kornia Jetson Benchmarking Discipline
Before ANY timing run
sudo nvpmodel -q # must say MAXN_SUPER (mode 2)
sudo nvpmodel -m 2 && sudo jetson_clocks # if not
Numbers taken in the default 15W mode are ~3× slower and not comparable to MAXN numbers. Never mix them in a before/after claim.
Reading PipelineTiming (crates/vrt/src/pipeline.rs)
| Field | Meaning | Use for |
|---|---|---|
source_ms |
Blocking wait for RTSP frame | Detecting camera-bound pipelines (≈ frame interval when GPU is fast) |
enqueue_ms |
CPU kernel-launch time | Should be <1ms; higher = hidden sync/alloc in a stage |
gpu_ms |
CUDA-event hardware time | The authoritative GPU cost. Use this for optimization claims |
sync_ms |
Wall-clock in cudaStreamSynchronize | gpu_ms + CPU wake-up jitter; don't quote as "GPU time" |
finalize_ms |
CPU postproc after sync | top-K / NMS cost |
A camera-paced pipeline shows large source_ms (~33ms at 30fps) — the GPU
is idle waiting. Real GPU headroom = gpu_ms, not 1000/fps.
Method
- Discard the first ~20 frames (TRT warm-up, CUDA context init, cache fill).
- Report averages over ≥100 frames AND peaks; examples already print both every 100 frames.
- For pure-GPU comparisons (engine A vs B), use
/usr/src/tensorrt/bin/trtexec --loadEngine=... --shapes=image:1x3xHxWwhich removes the camera entirely. - Watch thermals during long runs:
tegrastats— a thermally throttled run invalidates the numbers (look for falling GR3D_FREQ).
Reference numbers (MAXN_SUPER, this machine)
- XFeat backbone FP16 @ 640×640: ~3ms GPU
- Decode+VIC path: free of GPU (NVDEC + VIC fixed-function blocks)
- 1080p camera, VIC-resized to 1280×720, padded 1280×736: backbone ~10ms