Edge Workloads and Benchmarks are performance-optimized pipelines that utilizes the GStreamer multimedia framework and the Deep Learning Streamer (DL Streamer) for validating media and edge AI video analytics. The pipelines measure end-to-end throughput in frames per second (fps), pipeline stream density in fps, package power in watts, and workload efficiency in fps per package watt.
HEVC 1080p Video Decode (GPU Hardware-Accelerated) → Object Detection (GPU or NPU) → Object Tracking → 1-2x Object Classification (GPU or NPU)
| Config | Video | Detection | Classification |
|---|---|---|---|
| light | bears.h265 (2 obj/frame) | YOLOv11n (640x640) INT8 | ResNet‑50 (224x224) INT8 |
| medium | apple.h265 (1 obj/frame) | YOLOv5m (640x640) INT8 | ResNet‑50 + MobileNet‑V2 (224x224) INT8 |
| heavy | bears.h265 (2 obj/frame) | YOLOv11m (640x640) INT8 | ResNet‑50 + MobileNet‑V2 (224x224) INT8 |
Pipeline configurations include single-device pipelines (GPU or NPU only), pipelines with multiple devices (GPU Detect and NPU Classify), and multiple single-device pipelines running concurrently in separate processes (GPU only and NPU only, concurrently).
A GPU with video acceleration API (VA-API) media support that:
- Is validated on Ubuntu OS version 24.04.3 LTS with kernel version 6.16 and above
- Has Docker software installed and the user added in the docker group
- Has Integrated GPU
- Has NPU (optional)
- Docker software version 20.10 and above (installation guide)
- Python programming language version 3.8 and above with virtual environment (venv) support
- Network connectivity for model or media download
Breakdown of space required:
- Models: 230 MB
- COCO dataset: 950 MB
- CIFAR-100 dataset: 162 MB
- Videos: 1.9 GB
- Virtual Environment: 7.7 GB
Total space required: 10.9 GB
Optional ImageNet Dataset Download: 6.5GB. You will need to download this manually. See model-conversion.
Run the following commands to allow X server connection in the Docker container, so that the display pipeline sample can access the host's display:
xhost local:root
setfacl -m user:1000:r ~/.XauthorityThe Makefile automates the entire workflow. Use make help to display the following:
# Three-step setup
make prereqs # Install dependencies (params: INCLUDE_GPU=True INCLUDE_NPU=True).
make models # Download and convert models (params: IMAGENET_ROOT, optional)
make media # Download and transcode video files
# Run benchmarks
make benchmarks # Run all pipeline configurations (params: CORES={cores-to-pin-workload} DURATION={seconds})
sudo make benchmarks # Recommended: Adds power and efficiency metrics to report. Requires root permissions to read power sensors
# Generate results
make html-report # Generate HTML dashboard from benchmark results. Requires serve-report to view locally.
make serve-report # Host dashboard locally (params: PORT, default host: http://localhost:8000)
# Optional: display pipeline demo (requires display access permissions)
make display # Visualized pipeline demo (params: CONFIG={light,medium,heavy} DETECT={CPU,GPU,NPU} CLASSIFY={CPU,GPU,NPU} DURATION={seconds})
# Cleanup
make clean # Remove all results
make clean-all # Remove all generated collateral (models, media, results, and venv)- Prereqs:
make prereqs INCLUDE_GPU=True INCLUDE_NPU=True - Display:
make display CONFIG=light DETECT=GPU CLASSIFY=NPU DURATION=60 - Benchmarks:
sudo make benchmarks CORES=ecore DURATION=60
IMAGENET_ROOT- Path to pre-downloaded the ImageNet dataset for accuracy validation on ResNet architecture and MobileNet networks (see model-conversion)INCLUDE_GPU- Install GPU drivers during setupINCLUDE_NPU- Install NPU drivers during setup (requires reboot)DURATION- Duration to run pipeline in secondsCONFIG=light|medium|heavy- Pipeline configuration, tiered by compute complexityDETECT/CLASSIFY=CPU|GPU|NPU- Inference device assignmentCORES=pcore|ecore|lpecore- CPU core pinning based on core typePORT- HTTP server port for dashboard (default: 8000)
If you prefer step-by-step control:
Step 1. Prerequisites:
cd setup/
./install_prerequisites.sh
# Optional: --reinstall-gpu-driver=yes and/or --reinstall-npu-driver=yesStep 2. Models:
cd ../model-conversion/
./convert_models.sh
# Optional: -i "$HOME/datasets/imagenet-packages" for ImageNet quantizationStep 3. Media:
cd ../media-downloader/
./download_and_encode.shStep 4. Run benchmark:
cd ..
./benchmark_edge_pipelines.sh \
-p <light|medium|heavy> \
-n <num_streams> \
-b <batch_size> \
-d <DetectDevice> \
-c <ClassifyDevice> \
-i <duration_sec> \
-t <scheduling_core_type>
# Example command line
./benchmark_edge_pipelines.sh -p light -n 8 -b 8 -d GPU -c NPU -i 120 -t ecoreParameters:
-pPipeline config:light|medium|heavy(required)-nNumber of parallel streams (default: 1)-bBatch size (default: 1)-dDetection device:CPU|GPU|GPU.<idx>|NPU(default: CPU)-cClassification device:CPU|GPU|GPU.<idx>|NPU(default: CPU)-iDuration in seconds (default: 120)-tCPU core type for pinning, e.g.,"ecore"(optional)--concurrentEnable concurrent GPU or NPU execution mode (optional)
Note: Intel recommends the GPU or NPU for AI inference workloads.
Step 5. Display Results:
# Generate and view dashboard
python3 html/generate_report.py
cd html && python3 -m http.server 8000 # Access at http://localhost:8000Results are saved to the results/ folder, organized by execution mode:
*.log– Full GStreamer pipeline output (stdout or stderr)*.csv– Performance metrics (FPS, stream density, and configuration)
If you want to participate in the GitHub community for Edge Workloads and Benchmarks, you can contribute code, propose a design, download and try out a release, open an issue, benchmark application performance, and participate in Discussions.
To learn more, check out the following resources:
Before submitting a new report, check the existing issues to see if a similar one has been filed.
The Edge Workload and Benchmarks project is licensed under the APACHE 2.0 license.
