cuda-b200

What problem does it solve?

This Skill helps you author, debug, and performance-tune CUDA kernels specifically for NVIDIA B200 (compute capability 10.0) by enforcing a correctness-first, measurement-driven workflow and by separating generic Blackwell targets (sm_100) from architecture-specific ones (sm_100a).

Core Features & Use Cases

• Correctness-first kernel development using the right debugging loop with compute-sanitizer tools (memcheck, racecheck, synccheck, initcheck) and targeted device-side printf strategies.
• Blackwell-aware tuning decisions grounded in B200 facts like unified L1/texture/shared memory (256KB per SM), shared-memory carveout options, and cluster portability constraints.
• Profile-then-optimize performance workflow using nsys for hotspot discovery and ncu for kernel-level diagnosis (occupancy, stalls, coalescing, register pressure, shared-memory bottlenecks, TensorCore usage).
• Architecture-target selection and compatibility control between generic sm_100 / compute_100 and architecture-specific sm_100a / compute_100a, plus guidance on when PTX inclusion is needed.

Quick Start

Use the cuda-b200 skill to review your CUDA kernel for B200 by checking correctness, then running nsys and ncu to identify the main bottleneck before proposing a single measurable optimization change.