GloriousFlywheel Dawn WebGPU Proof 2026-04-22
Snapshot date: 2026-04-22
Owner surfaces:
#320/TIN-335#347/TIN-386#44
This note records the first bounded Dawn/WebGPU userspace proof on the shared
tinyland-nix-gpu lane.
Proof Surface
The proof is intentionally narrow and platform-owned:
- workflow:
Test ARC Runners Soak - job:
Test tinyland-nix-gpu on honey - branch proof run:
24786392621 - shared lane:
tinyland-nix-gpu - substrate:
honey
The goal is not to claim broad Dawn or WebGPU maturity. The goal is to prove
that the shared tinyland-nix-gpu lane can execute one honest non-software
WebGPU path after the earlier /dev/dri + Vulkan floor landed.
Implemented Contract
The landed proof adds:
- a checked-in smoke harness:
scripts/test-webgpu-smoke.py - explicit adapter inspection through
wgpu-py - software-renderer rejection for
llvmpipe,lavapipe,softpipe, andSwiftShader - one minimal compute submission path with deterministic output checking
- step-summary output that records backend and adapter identity for operators
The tinyland-nix-gpu soak job now proves both:
- Vulkan userspace on the shared lane
- one bounded Dawn/WebGPU userspace path on the same lane
Why This Counts
This counts as real bounded proof because:
- it runs on the GloriousFlywheel-owned shared ARC GPU lane, not only in a repo-scoped downstream workload
- it proves real adapter enumeration and rejects software-only fallback
- it completes a minimal compute submission path instead of only logging host metadata
- it leaves operator-visible backend and adapter evidence in the workflow summary
What It Does Not Yet Prove
This proof does not yet establish:
- a downstream default-branch Dawn/WebGPU canary repo
- a broader WebGPU product contract across multiple repos
- richer render-path coverage beyond the current bounded compute path
- a Kubernetes
nvidia.com/gpudevice-plugin contract
Those remain follow-ons after the first bounded shared-lane userspace proof.
Next Honest Move
After this proof exists, the next stronger GPU-runtime move is:
- use
#44to add at least one downstream default-branch canary ontinyland-nix-gpu, or - extend the shared-lane proof beyond the current bounded compute-only userspace path before broadening the public claim