RQ-001State: HypothesisConfidence: Low–mediumSovereign and Local AIModel Evaluation
Can Apple GPUs support persistent-kernel global synchronisation?
Hypothesis
A forward-progress-safe persistent kernel may be possible only if the runtime can bound residency, barriers and workgroup scheduling under realistic contention.
Evidence
- 577 barrier/token audit
- M1 Ultra vs M4 Max bandwidth traces
- Megakernel-lite stalls under long dependency chains
Experiments
- Barrier/token replay harness
- Occupancy cliff sweep
- Forward-progress watchdog
Results so far
No production-safe proof yet. The failure modes are reproducible enough to define the next test harness.
Current belief
Treat global synchronisation as unproven on Apple GPU until a watchdog-backed stress test survives pathological occupancy.