Skip to content

About EdgeBench

EdgeBench is a cloud-based platform designed to streamline the evaluation and benchmarking of AI/ML models on Analog Devices (ADI) embedded hardware. It eliminates the complexities traditionally associated with embedded AI development such as hardware setup, tools setup, and model deployment by offering a browser-accessible interface that enables remote compilation, deployment, and performance analysis.

At the heart of the platform is ADI's Remote Board Farm, a scalable and resilient infrastructure that provides secure access to a variety of ADI evaluation boards. With consistent lab conditions and 24/7 secure access, teams reduce evaluation delays and derisk Edge AI development.

EdgeBench allows users to run inference on actual hardware, gaining real-world insights into model performance. Users get hardware‑accurate latency, memory, and behavior metrics, enabling confident device selection and faster iteration.

Key features

  • Remote Hardware Access: Test models from anywhere—no manual setup or shipping required.

  • Zero‑setup, zero‑touch workflow: No firmware flashing, SDK installation, or physical bring‑up required.

  • ADI Model Zoo Integration: Access a curated library of pre-trained models or upload your own AI/ML models for testing.

  • Bring Your Own Model: Seamlessly upload supported models for remote evaluation and benchmarking.

  • Browser-Based Workflow: Compile, deploy, and test models directly through a web interface—no local installation needed.

  • Real Hardware, Real Results: No simulations—observe actual latency, memory usage, and inference behavior.

Use cases

  • AI Workload Validation: Assess whether ADI solutions meet your specific AI performance needs.

  • Model Benchmarking: Compare multiple models under consistent, real-world conditions.

  • Early Stage Exploration: Begin development and testing without the need to ship physical boards.

Supported devices

EdgeBench enables seamless model deployment across a variety of ADI digital devices. The ADI Remote Board Farm hosts and integrates evaluation boards with the client application.

Device specifications

Device Name Description Key Specifications
MAX78002 The MAX78002 is an advanced system-on-chip featuring an Arm Cortex-M4 with FPU CPU for efficient system control and with hardware-based convolutional neural network (CNN) accelerator optimized for deep learning tasks. • Arm Cortex-M4 processor (up to 120 MHz) and a RISC-V coprocessor (up to 60 MHz)

• 2.5 MB Flash, 384 KB SRAM, and 1.3 MB CNN data memory

• Support for 1-, 2-, 4-, and 8-bit weights with up to 2 MB weight storage

• Interfaces for camera (MIPI CSI-2, 12-bit parallel), audio (I²S), and SD/eMMC storage

• Built-in CNN Accelerator
MAX32690 The MAX32690 features an Arm Cortex-M4 with a floating-point unit (FPU) Microcontroller and Bluetooth LE 5, suitable for industrial sensors and health/medical wearables applications. • 120 MHz Arm Cortex-M4 Processor with FPU

• Ultra-Low-Power, 32-Bit RISC-V (RV32) Coprocessor

• 3.25 MB Flash and 1 MB SRAM, with external memory expansion support

• Integrated Bluetooth 5.2 Low Energy (LE) radio
ADSP-SC835 The ADSP-SC835 is a high-performance DSP, part of the SHARC-FX family, designed for demanding signal processing applications such as professional audio, automotive, and industrial systems. • SHARC-FX DSP core running up to 1 GHz

• Integrated Arm Cortex-M33 connectivity core at 400 MHz

• 256-bit SIMD vector architecture

• Up to 2 MB of on-chip L2 memory

• Built-in FIR and IIR accelerators

Supported AI models

EdgeBench supports two primary model sources:

  1. ADI Model Zoo: A curated repository of pre-trained models developed by Analog Devices, optimized for supported evaluation boards.

  2. Bring Your Own Model (BYOM): Upload and deploy custom AI pre-trained models using PyTorch, TensorFlow, and ONNX.

  • MacOS Sequoia 15.6 – Safari Version 18.6
  • Windows 11 – Chrome Version 139.0.7258.127 (64-bit)
  • Windows 11 – Edge Version 139.0.3405.86 (64-bit)
  • Ubuntu 22.04 – Mozilla Firefox Version 143.0

Limitations

  • Single Active Project: Users can work on only one active project at a time
  • Model Compatibility: Not all model architectures may be compatible with target hardware