CLC launch-queue drain

Resident CTAs start with one tile, take different amounts of time to process it, then try to steal the next not-yet-started tile with CLC. The red marker indicates a worker asks for more work after the queue (complete launch grid) is empty.

initial resident CTA tile CLC pull from queue empty pull / worker stops
queue slots: real tiles done: empty pulls: active SM:

Launch queue

Resident CTAs / SM workers

The noisy processing time is the point: faster workers return to the queue sooner, so SM ownership of tile IDs is intentionally not round-robin. Once the finite launch queue is empty, the next pull attempt fails and that worker stops.