many light rays per pixel and accumulates samples over time, stressing shader throughput, memory bandwidth, and scene traversal continuously.
- Why does the renderer appear frozen for a few seconds?
- Before rendering, the engine prepares the scene and shaders. On first load this can briefly make the tab unresponsive; this is expected, and the run resumes automatically.
- Is a browser path-tracing score comparable to native tools?
- It is directional, not absolute. Browser results differ from native applications and depend on backend, drivers, and browser state. We disclose the backend and sample count with every result.