Aspinity Takes on TinyML, Claiming the Industry’s First Fully Analog ML Chip

February 15, 2022 by Jake Hertz

What happens when you apply the merits of analog compute to ML? According to Aspinity, a chip that reduces system power by up to 95%.

In the pursuit of more energy-efficient computation for applications like TinyML, one of the more promising technologies being developed is analog computation. Analog, if done right, can lead to more power-efficient computation and, hence, huge energy savings for edge computing.

Today, analog compute company Aspinity has made headlines with the release of its new analog TinyML chip, the AML100.



Aspinity says the AML100 is "the world's first analog ML chip." 

The company claims its new product to be the industry’s first and only TinyML solution that operates “completely within the analog domain” with the results being energy savings for the edge. The company has a track record of trailblazing in the analog compute market, releasing an ML analog chip with “selective hearing” two years ago.

All About Circuits had the chance to talk with Tom Doyle, founder and CEO of Aspinity, to hear about the new product firsthand. 


Aspinity's Spin on Traditional Analog Compute

Most analog compute solutions today largely still live in the digital domain. The data (in weights) are stored in the digital domain but use DACs to convert back to analog for computation. 

When trying to achieve low power in an application such as always-on wake word detection, this continual conversion of data from native analog, to digital, back to analog, can limit the power savings offered by raw analog compute. 

“Lots of folks that leverage the term ‘analog’ and ‘analog computing’ are typically trying to solve a problem within the digital domain,” Doyle says. “If you read their websites and you dig a little deeper, you'll find out that actually they're digital domain processors, which means they're actually using digital data. And what they're trying to do with analog is to actually save power within their digital core by doing analog computations within memory.  That’s great, and we're all for it, but it’s very limited.”

Aspinity’s always-on architecture versus traditional computing

Aspinity’s always-on architecture vs. traditional computing. 


Aspinity takes a different approach with its AML100. Doyle explains, “Instead, we're able to move that machine learning capability from the digital domain directly to the analog domain. We're able to analyze that raw sensor data from natively analog, and then we're able to shut the ADC and the digital processor off.”

In this way, the AML100 keeps the data and performs the AI/ML compute entirely in the analog domain. This saves system power overall while minimizing total data quantity since engineers can now keep the digital components in low-power mode until important data are detected, thereby eliminating the power penalty of digitization, digital processing, and transmission of irrelevant data. 


Aspinity AML100 

Announced today, Aspinity’s new AML100 chip leverages this new architecture to provide significant power savings to customers.

The chip’s analog compute is based on an array of independent, configurable analog blocks (CABs), each of which is fully field-programmable within the software, allowing for a wide range of functions, sensor inputs, and applications. Deeper within these CABs, one will find arrays of analog non-volatile memory as well as analog signal processing blocks. The chip also leverages proprietary analog compression technology, allowing for preroll collection and accuracy for applications like wake-word detection. 

Block diagram of the AML100

Block diagram of the AML100. 


Aspinity also says that the chip supports a wide variety of model architectures and ML applications.

“In most of our testing on applications like wake word detection, we find that the AML100 uses probably a quarter of it’s available resources,” Doyle explains. “We’ve found that there's a large amount of leftover resources, indicating that we can realistically support a wide variety of models and new ML architectures if needed.”

Overall, the chip is said to consume less than 20uA while performing always-on sensing. The AML100 offers engineers a 95% reduction in always-on system power consumption, according to Aspinity. 


Cutting Data and Power

While other analog solutions exist on the market, Aspinity claims its approach is novel—which comes with measurable benefits. Purported to reduce data by up to 100x while saving power up to 95%, the AML100 may have a marked impact on edge computing and the world of TinyML.



All images used courtesy of Aspinity.