The center of gravity of AI technology is shifting from the cloud to the edge. This is not merely a technical change, but a trend that fundamentally alters product design methods and user experience across the entire industry.

According to the 'Edge AI Landscape' recently presented by STMicroelectronics, the technological options developers must consider are changing significantly as ultra-low-power AI based on MCUs and intelligent sensors enters the practical application stage.

Edge AI is no longer a future technology.

A structure that performs inference immediately at the point of data generation is becoming the standard, and balancing power, performance, and cost has emerged as a key challenge.

In the past, high-performance systems based on GPUs or MPUs were essential to implement AI functions, but now an era has dawned where even human detection, acoustic analysis, and simple vision models can be processed using only MCUs.

AI acceleration MCUs such as ST's STM32N6 provide 600 GOPS-class performance and can process even lightweight computer vision models.

Another important change is the role of the sensor.

While conventional sensors were limited to the role of collecting data, they are now evolving into 'intelligent sensors' equipped with MLC (Machine Learning Core) or ISPU (Intelligent Sensor Processing Unit) to perform AI inference on their own.

These sensors perform continuous event detection with ultra-low power of less than 1mW, significantly reducing the load on the MCU and dramatically extending battery life.

We have entered an era where the selection of sensors determines product competitiveness in battery-based products such as wearables, smart homes, and industrial IoT.

These technological changes present developers with three practical strategies.

The first strategy is to improve the functionality of existing products using AI. Since quality and accuracy can be enhanced solely through software without changing the hardware, the ROI is rapid.

Second is a cost reduction strategy. By replacing functions that previously required an MPU with MCU-based AI, BOMs and power consumption can be reduced. This provides a particularly strong competitive edge in mass-produced products.

Third, it is a strategy to create entirely new applications. By implementing features at the edge that were previously only possible via the cloud—such as posture and gesture recognition, biometric authentication, and object segmentation—new user experiences can be provided.

So, what should developers prepare? In this regard, ST offers the following advice.

First, we must move away from a 'cloud-centric mindset.' When designing models, edge optimization must be considered from the outset, and power, latency, and memory constraints should be prioritized.

Second, the roles of the MCU, NPU, and sensors must be clearly distinguished. Architecture design becomes much clearer when understanding the structure where the MCU is lightweight and control-centric, the NPU is for medium-difficulty vision and multimodal functions, and the sensor is for ultra-low-power event detection.

Third, sensor selection must be treated as a core element of product design. Support for MLC and ISPU is now an essential specification.

Fourth, you must actively participate in the TinyML ecosystem. TensorFlow Lite Micro, Edge Impulse, ST Neural-ART, etc., significantly improve development speed and model quality.

Edge AI is not just a technology trend, but a shift in the paradigm of product development.

Developers who understand the distributed AI architecture extending from the cloud to the edge and to sensors will lead the future.

We are not at the beginning of that change; we have already entered the stage of full-scale practical application.

ST stated, “What developers need is not to ‘wait’ for new technology, but to ‘utilize’ technology that has already arrived,” adding that “Edge AI is no longer an experiment, but a realistic technology that can be applied to products right now.”