Analog machine learning chip lowers ‘always-on’ power
The AML100 analog machine learning chip suitable for low-power event-detection ahead of analysis is now sampling from Aspinity Inc. (Pittsbugh, Pennsylvania).
Aspinity announced a partnership with Infineon in May 2020 and the availability of an evaluation kit in December 2020 (see Infineon gets ‘always-on’ with Aspinity deal).
The AML100 is now sampling with volume production due in 4Q22. Customers can evaluate the chip through the use of evaluation kits: EVK1 for glass break and T3/T4 alarm tone detection or EVK2 for voice detection.
The chip is a programmable filter and machine learning processor that operates in the analog domain. By operating adjacent to a microphone or other sensor and in the analog domain current consumption is down at 20 microamps and can reduce power consumption by 95 percent compared to systems that move data into the digital domain before performing pattern detection.
This allows for always-on devices for to be battery-powered. Events detected can include glass-break detection, voice detection, alarm detection but can be applied to other forms of pattern matching. Typical application areas include home and commercial security, predictive and preventative maintenance, and biomedical monitoring.
“The AML100 reduces always-on system power to under 100 microamps, and that unlocks the potential of thousands of new kinds of applications running on battery,” said Tom Doyle CEO and founder of Aspinity, in a statement.
Inside the AML100
The heart of the AML100 is an array of tens of independent, configurable analog blocks (CABs) that are software programmable to provide support functions, including sensor interfacing and machine learning.
These interface functions can include band-pass filter, low-pass filter, compression and so on to help isolate signals of interest. These functions and machine learning algorithms can be reprogrammed in the field with software updates or with new algorithms targeting other always-on applications.
The analog processor supports up to four analog sensors – microphones, accelerometers, etc. And it can do multiple inferences at the same time.
The chip is manufactured in a 0.35micron CMOS and operates with a Vdd of 3.3V with configuration data and weights stored in analog form on floating-gate non-volatile memory. It comes in a 7mm by 7mm 48-pin QFN package.
The architecture is scalable and future devices may support less or more sensors or come with different interfaces, Doyle said. More complex systems can be supported by the use of multiple AML100 devices.
Related links and articles:
Infineon gets ‘always-on’ with Aspinity deal
Analog, MEMS and sensor startups to follow in 2022
IP startup launches with ‘new analog’ approach
Agile Analog gains $19 million for global expansion
CEO interview: The importance of being agile
Analog ML voice eval kit for battery-operated devices