Browser and GPU support

Where it runs, what it uses when present, and what we test on.

The packages need WebGPU and nothing else. Optional features make them faster when the device has them. detectCaps(adapter) reports what a device offers, and every kernel compiles against that report.

What we test on

TestedChrome (stable and Canary) on an Apple M5 Pro (Metal) and on an NVIDIA RTX 3060 laptop (Windows, D3D12).
Should work, untestedEdge (same engine as Chrome); Safari 26 and Firefox 141+ on Windows ship WebGPU. Tell us what you find.
Correctness onlySwiftShader (Chrome's CPU WebGPU) in CI-like containers: right images, slow.

Optional features

FeatureUsed for
subgroupsFaster queue appends in the wavefront scheduler; parts of the research code.
shader-f16Half-precision storage in the real-time planes and the neural cache.
timestamp-queryTiming in the bench and the examples' stats.
chromium-experimental-sampling-resource-tableBindless textures (experimental; not usable in compute passes on Metal yet).

caps.tier is a summary label for reports: 0 is WebGPU baseline, 1 adds f16, timestamps, compressed textures and larger limits, 2 adds subgroups, 3 adds the resource table. Code never branches on the tier, only on the individual flags.

Things to know

  • Kernel size. A path tracer kernel with every material feature is about 50 KB of WGSL. Large kernels compile slowly in the Windows shader compiler (DXC). ptSceneOptions only compiles what the scene's materials use. ptwfAutoPolicy moves very large kernels to the wavefront scheduler, whose kernels are smaller.
  • Storage buffers. Chrome on Apple GPUs allows 10 storage buffers per shader stage whatever the device requests. The default scene-copy packing (packed) keeps one binding for the whole scene.
  • Binding size. A packed scene copy must fit one storage binding (128 MiB on most devices). That is about 1.5 million triangles with normals and positions; above that, upload with packing: 'split'.
  • Hidden tabs. Browsers throttle requestAnimationFrame in hidden tabs. A progressive render pauses there, which is usually what you want.