News and Events

The Digital Future Depends on Analog

For the last 50 years, digital processing capability has generally followed Moore’s Law with increasing chip densities and higher processing performance, delivering remarkable technological advancements that we leverage every day in all areas of our lives. But many argue that Moore’s Law has now run up against the limits of physics, slowing the progress in semiconductors that we’ve come to expect. This presents an especially difficult challenge at a time when we’re trying to move sophisticated data processing from the cloud into portable edge devices. If we’re ultimately going to achieve cloud-level data analytics at the battery-powered edge, we’ll need a new solution that doesn’t rely just on the digital performance improvements that accompany Moore’s Law, but that also leverages the many benefits of analog computing.

Recently, powerful deep learning processors, called tinyML (tiny machine learning) chips, have begun to provide a first step in enabling new levels of intelligence at the edge. While some of these chips use lower-power analog circuitry to perform select power-intensive computations as a way to reduce component power, they still operate within a standard digital system paradigm where all sensor data are digitized for processing even when most of that data are irrelevant to the task at hand. Because the system wastes power digitizing all that meaningless sensor data, it also unnecessarily drains the battery.

But what if we took a different approach with analog and used it to improve the system efficiency as a whole rather than just the efficiency of the tinyML chip? What if, instead of digitizing all sensor data, we were able to determine which sensor data were important prior to digitization?  That way, we could eliminate the power penalty of performing high-resolution digital analysis on irrelevant data altogether.

As it turns out, we can do that with analogML (analog machine learning), which combines the sophisticated inferencing functionality of a tinyML chip with ultra-low-power analog computing concepts. Aspinity’s analogML core is a fully analog inferencing solution that classifies raw, unstructured sensor data in the analog domain, without the need for power-hungry digitization and digital processors. By more quickly eliminating the irrelevant data at the very beginning of the signal chain, analogML enables a revolutionary system-level approach to low-power edge processing. 

The bottom line is that reaching the limits of Moore’s Law is not going to make or break our ability to move processing from the cloud to the battery-powered edge. But we will impede our progress if we try to solve the edge processing challenges at the component level. The ultimate solution has to address the overall system approach to data efficiency and not simply the performance of the digital processor – and for that we need to turn to the efficiency of analog.