Skip to main content

Performance and Metrics on Intel Devices

Benchmarks (Tiber AI PC)​

The following images show the model pipeline and component breakdowns used during benchmarking on the Tiber AI PC (Intel® Core™ Ultra 7 258V with Intel® Arc™ 140V GPU).

Pipeline

Hand Detector

Hand Landmarks

Gesture Embedder

Gesture Classifier

Test environment​

  • Machine: Tiber AI PC
  • CPU: Intel® Core™ Ultra 7 258V
  • GPU: Intel® Arc™ 140V
  • OS: Windows (PowerShell available)
  • OpenVINO: Tested with OpenVINO Toolkit (>= 2023.1 recommended)

How the models were tested​

We used the OpenVINO Benchmark Tool to measure throughput (FPS), latency (ms), and memory usage per model component. The general approach:

  1. Convert/prepare each model into OpenVINO IR or OpenVINO's format (if not already converted).
  2. Run the benchmark_app tool with the desired device (CPU/GPU) and performance counters enabled.
  3. Record average latency and throughput, and compare CPU vs GPU runs.

Expected metrics and interpretation​

  • Latency: Lower is better; report median and 90th percentile when possible.
  • Throughput (FPS): Higher is better; use this to determine real-time capability (e.g., 30 FPS target).
  • Device comparison: GPU runs should show lower latency or higher throughput for heavy models (landmarks, embedder) while the CPU may suffice for smaller models (classifier).