Reducing the Environmental Impact of AI

by Siyao Shao, Ph.D.

CTO, RECHO AI

One of the persistent topics in AI and machine learning is the obvious environmental impact of this technology. For example, 1.4 to 2% of the global electricity output is used to power the cloud. And the U.S. number is projected to reach 10% in 2028 . With more and more powerful hardware bringing technology to edge devices, the environmental burden is even more unsustainable. Over 78 million batteries are disposed every day to power edge IoT devices, and many of them have elevated power consumption due to added AI/ML capabilities. But with environmental mandates  to detect natural resource waste, chemical leakage, decreasing landfills, etc., AIoT devices are increasingly deployed. 

How do we break this vicious cycle? One may think that most of the energy consumed by AI is in running large models.  And this is true for the ChatGPT or Claude class of Large Language Models. However, for AIoT devices, most of the energy is actually consumed during the process of converting original signals to digital forms and preprocessing to make them more usable by AI models. Consequently, there are no devices that can operate always-active signals, especially audio recognition with sub-mW power consumption as a full system while being small enough to fit into water equipment, walls, motors and many other systems demanding a constant and accurate of the status monitoring. 

At RECHO AI, we intend to solve this problem by using physical reservoir computing technology and Micro-electro and Mechanical System (MEMS). Physical reservoir computing, as the name implies, uses the intrinsic structure of a non-linear physical systems as an untrained Deep Neural Network (DNN), or a reservoir, to emulate functions such as feature processing, memories, and non-linear activation of a DNN. MEMS are one of the most suitable physical reservoirs owing to their small form factor, tunable non-linearity, and wide dissemination in IoT applications. RECHO AI uses Hopf oscillation, one of very common non-linear oscillations in capacitance-based or piezoelectric-based MEMS sensors, as the reservoir computing mechanisms for audio and other time series signal processing. Note that these two types of MEMS sensors cover the majority of the commercial MEMS market for sensing, from audio to ultrasonic to accelerometers and pressure sensors.   

We are achieving the reservoir function for a number of audio related tasks through a product we call the Reservoir Processing Unit (RPU). The combination of the reservoir function along with the MEMS sensor allows us to detect and classify (infer) a number of audio related tasks such as glass breakage, water leakage, or abnormal sound detection. We chose the capacitance-based audio MEMS as our first market to tackle since it serves as the most mainstream and widest disseminated use cases of MEMS. We are active in research and development, and collaboration with academia to demonstrate the compatibility of both technologies and markets. Compared to incumbent technologies, we have eliminated the DSP and inner layers of DNN in the audio recognition pipeline, resulting in multiple orders of magnitude potential improvements in power consumption, latency and lifetime cost by without replacing batteries all in a package the size of a fingertip.

RECHO AI's technology can be integrated into walls, water infrastructure, industrial equipment, and many different places where no other technology can fit. While delivering intelligent insights of audio and other signals, it uses a fraction of the power consumed by devices that need to convert and process digital signals. Consequently, we can offer AI that can truly achieve sustainable AIoT. We can now open up new markets that need ultra low power AI!