Nine key tests for choosing a 3D engine for the web, built identically in each engine and measured frame by frame on WebGPU and WebGL2. Every scene renders the same geometry, lights and camera at 1920×1080 with 4× MSAA and ACES tone mapping; the harness owns the frame loop and times each engine through the same instrumentation.
Each test, path and backend either rendered a verified frame, uses the engine's default path because it has no separate optimized one, is not applicable because the engine cannot run it, or failed. Two marks per browser: WebGPU, then WebGL2.
CPU time per frame for each engine in each browser, at one fixed level per test, current backend and path. CPU time is the one metric every browser reports identically: Safari caps the frame rate at the display refresh.
WebGPU calls per frame recorded by the harness's probe (not the engine's own counters) for 1,000 animated meshes in a hierarchy, plus live GPU memory from every buffer and texture the engine allocated.
Shipped JavaScript (tree-shaken, minified, brotli) for the smallest lit-cube app and for a realistic one (glTF, animation, shadows, bloom), and the time from navigation to the first rendered frame.
Brotli-compressed. Total is every chunk the build emits; initial is what must download before the first frame; lazy chunks load on demand, and a page may never request some of them.
Frame 60 of each scene in Chrome on WebGPU, captured by the harness. Identical data in, so any visible difference is the engine's lighting model or defaults.
Lines of TypeScript each adapter needed to build the same nine tests, by file. Fewer lines for the same result means less to write and maintain; a missing file means the engine had no path for that test.