Speaker
Description
Evaluating the performance of GStreamer-based visual AI pipelines on heterogeneous hardware is a complex systems engineering challenge. Simple metrics like raw TOPS are insufficient for understanding real-world performance, which is heavily impacted by codec complexity, model architecture, and concurrent system resource contention across the CPU, GPU, and NPU.
This presentation will demonstrate how the benchmarking challenges have been directly addressed in the Visual Pipeline and Platform Evaluation Tool (ViPPET), an open, GStreamer-based utility.
Specifically, we will show how developers can use the tool to determine the maximum concurrent stream density for a given hardware.
It will cover four aspects of the tool for flexible initial benchmarking:
•Data Ingestion & Pipeline Customization - using custom media assets, image sets, and live camera feeds to test realistic input scenarios.
•Pre-built Vision AI Workloads - Leveraging optimized, out-of-the-box pipeline templates (e.g., Object Detection, Tracking, and Pose Estimation) as baseline references.
• Performance Analysis - Measuring true throughput (FPS), end-to-end pipeline latency, and correlating these with real-time CPU, GPU, and NPU utilization.
• Enhanced user experience, e.g. advanced view and simplified view on GStreamer pipeline.
Speaker Bio
Jan Gorecki is the Edge AI Product Manager at Intel, where he is responsible for DLStreamer, a GStreamer-based framework for media AI pipelines. Jan actively collaborates with ecosystem partners and enterprise customers to ensure that core Edge AI libraries are optimized and tailored to meet real-world developer needs.
| Duration of the talk |
|---|
