Lab-built artificial synapses switch in 26 nanoseconds

In a few minutes, a chip made by researchers in Suzhou, China, can switch its state in 26 nanoseconds and then dissolve in six minutes. The chip is a flexible array of 100 artificial synapses, the connections that let a brain store and change what it has learned. It exists so far in the laboratory.

The brain runs on roughly the energy of a refrigerator bulb, partly because memory and processing happen in the same place. Ordinary computers separate the two. Artificial intelligence, which needs billions of interconnected values at once, forces data to travel back and forth between them, and that traffic costs energy and time. Engineers have tried to copy the brain's layout with rigid devices such as memristors, usually built from silicon, hafnium oxide or titanium dioxide and layered with heavy metals like platinum or tungsten. Their metallic filaments grow along unpredictable paths, so no two devices on a chip behave quite alike, and their brittle materials cannot easily bend or sit against living tissue.

The Suzhou team's device replaces those materials with a hydrogen-bonded organic framework, called TCPP-HOF. Its building blocks are porphyrin molecules, which assemble into flat sheets stacked about 0.35 nanometers apart. Charges move through this ordered structure more easily than through a disordered organic film. A positive voltage traps electrons in the framework and lowers resistance; a negative voltage releases them and raises it again. That back-and-forth is what lets the device imitate a synapse strengthening or weakening over time. The team sandwiched this layer between two MXene electrodes on a flexible polyimine plastic, and deposited everything with liquid coating methods rather than the vacuum equipment that conventional chip manufacturing relies on. The result is a 10 x 10 array that can hold 10 distinct conductance levels.

The headline figure of 97.23% accuracy on handwritten digits needs a precise reading. The researchers fed their device measurements into a convolutional neural network trained on the standard MNIST dataset, and the network reached 97.23%, close to the 99.24% of the software reference. But that network represents about 120,000 synapses; the task was not run on the 100 physical devices themselves. The result shows the hardware's measured behaviour is usable for recognition. It does not yet show a chip of 100 synapses doing the work alone.

What changes, and for whom? The answer, for now, lies with research teams in flexible electronics, soft robotics and implantable medical devices, who need components that bend, sit on irregular surfaces and can be broken down without toxic residue. The sources give no timeline, no production cost and no plan for a larger array, and the report does not state how the bending test ended. The device is a promising demonstration, not a product. Its next test is whether these properties survive the step from one array to the large networks that real AI systems use.