Nsight Compute Basics for Kernel Authors

Nsight Compute Basics for Kernel Authors

Thesis

Nsight Compute (ncu) explains kernel slowness with occupancy, bandwidth, and stall metrics. ncu was not run here (no GPU).

Mental model

  ncu report → speed-of-light → stalls → fix memory first

Worked examples

ncu --kernel-name regex:my_gemm --set full -o gemm ./bench
ncu -i gemm.ncu-rep

Start with Speed of Light; then stall reasons.

Trap

Profiling debug builds or tiny launches. Noise dominates.

Boring rule

One ncu report before large kernel rewrites.

Try this

1. Compare ncu vs Nsight Systems.
2. Map Triton kernel names into reports.
3. Read NVIDIA Nsight Compute docs for your toolkit.

Sources

• NVIDIA Nsight Compute documentation