The always-on chip can detect wake words, sounds or vibrations at ultra-low power to activate various sensor systems.
Startup Aspinity’s first chip, an entirely analog sound, wake word and vibration detection chip, can be used to activate other systems. Now available, the AML100 employs a tinyML approach and consumes less than 20 µA of power to help boost battery life in devices requiring always-on AI operation.
All-analog processing eliminates power penalties associated with digitization and digital processing, as well as power consumption linked to downstream processing of irrelevant data.
“Two things are important to system power,” Aspinity CEO Tom Doyle told EE Times. “It’s not just the fact that we are looking at something with really low power. That’s a big driver, but we’re reducing the amount of data that moves from stage to stage.”
Configurable analog blocks
Aspinity’s reconfigurable analog modular processor (Ramp) architecture uses an array of configurable analog blocks (CABs). The blocks contain analog signal-processing elements along with Aspinity’s patented non-volatile memory that uses a floating-gate technique customized for accuracy in standard CMOS. The non-volatile memory could, for example, be used to store parameters for an analog filter or weights in a neural network, depending on block configuration.
All steps from sensor interfacing to pre-processing and feature extraction (which does not use ML) are performed by the CABs. The blocks are also used to compute neural network activations via analog compute in memory techniques, but without digital conversion. If required, the array also can support multiple signal paths.
The signal processing pipeline, including such steps as separating acoustic signals into frequency bands and extracting features such as zero-crossing rates, fits into the CAB array. Aspinity claims that approach is far more efficient than its digital equivalent, in which many neural network layers are dedicated to feature extraction.
Processed data is then passed to a neural network for decision making. The same CABs handle neural network processing.
“We are doing the matrix multiplication like everyone else,” said David Graham, Aspinity’s chief science officer. “The difference is that with other compute-in-memory schemes, they take an analog signal, convert it to digital then convert back to analog to do matrix multiplication.
“We never have to do those conversions,” Graham added. “We have circuitry that is able to do those that matrix multiplications straight up in the analog domain, on raw analog signals, outputting them as an analog signal for anything else that we would need.”
Non-volatile memory in each block is also used to tune individual circuits to account for device inconsistencies. “This is done when we have the algorithm, not before,” said CEO Doyle. “We understand the variation, but we deal with it when we need to deal with it – when we bring the application to bear.”
As circuit parameters are tuned to analog block non-volatile memory, they are also adjusted to account for device mismatch.
The AML100 supports up to four sensor inputs for applications such as three-axis accelerometers or IR sensors combined with audio sensors in a security system. Up to eight signal paths are currently supported. Future devices will scale down, perhaps to two sensor inputs, Doyle said.
Aspinity was founded in 2015 to commercialize research at the West Virginia University. The startup has since developed hardware while working with potential customers. A $2.9 million seed round in 2018 allowed Aspinity to deliver samples in 2019. A $5.3 million series A round followed in 2020.
Aspinity’s AML100 is currently sampling to customers, with volume production planned for the fourth quarter of 2022.
This article was originally published on EE Times.
Sally Ward-Foxton covers AI technology and related issues for EETimes.com and all aspects of the European industry for EE Times Europe magazine. Sally has spent more than 15 years writing about the electronics industry from London, UK. She has written for Electronic Design, ECN, Electronic Specifier: Design, Components in Electronics, and many more. She holds a Masters’ degree in Electrical and Electronic Engineering from the University of Cambridge.