Develop edge AI applications with ADI's MAX78002 MCU

The financial race by tech giants to commercialize generative artificial intelligence (GenAI) has somewhat obscured the amount of work being done on AI, especially at the edge of the network, where vendors are desperate for AI applications to run on iot devices, which are often limited by memory, bandwidth, and power consumption. A microcontroller (MCU) from Analog Devices, Inc. integrates a low-power convolutional neural network (CNN) accelerator to process AI inference on battery-powered devices, pushing the limits of edge processing.

Given that GenAI's main investment is to accumulate massive amounts of data and enhance processing power, which requires massive data centers and large amounts of power, Edge AI is to efficiently run data locally through models that can recognize objects, analyze medical images, and process car camera feedback to identify obstacles, pedestrians, and road signs. Thus achieve safe driving.

CNNS can process image data at the edge, detect anomalies and monitor the health of equipment on the factory floor. In addition, CNNS can be used in agriculture to detect pests and crop growth, processing images from drones, robots and smart cameras.

Optimized for deep CNNS

ADI's MAX78002 is an advanced, ultra-low-power system-on-chip that uses an Arm Cortex-M4 processor with a floating-point unit (FPU) and a hardware-based accelerator, and is optimized for deep CNNS and tasks requiring object recognition capabilities.

Neurons in a neural network are connected by weights (or parameters) to control their behavior. ADI's CNN engine has 2 MB of weight storage memory and can support 1, 2, 4 and 8-bit weights, as well as complex neural network models with up to 16 million weights. In this way, advanced AI applications can be implemented on edge devices, and since the CNN weight memory is based on SRAM, the model can be updated during runtime.

The CNN Accelerator offers programmable input image sizes up to 2048 x 2048 pixels, giving designers the flexibility to design applications that can handle high-resolution medical imaging or smaller input sizes on devices with limited resources.

The programmable network is up to 128 layers deep, so a balance can be struck between the expressiveness and efficiency of the application. In addition, the programmable per-layer network channel width of up to 1024 channels enables users to take advantage of wider channels to capture richer features, or use narrower widths to save memory and compute resources.

The MAX78002 supports a variety of high-speed, low-power communication interfaces, including I2S, MIPI CSI-2 serial camera, Parallel Camera (PCIF), and SD 3.0/SDIO 3.0/eMMC 4.51 secure digital interfaces. This makes the device ideal for a wide range of AI applications, including industrial sensors, process control, in-line quality assurance vision systems, portable medical diagnostic devices, factory robots, and drone navigation.

Power management is key

Ultra-low-power microcontrollers are critical for edge AI applications, especially when relying on battery-powered iot devices. ADI says the MAX78002 uses only a few microjoules to process AI inference.

This AI MCU has a built-in Single Inductor Multiple Output (SIMO) switching mode power supply (SMPS) that supports a range of supply voltages from 2.85V to 3.6V and can be adapted to a variety of power supplies. In addition, through this MCU, an external switch can be optionally controlled to provide dedicated power to the CNN from the outside.

The MAX78002's Power management Unit (PMU) intelligently and precisely controls power distribution between the CPU and peripheral circuits, enabling high-performance operation with minimal power consumption.

Single-chip power supply architecture can be powered by a single lithium battery. Users can voltage program the three step-down regulator outputs in SIMO to ensure optimal power efficiency. Because the MAX78002 requires only one inductor/capacitor, suppliers can reduce the bill of materials for circuit design.

The integrated Dynamic Voltage Regulation (DVS) controller adaptively adjusts voltage to reduce dynamic power consumption. By using a fixed high-speed oscillator and VCOREA power supply voltage, the DVS controller can run the Arm core at the lowest actual voltage, enabling product designers to strike a balance between performance requirements and power consumption limitations. Arm peripheral bus interface for control and status access.

The microcontroller core has 2.5MB of flash large on-chip system memory to ensure non-volatile storage of program and data memory, while its 384 KB built-in SRAM preserves application information in low-power form in all power modes except "power off".