Benchmarks

Forward-pass throughput of the main render paths, measured with examples/benchmark.py on an Apple-Silicon GPU. Reproduce locally with:

uv run python examples/benchmark.py            # plain table
uv run python examples/benchmark.py --markdown # the table below

All inputs are synthetic (no downloads); numbers are per-call averages after a warmup and will vary with your chip and thermal state.

Benchmark	Latency	Throughput
hard rasterizer · 20 480 faces @ 256px	4.3 ms	235 fps
hard rasterizer · 20 480 faces @ 512px	6.2 ms	162 fps
hard rasterizer · 20 480 faces @ 1024px	17.9 ms	56 fps
soft rasterizer · 20 480 faces @ 256px	68 300 ms	0.01 fps
gaussian splatting · 10 000 splats @ 512px	12.5 ms	80 fps
gaussian splatting · 100 000 splats @ 720px	100 ms	10 fps
vanilla NeRF · 4096 rays × 64 samples	152 ms	27k rays/s
hash-grid NeRF · 4096 rays × 64 samples	129 ms	32k rays/s

Reading the numbers

• Hard vs soft mesh rasterizer. The hard z-buffer rasterizer is four to five orders of magnitude faster than the soft one on a 20k-face mesh: it bins faces into 16×16 screen tiles (the same tiler the Gaussian Splatting path uses), so each pixel only tests the handful of faces in its tile, and renders 720p–1024px in real time. The soft path materializes per-face image buffers and is only for silhouette gradients.
• Gaussian Splatting runs interactively (the tile-based Metal kernels): real scenes with ~100k–400k splats render in real time at 512–720p.
• Hash-grid NeRF renders faster than the vanilla MLP NeRF and converges in far fewer iterations — it is the recommended NeRF path.

Scaling

Because faces are tile-binned, the hard rasterizer's per-pixel cost depends on the local face density rather than the total face count, so it stays fast as meshes grow and resolution increases (56 fps at 1024px on a 20k-face mesh). Very dense meshes that pack many triangles into one tile are the worst case; a depth-sorted early-out and finer tiles are the next refinements.