STAGE 3 OF 3

Path Tracing Benchmark

Stage 3 is the final and most GPU-intensive phase. It runs a real physically-based path tracer using the three-gpu-pathtracer engine with full BVH (Bounding Volume Hierarchy) acceleration. The benchmark scene — a BMW M4 Competition 2021 in a warehouse environment — accumulates rendered light samples in real time. The score is time-to-fidelity: how fast the GPU reaches 150 samples. Faster is better. There is no fixed time limit.

Score Formula

pt_samples_per_sec = 150 / elapsed_seconds_to_fidelity

A GPU that reaches 150 samples in 30 seconds scores 5.0 SPS. A GPU that takes 120 seconds scores 1.25 SPS. The score directly reflects the GPU's ray-tracing core throughput, VRAM bandwidth, and thermal stability under the most demanding workload in the benchmark suite.

Benchmark Phases

Asset streaming: BMW M4 GLB model and warehouse EXR environment load from CDN. Progress indicator shows percentage complete.
BVH construction: The path tracer builds an acceleration structure from the scene geometry. This phase takes 2–10 seconds depending on GPU driver and VRAM speed. The scene appears frozen during BVH build — this is expected.
Accumulation: The renderer fires rays per-pixel every frame. The scene starts noisy and progressively sharpens as samples accumulate. The ACCUMULATING LIGHT indicator shows live sample count and estimated time remaining.
MAXIMUM FIDELITY: At 150 samples, rendering freezes. A 5-second hold plays before advancing to results.

What Drives PT Performance

RT core speed — Dedicated ray-tracing hardware (NVIDIA RTX, AMD RDNA2+) dramatically accelerates BVH traversal. Software-only rasterization GPUs are far slower.
VRAM bandwidth — Each ray bounce reads texture data and BVH nodes from VRAM. Higher bandwidth = faster samples per second.
VRAM capacity — The warehouse + BMW scene requires ~3.5GB VRAM. Cards below 4GB will compress textures, reducing sample rate and consistency.
Driver maturity — WebGL path tracing relies on GLSL compute shaders. Updated drivers ship compiler improvements that can increase throughput 10–20%.
GPU thermals — Unlike the 60-second GPU compute test, PT runs until completion. Cards that throttle during the accumulation phase will see their sample rate drop mid-run, increasing total time and decreasing consistency score.

Consistency Score

Alongside samples-per-second, the benchmark records a PT consistency score (0–100). This measures the variance of per-sample delivery intervals. A score of 100 means every sample arrived at a perfectly steady rate. Low consistency (below 65) indicates VRAM contention — the GPU is stalling on memory reads between samples, producing the GHOST RIG class archetype.

After Stage 3

When MAXIMUM FIDELITY is reached, the test data is saved and you are automatically redirected to your results page. The Silicon Doctor analyzes your forensic profile and generates a personalized diagnosis with prescriptions.