Performance¶

Benchmark results and speedup analysis for RayON on the primary development machine (NVIDIA DGX Spark, 2026).

Test configuration¶

Latest benchmark results¶

Render times (seconds) for 720p at 1024 SPP across recent commits on main:

Renderer comparison¶

Measured on a single test scene (default scene, 720p, 256 SPP):

BVH acceleration¶

The BVH (Bounding Volume Hierarchy) provides the largest single speedup for scenes with many objects. Performance measured across three scene sizes at 128 SPP, 720p:

Sampling efficiency¶

The graph below shows how image quality (measured as RMS noise relative to the 4096 SPP reference) improves with sample count. The \(1/\sqrt{N}\) curve is the theoretical Monte Carlo convergence rate.

The \(\mathcal{O}(1/\sqrt{N})\) convergence is a fundamental property of Monte Carlo integration. Doubling sample count reduces noise by a factor of \(\sqrt{2} \approx 1.41\), meaning halving visible noise requires quadrupling the samples.

Reference timings, October 2026¶

Offline OptiX renders at 720p on the DGX Spark, measured during the development of the volume material (single run each, the GPU otherwise idle):

Running the benchmark yourself¶

cd /path/to/RayON
bash scripts/benchmark.sh --method cuda --sampler pcg
# Results appended to bench_results.csv

The script launches three sequential renders of the default scene, records GPU, resolution, samples, renderer method, sampler, commit hash, branch, and wall-clock render time to bench_results.csv. Supported methods are cuda (-m 2) and optix (-m 4), and the default benchmark sampler is pcg. You can override sampler with --sampler sobol.

timestamp,commit,branch,gpu,resolution,samples,method,sampler,run,time_s
2026-03-12T21:39:52+01:00,ce8002d-dirty,main,NVIDIA GB10,720,1024,cuda(2),pcg,1,1.651
2026-03-12T21:39:52+01:00,ce8002d-dirty,main,NVIDIA GB10,720,1024,cuda(2),pcg,2,1.787
2026-03-12T21:39:52+01:00,ce8002d-dirty,main,NVIDIA GB10,720,1024,cuda(2),pcg,3,1.652

Tuning tips¶

• Enable BVH for any scene with more than ~15 objects: use_bvh: true in YAML.
• Raise the Target FPS in the GUI's "Camera Motion" section for a very responsive orbit/pan experience — the batch size and the render resolution auto-scale down to keep up.
• Block size — the default 32×8 is tuned for warp alignment; register-cap and block-shape experiments (see ROADMAP.md) gave no consistent gain.
• Fast math — --use_fast_math is on in Release builds and off in Debug builds.
• Prefer OptiX (-m 4, -m 5) on RTX-class hardware: triangles and shadow rays run on the RT cores, and the interactive mode has the motion regulator.
• Keep an eye on the first run after a build: OptiX compiles its module on the first launch (a few seconds, then cached on disk), so time the second run.