Skip to content

Render-cost measurements

Recorded 2026-10-08, M1 Max/Godot 4.7.2. Lower DoF quality, unshadowed accents, and motion scaling reduced the isolated harness from 43.3 to 12.4 ms. The real 2560×1440 host improved from 32.4 to 22 ms.

Recorded result — 2026-10-08

typing frame, harness
before 43.3 ms
DoF quality 1 (project.godot) and unshadowed accents 26.4 ms
plus motion_render_scale 0.5 while the camera moves 12.4 ms

Motion scale 0.6 measured 16.7 ms; 0.55 measured 14.4 ms. Hiding scenery in the real host still cost 19.4 ms, so terminal/compositor work remained significant.

What was measured

The harness mirrored SceneryView: streamed scene, four sun cascades, 28 accents, SSAO, sky, typing lens, MSAA 4×, 3840×2160 target at scale 0.8. Metal GPU timestamps returned zero; these are uncapped frame times, not GPU timings. CPU was about 0.5 ms. Baselines alternated three times and medians were compared.

change, shipped baseline 37-41 ms delta
typing DoF off -13.6 ms, the same on an empty scene
DoF quality High -> Low -11.3 ms
28 accent lights removed -14.1 ms
28 accent lights unshadowed -2.4 ms
MSAA off -2.0 ms
SSAO off about -1 ms
scenery hidden -19.8 ms

Interpretation

Finding Evidence
Depth of field dominated Full-resolution capped 32 px bokeh, quality-3 24-tap passes
Accents cost more than shadows Removing lights saved 14.1 ms; removing shadows 2.4 ms
Shadow redraw could persist Atlas-size mismatch failed to synchronize slot versions after changes
Geometry alone was cheap Unshaded scene matched an empty frame

The bokeh-shape setting had merged into a comment, leaving hexagonal shape. Uniform atlas slot sizes avoid the observed perpetual shadow redraw (~2.45M triangles/frame).

Method, and two ways of getting it wrong

  • Use separate processes for pipeline/material changes; runtime specialization polluted one nosky measurement to 72 ms.
  • Pair visibility-only changes inside one process using --rt=ab_*: 120 frames before/after. Thermal baseline drift was 44–68 ms across longer runs.
  • Use --rendertest --rt=CASE --fullscreen; the tiling manager ignored resolution.

Open

The recorded follow-ups were static-light baking, non-scenery profiling, and per-world render settings. The screen cap also multiplied already-pixel macOS screen size by scale. These dated observations need fresh measurements before drawing current performance conclusions.