GloriousFlywheel Honey GPU Lane Implementation 2026-04-22

Snapshot date: 2026-04-22

Owner surfaces:

• #320 / TIN-335
• #342 / TIN-379
• #44

This note records the first real shared GPU lane implementation after the bounded repo-scoped cmux floor.

Current Constraint

The live honey cluster does not currently advertise a Kubernetes GPU device plugin contract:

• no node currently has an accelerator=nvidia label
• no node currently reports allocatable nvidia.com/gpu

That means the first honest shared GPU lane cannot be framed as a Kubernetes nvidia.com/gpu lane.

Live Host Floor

The current shared GPU floor is the host hardware on honey itself:

• host honey exposes /dev/dri/card0 and /dev/dri/renderD128
• host honey reports an AMD Navi 48 GPU through lspci
• host honey Vulkan userspace sees:
  • deviceType = PHYSICAL_DEVICE_TYPE_DISCRETE_GPU
  • deviceName = AMD Radeon Graphics (RADV GFX1201)
  • driverName = radv

That is enough to justify one bounded host-device ARC lane.

Implemented Shared Lane

The first shared ARC GPU lane is:

• runner label: tinyland-nix-gpu
• stack owner: tofu/stacks/arc-runners/honey.tfvars
• execution model: Nix runner on honey
• device contract: host /dev/dri mounted into the runner container
• runner security contract: root runner with RUNNER_ALLOW_RUNASROOT=1
• proof workflow: Test ARC Runners Soak
• proof job: Test tinyland-nix-gpu on honey

Why This Counts

This implementation counts as real shared proof because:

• it is GloriousFlywheel-owned ARC config, not only repo-local CI
• it uses the actual current GPU interface on honey
• it proves device visibility and Vulkan userspace from inside the runner pod
• it does not overclaim a broader Dawn / WebGPU or Kubernetes GPU-operator contract than the platform currently has

What It Still Does Not Prove

This lane does not yet prove:

• a cluster-wide nvidia.com/gpu scheduling contract
• broader Dawn runtime correctness
• broader WebGPU runtime correctness
• downstream default-branch adoption on the shared lane

Those are follow-ons after the bounded shared tinyland-nix-gpu floor.

Next Honest GPU Step

After this lane exists, the next GPU-specific product move is:

• use #44 to define one explicit Dawn / WebGPU proof surface and at least one downstream default-branch canary on the shared lane

That keeps the shared-lane implementation and broader runtime API claims properly separated.