SDK Version 1.5

SDK 1.5.0 brings a wide range of enhancements to quantization, pipeline efficiency, GStreamer integration, and developer experience. This release introduces new post-training quantization features, automatic mixed-precision support, and updates to our machine learning acceleration framework. Additionally, pipeline improvements have been made for better performance, flexibility, and out-of-the-box usability.

New Features & Enhancements

Category	Availability	Feature Name	Description
Quantization	GA	Post-Training Quantization (PTQ) - Channel Equalization	Improves quantization accuracy by balancing per-channel distributions. Validated on Densenet-121.
Quantization	Alpha	Automatic Mixed-Precision Quantization	Optimizes performance using INT8 and FP16. Uses binary search for best trade-off. Currently CPU-only, GPU support planned.
Quantization	Beta	Quantization-Aware Training (QAT)	Provides high accuracy for Yolo models (Yolo v8). Requires additional tuning for transformers.
MLA Acceleration	GA	Extended MLA Operator Support	Added support for Argmax and Resize. Bilinear interpolation supports up to 63x scaling. Enhanced non-vectorized Argmax.
MLA Acceleration	GA	MLA Runtime Resiliency & Reliability	Improved execution consistency and error handling.
GStreamer	GA	PyGst Plugin for GStreamer	New plugin for integrating Python post-processing in GStreamer. Enables seamless Python-to-GStreamer transition. YoloV8 pipeline included.
GStreamer	GA	Out-of-the-Box Experience Enhancements	Redesigned services using ports 8001-8004 for compatibility. PCIe-based pipelines moved to gst-App-Code. New GStreamer workflow tools.
Pipeline Optimization	GA	Improved PeopleTracker Pipeline Efficiency	Integrated Centernet + ReID + Tracker updates for improved tracking.
Pipeline Optimization	Beta	Pipeline Accuracy Measurement Framework	Supports PeopleTracker, YoloV8n, Baxter pipeline, YoloV7, DETR, and EffDet. Pending support for Yolov5, PeopleDetector, and OpenPose.
Pipeline Optimization	GA	Ethernet and PCIe Enhancements	Simplified Ethernet pipeline execution via Dispatcher + Config Manager. gstBufferPool integration for efficiency. Ethernet over PCIe fully supported.
SDK & CLI	GA	SDK & CLI Enhancements	Install script now includes dependency checks. Improved GStreamer plugin documentation. Standardized naming conventions. New runtime profiler and tracer.
SoC Software	GA	SoC Software Updates	Added H.265 encoding support for TVS. Enhanced file transfer over PCIe. Common C++ APIs for EV74 and MLA. Memory leak detection via Valgrind.
Deployment	GA	MPK Commands & Deployment Enhancements	MPK commands now support Ethernet & PCIe. New CLI tools for app deployment.Virtual Ethernet support in PCIe mode.

Performance Validations

Type	Model Name	FPS	Explanation
PCIe	MaskRCNN	20	Achieved 20 FPS performance under PCIe deployment.
PCIe	OpenPose	45	Performance measured at 45 FPS on PCIe.
PCIe	YOLOv5	20	Running at 20 FPS over PCIe.
PCIe	YOLOv7	20	Running at 20 FPS under PCIe deployment.
Standalone	YoloV8	80 (20*4)	Runs at 20 FPS per camera, tested with 4 cameras.
Standalone	DETR	30	Achieved 30 FPS over Ethernet.
Standalone	OpenPose	50	Optimized to run at 50 FPS over Ethernet.
Standalone	YOLOv5	20	Achieved 20 FPS performance over Ethernet.
Standalone	YOLOv7	20	Achieved 20 FPS over Ethernet.
Standalone	YOLOv7 (4-Camera)	12 (per cam)	Each of the 4 cameras operates at 12 FPS.
Standalone	PeopleTracker	30	Achieved 30 FPS under Ethernet.
Standalone	EfficientDet	20	Running at 20 FPS over Ethernet.
Standalone	PeopleDetector (4-Camera)	20 (per cam)	Each camera runs at 20 FPS under Ethernet.

Bug Fixes

• Improved system stability for long-duration ML workloads.

• Optimized quantization evaluation for faster execution.

Known Issues

• Some pipeline accuracy measurement tools are still in progress.

• Certain Ethernet pipelines experience occasional memory issues.

• YoloV8_4cams pipeline exhibits instability during long-term testing.