NVIDIA Nsight Systems profiles of FastVideo running Wan-AI/Wan2.1-T2V-1.3B-Diffusers text-to-video inference on a single NVIDIA H100 SXM with sequence-parallel size 1.

The dataset contains two profiles of the same workload on the same machine, differing only in the attention backend:

	Attention backend	Kernel observed	File
FA2 baseline	FlashAttention-2	`flash_fwd_kernel<...>`	`profiles/perf_h100_sp1.sqlite`
FA3 (Hopper)	FlashAttention-3	`cutlass::device_kernel<flash::enable_sm90<flash::FlashAttnFwdSm90<...>>>`	`profiles/perf_h100_sp1_fa3.sqlite`

Headline result

FA3 cuts steady-state Wan2.1-T2V-1.3B inference from 182.5 s → 114.7 s per video sample on H100 SXM (1.59×, –37 %). All of the savings come from the attention kernel (–46.5 % per call), driven by three Hopper-specific techniques the FA3 kernel uses and FA2 cannot:

Persistent kernel — grid = 132 = #SMs (vs FA2's 6,756 blocks).
Warpgroup specialization — 384 threads/block = 3 warpgroups × 128 (producer + 2 consumers); FA2 launches 128 threads/block.
211 KiB SMEM tile — uses Hopper's 228 KiB/SM SMEM budget; FA2 sticks to 64 KiB to stay compatible with Ampere.

What's in this dataset

fastvideo-wan-h100-sp1-nsys/
├── README.md                        ← you are here (dataset card)
├── analysis_fa2.md                  ← FA2 profile standalone deep-dive (§1–§10)
├── analysis_fa3.md                  ← FA3 profile standalone deep-dive (§1–§10)
├── analysis_comparison.md           ← FA2 ↔ FA3 head-to-head + relationship
│
├── profiles/
│   ├── perf_h100_sp1.sqlite         ← FA2 baseline (157 MiB)
│   └── perf_h100_sp1_fa3.sqlite     ← FA3 Hopper      (180 MiB)
│
├── plots/
│   ├── README.md                    ← which plot supports which §
│   ├── make_plots.py                ← reproduction script (stdlib + matplotlib)
│   ├── 1_denoising_wall_time.png            (per-sample, FA2 vs FA3)
│   ├── 2_attention_kernel_distribution.png  (bimodal histogram, both)
│   ├── 3_kernel_category_breakdown.png      (7-category stacked bars)
│   ├── 4_walltime_stack.png                 (compute vs idle stack)
│   ├── 5_attention_trace.png                (10,800 kernels scatter)
│   └── 6_kernel_launch_params.png           (grid/block/regs/SMEM/time)
│
├── skill_outputs/                   ← every nsys-ai skill, pre-computed
│   ├── fa2/                         (20 JSON files)
│   ├── fa3/                         (20 JSON files)
│   ├── diff_fa2_to_fa3.md           (nsys-ai diff, markdown)
│   └── diff_fa2_to_fa3.json         (nsys-ai diff, JSON)
│
└── raw_queries/                     ← supporting raw SQL extractions used in the writeup
    ├── 01_row_counts.json
    ├── 02_profile_span.json
    ├── 03_attention_kernels.json
    ├── 04_nvtx_regions.json
    └── 05_attention_bimodal.json

The .nsys-rep files (NVIDIA's native format) are intentionally not included — the exported .sqlite files contain all the same trace data in a queryable form. Re-exporting from .nsys-rep is trivial via nsys export --type=sqlite.

How to use it

Open in the browser (interactive timeline)

pip install nsys-ai
nsys-ai timeline-web profiles/perf_h100_sp1.sqlite       # FA2
nsys-ai timeline-web profiles/perf_h100_sp1_fa3.sqlite   # FA3

Reproduce a skill output

# Any of the 20 skills in skill_outputs/{fa2,fa3}/
nsys-ai skill run profile_health_manifest profiles/perf_h100_sp1.sqlite --format json
nsys-ai skill run overlap_breakdown        profiles/perf_h100_sp1_fa3.sqlite --format json
nsys-ai skill list   # see all 36 available skills

Reproduce the FA2 → FA3 diff

nsys-ai diff profiles/perf_h100_sp1.sqlite profiles/perf_h100_sp1_fa3.sqlite \
        --format markdown --limit 30 -o diff_fa2_to_fa3.md

