Skip to content

Key Concepts

Welcome to the EdgeBench Application Overview. This section is designed to help users understand the core concepts and features of the EdgeBench web application. It provides contextual information to support efficient navigation and utilization of the platform.

User

The User is the central actor in the EdgeBench, responsible for initiating and managing the entire workflow. Users create or import projects, upload models and input data, configure builds, and trigger deployment & inference requests.

Workspace

The Workspace functions as a central hub and dedicated personal storage area within the platform, allowing users to manage their AI/ML projects and models. Each account is provided with 10 GB of cloud storage, which can be used to store model files, project configurations, project builds, and related resources.

For procedures on how to access or add projects and models, go to Navigating the Workspace section.

Project

A Project is the foundational unit within the EdgeBench application, acting as a structured container that organizes and encapsulates both Model and Build. It defines the scope and context of the build, enabling streamlined management and deployment of an AI model.

Each Project includes key attributes such as:

  • Name – A unique identifier for the project.
  • Framework – Specifies the underlying model training framework (TensorFlow, PyTorch, ONNX).
  • Task Type – Indicates the nature of the AI task (e.g., image classification, object detection).

Supported Upload Formats:

  • ZIP
  • TAR.GZ

Model

A Model is the core element used in builds and inference. It includes the following key attributes:

  • Framework – The AI framework used
  • Name – Model identifier
  • Task Type – Defines the model’s purpose

Models can be selected from the model zoo or uploaded via BYOM.

Model zoo

The model zoo is a curated catalog of pre-trained AI models optimized for edge deployment on ADI hardware. Covering domains such as audio, sensor, and vision, these models feature compact, efficient architecture designed for real-time, low-power applications. Each entry includes pre-trained files, sample scripts, benchmark data, and quantized versions, enabling streamlined evaluation and integration into edge AI deployment workflows.

To learn how to access and use these models, see the Exploring the ADI Model Zoo section.

BYOM

The Bring Your Own Model (BYOM) feature allows users to upload and deploy their own pre-trained AI models. Supported frameworks and formats include:

Framework Required Files Remarks
PyTorch - Single Model File: .pt or .pth (complete model with architecture)
- input_shape.txt
- (Optional) model.py (model definition)		 Input shape file must be in one of the formats:
tuple:(1,3,224,224),
array: [1,3,224,224], or
csv: 1,3,224,224
TensorFlow .h5, .pb, or SAVED_MODEL Complete model with architecture
ONNX .onnx file Complete model with architecture

See Workspace Tasks > Uploading a Model (BYOM) for complete details on Supported Formats.

Models uploaded through BYOM are private and accessible only to the individual who uploaded them. They are not visible or available to other users on the platform. For publicly accessible models, users are encouraged to explore ADI's Model Zoo.

Disclaimer: Please note that not all BYOM models may successfully run on the target device. Models that are too large or contain unsupported operations may fail during compilation or during runtime.

Device

EdgeBench supports seamless model deployment across a range of remote devices. ADI evaluation boards such as MAX78002EVKIT, MAX32690EVKIT, and ADSPSC835W-EV-SOM are currently supported. These boards are all hosted within the ADI Remote Board Farm — a secure, web-connected infrastructure that enables real-time access to physical hardware through a browser-based interface. This setup eliminates the need for manual configuration or on-site devices, streamlining development and accelerating validation.

EdgeBench users benefit from 24/7 access to shared hardware via the Remote Board Farm, enabling continuous development and testing without the need for physical board shipments or setup.

Build

A Build represents a deployable or executable unit within a project. It is generated by compiling an AI model for a specific target device. A build offers several configuration options, allowing a user to tweak and experiment. Options include:

  • Device settings
  • Compilation optimization flags
  • Quantization
  • Kernel options

Build version

A Build Version is a uniquely identifiable snapshot of a build that captures the AI model, the target device, configuration, and compilation settings at the time of an inference run. Each build version includes metadata such as:

  • Build Status - Represents the current progress of a build. It indicates whether the build has been completed successfully, encountered an error, or is still ongoing. This status may also show if the build is still in an intermediate step, such as compilation.
  • Last Updated Timestamp - Shows the most recent time the build was modified or executed.
  • File Size - Displays the total size consumed by the build.
  • Compilation Settings - Lists the parameters used during model compilation.

Data input

Data input refers to the data used for inference. Supported formats include image, audio, and time-series files.

Data processing script

Data Processing Scripts are Python codes used to transform raw data into tensor data, and vice versa. If using the BYOM workflow, the user is required to provide these scripts into the platform. For ADI Models, this is already handled by EdgeBench.

The platform expects two kinds of data processing scripts:

  • Pre-Processing Script – Transforms the end-user's uploaded data (image, audio, csv) into tensor data that aligns with the model's expected input tensor shape (required for BYOM).
  • Post-Processing Script – Transforms tensor data from the model output back into raw data format (image, audio, csv) that can be easily consumed by the end-user (required for BYOM).

Result

Results are the outputs generated from inference runs, and are presented with the following details:

  • Timestamp – Datetime when the inference was executed
  • Batch – Input dataset that was sent as group for inference execution
  • Hardware Performance Metrics
  • Inference latency
  • Memory usage
  • Power consumption

Inference history

The Inference History logs the outcomes of model evaluations, enabling users to monitor and analyze performance over time.

It includes:

  • Build Version – Identifies which compiled model was evaluated
  • Inference Run Status – Indicates success, failure, or in-progress state
  • Timestamp – When the inference was last executed or updated
  • Device Target – Specifies the ADI hardware used for the run
  • Compilation Settings – Parameters used during model build and deployment