STMicroelectronics Adds TinyML Capabilities to Its Latest "Smart" MEMS Accelerometers
With dedicated finite state machine and machine learning cores, STMicro's latest sensors aim to offload work from the host processor.
STMicroelectronics has announced the launch of its first "AI-enhanced smart accelerometers," a pair of MEMS-based sensors packing finite state machine (FMC) and machine learning core (MLC) hardware for on-device tinyML workloads.
Based on STMicro's third-generation MEMS sensor platform, the new LIS2DUX12 and LIS2DUXS12 accelerometers aim to offload machine learning and artificial intelligence tasks from a host processor to directly on-sensor. The secret: a triple-core design which augments an adaptive self-configuration (ASC) management core with a machine learning core and a finite state machine.
With the MLC, the sensors are able to run machine larning workloads like activity detection directly on-sensor, sending the result to a host processor rather than sending raw data and expecting the host to do all the work. The finite state machine, meanwhile, works in the background to enhance the sensor's movement recognition capabilities.
With the ASC core adjusting the settings of the sensor on-the-fly to minimize power draw, STMicro claims the result as a way to "make each milliampere count" and reduce load on the host processor. The top-end LIS2DUXS12, meanwhile, includes STmicro's QVar sensing channel, which monitors the ambient electrostatic environment for presence and proximity detection — allowing the sensor to act as a touch-sensitive input, detect liquids, or even monitor a person's heart rate.
These aren't the first sensors to which STMicro has added tinyML capabilities. Earlier this month the company unveiled the iNEMO LSM6DSV16BX system-in-package (SIP), which offered Bluetooth audio support and bone-conduction audio combined with an inertial measurement unit (IMU) with a finite state machine and machine learning core for on-device gesture recognition and more.
The LIS2DUX12 and LIS2DUXS12 are now available, priced at $1.38 and $1.43 each in 1,000-unit tray quantities. For those on a budget, the company has also launched a cut-down version dubbed the LIS2DU12, lacking the machine learning core and QVar support, at $1.20 in the same quantities. More information is available on the STMicro website.