Devices that must always remain on must conserve power in every way possible to extend their battery life. Their design starts with the lowest possible system power and every mode of operation must consume the bare minimum power necessary for operation. Now, with AML100, an analog machine learning or ML chip from Aspinity, it is possible to cut down the system power by up to 95%, even when the system always remains on. AML100 consumes less than 100 µA of always-on system power. This opens new types of products for biometric monitoring, preventive and predictive maintenance, commercial and home security, and voice-first systems, all of which are systems that continuously must remain switched on.
The movement of data to and from a system consumes power. One of the most effective ways of reducing power consumption is, therefore, minimizing the amount and movement of data through a system. The AML100 transfers the machine learning workload to the analog domain where it consumes ultra-low levels of power. The chip determines the relevancy of data with highly accurate and near-zero power. By intelligently reducing the data at the sensor, while it is still in the analog mode, the tiny ML chip keeps its digital components in low-power mode. Only when it detects important data, does the chip allows the analog data to enter the digital domain. This eliminates the extra power consumption in digitizing, processing, and transmission of irrelevant analog data.
The AML100 consists of an array of independent analog blocks configurable to be fully programmable with software. This allows the chip to support a wide range of functions that include sensor interfacing and machine learning. It is possible to program the device in the field, using software updates, or with newer algorithms that target other always-on applications. When it is in always-sensing mode, the chip consumes a paltry 20 µA, and it can support four analog sensors in different combinations like accelerometers, microphones, and so on.
At present, Aspinity is producing the AML100 chip in sampling numbers for key customers. The chip has dimensions of 7 x 7 mm and is housed in a 48-pin QFN package. Aspinity has slated the volume production of this chip for the fourth quarter of 2022 and is presently offering two evaluation kits with software. One of the kits is for glass breakage and T3/T4 alarm tone detection, while the other is for voice detection with preroll collection and delivery. Other kits with software for other applications are also available from Aspinity on request.
AML100 is the first product in the AnalogML family from Aspinity. It detects sensor-driven events from raw, analog sensors by classifying the data. It allows developers to design edge-processing devices with significantly low power consumption, those that are always on. The device has a unique RAMP or Reconfigurable Analog Modular Processor technology platform that allows the AML100 to reduce the always-on system power by more than 95%. This enables designers to build ultra-low power always-on solutions with edge-processing techniques for biomedical monitoring, predictive and preventive maintenance for industrial equipment, acoustic event monitoring applications, and voice-driven systems.