Resistance Is Useful

If you are going to innovate, you have to be willing to think outside the box. Even very effective and well-established technologies, like the von Neumann architecture underlying the vast majority of today's digital computing systems, need to be reevaluated when new applications come along that push them beyond their limits. Without question, modern generative artificial intelligence (AI) tools are one such application.

Having separate processing and memory units is neither fast nor energy efficient when massive amounts of stored data need to continually zip between RAM and the CPU, which is exactly what happens when large AI models run on traditional computing hardware. The fact that the brain — which neural networks seek to emulate — operates nothing like a computer with a von Neumann architecture is another big hint that we are on the wrong track. Without major innovation in this area, forward progress in AI is bound to hit a wall sooner rather than later.

An idea that was recently proposed by a group of researchers at the University of Michigan may help to keep the good times rolling. Using an experimental technology that blends data storage and processing in the same unit, they have developed a brain-like computer that is far more energy efficient than traditional computers. In fact, they showed that their system can run inferences on a small neural network using just 12.5 microwatts of power — roughly 0.25% of the energy required by other existing hardware options.

The team’s work makes use of memristor-based computing, a method that mimics the way biological neurons process information. Memristors, or memory resistors, are electronic components that can store and process information in the same physical location. Unlike traditional computing, which constantly transfers data between memory and processing units, memristors inherently retain information in their electrical resistance, reducing power consumption and improving efficiency.

The researchers built their memristor circuits using a technique called rubbing-induced site-selective deposition. This method allowed them to precisely control the arrangement of memristors on a silicon chip, which made it possible to create a highly efficient computing system.

By integrating these circuits into a reservoir computing network — a machine learning framework designed for time-series data — the researchers achieved real-time robotic control while drastically reducing energy consumption. Their system effectively emulated traditional control algorithms such as proportional-integral-derivative controllers, but at a fraction of the power consumption.

The technology was demonstrated in two applications — a rolling robot that tracked a moving target and a system that controlled a drone motor to keep a lever arm balanced. An Arduino Nano microcontroller board was used to read the data produced by the memristor networks. In both cases, the memristor-based controller performed just as well as conventional digital controllers, but with a fraction of the power consumption.

With the growing demand for AI and autonomous systems, reducing power consumption is a major concern. Large-scale AI models already require vast amounts of energy, and future applications — ranging from smart cities to autonomous vehicles — will need even more efficient computing solutions. Memristor-based computing could provide a path forward, allowing AI to operate at the edge — closer to where data is generated — without relying on massive cloud-based data centers.