Blog / Technical / Tutorial 25: CUDA Graphs

Tutorial 25: CUDA Graphs

See `src/tutorials/25_cuda_graphs/main.cu` for detailed inline comments explaining each CUDA concept, all API calls, and the ablation experiments.

Shreedhar Kodate · 19 November 2025 · 1 min read
cuda-programmingdocs

Concept

See src/tutorials/25_cuda_graphs/main.cu for detailed inline comments explaining each CUDA concept, all API calls, and the ablation experiments.

Key APIs

Refer to the source file for full API usage and inline explanations.

Execution Flow

flowchart TD A[main: load config + init logger] --> B[cudaSetDevice + print_device_info] B --> C[Allocate host and device memory] C --> D[Run kernel variants / ablations] D --> E[CUDA event timing per variant] E --> F[Log results: time, bandwidth, correctness] F --> G[Cleanup: cudaFree + handle destroy] G --> H[Exit 0]

Ablation Results (placeholder — fill after running on H200)

Variant Mean Time (ms) Bandwidth / GFLOPS Notes
TBD TBD TBD TBD

What to Observe in Nsight Compute

Run:

bash
ncu --set full ./build/bin/tutorial_25_cuda_graphs

Key metrics:

  • sm__throughput.avg.pct_of_peak_sustained_active — SM utilization
  • dram__bytes.sum — HBM traffic
  • smsp__warp_issue_stalled_long_scoreboard_per_warp — memory stalls
  • l1tex__t_bytes_pipe_lsu_mem_global_op_ld.sum — L1/global load bytes