cuda-b200

To optimize CUDA kernels for NVIDIA B200, enforce a correctness-first workflow using compute-sanitizer, then profile with nsys to discover hotspots and use ncu for kernel-level diagnosis before applying Blackwell-specific tuning constraints like unified L1/texture/shared memory carveouts.

The sm_100 and compute_100 targets represent generic Blackwell behavior, while sm_100a and compute_100a are architecture-specific targets for B200 that enable features not guaranteed across all sm_100 devices, requiring careful compatibility control and PTX inclusion when needed.

Debug CUDA kernel correctness issues using compute-sanitizer tools including memcheck, racecheck, synccheck, and initcheck, combined with targeted device-side printf strategies to validate memory access, thread synchronization, and initialization before profiling.

Profile Tensor Core usage on B200 GPUs by running nsys for initial hotspot discovery, then use ncu for detailed kernel-level diagnosis of occupancy, stalls, coalescing, register pressure, shared-memory bottlenecks, and TensorCore utilization.

NVIDIA B200 supports cluster portability but requires separating generic sm_100 targets from architecture-specific sm_100a targets, with shared-memory carveout options and cluster constraints factored into tuning decisions to ensure compatibility.

CUDA kernels may fail for B200 compute capability 10.0 if the wrong nvcc target is selected; use generic sm_100 or compute_100 for broad compatibility, or sm_100a for architecture-specific features, ensuring proper PTX inclusion when required.