VisionFlow
NVIDIA worker disconnected

GPU VIDEO INFERENCE LAB · ITERATION 01

See what every frame costs.

Upload a short video, inspect how compressed bytes become frames, and trace those frames through decoding, inference, tracking, analytics, and encoding.

00:00:00
sample-vision-feed.mp41920 × 108030 FPS
GPU
Real inference requires the NVIDIA worker

This browser version can inspect and decode your video. Detection, DeepStream measurements, and TensorRT results begin in Iteration 2 when a Linux NVIDIA worker is connected.

SIMULATION MODE

VIDEO OUTPUT

General object detection

Sample simulation
SAMPLE · SIMULATED DETECTION
FRAME 000001
LIVE · 00:00:00
00:00:08.184FRAME3 objects · 6.2 KB metadata

MEDIA INSPECTOR

From compressed bytes to pixels

Sample payload · 18.4 MB
ContainerMP4
Video codecH.264*
Compressed input4.9 Mb/s
Decoded frame5.93 MB
FILE HEADER · HEXOFFSET 0x00000000
00 00 00 20 66 74 79 70 69 73 6F 6D 00 00 02 00
69 73 6F 6D 69 73 6F 32 61 76 63 31 6D 70 34 31
SELECTED BYTES · BINARY8 BITS / BYTE
00000000 00000000 00000000 00100000
01100110 01110100 01111001 01110000
01Read bytesCompressed file data
02Demux containerSeparate video and audio
03Parse codecPackets and keyframes
04DecodeNVDEC or CPU
05Raw frameNV12 / RGBA pixels

* Browser APIs expose the container and dimensions reliably. Exact codec identification is estimated until the GStreamer media probe is connected.

FRAME TRACE

One frame, end to end

Select a stage to learn what runs there and why.

OPTIMIZATION LAB

Change the workload

WHY THIS CHANGED

Balanced for real-time video

92

FP16 cuts inference memory and compute while preserving strong detection quality. Hardware decode and zero-copy keep frame data on the GPU, avoiding a costly round trip through system memory.

MEASURE, THEN OPTIMIZE

Baseline vs. current configuration

Projected values use a calibrated simulation until the NVIDIA worker is connected.

LEARNING GUIDE

How to use this lab

  1. Observe the baseline. Watch latency, utilization, and dropped frames with one stream.
  2. Trace a frame. Select pipeline stages to separate CPU orchestration from GPU-intensive work.
  3. Create pressure. Increase streams, resolution, or model precision and identify the first bottleneck.
  4. Optimize deliberately. Change one control at a time and read why the result moved.

Iteration 2 is the core integration: it connects these controls to a Linux NVIDIA worker running GStreamer, DeepStream, and TensorRT. Until then, uploaded videos are decoded locally and all performance numbers are clearly marked as projections.