Researchers at the University of Tokyo have built a magnetic switching device that operates up to 1,000 times faster than the fastest AI accelerators available today, and it barely generates any heat. The work, published in Science, demonstrates a way to flip a binary magnetic state at picosecond speed without relying on the heat-heavy behavior that makes current faster chips difficult to scale.

Why the Speed Gap Matters

The device uses an antiferromagnet layer to achieve switching speeds measured in picoseconds. The scale at which electrons actually move inside a chip. Today's AI accelerators push data through transistors at nanosecond speeds, roughly 1,000 times slower. The gap compounds across billions of operations because AI workloads spend most of their time moving data between memory and compute.

Every nanosecond saved at the switching level translates to real performance gains at the system level. That is why the Tokyo result is getting attention from hardware researchers who have been watching the power wall problem get worse each generation.

The Power Wall Problem

Chip makers have kept scaling performance by cramming more transistors onto each generation, but the energy cost has become a practical ceiling. Data centers now dedicate as much power to cooling as they do to actual computation. The Tokyo device sidesteps this by using magnetic state changes that do not depend on electron movement through semiconductor material, which is what generates most of the heat in conventional chips.

The antiferromagnet approach matters here. Antiferromagnets have magnetic moments that cancel each other out internally. That internal cancellation means they do not produce the magnetic field interference that limits how small conventional magnetic memory can get. It also means signals do not bleed between adjacent bits, which lets the material store data more densely without crosstalk.

How the Switching Works

The research team used laser pulses to trigger the magnetic switching. The laser drives the state change without making physical contact with the magnetic material, which eliminates the resistive heating that plagues electrical switching methods. Early tests suggest the device could switch using roughly 10 times less energy than a comparable transistor operating at the same speed.

The antiferromagnet layer stores data in its magnetic orientation. To read or write, the laser hits a specific spot and flips the orientation. Because the laser does not need to push current through a wire, the energy overhead is dramatically lower.

From Lab to Real Hardware

There is a significant gap between demonstration and commercial silicon. The current device works as a proof of concept at the materials level, not as a drop-in replacement for existing chip architectures. The team estimates it will take years of engineering work to figure out how to:

  • Manufacture the components at scale
  • Integrate them with existing semiconductor processes
  • Package them in ways that work inside a real computing system

For AI infrastructure operators, the research offers a glimpse at where chip architecture is heading when current silicon approaches its practical limits. The question is not whether the current approach will run out of headroom, but when. This work suggests there are paths around the power wall that do not require abandoning magnetic storage principles entirely.

The team published their findings in Science under the title "Picosecond ultralow-power switching device based on an antiferromagnet." The research was conducted at the University of Tokyo's Department of Electrical and Electronic Engineering.

Want to stay ahead in the tech industry? Visit XerAds for proven product and growth strategies that drive real results in competitive markets.