MIT Achieves Breakthrough by Fitting Tens of Thousands of Artificial Brain Synapses on One Chip

Someday, we may carry compact AI brains that don’t rely on supercomputers, the internet, or the cloud. MIT researchers have made significant strides towards this goal with their innovative "brain-on-a-chip" design. This cutting-edge chip, smaller than a piece of confetti, houses tens of thousands of artificial brain synapses, called memristors.

In a study published in Nature Nanotechnology, the team demonstrated the chip's capabilities by successfully memorizing and recreating a grayscale image of Captain America’s shield, and adeptly altering an image of MIT’s Killian Court by sharpening and blurring it. Although these tests may appear minor, the researchers believe this chip design could pave the way for small, portable AI devices capable of performing complex computations currently reserved for supercomputers.

Jeehwan Kim, associate professor of mechanical engineering at MIT, remarks, “So far, artificial synapse networks have existed primarily as software. We aim to build tangible neural network hardware for portable AI systems.” He envisions a scenario where a neuromorphic device is connected to a car's camera, enabling it to instantly recognize signals and objects, all without an internet connection.

MIT isn't alone in the pursuit of neuromorphic chips; tech giants like Apple, Google, Microsoft, and NVIDIA are also developing their own machine learning hardware. Intel's Lohi chip, for example, mimics the brain with 1,024 artificial neurons. However, most existing memristors use silver, while Kim's team discovered that by using silver and copper alloys along with silicon, they could fabricate each memristor more efficiently. This innovation led to the creation of a millimeter-square silicon chip containing tens of thousands of memristors.

Kim emphasizes the importance of their work, stating, “We’re using artificial synapses to conduct real inference tests. Our goal is to expand this technology to enable larger-scale arrays for image recognition tasks. One day, you could carry around these artificial brains to perform such tasks independently of supercomputers, the internet, or the cloud.”