ST Aims to Pack a Punch With DSP, AI, and a MEMS Sensor All on the Same Silicon

February 21, 2022 by Abdulwaliy Oyekunle

In a bid to push forward innovation in micro-electro-mechanical systems (MEMS), STMicroelectronics has launched a power-efficient system that combines signal processing, AI, and MEMS sensors all-in-one.

By integrating both mechanical and electrical functions, MEMS technology incorporates miniaturized electromechanical components such as microsensors and microactuators. 

These devices are often fabricated using several industrial processes, including large-scale integration (VLSI) technology, micromachining, IC process sequences, etc.


An example of a MEMS sensor.

An example of a MEMS sensor. Image used courtesy of Bosch


Recently, to keep pushing the boundaries of MEMS technology, STMicroelectronics has developed a MEMS sensor that integrates digital signal processing (DSP) and AI algorithms on the same silicon.

In this article, let's explore how AI might be beneficial in MEMS devices, the challenges, and ST's latest release. 


How Might AI Scale on MEMS?

Researchers and designers are working round the clock to integrate AI into MEMS to improve performance and expand the use cases of MEMS devices. 

With this in mind, novel AI-MEMS architectures are emerging. 

One such architecture utilizes a resonator with nonlinear dynamics to achieve machine learning (ML) processing in the mechanical domain.


Example MEMS device from the University of Sherbrooke. Image used courtesy of Dion et al


In 2018, researchers at the University of Sherbrooke, Canada, hit a milestone by introducing reservoir computing that allows MEMS oscillators to make time-series predictions and spoken word classification. 

The researchers exploited the nonlinear dynamics of silicon beam, which oscillates in space at widths 20 times thinner than a human hair. 

It is said that the results from this oscillation are used to construct a virtual neural network that projects the input signal into the higher dimensional space required for neural network computing. 

Such AI-MEMS eases mechanical functionalities in robots using accelerometers, which could generate a control signal for the robots.

In general, manufacturing scalable AI-MEMS architectures accelerates the versatility of MEMS devices and signal processing. 

Apart from delivering improved performance, it can also eliminate the use of external microprocessors and field-programmable gate arrays (FPGA).


Challenges for AI on MEMS

When it comes to challenges, MEMS designers face several pitfalls and design constraints when fabricating MEMS devices

Acquiring real-world data in smart MEMS sensors involves employing a higher-resolution analog-digital converter. As such, a 10-bit resolution analog-digital converter might not be helpful in specific applications such as health monitoring. 

In addition, when processing data for data transmission, a limited bandwidth could pose challenges and truncate the data processing.

Designers also face challenges when fabricating state-of-the-art MEMS sensors integrating ML algorithms. 

MEMS sensors that employ classification algorithms such as support vector machines (SVM) require large memory to store a large chunk of real-life datasets.

Despite the challenges of incorporating AI and ML into a MEMS device, ST hopes to make it easier. 


Intelligent Sensor Processing Unit Meets MEMS Sensors

To surmount all the potential challenges associated with fabricating AI on MEMS, ST has introduced an intelligent sensor processing unit (ISPU) that integrates a DSP with a MEMS sensor on an IC. 

The programmable DSP features a single-cycle 16-bit multiplier and could be comfortably operated from 16-bit variable-length instructions. It also includes a full-precision floating-point unit.


Overview of ST's ISPU.

Overview of ST's ISPU. Image used courtesy of STMicroelectronics


The ISPU facilitates full- to single-bit-precision neural networks in quantized AI sensors. 

With AI algorithms running on the DSP, the ISPU accelerates accuracy and efficiency in activity recognition and anomaly detection tasks by analyzing inertial data. 

Furthermore, the ISPU supports AI computing in the Edge, which allows the development of algorithms for the MEMS sensor using AI commercial models while maximizing an ultra-low power consumption. 

ST also claims that the product promises to cut power by up to 80% while reducing size over system-in-package devices. 

Speaking on the product, Andrea Onetti, Executive Vice President, MEMS Sub-Group, STMicroelectronics remarked that the new era—the "Onlife Era"—aims to advance sensor features to speed decision-making by reducing data transfers, enhancing privacy by keeping data local, while reducing the size and power consumption, which cuts costs.

All in all, while achieving maximum privacy, the Onlife Era aims to introduce MEMS devices that could sense real-life data, process complex AI algorithms, and take intelligent real-time action.