Skip to content

1.0.0 Release notes

About this release

EdgeBench 1.0.0 is the first release of Analog Devices’ cloud-based Edge AI benchmarking platform. This release introduces a browser-accessible workflow that enables users to evaluate, compile, deploy, and benchmark AI/ML models on real ADI embedded hardware hosted in the ADI Remote Board Farm.

What’s new

Platform capabilities

  • Cloud-Based Edge AI Evaluation
    Evaluate and benchmark AI/ML models directly on ADI embedded devices through a web browser—no local toolchains or hardware required.

  • Remote Board Farm Integration
    Secure, 24/7 access to ADI evaluation boards hosted in the Remote Board Farm, enabling hardware-accurate inference results under consistent test conditions.

  • End-to-End Guided Workflow
    A step-by-step build process covering project creation, device selection, compilation, inference, and results analysis.

Model support

ADI Model Zoo Integration

Access a curated collection of ADI-optimized, pre-trained AI models spanning vision, audio, and time-series applications. Models available in EdgeBench are also accessible through Analog Devices' public GitHub repository: ADI Model Zoo

Bring Your Own Model (BYOM)

Upload and evaluate custom pre-trained models using:

  • PyTorch
  • TensorFlow
  • Open Neural Network Exchange (ONNX)

Quantization and Optimization Options

Configure compilation with int8, int16, float32, and select optimization flags for speed, power, or kernel selection.

Supported devices

EdgeBench 1.0.0 supports deployment and inference on the following ADI platforms hosted in the Remote Board Farm:

  • MAX78002–AI accelerator SoC with convolutional neural network (CNN) hardware
  • MAX32690–Arm Cortex‑M4 MCU with Bluetooth LE and RISC‑V coprocessor
  • ADSP-SC835–SHARC‑FX DSP with Arm Cortex‑M33 connectivity core

Data and inference

Input Data Support

  • Vision: PNG, JPEG
  • Audio: WAV, MP3
  • Time-series: CSV

Pre- and Post-Processing Scripts

  • Python-based scripting environment (required for BYOM)
  • Jupyter-like editor with test and validation support

Inference Results and Metrics

  • Hardware-accurate latency
  • Memory usage
  • Power consumption
  • Batch-based inference tracking and history

User experience

Workspace Management

  • Projects, builds, and models organized in a centralized workspace
  • 10 GB personal cloud storage per user

Search, Filter, and Sort

  • Quickly locate projects and models using metadata, framework, task type, and size filters

Light/Dark Theme Support

  • Toggle UI themes using the top-bar controls

Host environment support

EdgeBench 1.0.0 is accessible through modern web browsers on the following platforms:

  • macOS: Safari
  • Windows 11: Chrome, Microsoft Edge
  • Ubuntu 22.04: Firefox

Known limitations

Single active project

Users may work on only one active project at a time.

Model compatibility

Not all user‑supplied models may successfully compile or deploy due to size, architecture, unsupported operations, or framework limitations.

BYOM constraints

Users can upload and deploy custom‑trained models to EdgeBench using the BYOM feature. However, successful compilation and flashing depend on adhering to the following constraints.

Embedded resource limitations
  • Remote boards are memory‑constrained edge devices; large models may fail compilation or deployment.
  • Models should be trained with embedded constraints in mind (for example, smaller input resolutions, lightweight architectures).
  • High‑resolution inputs are not recommended.
Input/output tensor size
  • The target firmware uses nanopb (Protocol Buffers for embedded systems) to transmit model inputs and outputs.
  • Nanopb supports messages up to 64 KB by default.

  • Models with input or output tensors exceeding this limit will fail during firmware compilation—even if the model itself is valid.

  • int8 quantization is strongly recommended to reduce tensor size.

Framework and export requirements

Supported frameworks for BYOM:

  • PyTorch
  • TensorFlow
  • ONNX (opset ≤ 13)

Additional requirements:

  • Models must be exported with complete architecture and trained weights.
  • Checkpoint‑only or partial exports are not supported.
  • ONNX models must include all required operators and parameters; missing or unsupported operators may cause build failures.
Dependency compatibility

Best results are achieved when training and exporting models using the following versions:

  • Python ≥ 3.10
  • TensorFlow ≥ 2.12
  • PyTorch ≥ 2.1.2
  • ONNX ≤ 1.13.0
Metadata requirements
  • Certain tasks require additional metadata files for correct output interpretation (for example, classes.txt, sampling_rate.txt).
  • Missing, malformed, or mismatched metadata may result in build errors or incorrect inference results.

Support and resources

This release is governed by the EdgeBench Terms of Use and applicable Analog Devices licensing agreements.