AmorisData and validation
These results describe particular workloads and builds. They are not promises about every scene or a general ranking against other engines.
High-detail native rendering
The complete Bistro exterior contains 2,829,226 triangles, 1,296 mesh objects and 132 authored materials. Its 405 embedded source images decode to 3.76 GB of RGBA8 before the renderer's 1K conversion. The imported runtime has 1,591 primitive instances and 133 material slots including its default material. At the recorded street-side camera, the same 1080p render-and-wait method measured 12.02 ms mean, 11.94 ms p50, 13.00 ms p95, and 1.827 ms CPU encode mean over 300 frames. Raw Bistro result.
Recorded 5 October 2026, Apple M5 / Metal, 1920 × 1080. Each CC0 Flight Helmet contributes 94,722 triangles and six materials. Geometry and textures are shared; the scene repeats the complete model rather than substituting a primitive. Directional shadows and normal PBR/postprocessing passes remain enabled.
| Helmets | Scene triangles before culling | Primitive instances | Completed frame mean | p50 | p95 | CPU encode mean |
|---|---|---|---|---|---|---|
| 64 | 6,062,210 | 385 | 10.52 ms | 10.48 ms | 11.48 ms | 0.317 ms |
| 256 | 24,248,834 | 1,537 | 28.81 ms | 28.83 ms | 29.65 ms | 0.407 ms |
| 1,024 | 96,995,330 | 6,145 | 82.43 ms | 82.41 ms | 83.66 ms | 0.454 ms |
The benchmark waits for zero pending assets, completes 60 loaded warm-up frames, then measures 300 frames. Each frame renders offscreen and waits for GPU completion. This wall time includes the complete static rendering workload and command submission; it excludes game simulation, window presentation, pixel readback and PNG encoding. It is not an interactive game's FPS. The scene triangle counts exclude shadow-pass multiplication and precede culling. An untimed ID pass confirmed visible pixels from all 64 and 256 helmets, and 1,021 of 1,024 helmets (others were occluded), plus the ground.
64-instance JSON, 256-instance JSON, 1,024-instance JSON. These raw reports also include pass timings, material counts and source-image dimensions. The existing GPU profiler covers selected passes and omits postprocessing; Metal pass boundaries can omit tile work. Its sum is not a whole-frame GPU time and is not used for the chart above. The renderer currently downsamples maps into 1024 × 1024 layers.
Reproduce the workload:
python3 tools/showcase_assets.py --projects
cargo run --release -p pocket-render --example showcase_bench -- \
site/demos/stress-1024 --width 1920 --height 1080 --frames 300 --warm 60
The asset fetch verifies pinned hashes. See asset credits for licenses, source authors, conversion details, and current material limitations.
Native and browser validation
The imported implementation was checked on 5 October 2026:
| Check | Result |
|---|---|
| Native release build | Passed |
| Native tests | 383 passed, 0 failed, 0 ignored in 78 suites; ignored tests included |
| Dependency boundaries | Passed across 19 workspace crates and 2 targets |
| wasm compilation | Game crates and pocket-web passed |
| Script types | Passed for 2 projects |
| Determinism | Passed, including a native debug/release comparison |
| Fork, replay, and reload | Passed on the configured sample workloads |
| Browser/native hash chain | Passed in headless Chrome, isolated and non-isolated, and after wasm-opt -O3 |
| Editor | Typecheck and production build passed |
The full cargo xtask check verdict remains Fail: the source has Rust formatting,
Markdown wrapping, and Clippy diagnostics. Its contract conformance and performance
budget steps are still unimplemented and report Skipped. The test count above does
not mean the full check is green.
Web rendering study
Recorded 4 October 2026 on Apple M5, Chrome 152 headless, with a 1280 × 577 canvas at DPR 1. Results are the means of the last 300 frames after warm-up, without a frame-rate cap.
The sphere layout follows Bevy's many_cubes: one mesh, one material, a directional light,
and 4× MSAA. three.js r186 uses WebGPURenderer and an InstancedMesh. The Amoris renderer
uses per-instance GPU culling, indirect draws, clustered PBR, atmosphere lighting, bloom,
and AgX. The two paths perform different work.
| Cubes | three.js frame interval | three.js GPU | Amoris frame interval | Amoris GPU |
|---|---|---|---|---|
| 100,000 | 1.00 ms | 2.45 ms | 0.95 ms | 0.71 ms |
| 400,000 | 4.17 ms | 2.36 ms | 2.36 ms | 1.81 ms |
| 1,600,000 | 24.8 ms | 21.9 ms | 15.0 ms | 14.7 ms |
The frame interval measures submission pace; timestamp queries measure GPU work. The original web study records methodology, module sizes, and the open gap between native and web GPU timing.
Physics is a separate study
The current physics backend is Rapier 3D with enhanced determinism. A separate benchmark compares its feature configurations with Jolt 5.6 on equivalent Jolt PerformanceTest scenes, and studies snapshots and forks. The source explicitly distinguishes results from a future backend recommendation; the engine has not switched to Jolt.
See the physics study for native, wasm, SIMD, and deterministic configurations.
Reproduce the checks
cargo xtask check --json
./target/release/pocket check samples/sailing --json
cd editor && bun run typecheck && bun run build
Keep CARGO_TARGET_DIR inside the checkout. Never set POCKET_BLESS while validating.
Media capture is separate from these performance measurements: the demonstration clips
use recorded frames and fixed simulation steps.