15. Device placement: compiler cost model, explicit, or overridable¶
Date: 2026-11-20
Status: Accepted
Context¶
Work may run on CPU cores, the GPU, or eventually the Neural Engine, and authors sometimes know better than a cost model.
Decision¶
Three forms are supported: @place(auto) uses the compiler's cost model; @place(gpu) or @place(cpu) fixes the device; @place(auto, prefer = gpu) keeps the cost model but overrides it. Placement must never change the proved result. CPU and GPU (Metal) are first targets; the Neural Engine is reachable only through Core ML, so it is a later, optional target for fixed operations.
Consequences¶
- Placement interacts with float rules (ADR 0012).