Researchers at the University of Tokyo say they have demonstrated a non-volatile magnetic switching device capable of flipping states in just 40 picoseconds while consuming unusually little power and generating far less heat than many previous ultrafast switching approaches — potentially addressing one of the biggest problems facing modern AI hardware: the enormous energy and cooling demands created by moving and storing data.

The researchers built the device using an antiferromagnetic material called manganese-tin (Mn3Sn), then showed that ultrashort electrical pulses could reliably switch its magnetic state while retaining the stored information after power removal. They also demonstrated similar switching using ultrafast photocurrent pulses generated from a telecom-band laser and photodiode, effectively converting optical signals directly into memory-writing electrical pulses.

At its most fundamental level, modern computing is really the science of switching physical states. Every operation inside a computer — whether running a game, training an AI model, opening a browser tab, or loading a file from storage — ultimately involves billions or trillions of tiny physical state changes. Transistors switch on and off, memory cells charge and discharge, cache states update, data moves through interconnects, and storage cells trap or release electrons.

New Cambridge human brain-inspired chip could slash AI energy use NEO Semiconductor 3D X-DRAM NEO Semiconductor's revolutionary 3D X-DRAM for AI processors has passed proof-of-concept validation The project logo for TailSlayer, the software described in the article. Ambitious hacker reduces worst-case memory latency by up to 93%, but with severe downsides Those switching events are what physically represent binary information. The problem is that switching states requires energy, and almost all of that energy eventually becomes heat. That reality is becoming increasingly problematic in the AI era. Modern AI accelerators process enormous volumes of data. But much of their power consumption comes not just from computation itself, but from constantly moving and refreshing information between caches, memory, storage, and interconnects. As GPU clusters scale to hundreds of thousands of accelerators, power delivery and cooling are becoming some of the industry's biggest bottlenecks.

Current memory technologies all handle switching differently, but each comes with major tradeoffs. DRAM — the main system memory used in PCs, servers, and GPUs — stores information as electrical charge inside tiny capacitors. A charged capacitor represents one state, while a discharged capacitor represents another. However, those capacitors constantly leak charge, meaning the system must repeatedly refresh the memory cells thousands of times per second simply to preserve data. That constant re-switching consumes significant power and generates heat, even when systems are relatively idle.

Flash memory used in SSDs avoids that problem by trapping electrons in floating-gate structures, which retain data without continuous power. On the other hand, changing those states is slower and more energy-intensive, making flash unsuitable for high-speed working memory.

SRAM, used inside CPU caches, achieves extremely fast switching using transistor feedback circuits that continuously maintain state. But SRAM consumes significant chip area and power, making it expensive and difficult to scale to large capacities.

The industry has spent decades searching for a kind of "universal memory" that could combine the speed of SRAM, the density of DRAM, the persistence of flash, and low power consumption. That challenge becomes even harder at ultrafast timescales, where many experimental switching technologies partially rely on brute-force heating to destabilize and flip states rapidly.

The faster the switching, the more severe the thermal problem often becomes. Several previously demonstrated picosecond-scale switching approaches cited in the paper involve temperature rises of several hundred Kelvin during operation.

The Tokyo researchers are instead pursuing a radically different switching mechanism through a field known as spintronics. Instead of storing information as an electrical charge, spintronic devices store information using magnetic states.

40 years ago we entered the megabit memory era with IBM’s DRAM breakthrough Chips

SoftBank subsidiary working with Intel to develop radical new ZAM memory is now receiving Japanese gov't subsidies

Samsung DRAM

Samsung and SK hynix warn AI-driven memory shortages could last until 2027 and beyond, as HBM demand explodes

Conventional magnetic memories typically use ferromagnets — materials such as iron, cobalt, or nickel in which magnetic moments align in the same direction.

The new device instead uses an antiferromagnetic material called Mn3Sn, where neighboring magnetic moments largely cancel one another out.

Researchers are interested in antiferromagnets because they can potentially switch much faster, resist magnetic interference more effectively, and scale to smaller dimensions without generating large stray magnetic fields.

The researchers fabricated layered Mn3Sn/Ta structures on silicon substrates and then used ultrafast electrical pulses to flip the material between two stable magnetic configurations, representing binary states.

Crucially, the switching mechanism is not primarily based on heating the material. Instead, the pulses generate what is known as spin-orbit torque — a process that transfers angular momentum directly into the magnetic structure itself, flipping the magnetic state without requiring extreme temperature spikes.

That distinction is the paper's central claim. The research is not merely about creating a new kind of memory, but about finding a potentially more energy-efficient way to switch digital states themselves. Currently, almost all electrical energy consumed by computing hardware eventually becomes heat. Modern AI infrastructure is already hitting serious power and cooling limits as GPU clusters scale to hundreds of thousands of accelerators.

The team's device reportedly achieved switching in just 40 picoseconds — roughly 1,000 times faster than typical nanosecond-scale memory switching. Normally, pushing switching speeds into the picosecond regime causes heat generation to spike dramatically, as systems often rely partly on intense transient heating to destabilize states quickly enough for reversal.

However, simulations in one device configuration showed temperature rises of only about 8 K (14.4°F) during switching, supporting the researchers' claim that the mechanism relies primarily on direct angular-momentum transfer rather than brute-force thermal switching. This also confirms that the Mn3Sn device may avoid much of the heat problem that has plagued earlier ultrafast memory research.

The optical switching demonstration may also prove important for future data-center architectures. The researchers generated 60-picosecond photocurrent pulses using a telecom-band laser and photodiode, then used those pulses to switch the device's magnetic state.

That could eventually align with broader industry efforts toward optical interconnects and silicon photonics, where hyperscalers are increasingly seeking ways to move information using light rather than conventional electrical signaling.

If technologies like this ever become commercially viable, they could theoretically reduce memory refresh overhead, lower cooling requirements, reduce idle power draw, and potentially blur the distinction between memory and storage. For personal computing, that could someday translate into systems that retain working memory contents without standby power, resume instantly, and generate less heat. For hyperscale AI infrastructure, the implications would center more around power efficiency and cooling reduction across massive GPU clusters.

For now, however, the technology remains firmly experimental. The current devices are tiny laboratory structures rather than manufacturable memory chips, and the paper notes that the present implementation still requires an external bias magnetic field for deterministic switching — a major practical limitation for commercial hardware.

Manufacturing scalability, endurance validation, cost competitiveness, and integration with existing CMOS manufacturing processes also remain unresolved. The history of computing is full of promising "next-generation memory" technologies that never displaced mature DRAM or NAND ecosystems. Even so, the work highlights the growing reality in the computing industry that future performance gains may depend less on shrinking transistors and more on reducing the energy required to physically switch, move, and store information.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved