Advances in TinyML and generative platforms are significantly lowering the barrier to MCU-based AI development.

At the '2025 e4ds Tech Day' event held on September 9, MDS Tech's Taejun Park presented 'Development Solutions to Break Down Entry Barriers to On-Device AI', and explained in detail the definition, advantages, market outlook, and changes in the development environment of on-device AI.

Cloud-based AI is suitable for large-scale computations, as it transmits data collected from devices to a cloud server for processing and then returns the results to the device. However, it requires a stable network connection and has limitations in terms of real-time performance and privacy protection.

On the other hand, on-device AI is an AI model that is installed inside the device and performs analysis and inference locally without sending data externally.

Park Tae-jun emphasized, “On-device AI can perform real-time analysis without an internet connection, and it is optimized for battery-powered devices as it operates at low power and low cost.”r />
In particular, it is said to be strong in real-time processing by utilizing dedicated hardware such as MPU (Machine Processing Unit) and DSP (Digital Signal Processor), and is more competitive than cloud-based AI in terms of security and strategy.

The market outlook is also bright.

“The on-device AI market is expected to grow by more than 20% annually to reach approximately $39.5 billion by 2033,” said Park Tae-jun, adding, “The shift from cloud-centric AI to on-device-centric AI is accelerating.”

In fact, on-device AI is being applied to various products, such as Samsung's Galaxy S25 and Incheon Airport's guide robot, and Apple's delay in implementing AI functions is a clear example of the importance of technological competitiveness.

On the other hand, there are high barriers to entry in on-device AI development.

Representative Park Tae-jun cited complex and isolated development environments, the difficulty of creating lightweight models, the burden of hardware-specific code optimization, and the complexity of model updates as key challenges.

In particular, while there was a strong perception that high-performance hardware and large-scale infrastructure were essential in the past, it is noteworthy that on-device AI implementation has recently become possible even on 32-bit MCUs (Microcontroller Units).

At its core is "TinyML," an ultra-small machine learning technology capable of executing AI models with memory ranging from tens of kJ to several kJ and power consumption in the milliwatt range. It is suitable for MCU environments because it reduces memory usage based on integer operations and enables local inference without an Internet connection.

Mr. Park introduced major TinyML generation platforms such as Edge Impulse, SensiML, Neuton, and DeepCraft, and emphasized the market interest by showing the background of their acquisition by chip manufacturers.

These platforms provide a one-stop solution for everything from data collection to labeling, feature extraction, training, validation, optimization, and deployment, making it easy to develop models without AI expertise.

In particular, ARM's development environment, 'Keil MDK', supports more than 11,000 Cortex-M-based MCUs and enables efficient development through the CMSIS-NN neural network library and high-performance ARM compiler.

Representative Park Tae-jun explained, “Compared to GCC, the ARM compiler is 20% faster in execution speed and 5% smaller in code size, so we can expect to see a reduction in mass production costs.”

He also said, “Now, you can automatically create and deploy models by simply defining data collection without having to implement complex algorithms yourself,” adding, “This is the beginning of an era where developers can implement AI more quickly and efficiently.”