How Edge Artificial Intelligence (AI) Chips Will Take IoT Devices to the Next Level


Edge artificial intelligence (AI) chips will help IoT devices run faster and smarter in 2022

In recent years, edge computing has grown in popularity to provide IoT (Internet of Things) devices and AI (Artificial Intelligence) applications with valuable sensor information quickly and efficiently. But, to effectively implement these innovative technologies at scale, IC manufacturers and researchers must first create new, specialized chips that can support their compute-intensive demands.

The Chinese startup Reexen Technology was established in 2018 by Dr. Hongjie LIU, graduate of ETH Zurich. It has now established itself as a world leader in the field of edge-AI ASICs (Application-Specific Integrated Circuits) for medical, industrial and consumer markets.

Although they have a significant impact on our daily lives, it can be difficult to navigate the wide range of terms used to describe the latest technical trends, such as “edge-AI application-specific integrated circuits” Where “DNN Features Embedded in Low-Power Internet of Things (IoT) Sensors.” This article includes two main objectives to solve this problem. The first objective is to provide a brief overview of key concepts related to the emerging field of AIoT (“Artificial Intelligence of Objects”), which encompasses many of the terms listed above. The second objective is to provide practical examples of how these technologies are implemented in the real world, using the work of Reexen.

IoT (Internet of Things), Edge Computing and AI (Artificial Intelligence)

The IoT has become one of the most promising new paradigms of the past decade. In a broad sense, “it is simply a network of intelligent objects capable of automatically organizing and sharing information, resources and data. They can also make decisions and react to changes in the environment. This widely publicized idea promises to bring everything in our world together under a single infrastructure, allowing us to communicate and connect with anyone from anywhere in the world. This has led to the proliferation and development of many “smart devices” for many industries including energy, industrial manufacturing, urban planning, healthcare, and more.

Although there are many definitions of what makes an object “smart”, the most important aspects include its ability to gather information about its surroundings through embedded sensors. This information must be analyzed quickly. However, large datasets can be generated quickly, especially when many sensors are connected within an IoT network. This raises the question of what type of computing is most appropriate for the job.

Cloud computing is the most popular option. It outsources the task of managing, processing, and storing data to a network of remote servers located on the Internet rather than a personal computer or local server. However, while this strategy is suitable for specific IoT sectors, it also has several drawbacks, including reduced bandwidth, increased latency, privacy issues, and the possibility of data loss.

Therefore, edge computing has emerged as a promising option for time-sensitive applications in which data is analyzed and processed by small computing devices located close to the data source, i.e. the sensors. These “advanced devices” can open up a wide range of applications that use AI, which has resulted in the development of a new field known as AIoT (Artificial Intelligence of Things). This could be a game-changer as industry experts and researchers predict that AI systems in objects will soon be able to not only identify failures and events, but also gather the necessary information and take the right decisions based on this data – all without the need for a human. intervention.

Despite significant advances in this area, several IoT sensors still use traditional processor chips. These chips are not well suited to implementing many of today’s computationally intensive algorithmic programs that sensors need to run at the edge. For example, these include DNNs (deep neural networks) and cutting-edge machine learning algorithms responsible for many recent breakthroughs in artificial intelligence like DeepMind’s AlphaGo. As a result, significant efforts are currently being made to build new ASICs, which, as their names suggest, are primarily designed for a specific application or task. This is where companies like Reexen Technology are trying to create creative solutions to deploy cutting-edge AI technology at scale.

Reexen technology and neuromorphic engineering

As Dr. Hongjie Liu explains, Reexen Company is involved in neuromorphic engineering, sometimes known as neuromorphic computing. This technique aims to mimic the neural operations and structure of the human brain with hardware and software. Reexen’s goal is to mimic the functioning of the brain, eyes, and cochlea in our ears. This is also called “neuromorphic processing and sensing”. They are currently developing “mixed sign inference detection or memory computation” solutions.

In addition, mixed-signal in-memory computer circuits address latency and power consumption issues in A/D (analog-to-digital) conversion and data-intensive DSPs (digital signal processors) in two ways. First, unlike traditional CPUs (Central Processing Units) or GPUs (Graphics Processing Units), which can only process data in the “computer-readable” domain, mixed-signal computing circuits can process sensory signals directly in analog and digital domains. . Second, by now integrating compute cells into memory cells, in-memory processing solutions can address flaws in the “von Neumann” architecture of traditional computers, which dramatically increase the energy and time required to transfer data. data from memory to central processing unit for calculation.

On the other hand, inference sensing means that inputs generated from the physical world are processed and transformed on the sensor side rather than on a mainframe or in the cloud, which is beneficial for a variety of applications, including including headphones, smart watches, smart IoT gadgets, etc.

In this case, Reexen Company collaborated with the leading manufacturer of Micro-Electro-Mechanical Systems (MEMS) microphones to place its innovative audio processing chip inside the MEMS sensor itself, allowing the microphone to use detection of key words. This is crucial for many voice recognition apps, including “Hey Alexa”, “Okay Google” or “Hey Siri”. These words allow digital assistants to respond to user queries. Reexen technology is currently working on a vision processing chip mainly used in AR/VR glasses or smartphones.


In summary, edge-based computing has proven to be an attractive solution for IoT devices that provide high-quality and actionable sensor information. It can also save time and reduce energy.

However, industry leaders and researchers have been working together to create new chips capable of performing more demanding machine learning tasks on real-time devices, either fully or using a hybrid strategy.

Reexen Technology is a Chinese startup that develops “mixed-signal in-memory computing” and “inference detection” solutions. These solutions aim to mimic the neural functioning and structure of the human brain. For example, it led to the creation of an innovative audio processing chip used to build MEMS microphones with built-in keyword detection functions. It is also used to develop a vision processing chip for AR/VR glasses and smartphones.

Alexis Gutzmann
Previous Juniper Networks Unveils New Cloud Metro Solution to Improve Service Provider Productivity
Next 5 digital solutions for a greener Europe