GloriousFlywheel Shared GPU Lane 2026-04-22

GloriousFlywheel Shared GPU Lane 2026-04-22

Snapshot date: 2026-04-22

Owner surfaces:

  • #320 / TIN-335
  • #340 / TIN-378
  • #338 / TIN-377
  • #44

This note records the next explicit GPU move after the bounded cmux proof floor landed.

The goal is not to claim broad GPU maturity. The goal is to say what the first shared GPU lane should be before GloriousFlywheel broadens into Dawn, WebGPU, or generalized shared-GPU claims.

Current Floor

The current GPU floor is already real:

  • repo: Jesssullivan/cmux
  • workflow: GPU Smoke Test (Self-Hosted)
  • run: 24756928163 on main
  • job: GPU smoke test (cmux-nix) 72432003462

That proves one bounded repo-scoped GitHub-first GPU path. It does not prove a shared platform-owned GPU lane.

Current Gap

GloriousFlywheel still has no ARC-owned GPU lane in:

  • tofu/stacks/arc-runners
  • tofu/modules/arc-runner

The repo still carries legacy GitLab-shaped GPU module notes, but there is no named shared honey GPU lane with a current default-branch workload canary and operator-owned lifecycle contract.

Next Shared Lane

The next honest product step is:

  • one named shared GPU lane on honey
  • one real GitHub Actions canary workload on that lane
  • explicit ownership for bootstrap, cache behavior, trace extraction, queue behavior, and cleanup

This should stay narrower than the broader Dawn / WebGPU backlog. The first shared lane is the promotion step that should come before generalized GPU maturity language.

Why This Is Separate From #44

#44 remains broader than the first shared lane. It still covers:

  • Dawn-specific follow-on work
  • WebGPU-specific follow-on work
  • local NVIDIA-fabric or richer shared-GPU backend ideas

Those are follow-ons after the first shared honey GPU lane is named.

Definition Of Done

For this shared-lane promotion slice, done should mean:

  • one explicit shared honey GPU lane is named
  • one real canary workload is attached to that lane
  • bootstrap, cache, trace-extraction, and cleanup behavior are documented
  • public docs distinguish the shared-lane promotion from the broader Dawn / WebGPU backlog

Current Boundary

The honest current boundary is:

  • bounded repo-scoped GPU proof exists today through cmux
  • the first shared honey GPU lane is the next GPU productization step
  • broader Dawn / WebGPU maturity is still after that step, not before it

GloriousFlywheel