Benchmarking

ViPPET provides two benchmarking modes for evaluating AI inference pipeline performance on Intel® hardware:

• Performance Testing — Run one or more pipelines with a fixed number of streams and measure throughput (FPS), system utilization, and optionally latency in real time.

• Density Testing — Automatically find the maximum number of concurrent streams that maintain a target FPS floor, using an exponential-growth and binary-search algorithm. Density testing supports two modes:

  • Classic mode — All pipelines share a single search variable (total stream count) distributed according to per-pipeline stream_rate ratios.

  • Mixed mode — Exactly two pipelines: one is pinned to a fixed stream count, the other is incremented by the same algorithm. See Stream Density Testing for details.

Both modes share a common execution configuration and report results through the same job management system.

For detailed instructions, see:

• Performance Testing

• Stream Density Testing

• Managing Jobs

Key concepts

Concept	Description
Total FPS	Aggregate frames per second across all active streams
Per Stream FPS	Average FPS per individual stream (Total FPS ÷ stream count)
FPS Floor	Minimum acceptable per-stream FPS used as the pass/fail threshold in density testing
Stream Rate	Percentage of total streams allocated to each pipeline in classic density mode (must sum to 100%)
Fixed Streams	Pinned stream count for one pipeline in mixed density mode; the other pipeline is incremented by the search
Output Mode	Controls whether output video is saved to file, streamed live, or disabled
Latency Metrics	Optional end-to-end pipeline latency measurement (avg/min/max) reported per interval

Workflow overview

1. Configure — Select pipelines, set stream counts or density parameters, and choose output options.

2. Run — Start the test; ViPPET creates a job and begins executing pipelines as subprocesses.

3. Monitor — Real-time system metrics (CPU, GPU, NPU, memory, power) and pipeline metrics (FPS) are displayed in the dashboard while the job runs.

4. Review results — When the job completes, view final FPS and output videos in the job detail view.