Module chip 

The GPU chip / family knob — which NVIDIA silicon the CUDA backend targets.

This is the CUDA analog of the Tenstorrent backend’s per-part crate::spatial::target::SpatialTarget (wormhole / blackhole): a described target part whose SM architecture, shared-memory budget, and realized numeric formats make the fastest kernel chip-specific. The SAME recognized LOGOS tensor op lowers to a kernel tuned for the local RTX 3060 (Ampere, sm_86) — where we prove correctness — or for a cloud B200 (Blackwell, sm_100), the target we submit to NVIDIA’s SOL-ExecBench against.

Choosing the chip is a genuine trade-off (hence a knob, not a constant): it trades portability against hardware exploitation. A kernel compiled for sm_100 won’t run on an sm_86 host; a kernel compiled for sm_86 leaves the B200’s 5th-gen tensor cores, fp8/nvfp4, and larger shared memory unused.

Structs

CudaChip

A described CUDA target part — the hardware limits and realized numeric formats that make the fastest kernel chip-specific, plus the SM arch its kernels are compiled for. Construct with a named part (CudaChip::rtx3060, CudaChip::b200, …); the emitter reads it to pick the arch flag, the shared-memory budget, and which precisions are legal.

Enums

GpuFamily

An NVIDIA GPU architecture family — fixes the tensor-core generation and which numeric formats exist.