Overview

General Description

Artificial intelligence (AI) requires extreme computational horsepower, but Analog Devices is cutting the power cord from AI insights. The MAX78002 is a new breed of AI microcontroller that enables neural networks to execute at ultra-low power and live at the edge of the IoT. This device combines the most energy-efficient AI processing with Analog Devices' proven ultra-low-power microcontrollers. Our hardware-based CNN accelerator enables battery-powered applications to execute AI inferences while expending only millijoules of energy.

The MAX78002 is an advanced system-on-chip featuring an Arm® Cortex®-M4 with FPU CPU for efficient system control with an ultra-low-power deep neural-network accelerator. The CNN engine has a weight storage memory of 2MB, and can support 1-, 2-, 4-, and 8-bit weights (supporting networks of up to 16 million weights). The CNN weight memory is SRAM-based so that AI network updates can be made on the fly. The CNN engine also has 1.3MB of data memory. The CNN architecture is highly flexible, allowing networks to be trained in conventional toolsets like PyTorch® and TensorFlow®, then converted for execution on the MAX78002 using tools provided by Analog Devices.

In addition to the memory in the CNN engine, the MAX78002 has large on-chip system memory for the microcontroller core with 2.5MB flash and up to 384KB SRAM. Multiple high-speed and low-power communications interfaces are supported, including I2S, MIPI® CSI-2® serial camera, parallel camera (PCIF), and SD 3.0/SDIO 3.0/eMMC 4.51 secure digital.

The device is available in a 144 CSBGA, 12mm x 12mm, 0.8mm pitch package.
Simplified Block Diagram
Benefits and Features
  • Dual-Core, Low-Power Microcontroller
    • Arm Cortex-M4 Processor with FPU up to 120MHz
    • 2.5MB Flash, 64KB ROM, and 384KB SRAM
    • Optimized Performance with 16KB Instruction Cache
    • Optional Error Correction Code (ECC SEC-DED) for SRAM
    • 32-Bit RISC-V Coprocessor up to 60MHz
    • Up to 60 General-Purpose I/O Pins
    • MIPI Camera Serial Interface 2 (MIPI CSI-2) Controller V2.1 – Support for Two Data Lanes
    • 12-Bit Parallel Camera Interface
    • I2S Controller/Target for Digital Audio Interface
    • Secure Digital Interface Supports SD 3.0/SDIO 3.0/eMMC 4.51
  • Convolutional Neural Network (CNN) Accelerator
    • Highly Optimized for Deep CNNs
    • ​​2 Million 8-Bit Weight Capacity with 1-, 2-, 4-, and 8-bit Weights
    • 1.3MB CNN Data Memory
    • Programmable Input Image Size up to 2048 x 2048 Pixels
    • Programmable Network Depth up to 128 Layers
    • Programmable per Layer Network Channel Widths up to 1024 Channels
    • 1- and 2-Dimensional Convolution Processing
    • Capable of Processing VGA Images at 30fps
  • Power Management for Extending Battery Life
    • Integrated Single-Inductor Multiple-Output (SIMO) Switch-Mode Power Supply (SMPS)
    • 2.85V to 3.6V Supply Voltage Range
    • Support of Optional External Auxiliary CNN Power Supply
    • Dynamic Voltage Scaling Minimizes Active Core Power Consumption
    • 23.9μA/MHz While Loop Execution at 3.3V from Cache (CM4 only)
    • Selectable SRAM Retention in Low-Power Modes with Real-Time Clock (RTC) Enabled
  • Security and Integrity
    • Available Secure Boot
    • AES 128/192/256 Hardware Acceleration Engine
    • True Random Number Generator (TRNG) Seed Generator
Applications
  • Factory Robot and Drone Navigation
  • Industrial Sensors and Process Control
  • Inline Quality Assurance Vision Systems
  • Smart Security Cameras
  • Portable Medical Diagnostics Equipment

Arm and Cortex are registered trademarks of Arm Limited.

Coremark is a registered trademark of EEMBC.

CSI-2 and MIPI are registered trademarks of MIPI Alliance, Inc.

PyTorch is a registered trademark of Facebook, Inc.

TensorFlow is a registered trademark of Google, Inc.


19-101571; Rev 0; 6/22