Godot 4.7.1 p0014 published Forward Mobile WebGPU beta

A published Godot game running through WebGPU.

The initial backend lineage is David Walter's MIT-licensed dwalter/godotwebgpu work. Studio Foundation maintains the Godot 4.7.1 port, scoped patches, renderer fixes, build/export tooling, validation, release evidence, MCP tooling, and AI-native distribution layer. Official Godot remains upstream.

A Roman colosseum with crowded stands and chariots rendered by Godot through WebGPU

Run and inspect

WebGPU support: checking…

Enter the colosseum Minimal showcase Repository p0014 templates Evidence matrix

Initial load downloads a ~45.8 MB wasm (~12.0 MB over GitHub Pages gzip). In the recorded 12 Mbps p0014 run, frames appeared around 13 seconds and pipeline construction occupied about 2 seconds of engine startup.

Published release versus current main

StatePatchesRendererArtifactsHardware status
Published godot-4.7.1-webgpu-p0014 0001–0014 WebGPU Forward Mobile Release/debug downloads available Minimal scene and three game scenes rendered on an NVIDIA Tesla P40 with 0 GPUValidationError
Current main 0001–0022 Forward Mobile plus unfinished Forward+ investigation No p0022 templates published Forward+ was hardware-run on the P40; 18 validation errors remain and no frame renders

Patches 0015–0022 are not in the p0014 downloads. Forward+ hardware execution is investigation evidence, not a successful-render claim.

What the published demo proves

WebGPU support is beta. Safari/iOS, native mobile devices, and non-NVIDIA GPU vendors remain unverified. The p0014 Jolt web build also has a known concave-collider regression documented in the evidence matrix.

Reproduce it

The engine integration, minimal scene, and Chariot source are available in the repository. Chariot is public for reproducibility but remains all rights reserved under games/LICENSE.

Download both p0014 templates, verify their SHA-256 values from the evidence matrix, and export with just export-browser-webgpu. That command applies the WebGPU shell handoff the official editor cannot emit for these templates.

AI-native positioning

“AI-native” describes how Studio Foundation is operated: an MCP server, agent workflows, Blender asset tooling, and public validation commands. It is separate from the renderer claims above, which are scoped to named artifacts, hardware, scenes, and measured caveats.