New Brain-Like Chip Fires Like a Neuron and Sips Almost No Power

A team of researchers has built a single artificial brain cell — a neuron — out of hardware, and it behaves a lot more like the real thing than most chips do. Real neurons don't just switch on and off; they fire in rhythms. Some fire in quick bursts, some fire steadily, some stay quiet until pushed. This new device reproduces five of those firing styles, which is unusual for a chip.

The trick is a tiny magnetic component called a magnetic tunnel junction (think of it as a switch flipped by magnetism instead of electricity) paired with ordinary chip circuitry. A small change in voltage reshapes how easily the magnetic switch fires, which is what creates the different rhythms. The researchers used GlobalFoundries' 22-nanometer manufacturing process — a real commercial production line, not a lab one-off — suggesting this could actually be manufactured at scale someday.

Here's the number that matters: each firing event uses about 145.44 femt ojoules of energy. A femt ojoule is so small it's hard to picture — a single LED bulb uses trillions of times more per second. In simulations, these smarter firing patterns also cut the total number of signals needed by up to 88.6% while keeping the same accuracy on classification tasks. Fewer signals, less energy, same results.

Why should you care? AI's biggest hidden cost is electricity. Data centers are straining power grids, and smartwatches, earbuds and sensors all die too fast because AI drains batteries. Chips that mimic how brains actually work — efficient, event-driven, quiet until needed — could ease both problems. The honest catch: this is measured component data plus computer simulation, not a finished product. A working commercial chip is likely years away, and tiny energy savings per neuron don't automatically scale to whole systems.