Click here to ask about the production status of specific part numbers. MAX78000 Artificial Intelligence Microcontroller with Ultra- Low-Power Convolutional Neural Network Accelerator General Description Benefits and Features Artificial intelligence (AI) requires extreme computational Dual Core Ultra-Low-Power Microcontroller horsepower, but Maxim is cutting the power cord from Arm Cortex-M4 Processor with FPU up to 100MHz AI insights. The MAX78000 is a new breed of AI mi- 512KB Flash and 128KB SRAM crocontroller built to enable neural networks to execute Optimized Performance with 16KB Instruction at ultra-low power and live at the edge of the IoT. This Cache product combines the most energy-efficient AI processing Optional Error Correction Code (ECC-SEC-DED) with Maxim s proven ultra-low power microcontrollers. Our for SRAM hardware-based convolutional neural network (CNN) ac- 32-Bit RISC-V Coprocessor up to 60MHz celerator enables battery-powered applications to execute Up to 52 General-Purpose I/O Pins AI inferences while spending only microjoules of energy. 12-Bit Parallel Camera Interface 2 One I S Master/Slave for Digital Audio Interface The MAX78000 is an advanced system-on-chip featuring an Arm Cortex -M4 with FPU CPU for efficient system Neural Network Accelerator control with an ultra-low-power deep neural network accel- Highly Optimized for Deep Convolutional Neural erator. The CNN engine has a weight storage memory of Networks 442KB, and can support 1-, 2-, 4-, and 8-bit weights (sup- 442k 8-Bit Weight Capacity with 1,2,4,8-Bit Weights porting networks of up to 3.5 million weights). The CNN Programmable Input Image Size up to 1024 x 1024 weight memory is SRAM-based, so AI network updates pixels can be made on the fly. The CNN engine also has 512KB Programmable Network Depth up to 64 Layers of data memory. The CNN architecture is highly flexible, Programmable per Layer Network Channel Widths allowing networks to be trained in conventional toolsets up to 1024 Channels like PyTorch and TensorFlow , then converted for exe- 1 and 2 Dimensional Convolution Processing cution on the MAX78000 using tools provided by Maxim. Streaming Mode Flexibility to Support Other Network Types, In addition to the memory in the CNN engine, the Including MLP and Recurrent Neural Networks MAX78000 has large on-chip system memory for the mi- crocontroller core, with 512KB flash and up to 128KB Power Management Maximizes Operating Time for SRAM. Multiple high-speed and low-power communica- Battery Applications 2 tions interfaces are supported, including I S and a parallel Integrated Single-Inductor Multiple-Output (SIMO) camera interface (PCIF). Switch-Mode Power Supply (SMPS) 2.0V to 3.6V SIMO Supply Voltage Range The device is available in a 81-pin CTBGA (8mm x 8mm, Dynamic Voltage Scaling Minimizes Active Core 0.8mm pitch) package. Power Consumption 22.2A/MHz While Loop Execution at 3.0V from Applications Cache (CM4 Only) Object Detection and Classification Selectable SRAM Retention in Low-Power Modes Audio Processing: Multi-Keyword Recognition, Sound with Real-Time Clock (RTC) Enabled Classification, Noise Cancellation Security and Integrity Facial Recognition Available Secure Boot Time-Series Data Processing: Heart Rate/Health AES 128/192/256 Hardware Acceleration Engine Signal Analysis, Multi-Sensor Analysis, Predictive True Random Number Generator (TRNG) Seed Maintenance Generator Ordering Information appears at end of data sheet. Arm and Cortex are registered trademarks of Arm Limited (or its subsidiaries) in the US and/or elsewhere. CoreMark is a registered trademark of the Embedded Microprocessor Benchmark Consortium. Motorola is a registered trademark of Motorola Trademark Holdings, LLC. PyTorch is a trademark of Facebook, Inc. TensorFlow is a trademark of Google, Inc. 19-100868 Rev 1 5/21MAX78000 Artificial Intelligence Microcontroller with Ultra-Low- Power Convolutional Neural Network Accelerator Simplified Block Diagram www.maximintegrated.com Maxim Integrated 2