5 Myths About AnalogML


Because Aspinity’s analogML™ (analog machine learning) core is so unique, it’s not surprising that there are so many questions about exactly what it is and what it does. Over the next couple of months, we invite you to join us as we address some of the top questions, and in some cases misconceptions, about analogML in our Blog series: 5 Myths about AnalogML.

First up: Myth #1: AnalogML is the same as analog-in-memory computing.

These days, it’s becoming all too common for machine learning (ML) processor companies to talk about “analog in-memory computing” as a way to save power in always-listening devices. But is this the same as what Aspinity’s doing? The answer is simple: It’s definitely not.

While leveraging the inherent low power of analog circuitry can definitely save power—and is sometimes referred to “analog computing”—it’s very different from the analog in-memory computing touted by other ML processor companies. With the terms so close, it’s easy to see why the marketplace might confuse the terms, but the truth is that only Aspinity provides a complete analog processing solution: our analogML™ core.

Analog in-memory computing celebrates the fact that analog is more efficient than digital for some of the most important functions of a neural network. While an inferencing chip using analog in-memory computing actually converts sensor data from digital to analog within the chip, it’s only computing multiply-accumulate (MAC) functions using less power. But these ML chips aren’t really analog processors. They’re typical clocked processors that still operate primarily within a traditional digital paradigm that requires the digitization of all analog sensor data before processing. In fact, a chip that uses analog in-memory computing actually requires three separate data conversions prior to knowing the importance of the data. Sensor data are immediately converted to digital for initial processing, then converted to analog within the chip for the MAC functions, and then are again converted back to digital within the chip for the additional digital processing required for inferencing, classification, and other functions. So, plenty of data conversion but not much analog processing.

While in-memory computing may reduce the power of an individual inferencing chip, this technique is very different from the lower-power system solution that’s achieved using Aspinity’s analogML core.

The analogML core operates completely within the analog domain, with no clock required, and uses raw analog sensor data for inference and classification before digitizing any data. In Aspinity’s analyze-first configuration, the analogML core determines the importance of data prior to spending any power on even a single data conversion. The analogML core keeps the digital system off unless relevant data are detected. In some applications, such as glass break detection, where the event might happen once every ten years, minimizing the always-on system power using the analogML core can extend battery life by years, enabling long-lasting remote applications that could never be achieved using analog in-memory computing. Watch the video.