FastCuda

Use when creating or revising CUDA benchmark runners and result artifacts.

Use when CUDA availability, compiler setup, or profiler visibility may be broken or inconsistent.

Use when designing or reviewing handwritten CUDA FlashAttention kernels.

Use when designing or reviewing handwritten CUDA GEMM kernels.

Use for kernel-level analysis with Nsight Compute after a benchmark is reproducible.

Use for stream overlap, launch overhead, and end-to-end CUDA timeline analysis with Nsight Systems.

Use when deciding whether a CUDA kernel is likely memory-bound or compute-bound.