Welcome back to our 5 Myths about AnalogML™ blog series, which addresses the most common misconceptions about Aspinity’s analogML™ core.
We want our always-on electronic devices to enhance our lives and make it easier. We tell them to play music or order a pizza, protect our home from intruders or water leaks, tell us if our heart rate is too high or our equipment is starting to break down. That’s why we now have hundreds of millions of battery-powered always-on devices that are continuously gathering vast amounts of data as they sense the world around them.
These always-on devices generally require a high-performance DSP or MCU in the system to perform high-resolution analysis of the data they’re monitoring, and control the sending of data to the next stage of the device or to the cloud. But because DSPs and MCUs are digital processors, they indiscriminately digitize all sensor data immediately upon entry into the system—even if that data are irrelevant to the task at hand—an approach that wastes significant amounts of power.
Given their essential role in signal processing, however, we need DSPs and MCUs. The challenge is to get them to work more efficiently.
The analogML core doesn’t replace the DSP or MCU. It works with them.
Aspinity’s system-level solution to saving power leverages the fact that all sensor data are analog, so analyzing the data in the analog domain, at the start of the signal chain before the ADC—makes the system more power-efficient.
How does it work?
Acting like an intelligent gatekeeper, Aspinity’s analogML core determines data relevance and keeps the digital system (ADC/DSP/MCU/other) asleep until important analog data are detected, and further analysis is required. See fig.1.
In contrast to older “digitize-first” architectures, using the analogML core in a newer “analyze-first” architecture can save massive amounts of power, especially in devices where the event of interest happens very rarely. In a voice-first device, for example, the only sound that can possibly contain a keyword is voice. Since voice only happens around 20% of the time, the analogML core—which draws only 10-25uA in always-listening mode—can keep the digital system—which draws a whopping 3-10mA in always-listening mode,—asleep for the other 80% of the time. This cascaded always-sensing architecture eliminates the inefficiency of high-power digital processors analyzing all of the data before determining whether they’re important. Designers who adopt this analogML model can increase battery life by 10x or more in their portable, always-on sensing devices.
No trade-offs
Most of the time when we’re designing a battery-powered device, we weigh the trade-off between reduced features and performance. This is where the analogML core really benefits system designers. By keeping the digital system turned off for most of the time, designers can choose higher-performance options for digital processors, which would otherwise be out of the question for a battery-powered system. This system level power efficiency ultimately paves the way towards completely localized voice enabled devices that do not require a network connection. While it might seem like adding another chip to the system is extra work, the analogML core actually unburdens the digital components in the system—and frees the system designer to make better choices.
