Introduction to Edgematic

Edgematic is a cutting-edge UI-based development platform design for Edge AI applications. It empowers developers to streamline the creation and deployment of Edge AI applications with minimal coding requirements. Built on Palette, Edgematic provides a low/no-code environment to construct and evaluate optimized pipelines. It uses GStreamer alongside SiMa’s engine for model quantization, compilation, and deployment, making it an essential tool for accelerating Edge AI workflows.

This document is a comprehensive resource for understanding, exploring, and maximizing the potential of Edgematic, a cutting-edge UI-based development platform designed for Edge AI applications. Whether you are a seasoned developer or new to Edge AI, this guide will help you navigate through the platform’s capabilities, ensuring a seamless experience as you build, test, and deploy advanced machine learning applications.

Typical users include:

AI developers looking to optimize their models for edge deployment.
Data scientists seeking a user-friendly platform for model evaluation and quantization.
System integrators aiming to create scalable, efficient AI solutions.
Anyone interested in exploring the intersection of AI, edge devices, and low-latency applications.

Key Features

Edgematic offers a range of features to simplify your Edge AI development process:

Pre-Created Pipelines: Test and evaluate pipelines built for real-world use cases.
Model Import and Evaluation: Bring your own models to obtain performance metrics like FPS and FPS/W.
Seamless Integration: Drag and drop compiled models from Palette into Edgematic for automatic pre/post-processing optimization.
Future Enhancements: Look forward to features like application development, local deployment, online model quantization and compilation, and MLSoC device management.

Benefits

Edgematic saves you time and effort by automating complex processes, providing a streamlined environment for deploying highly optimized Edge AI applications. By leveraging its capabilities, you can focus on innovation and efficiency without worrying about the intricacies of edge hardware configurations or low-level software development.

Getting Started

In the sections ahead, you will find step-by-step instructions, examples, and best practices to help you get started with Edgematic. Whether you want to evaluate existing pipelines, deploy a custom model, or build a comprehensive application, this guide will provide the insights you need to succeed.

Let us begin your journey to transforming Edge AI development with Edgematic!

Layout

The Project menu allows user to do the following:

New / Open Recent - Create a new project or open a recently used project.
Download - Download the project. You can use the downloaded project to deploy the project/pipeline into a local device.
Builds - You can choose to re-deploy one of the builds that you built in the past.

Displays the file structure of the project, organizing files, folders, and dependencies for easy navigation.

Key Elements

Project Files:
Contains project definitions, dependency information, and configuration files.
Plugins:
Includes available modules such as capsfilter, h264parse, and nmsyolov5 that can be used within your pipeline.
Resources:
Stores additional assets, including application.json.

Users can utilize the search bar to quickly locate files within the project structure.

The Canvas is the visual interface for building and editing application pipelines.

Features

Pipeline Blocks:
Drag and drop blocks such as queue2, overlay_v2, and udp_sink to configure workflows.
Connections:
Link blocks to define the data flow between modules.
Zoom and Pan Tools:
Adjust the view for detailed editing or a comprehensive overview.

The canvas provides a graphical representation of the processing pipeline, enabling intuitive design and configuration.

The Toolbar offers quick access to application controls and debugging tools.

Options Include

Board Status Indicator: The circle on the far left, next to the board name, indicates whether the board is powered on or off.

Connection Type: Displays how the board is connected to the server: Eth for Ethernet/networking connections or PCIe for PCI Express connections.

Board Name: Shows the device name, e.g., stage-board01.

Power Metrics: The lightning bolt icon indicates whether the board is connected to a power measurement station. This is used when users request model KPIs.

Allocation Clock: The green circle with a clock icon shows the remaining session time. Each user session is limited to one hour.

Play Button: Deploys the current application on the canvas to the specified device.

Stop Button: Stops the application currently running on the specified device.

Profile Button: Displays profile information and provides an option for users to sign out.

The Catalog offers quick access to applications, models and plugins validated and optimized by SiMa’s engineers for generic purposes. You can drag-and-drop these resources to start building your own projects.

Categories:

Applications: Pre-configured applications like yolo_v5_ethernet and people_tracking. The application catalog is divided into:

Developer Community
Within Developer Community you can find the same applications that you can find in the Developer Community Tab on the landing page.

SiMa
Within SiMa you can find the same Applications that you can find in the Demo Tab on the landing page.
Models: The model catalog has all the compiled models from our internal Model Zoo. These models offer the maximum FPS and Accuracy that the SiMa team was able to achieve. All the models have been trained in generic Dataset like COCO and Imagenet.
Plugins: The plugin catalog offers both sima specific plugins like how to do inference on our devices or pre and post processing. We also offer open source plugins like rtspsrc or udpsink.

The visualization toggle provides several key functions, allowing users to access application and model KPIs, system and device logs, as well as remote device terminal access.

Application KPI

The Application KPI section displays multiple performance metrics. The center of the visualization window shows plugin latency across different processing units, where each color represents a specific processor:

Light blue: CVU (DSP EV74)
Purple: APU (CPU ARM A65)
Yellow: MLA (SiMa’a Machine Learning Accelerator)

Displayed metrics include:

Full Pipeline Throughput: Current pipeline performance in frames per second (FPS), typically limited by the input speed.
MLA Throughput: Maximum throughput achieved by the Machine Learning Accelerator for the pipeline.
APU Throughput: Maximum throughput achieved by the CPU ARM A65 for the pipeline.
CVU Throughput: Maximum throughput achieved by the DSP EV74 for the pipeline.

Model KPI

The Model KPI section provides detailed model performance information. The central panel displays model latency in microseconds (µs), while the right panel presents the following metrics:

Model Throughput: Current model performance in frames per second.
Efficiency: Model throughput divided by the Model Execution Power.
Model Execution Power: Calculated as the Board Running Power minus the Board Idle Power.
Board Running Power: Average total board power (in watts) while running the model over a 30-second interval.
Board Idle Power: Average total board power (in watts) before running the model.

Logs

The Logs section provides two types of logs: SYSTEM and DEVICE.

SYSTEM: Logs generated by the Edgematic SYSTEM, including compilation logs during application creation. These logs are useful for identifying code or configuration errors within different plugins.
DEVICE: Logs generated by the device after pipeline deployment. These include runtime errors, GStreamer deployment errors, and any log messages generated within custom plugin code.

Terminal

The Terminal section allows users to execute remote commands on the device.

Currently, the supported commands include:

ps
top

These commands enable users to monitor running processes. Additional commands may be supported upon request.