MAX78000 Low-Power CNN AI Microcontroller

Overview

MAX78000 is an AI microcontroller from Maxim that reduces power consumption for on-device neural network inference. It integrates a hardware convolutional neural network (CNN) accelerator with a proven low-power microcontroller architecture. The CNN accelerator enables edge AI inference in battery-powered applications with energy per inference in the microjoule range.

The device is a system-on-chip that combines an Arm Cortex-M4 core with FPU for system control and a ultra-low-power deep neural network accelerator. The CNN engine includes a 442 KB weight memory supporting 1-, 2-, 4-, and 8-bit weights (up to 3.5 million weights). The weight memory is SRAM-based, allowing immediate updates to AI networks. The CNN engine also integrates 512 KB of data memory. The CNN architecture supports training with standard toolchains such as PyTorch and TensorFlow, with conversion tools provided by Maxim to deploy networks on the MAX78000.

In addition to the CNN memories, the MAX78000 provides microcontroller-class on-chip memory with 512 KB flash and up to 128 KB SRAM, and multiple high-speed and low-power interfaces including I2S and a parallel camera interface (PCIF).

The package is an 81-pin CTBGA (8 mm x 8 mm, 0.8 mm pitch).

Applications

Audio processing: multi-keyword spotting, sound classification, noise reduction
Face recognition
Object detection and classification
Time-series data processing: heart-rate and vital-sign analysis, multi-sensor analysis, predictive maintenance

Key Features

Dual-core ultra-low-power microcontroller
Arm Cortex-M4 processor with FPU, up to 100 MHz
512 KB flash and 128 KB SRAM
16 KB instruction cache to help optimize performance
Optional ECC (SEC-DED) for SRAM
32-bit RISC-V coprocessor, up to 60 MHz
Up to 52 general-purpose I/O pins
12-bit parallel camera interface
One I2S master/slave for digital audio interface

Neural Network Accelerator

Highly optimized for deep convolutional neural networks
442 KB of 8-bit weight capacity, with support for 1-, 2-, 4-, and 8-bit weights
Programmable input image size up to 1024 x 1024 pixels
Programmable network depth up to 64 layers
Programmable per-layer channel width up to 1024 channels
Support for 1D and 2D convolution processing
Stream mode operation
Flexible support for other network types including MLP and recurrent neural networks

Power Management for Battery Applications

Integrated switch-mode power supply (SMPS) with single-inductor multiple-output (SIMO)
SIMO supply voltage range: 2.0 V to 3.6 V
Dynamic voltage regulation to minimize core power consumption
At 3.0 V, current when executing a while loop from cache is 22.2 μA/MHz (CM4 core only)
Supports SRAM data retention in low-power modes when the real-time clock (RTC) is enabled