# Profiling Examples

## Nsight Compute

```bash
./tools/profile_ncu.sh python bench/bench_vector_add.py --device cuda --mode triton
./tools/profile_ncu.sh python bench/bench_softmax.py --device cuda --mode torch
```

## Nsight Systems

```bash
./tools/profile_nsys.sh python bench/bench_matmul.py --device cuda --mode triton
./tools/profile_nsys.sh python bench/bench_attention.py --device cuda --mode torch
```

## First Things To Inspect

- median runtime from the benchmark harness
- whether warmup was excluded
- whether kernels overlap or serialize
- whether memory throughput is near a practical ceiling
- whether a kernel launch is tiny enough that launch overhead matters