TinyML is a miniature or small-scale artificial intelligence technology that can run on resource-constrained microcontrollers (MCUs) with features such as low latency, low power consumption, and low cost. It can perform inference tasks in AI such as keyword detection, anomaly detection, and object recognition.
MCU Manufacturers Merging with AI Companies to Layout TinyML
In May 2021, STMicroelectronics acquired the edge AI software specialist Cartesiam.
Cartesiam is a software company specializing in the development of artificial intelligence (AI) development tools, enabling Arm®-based microcontrollers to have machine learning and inference capabilities. The number of products using Cartesiam technology has reached billions. The company’s development team consists of data scientists and embedded signal processing experts, with extensive experience in developing both standard and custom solutions. NanoEdge™ AI Studio is its patented flagship solution, allowing embedded system designers without AI knowledge to quickly develop dedicated software libraries and directly integrate machine learning algorithms into various application systems. Devices containing Cartesiam technology have already been mass-produced globally, including IoT devices, home appliances, and industrial equipment.
In July 2022, Renesas Electronics acquired the outstanding supplier of embedded AI solutions, Reality Analytics, Inc. (Reality AI), making it an indirect wholly-owned subsidiary of Renesas Electronics.
Reality AI is a U.S. company providing a wide range of embedded AI and TinyML solutions for advanced non-visual sensing in automotive, industrial, and consumer products. The combination of Reality AI’s superior AI inference technology with Renesas Electronics’ extensive MCU and MPU product offerings will enable seamless integration of machine learning and signal processing. This acquisition will allow Renesas Electronics to expand its toolkit and software products for AI applications and enhance its ability to offer highly optimized hardware-software integrated endpoint solutions.
In May 2023, Infineon announced it had acquired the TinyML startup Imagimob.
Imagimob is dedicated to providing an end-to-end development platform for machine learning on edge devices. Imagimob’s platform supports various use cases such as audio event detection, voice control, predictive maintenance, gesture recognition, signal classification, and material detection, further expanding Infineon’s hardware and software ecosystem. The merger will combine the expertise of both parties and apply it to a complete sensor product portfolio, allowing them to provide a unified user experience across products for existing customers, facilitating the rapid deployment of powerful solutions, and accelerating the further popularization of TinyML across all applications and fields.
In August 2023, Nordic Semiconductor announced the acquisition of the U.S. AI and machine learning company Atlazo.
Nordic plans to incorporate Atlazo’s ultra-low-power AI/ML processor technology into future SoCs, enhancing Nordic’s business and technology products across many vertical markets. With the integration of Atlazo’s sensor technology in health applications, Nordic can better serve the expanding smart health market, including optical heart monitoring, continuous glucose monitoring, and other wearable technologies.
Additionally, beyond directly acquiring TinyML companies, there are enhancements in products and tools for TinyML applications. For instance, ST will launch the STM32N6 integrated with an NPU, NXP has introduced the MCX product line integrated with a DSP coprocessor and a neural processing unit (NPU), and Microchip provides the MPLAB machine learning development kit to enhance the machine learning capabilities of its products.
Pathways to Layout TinyML
From the current application situation of various microcontroller (MCU) manufacturers and the industry, the pathways to layout TinyML are roughly as follows:
-
TinyML tools
-
Integration of Neural Processing Unit (NPU) coprocessors
-
TinyML platforms
Compared to large language models, TinyML is a small or extremely small-scale machine learning. TinyML itself is a software algorithm, a computational processing method for data analysis and processing, performing inference by constructing machine learning models, and there are also some general-purpose TinyML machine learning models in the industry. TinyML requires software and tools to analyze and process the collected data, perform inference and validation, etc. After continuously optimizing the model, it needs to be converted into firmware that can run on microcontrollers (MCUs) with weaker performance and fewer resources. TinyML tools facilitate the development and application of artificial intelligence, making the development of complex AI software simple and fast.
An NPU is a neural network unit dedicated to neural network computations. As the demand for computational resources in AI applications grows and tasks become increasingly complex, the computational requirements for neural networks also rise. The ALU (Arithmetic Logic Unit) in microcontrollers (MCUs) is insufficient to support the complex computational demands of neural networks. By hardware implementing and integrating the NPU into microcontrollers (MCUs) as a coprocessor, the computational efficiency of AI applications can be significantly enhanced. The NPU is essentially a dedicated hardware computing unit.
Generally, products designed and developed based on microcontrollers (MCUs) can use their firmware for several years, or even until the end of their lifecycle. Once the firmware is verified to be stable, it typically does not require further changes. As the deployment scale of products continues to expand and products undergo continual iterations and updates, it is necessary to update the firmware in microcontrollers (MCUs). A platform can manage and maintain devices effectively. TinyML platforms not only need to iterate and upgrade firmware but can also collect data, continuously train machine learning models, and optimize models, improving model accuracy. On the other hand, the local computational resources of microcontrollers (MCUs) are insufficient to support the storage of large amounts of data and tasks such as model training and optimization. TinyML requires a high-performance computing platform to support the continuous iteration of AI products.
Conclusion
From intelligent control of devices to autonomous intelligence, market demand is continuously driving the innovation and development of technology and products. Cloud-based artificial intelligence will penetrate to the edge, making AI applications more grounded and forming distributed computing for AI that collaborates between cloud and edge, empowering various industries. TinyML is typically applied to edge or extreme edge sensors or devices, which can not only process collected data but also perform simple inference, making sensors or devices more autonomous and intelligent.
The operation of TinyML artificial intelligence algorithms relies on computing hardware – microcontrollers (MCUs). By leveraging TinyML tools, NPU hardware acceleration, and platform-based deployment, the computational efficiency of microcontroller (MCU) products is greatly enhanced, opening new development markets for AI applications. The joint participation of numerous microcontroller manufacturers also provides the market with more innovative products or solutions. This will bring new development opportunities for microcontrollers (MCUs) in the era of ‘compatible machines’ and is bound to open a new track.
-
Artificial Intelligence | Introduction to TensorFlow
-
Artificial Intelligence | Introduction to AI for Microcontrollers
-
Artificial Intelligence | Introduction to Keras
-
Artificial Intelligence | Introduction to ONNX
-
Artificial Intelligence | Arm White Paper ‘Machine Learning on Arm Cortex-M Microcontrollers’
-
Artificial Intelligence | Introduction to PyTorch
-
Artificial Intelligence | Introduction to Scikit-learn
-
Artificial Intelligence | ABIresearch ‘TinyML: The Next Big Opportunity in Tech’