DOCUMENTATION / experiments/racing/RESULTS.md

Racing cartridge comparison

Measured on 2026-10-19 with an Intel Core i7-11700K, Windows x64, Chrome 155.0.8059.41, Odin dev-2026-10 (84bc3fc), Wasmtime 49.0.3 and the pinned Luau Wasm executor. Both cartridges use the same Odin player and shared renderer at 320×180. Native uses the normal fuel-metered execution policy.

Host	Cartridge	Mean CPU frame	p95 CPU frame	Startup, cache bypassed	Startup, executor cached
Native	Odin	0.046 ms	0.078 ms	50.2 ms	—
Native	Luau	0.212 ms	0.516 ms	168.9 ms	11.7 ms
Browser	Odin	0.093 ms	0.200 ms	7.7 ms	—
Browser	Luau	0.231 ms	0.500 ms	30.5 ms	9.2 ms

The frame figures use the cached Luau samples; bypassing its code cache produced similar steady-state frame times. Odin is approximately 4.6× faster on native and 2.5× faster in the browser in this workload. Both CPU workloads are well below a 16.67 ms tick on this desktop. This is not a measured end-to-end FPS rate, and does not establish performance on mobile or low-power devices.

Where the time goes

Host / cartridge	Simulation	Drawing / submission	Shared rasterization	Pixel copy, additional
Native / Odin	0.0003 ms	0.0121 ms	0.0337 ms	—
Native / Luau, cached	0.0027 ms	0.1751 ms	0.0337 ms	—
Browser / Odin	0.0006 ms	0.0125 ms	0.0803 ms	0.0741 ms
Browser / Luau, cached	0.0029 ms	0.1487 ms	0.0797 ms	0.0725 ms

Most of the difference comes from generating drawing commands. The rasterizer's cost is nearly identical within each host, as expected from the shared path. Luau still calls its C graphics bindings for individual commands while building the batch inside Wasm. The simulation itself is inexpensive in both languages. Browser timer granularity limits interpretation of the smallest phase values.

Memory and size

Item	Odin cartridge	Luau cartridge + executor
Guest linear memory	1.5625 MiB	2 MiB
Peak Luau allocator usage, inside guest memory	—	670 KiB
Native process working set, median	22.1 MiB	17.2 MiB with executor cache
Cartridge bytes	160,354 Wasm	10,269 bundled source
Player-provided executor bytes	—	711,359 Wasm

The browser's shared Odin renderer reserves 3.125 MiB of additional linear memory for either language. The native renderer lives inside the process working set. Active low-resolution pixel transfers are 225 KiB/frame; the renderer reserves the maximum framebuffer so switching modes does not require reallocating it. Guest memory figures exclude compiled code, JS/browser objects, stack reservations outside the guest and GPU/display buffers. Working-set figures are process-wide and do not imply that Luau uses less guest memory.

Method and limitations

Three serial samples per variant; 720 frames per replay; two warmup and five measured passes per sample, totaling 10,800 measured frames per row.
Native compiles Odin's module on each process launch. Only the Luau executor is cached by the current player. Its cached startup is therefore not an equal-cache language comparison; the bypassed startup columns are the useful baseline.
Browser cache hits reuse a compiled executor in the owning page and instantiate fresh guest memory in a new Worker. Cache bypass does not clear OS file caches or the browser engine's internal compilation caches. Browser timings use a local HTTP server, not the public tunnel.
Both source compilation and game initialization are included on every Luau load. The cache stores executor code, not Luau game state or precompiled cartridge code.
CPU frame time includes update, command submission and software rasterization. It excludes hashes, native display/audio, browser canvas painting and scheduling. Browser framebuffer copies are measured separately, not included in the table.
This road-and-billboard workload does not predict the performance of a future mesh-based 3D renderer, large RPG, pathfinding system or larger asset set.

Full race and handling replays match across languages and hosts. See the demo guide for checks and reproduction commands. The portable measurement summary is checked in; raw samples and traces remain under build/racing-player/benchmark/.