Analog Devices' AI MCU features low-power convolutional Neural network (CNN) accelerators

Analog Devices' MAX78002 is an AI MCU that enables neural networks to execute at ultra-low power, at the edge of the Internet of Things. The device combines the most energy-efficient AI processing with Analog Devices' proven ultra-low power MCU. Hardware-based CNN accelerators enable battery-powered applications to perform AI reasoning while consuming only nanojoules of energy.

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

The MAX78002 Evaluation Kit (MAX78002EVKIT#) provides the platform and tools to leverage device capabilities to build next-generation AI products in a 144 CSBGA, 12 mm x 12 mm, 0.8mm pitch package. In addition to the memory for the CNN engine, the MAX78002 has large system-on-chip memory for the microcontroller kernel, with 2.5 MB of flash memory and up to 384 KB of SRAM. Supports a variety of high-speed and low-power communication interfaces, including I2S, MIPI CSI-2® serial cameras, parallel cameras (PCIF), and SD 3.0/SDIO 3.0/eMMC 4.51 SD cards.

Features

● Dual-core, low-power MCU
-   Arm Cortex-M4 processor with up to 120 MHz FPU
-   2.5 MB flash memory, 64 KB ROM, and 384 KB SRAM
-   Optimize performance with 16 KB instruction cache
-   Optional error correction codes for SRAM (ECC SEC-DED)
-   32-bit RISC-V coprocessor of up to 60 MHz
-   Up to 60 universal I/O pins
-   MIPI Camera Serial Interface 2 (MIPI CSI-2) Controller V2.1 - Supports two data channels
-   12-bit parallel camera interface
-   I2S controller for digital audio interface
-   The secure digital interface supports SD 3.0, SDIO 3.0, and eMMC 4.51

● Convolutional Neural Network (CNN) accelerator
-   Highly optimized for deep CNN
-   2 million 8-bit weight capacity, with 1, 2, 4 and 8-bit weights
-   1.3 MB CNN data storage
-   Programmable input image size Max 2048 pixels x 2048 pixels
-   The programmable network has a maximum depth of 128 layers
-   The programmable network channel width of each layer is up to 1024 channels
-   Convolution processing in one and two dimensions
-   Capable of processing VGA images at 30 fps

● Power management to extend battery life
-   Integrated single-inductor Multi-Output (SIMO) Switch-mode Power Supply (SMPS)
-   2.85V to 3.6V power supply voltage range
-   Optional external auxiliary CNN power supply is supported
-   Dynamic voltage regulation minimizes active kernel power consumption
-   Power consumption 23.9μA /MHz at 3.3V from cache loop (CM4 only)
-   SRAM hold is optional in low power mode with Real-time Clock (RTC) enabled

● Security and integrity
-   Provide safe startup
-   AES 128/192/256 Hardware acceleration engine
-   True Random Number Generator (TRNG) seed generator

Applications

● Factory robots and drone navigation
● Industrial sensors and process control
● On-line quality assurance vision system
● Intelligent security camera
● Portable medical diagnostic equipment