Reproduce the plots

cd plots && python3 make_plots.py

Query the SQLite directly

sqlite3 profiles/perf_h100_sp1.sqlite \
  "SELECT s.value, COUNT(*), SUM(end-start)/1e6 AS ms
   FROM CUPTI_ACTIVITY_KIND_KERNEL k JOIN StringIds s ON s.id=k.shortName
   GROUP BY s.value ORDER BY ms DESC LIMIT 10;"

How the profiles were captured

Run config captured (from profiles/perf_wan-t2v-1.3b-h100-sp1_*.json):

benchmark_id          wan-t2v-1.3b-h100-sp1
device                NVIDIA H100 80GB HBM3 SXM5
num_gpus              1   (sp_size=1, tp_size=1)
num_warmup_runs       1
num_measurement_runs  2   → 3 DenoisingStage instances in each profile
num_inference_steps   30
guidance_scale        > 1.0  (CFG enabled → 2 forward passes per step)
  ⇒ 30 steps × 2 CFG forwards × 30 transformer blocks × 2 attn calls/block
    = 10,800 attention kernel invocations per profile (matches measured count)

# Profiling command (representative — exact run config in profiles/perf_*.json)
nsys profile \
    --trace=cuda,nvtx,osrt,cublas,cudnn \
    --capture-range=cudaProfilerApi \
    --cuda-event-trace=true \
    --output=perf_h100_sp1.nsys-rep \
    pytest fastvideo/tests/performance/test_inference_performance.py \
        -k wan-t2v-1.3b-h100-sp1
    # benchmark config: .buildkite/performance-benchmarks/tests/wan-t2v-1.3b-h100-sp1.json
    # (sp_size=1, num_inference_steps=30, guidance_scale>1 → 2× CFG forwards)
    # FA3 run swaps the attention backend env var to flash_attn_3 before pytest

# Export to SQLite for nsys-ai consumption
nsys export --type=sqlite --output=perf_h100_sp1.sqlite perf_h100_sp1.nsys-rep

Hardware

GPU              NVIDIA H100 SXM5
Compute capability  9.0 (Hopper)
SMs                132
Peak FP16/BF16 TC  989 TFLOPS  (vectorised, BF16 inputs · FP32 accumulator)
HBM3 capacity       80 GiB
HBM3 bandwidth      3.35 TB/s
SMEM / SM (max)     228 KiB
Host PCIe           Gen5 ×16

Software

FastVideo                main (2026-05 snapshot)
Wan model               Wan-AI/Wan2.1-T2V-1.3B-Diffusers (HF Hub)
PyTorch                 2.5+ with torch.compile / Inductor
flash-attn-2            via FastVideo's default attention path
flash-attn-3            via FastVideo --attn_backend flash_attn_3
nsys                    NVIDIA Nsight Systems CLI
nsys-ai (analysis tool) https://pypi.org/project/nsys-ai/

Related dataset

rich7421/fastvideo-wan-l40s-nsys — same workload (FastVideo Wan2.1-T2V-1.3B) on 4× L40S with sequence parallelism. On L40S (Ada, no Hopper TMA/wgmma), FA3 cannot be used at all, and the dominant cost is NCCL all-to-all (PCIe Gen4 × no NVLink). The two datasets together cover the hardware sensitivity of the FastVideo inference path.

Citation

If you use this dataset in a paper or blog post, please cite:

@misc{fastvideo-wan-h100-sp1-nsys-2026,
  title  = {FastVideo Wan2.1-T2V-1.3B — H100 SXM SP=1 Nsight Systems profiles
            (FlashAttention-2 vs FlashAttention-3)},
  author = {rich7421},
  year   = {2026},
  url    = {https://huggingface.co/datasets/rich7421/fastvideo-wan-h100-sp1-nsys},
  note   = {Profiles analysed with nsys-ai (https://pypi.org/project/nsys-ai/)}
}

License

The profile data is released under CC-BY-4.0. No model weights, training data, or proprietary information are included — the profiles capture only GPU activity timing, kernel names, NVTX annotations, and memory transfer metadata from inference runs.

Downloads last month: 163

Total file size:

342 MB