A Vision Camera That Runs on a Microcontroller

CamThink has launched the NeoEyes NE302, a 42 by 42 by 20 millimetre Wi-Fi 6 vision camera built on STMicroelectronics' STM32N6, pairing an Arm Cortex-M55 at 800 MHz with the ST Neural-ART NPU running at 1 GHz for 600 GOPS at an efficiency of 3 TOPS per watt, fanless. It adds a Neo-Chrom 2.5D GPU with Chrom-ART DMA2D acceleration, hardware H.264 and JPEG encoders to 1080p30, and an ISP with demosaicing and auto white balance. Memory is 4.2 MB of SRAM, 32 MB of PSRAM and 64 MB of SPI flash plus a microSD slot. The sensor is a 4-megapixel OS04C10 on MIPI CSI-2 with a selectable 88-degree wide or 137-degree ultra-wide field of view, an M12 lens mount and a white LED fill light. Connectivity is 2.4 GHz Wi-Fi 6 and Bluetooth 5.3 LE on an external SMA antenna at 3 to 4 dBi. It runs from 5 V USB-C, operates from −20 to 50 °C, and costs $99 at launch against a $109 list price, with code, schematics, layouts and documentation published on GitHub.

What is genuinely new here is the class of device, not the product. Until recently, running a useful neural network on image data required an applications processor — a Linux system with hundreds of megabytes of RAM, a boot sequence, a filesystem and everything that follows from those. The STM32N6 puts a competent NPU behind a microcontroller, which means the whole thing boots in milliseconds, has deterministic timing, draws microcontroller-class power and has no operating system to maintain, patch or have a vulnerability in.

For industrial use those properties matter more than the TOPS figure. A device that wakes, classifies and sleeps is a fundamentally different power and reliability proposition from a small Linux computer doing the same job, and the absence of an OS removes an entire category of lifecycle obligation — no kernel updates, no distribution end-of-life, no root filesystem to corrupt on power loss. That is the argument for MCU-class vision in fixed installations: presence detection, counting, a gauge reader, a safety interlock confirming a guard is closed, a quality check on a simple feature.

The open documentation is the part that makes this useful as an evaluation platform rather than a product. Published schematics and layouts mean a team can validate the concept on a $99 device and then design their own board around the same silicon with the reference in hand, which compresses the usual gap between a demo and a product. The limits are equally clear and worth stating: 600 GOPS handles small classification and detection models at modest resolution, not a general-purpose detector on 4-megapixel frames, and the −20 to 50 °C range and USB-C power place this as a development and light-duty device rather than something to bolt inside a machine.