Modern machine vision systems are gaining autonomy: engineers have managed to implement an image recognition algorithm directly on a miniature chip, freeing the camera from having to send data to a bulky computer. The working principle resembles the strategy of an exoplanet-hunting telescope: instead of passively scanning the entire field of view, the camera selectively focuses on significant objects — moving targets, contours, faces. The key processing element is the convolution, a mathematical operation underlying convolutional neural networks. Imagine a stencil you place over a photo: it doesn't just highlight edges but computes how much each image patch resembles the sought pattern — be it a vertical line or an arc. This "stencil" (convolution filter) slides across the frame, multiplying its values by pixel intensities and summing the results, creating a feature map. This hardware implementation bypasses the traditional step of decomposing a color image into separate color channels, which requires sequential processing and slows down the pipeline. As a result, power consumption drops by a factor of three, while recognition accuracy loss is only 5%—a trade-off justified for mobile devices. The specialized processor, whose architecture is tailored exclusively for computer vision tasks, achieves a speed of 34 frames per second — faster than a single human blink (about 100–150 milliseconds). The energy cost per frame is comparable to a single mouse click, paving the way for radically new scenarios: drones that can patrol for hours without recharging, and robots that react to environmental changes with minimal latency.